July 2024
Uni e sidade do Minho
Escola de Engenha ia
Filipe Alexand e de Sousa Pe ei a
De elopmen o an in elligen compu e
ision sys em o iden i ica ion,
cha ac e iza ion and analysis o ya n
quali y
De elopmen o an in elligen compu e ision sys em o iden i ica ion, cha ac e iza ion
and analysis o ya n quali y
Filipe Alexand e de Sousa Pe ei a
UMinho | 2024
Filipe Alexand e de Sousa Pe ei a
De elopmen o an in elligen compu e
ision sys em o iden i ica ion,
cha ac e iza ion and analysis o ya n quali y
PhD Thesis
Doc o al P og am in Elec onics and Compu e s Enginee ing
specializing in Con ol, Au oma ion and Robo ics
Wo k conduc ed unde he guidance o
P o esso Doc o Filomena Ma ia da Rocha Menezes
de Oli ei a Soa es
P o esso Doc o Ví o Hugo Mendes da Cos a
Ca alho
P o esso Doc o Rosa Ma ia Cas o Fe nandes
Vasconcelos
July 2024
i
DIREITOS DE AUTOR E CONDIÇÕES DE UTILIZAÇÃO DO TRABALHO POR TERCEIROS
Es e é um abalho académico que pode se u ilizado po e cei os desde que espei adas as eg as e
boas p á icas in e nacionalmen e acei es, no que conce ne aos di ei os de au o e di ei os conexos.
Assim, o p esen e abalho pode se u ilizado nos e mos p e is os na licença abaixo indicada.
Caso o u ilizado necessi e de pe missão pa a pode aze um uso do abalho em condições não
p e is as no licenciamen o indicado, de e á con ac a o au o , a a és do Reposi ó iUM da Uni e sidade
do Minho.
Licença concedida aos u ilizado es des e abalho
A ibuição
CC BY
h ps://c ea i ecommons.o g/licenses/by/4.0/
ii
ACKNOWLEDGEMENTS
I am using his oppo uni y o exp ess my g a i ude o e e yone who suppo ed me h oughou he
de elopmen and w i ing o he p esen wo k. I am hank ul o hei aspi ing guidance, in aluably
cons uc i e c i icism, and iendly ad ice du ing he p ojec wo k.
In pa icula , I would like o gi e my hanks o:
o My supe iso P o esso Doc o Filomena Soa es, o he uncondi ional suppo , guidance, and
eedom o pu sue my ideas.
o My ad iso P o esso Doc o Vi o Ca alho, o his help, ad ice, supe ision, and excellen
eedback. He cons an ly allowed his hesis o be my own wo k bu guided me in he igh di ec ion
whene e necessa y. All he ad ice and guidance he ga e me was in aluable and I since ely
hank him.
o My co-supe iso P o esso Doc o Rosa Vasconcelos, o he help, ad ice, supe ision, and
excellen eedback. All he ad ice she ga e me was e y impo an .
o My o me s uden and colleague Leand o Mon eneg o o his aluable knowledge in machine
lea ning. I am e e nally g a e ul o his aluable obse a ions ega ding he wo k ca ied ou .
o P o esso Doc o José Mendes Machado o his help, ad ice, supe ision, and excellen eedback.
o P o esso Doc o Manuel Ca los Felguei as o his ision and en husiasm, and o sha ing his
opinion on aspec s ela ed o he opic.
o My pa en s o p o iding uncondi ional suppo and con inued encou agemen h oughou my
yea s.
o My wi e and child en o hei uncondi ional suppo in he s udy and also in he esea ch and
w i ing p ocess o he Thesis. This achie emen would no ha e been possible wi hou hem.
FUNDING
This wo k has been suppo ed by FCT –
Fundação pa a a Ciência e Tecnologia
in he scope o he p ojec :
UIDB/04077/2020 and UIDB/00319/2020.

iii
STATEMENT OF INTEGRITY
I he eby decla e ha ing conduc ed his academic wo k wi h in eg i y. I con i m ha I ha e no used
plagia ism o any o m o undue use o in o ma ion o alsi ica ion o esul s along he p ocess leading o
i s elabo a ion.
I u he decla e ha I ha e ully acknowledged he Code o E hical Conduc o he Uni e si y o Minho.
Uni e si y o Minho, 30 h July 2024
Full name: Filipe Alexand e de Sousa Pe ei a
Signa u e:
____________________________________________________________________
i
RESUMO
Po ugal é um país com uma o e adição da indús ia êx il e, mesmo no con ex o de mudanças no
me cado, as emp esas po uguesas con inuam a pe segui o ní el de excelência que lhes pe mi e
sob e i e num mundo mais complexo e exigen e.
Na indús ia êx il, a qualidade do p odu o inal es á di e amen e elacionada com a qualidade do io e,
po an o, é essencial aze uma a aliação p ecisa das ca ac e ís icas de aco do com ce os pa âme os
p ede inidos. Há uma e olução dos disposi i os que a aliam a qualidade do io, no en an o, ainda êm
limi ações, como al o cus o, dimensão e peso, assim como esolução e p ecisão limi adas na
de e minação de ce os pa âme os do io.
O obje i o p incipal des a ese é desen ol e algo i mos de
deep lea ning
pa a iden i ica e ca ac e iza
a pilosidade do io, além de c ia algo i mos pa a ca ac e iza e analisa ou os pa âme os de qualidade
do io usando isão compu acional. A es a égia oi, em p imei o luga , p oje a um p o ó ipo meca ónico
que pe mi isse a cap u a di e a de imagens ou ídeos de al a qualidade do en olamen o do io, e ambém
uma análise e classi icação das pilosidades do io. O p o ó ipo pe mi e ob e ou as ca ac e ís icas
ine en es à análise da qualidade do io, como: massa linea , diâme o, olume, di eção da o ção, passo
da o ção, des io médio de massa, coe icien e de a iação, coe icien e de pilosidade, des io médio de
pilosidade e des io pad ão. Es a ese de dou o amen o in oduz, como uma das p incipais con ibuições,
uma no a abo dagem de
deep lea ning
u ilizando um algo i mo o imizado baseado no YOLO 5s6 (You
only look once) pa a ca ac e iza di e en es ipos de pilosidade do io. Os esul ados mos am que o
algo i mo p opos o melho a signi ica i amen e o desempenho do modelo, com um aumen o de 5-6% na
mé ica mAP0.5 (
mean a e age p ecision a 0.5 in e sec ion o e union
(IoU)) e um aumen o de 11-12%
na mé ica mAP0.5:0.95 em compa ação com o algo i mo YOLO 5s6 pad ão. A abo dagem melho a
e e i amen e odas as mé icas analisadas pa a a ca ac e ização da pilosidade do io. A implemen ação
bem-sucedida des e abalho pode aumen a a e iciência p odu i a da indús ia êx il e con ibui pa a a
c iação de p odu os de al o alo ac escen ado.
PALAVRAS-CHAVE:
Deep Lea ning
, De ei os do Fio, Pilosidades, P ocessamen o de Imagem, Qualidade do
Fio.
ABSTRACT
Po ugal has a s ong adi ion in he ex ile indus y, e en unde he con ex o ma ke and demands
changes, he Po uguese companies con inue pu suing he excellence le el ha makes hem su i e in a
mo e complex and challenging wo ld.
In he ex ile indus y, he quali y o he inal p oduc is di ec ly ela ed o he quali y o he ya n and
he e o e i is essen ial o make an accu a e assessmen o he ya n cha ac e is ics, acco ding o ce ain
p e-es ablished pa ame e s. The e is a signi ican e olu ion o he de ices ha e alua e he quali y o he
ya n, howe e , hese de ices s ill ha e se e al limi a ions such as high cos , la ge dimension, and weigh ,
as well as limi ed esolu ion and p ecision in de e mining ce ain pa ame e s o he ya n.
The main goal o his hesis is o de elop deep lea ning algo i hms o iden i y and cha ac e ize ya n
hai iness, as well as o c ea e algo i hms o cha ac e izing and analyzing o he ya n quali y pa ame e s
using compu e ision. The s a egy was, i s , o design a mecha onic p o o ype ha allows o he di ec
cap u e o high-quali y images o ideos o ya n winding, and also o he analysis and classi ica ion o
ya n hai iness.
I also allows o ob ain o he cha ac e is ics inhe en o he analysis o he ya n quali y, such as: linea
mass, diame e , olume, wis di ec ion, wis s ep, a e age mass de ia ion, coe icien o a ia ion,
hai iness coe icien , a e age hai iness de ia ion and s anda d de ia ion.
This hesis in oduces, as one o he main achie emen s, a no el deep lea ning app oach using an
op imized algo i hm based on YOLO 5s6 (You only look once) o cha ac e ize di e en ypes o ya n
hai iness. The esul s show ha he p oposed algo i hm signi ican ly imp o es he model pe o mance,
wi h a 5-6% inc ease in mAP0.5 (mean a e age p ecision a 0.5 in e sec ion o e union (IoU)) me ic and
an 11-12% inc ease in he mAP0.5:0.95 me ic compa ed o he s anda d YOLO 5s6 algo i hm. The
app oach e ec i ely enhances all analyzed me ics o ya n hai iness cha ac e iza ion. The success ul
implemen a ion o his wo k can inc ease he p oduc i e e iciency o he ex ile indus y and con ibu e
o he de elopmen o high added alue p oduc s.
KEYWORDS: Deep Lea ning, Faul s, Hai iness, Image P ocessing, Ya n Quali y.
i
LIST OF CONTENTS
Acknowledgemen s .............................................................................................................................. ii
Resumo.............................................................................................................................................. i
Abs ac ...............................................................................................................................................
Lis o Con en s .................................................................................................................................. i
Lis o Figu es ..................................................................................................................................... xi
Lis o Tables ..................................................................................................................................... xix
Abb e ia ions ..................................................................................................................................... xxi
1. In oduc ion ................................................................................................................................ 1
1.1 F amewo k .......................................................................................................................... 2
1.2 Objec i es ........................................................................................................................... 6
1.3 De elopmen me hodology .................................................................................................. 7
1.4 Main con ibu ions o he wo k ............................................................................................. 8
1.5 Thesis S uc u e ................................................................................................................ 13
2. Theo e ical Concep s ................................................................................................................ 15
2.1 Ya n cha ac e iza ion ......................................................................................................... 16
2.1.1 Linea mass and diame e ......................................................................................... 16
2.1.2 Speci ic olume ......................................................................................................... 17
2.1.3 Hai iness ................................................................................................................... 17
2.1.4 Ya n wis , wis o ien a ion, and wis s ep ................................................................. 18
2.1.5 Types o aul s in ya n ................................................................................................ 19
2.2 S a is ical Pa ame e s o Ya n ............................................................................................ 20
2.2.1 Mean De ia ion o Mass ............................................................................................. 20
2.2.2 Coe icien o Va ia ion ............................................................................................... 20
2.2.3 Hai iness Coe icien .................................................................................................. 20
2.2.4 Hai iness Mean De ia ion .......................................................................................... 21
2.2.5 Hai iness S anda d De ia ion ..................................................................................... 21
2.3 Spec al Analysis ............................................................................................................... 21
2.4 Image P ocessing & Compu e Vision ................................................................................ 22
2.4.1 Uni a y elemen and colo scales ............................................................................... 22
xiii
Figu e 59 –Con usion Ma ix o Bina y Classi ica ion [63]. ................................................................ 58
Figu e 60 – An example o he p ecision- ecall cu e [64]. ................................................................. 59
Figu e 61 – An example o he p ecision- ecall cu e [65]. ................................................................. 60
Figu e 62 – An example o he p ecision- ecall cu e [69]. ................................................................. 62
Figu e 63 – Sub- a ian s o YOLO 5 [56]. ......................................................................................... 64
Figu e 64 – Ne wo k s uc u e o YOLO 5s6 - (A): The comple e s uc u e wi h backbone, neck, and head
modules. (B, C): Two unique a ia ions o CSP blocks (C3); (D): CBS - Con olu ional Ba ch No maliza ion
laye and SILU (sigmoid linea uni s) ac i a ion unc ion; (E): O he blocks wi h CBS; (F) - Two dis inc
bo leneck blocks] [76]. .................................................................................................................... 66
Figu e 65 – C3 Module om YOLO 5s6 [76]. ................................................................................... 67
Figu e 66 – SPPF Module om YOLO 5s6 [76]. ................................................................................ 67
Figu e 67 – Schema ic ep esen a ion o he ac i a ion unc ion [79]. ................................................ 68
Figu e 68 - SiLU ac i a ion unc ion [79]
.
.......................................................................................... 69
Figu e 69 - Ac i a ion unc ion GeLU [80]. ......................................................................................... 71
Figu e 70 - Compa ison o ac i a ion unc ions [81]. .......................................................................... 72
Figu e 71 – YOLO 5s6 Image Augmen a ion [82]. ............................................................................. 73
Figu e 72 - Flowcha o he me hod o de ec ing ya n pa ame e s and e alua ing ya n quali y [83]. .. 77
Figu e 73 - Sys em de ice o de ec ya n pa ame e s [83]. ................................................................ 77
Figu e 74 - a) New o eg ound image, b) ya n co e, c) hai iness [83]. ................................................ 78
Figu e 75 - An example o a ya n image wi h a nep de ec ob ained using he image acquisi ion sys em
[87]. ................................................................................................................................................. 79
Figu e 76 - An example o a ya n image wi h a nep de ec ob ained using he image acquisi ion sys em
[87]. ................................................................................................................................................. 80
Figu e 77 - Con usion ma ices ha o e insigh s in o he pe o mance o he model sys em [87]. ..... 81
Figu e 78 - Co on ya ns: he a chi ec u e o neu al ne wo k ac ion p ope ies [89]. ......................... 83
Figu e 79 - Blended ya ns: he a chi ec u e o neu al ne wo k ac ion p ope ies [89]. ....................... 83
Figu e 80 - Blended ya ns: he s uc u e o neu al ne wo k ac ion p ope ies [91]. ........................... 84
Figu e 81 - The s uc u al composi ion o he neu al ne wo k, including i s enaci y and elonga ion
pe cen ages: (a) Ring ya n sys em (le ), (b) compac ya n sys em ( igh ) [91]. ................................... 84
Figu e 82 - The CVm% neu al ne wo k a chi ec u e o ing-spun and compac ya n sys ems [91] ....... 85
Figu e 83 - The neu al ne wo k a chi ec u es o impe ec ions: a) Ring-spun ya n sys em, b) Compac
ya n sys em [91]. ............................................................................................................................. 85

xi
Figu e 84 – a) A ya n in g ayscale (le ); b) Image a e linea iza ion using he OTSU algo i hm ( igh ) [97].
........................................................................................................................................................ 92
Figu e 85 - a) G ay scale image (le ); b) Image wi h Gaussian il e ( igh ) [97]. ................................. 93
Figu e 86 - a) Image p io o he applica ion o he opening mo phological ope a ion (le ); b) Image
subsequen o he implemen a ion o he opening mo phological ope a ion ( igh ) [97]. ..................... 93
Figu e 87 - a) Image a e he closing p ocess (le ); b) Isola ed co e image ( igh ) [97]. ..................... 94
Figu e 88 - Isola ed ou lines o he ya n. ............................................................................................ 96
Figu e 89 - Con ou s o he isola ed co e. .......................................................................................... 96
Figu e 90 - Con ou s o he isola ed hai iness. ................................................................................... 97
Figu e 91 - Sepa a ion o pa icles in single cable o ya n. ................................................................. 98
Figu e 92 - Sepa a ion o pa icles in Sepa a ion o pa icles in 2 cables o olded ya ns. .................... 98
Figu e 93 - Image a e he opening p ocess. ..................................................................................... 98
Figu e 94 - Image o isola ed pa icles. .............................................................................................. 99
Figu e 95 – S Viewe so wa e execu able [92]. ............................................................................... 100
Figu e 96 –A ailable pa ame e ca ego ies in S Viewe [92]. ............................................................ 100
Figu e 97 – OMRON STC-MBS163U3V Came a [92]. ...................................................................... 101
Figu e 98 – Lens se al eady equipped wi h a ya n holde [92]. ....................................................... 101
Figu e 99 – Use in e ace [92]. ...................................................................................................... 102
Figu e 100 – Algo i hm Analysis Me hodology Flowcha [92]. ......................................................... 105
Figu e 101 - Imp o ed a chi ec u e based on YOLO 5s6 – YOLO 5s6-Hai iness [76]. ...................... 109
Figu e 102 - The CBG module in he op imized YOLO 5s6-Hai iness [104]. ..................................... 110
Figu e 103 - The CBG module employed in a ious o he blocks [104]. ........................................... 110
Figu e 104 - The CBG module is used in wo di e en ypes o bo leneck blocks [104]. ................... 110
Figu e 105 - Two dis inc ypes o CSP blocks (C3) [104]. ............................................................... 111
Figu e 106 - Di e ence be ween C3_x ( op) and C2 module (bo om) [104]. ................................... 113
Figu e 107 - The a chi ec u e o he Bo -T ans o me block. (a) Bo leneckT ans o me *x signi ies ha a
o al o x Bo leneck T ans o me blocks a e s acked, and each Bo leneck T ans o me is depic ed in (b)
[104].............................................................................................................................................. 115
Figu e 108 - MHSA Laye used in he Bo - ans o me Block [106]. ................................................. 116
Figu e 109 – V-Model o he Design o a Mecha onic Sys em - VDI 2206 [110]. ............................. 125
Figu e 110 – a) O e end wi hd awal; b) Side wi hd awal [97]. ........................................................ 131
x
Figu e 111 – a) Ya n ballooning ep esen a ion; b) Ya n guide posi ion abo e conical package apex
ep esen a ion [97]. ........................................................................................................................ 131
Figu e 112 – a) Pa allel package; b) C oss wound package [97]. ..................................................... 132
Figu e 113 – Random winding: a) g oo ed d um; b) ya n guide [97]................................................ 133
Figu e 114 – Illus a ion o p ecision winding [97]. .......................................................................... 133
Figu e 115 – S ep-p ecision winding and depic ion o he luc ua ion in coil angle ac oss phases [97].
...................................................................................................................................................... 134
Figu e 116 – Illus a ion o a lexible winding sys em [97]. ............................................................... 135
Figu e 117 – Di e en ypes o ce amic ya n guides [97]. ............................................................... 136
Figu e 118 – Ya n guide olle [97]. ................................................................................................ 136
Figu e 119 – Me hods o applying ension o a ya n [97]. ................................................................ 137
Figu e 120 – Ya n ensione s: a) addi i e; b) mul iplica i e; c) mul iplica i e sel -adjus ing (le e ension);
d) mul iplica i e sel -adjus ing (ball ension) [97]. ............................................................................ 138
Figu e 121 – Au oma ic ension con ol me hod employing a ya n posi i e o e eed sys em [97]. .... 139
Figu e 122 – F on al schema ic iew o he machine: 1) Box; 2) Bobbin suppo ; 4) Guide suppo ; 6)
Bobbin; 9) Fi s posi i e ya n eede (PYF); 11) Second PYF; 12) PYF suppo ; 13) Too hed bel ; 14)
Too hed bel ; 19) Came a; 20) Se omo o suppo ; 21) Ring ligh ; 22) Guide suppo on lens; 24) Lens;
26) Came a suppo ; 27) Sliding base se omo o ; 28) Sliding base PYF; 31) Suppo block; 33) G oo ed
d um; 34) Fla bel ; 38) G oo ed d um suppo ; 40) Fla pulley; 43) S uc u al beam; 47) Lens suppo ;
49) Ya n guide; 50) Guidance subsys em co e senso [97]. ........................................................... 141
Figu e 123 – Pa h o he ya n du ing winding [97]. ......................................................................... 142
Figu e 124 – Cinema ic chain: 1) Too hed bel and pulley se be ween se omo o and second PYF; 2)
Too hed bel and pulley se be ween i s and second PYF; 3) Fla bel and pulley se be ween se omo o
and winding subsys em; A Sliding suppo o i s PYF; B Sliding suppo o se omo o [97]. ........... 143
Figu e 125 – Au oma ion sys em ep esen a ion: 1) 2 pole ci cui b eake ; 2) 1 pole ci cui b eake (× 2);
3) Powe socke o ya n illumina ion; 4) Powe supply (24 V DC); 5) Se omo o d i e ; 6) Se omo o ; 7)
Posi i e ya n eede 1; 8) ya n unwinding senso elay (24 V DC); 9) Ya n guidance senso elay; 10)
Posi i e ya n eede 2; 11) Ya n winding senso elay; 12) Guidance subsys em co e senso ; 13)
Eme gency bu on; 14) PLC; 15) Local HMI (Human-Machine In e ace); 16) Local Compu e ; 17) NI OPC
Se e ; 18) LabVIEW [97]. .............................................................................................................. 144
Figu e 126 – a) Local HMI loca ion (le ); b) Local HMI sc een o ya n quali y es de ined by ime and
eloci y ( igh ) [97]. ........................................................................................................................ 145
x i
Figu e 127 – Ya n quali y es sa ed da a [97]. ............................................................................... 146
Figu e 128 – Kinema ic chains ((A) and (B) a e Se omo o s [92]. .................................................. 146
Figu e 129 – Ope a ing Modes Coo dina ion Flowcha . .................................................................. 150
Figu e 130 – Tes p epa a ion mode lowcha . ............................................................................... 153
Figu e 131 – Flowcha o he es mode (pa 1). ........................................................................... 155
Figu e 132 – Flowcha o he es mode (pa 2). ........................................................................... 156
Figu e 133 – Flowcha o he es e mina ion mode. ..................................................................... 158
Figu e 134 – Flowcha o he s op mode (pa 1). .......................................................................... 159
Figu e 135 – Flowcha o he s op mode (pa 2). .......................................................................... 160
Figu e 136 – Flowcha o he manual mode. .................................................................................. 162
Figu e 137 – Ou line o he p ocess o ansla ing G a ce in o Ladde Diag am. .............................. 164
Figu e 138 - Indus ial cloud Talk2M wi h Flexy 205 [120]............................................................... 165
Figu e 139 - P o o ype’s mechanical sys em [97]. ........................................................................... 167
Figu e 140 - Au oma ion sys em’s ha dwa e [97]. ........................................................................... 167
Figu e 141 - Th ee ypes o ya n es ed [92]. ................................................................................... 174
Figu e 142 - Con ou s o he ya n ha ha e been isola ed (Co on ya n wi h an a e age linea mass o
56.4 ex) [92]. ................................................................................................................................. 177
Figu e 143 - Image o pu ple h ead wi h isola ed pa icles. ............................................................. 178
Figu e 144 - Mass spec og am by p o o ype [92]. .......................................................................... 183
Figu e 145 - Mass spec og am g aph gene a ed by he USTER TESTER 3 [92]. .............................. 183
Figu e 146 – Hai iness spec og am by p o o ype [92]. ................................................................... 184
Figu e 147 - Hai iness spec og am g aph gene a ed by he USTER TESTER 3 [92]. ........................ 184
Figu e 148 - Anno a ions in a ya n image we e made using LabelME, wi h loop ibe s ma ked in g een
and p o uding ibe s ma ked in ed. The image is o co on ya n wi h an a e age linea mass o 56.4 ex
and a magni ica ion ac o o 20x. ................................................................................................... 185
Figu e 149 - Visualiza ion o he da ase includes: (a) The numbe o anno a ions pe class; (b)
Visualiza ion o he loca ion and size o each bounding box; (c) The s a is ical dis ibu ion o he bounding
box posi ions; (d) The s a is ical dis ibu ion o he bounding box sizes. ............................................ 187
Figu e 150 - An example o a ya n image wi h he p esence o loop ibe s and a con using appea ance.
This image is o co on ya n wi h an a e age linea mass o 56.4 ex, aken wi h a magni ica ion ac o o
20x. ............................................................................................................................................... 188
x ii
Figu e 151 - Da a augmen a ion o he da ase used in he s udy (Co on ya n wi h an a e age linea mass
o 56.4 ex, wi h a magni ica ion ac o o 20x). ............................................................................... 189
Figu e 152 – Numbe o epochs - mAP_0.5. ................................................................................... 191
Figu e 153 – Numbe o epochs - mAP_0.5:0.95. ........................................................................... 192
Figu e 154 – Numbe o epochs - p ecision. .................................................................................... 192
Figu e 155 – Numbe o epochs - ecall. ......................................................................................... 193
Figu e 156 – Va ious alues o lea ning a e in Yolo 5s6 – Hai iness Imp o ed - mAP_0.5. ............. 195
Figu e 157 – Va ious alues o lea ning a e in Yolo 5s6 – Hai iness Imp o ed - mAP_0.5:0.95. ..... 195
Figu e 158 – Va ious alues o lea ning a e in Yolo 5s6 – Hai iness Imp o ed - mAP_0.5. ............. 196
Figu e 159 – Va ious alues o lea ning a e in Yolo 5s6 – Hai iness Imp o ed - mAP_0.5. ............. 196
Figu e 160 - Pe o mance me ics o he YOLO 5s6-Hai iness app oach. ........................................ 199
Figu e 161 - Pe o mance me ics o he p oposed YOLO 5s6-Hai iness app oach. ......................... 200
Figu e 162 - Pe o mance me ics o he p oposed YOLO 5s6-Hai iness app oach. ......................... 202
Figu e 163 - P oposed YOLO 5s6-Hai iness a chi ec u e app oach. ................................................. 211
Figu e 164 - Tes esul s compa ing he p oposed op imized YOLO 5s6-Hai iness ( igh image) wi h he
de aul YOLO 5s6 (le image) a a 0.2 con idence h eshold (Co on ya n wi h an a e age linea mass o
56.4 ex, magni ied by a ac o o 20x). ........................................................................................... 212
Figu e 165 - Tes esul s compa ing he p oposed op imized YOLO 5s6-Hai iness ( igh image) wi h he
de aul YOLO 5s6 (le image) a a 0.3 con idence h eshold (Co on ya n wi h an a e age linea mass o
56.4 ex, magni ied by a ac o o 20x). ........................................................................................... 213
Figu e 166 - Tes esul s compa ing he p oposed op imized YOLO 5s6-Hai iness ( igh image) wi h he
de aul YOLO 5s6 (le image) a a 0.5 con idence h eshold (Co on ya n wi h an a e age linea mass o
56.4 ex, magni ied by a ac o o 20x). ........................................................................................... 213
Figu e A.1-1 - Field o View Visualiza ion [128] ................................................................................ 241
Figu e A.2-1 - Highe Hie a chy G a ce - Coo dina ion o Ope a ing Modes [128] ............................. 243
Figu e A.2-2 - G a ce A – Pa 1 ..................................................................................................... 244
Figu e A.2-3 - G a ce A – Pa 2 ..................................................................................................... 245
Figu e A.2-4 - G a ce A – Pa 3 ..................................................................................................... 245
Figu e A.2-5 - G a ce B .................................................................................................................. 246
Figu e A.2-6 - G a ce A - Tes Mode ................................................................................................ 246
Figu e A.2-7 - G a ce B - Tes Mode – Pa 1 .................................................................................. 247
Figu e A.2-8 - G a ce B - Tes Mode – Pa 2 .................................................................................. 247
x iii
Figu e A.2-9 - G a ce C - Tes Mode ................................................................................................ 248
Figu e A.2-10 - G a ce o Tes Te mina ion Mode – Pa 1 .............................................................. 249
Figu e A.2-11 - G a ce o Tes Te mina ion Mode – Pa 2 .............................................................. 250
Figu e A.2-12 - G a ce o Tes Te mina ion Mode – Pa 3 .............................................................. 251
Figu e A.2-13 - G a ce o S op Mode – Pa 1 ................................................................................. 252
Figu e A.2-14 - G a ce o S op Mode – Pa 2 ................................................................................. 253
Figu e A.2-15 - Manual Mode G a ce – G a ce A – Pa 1 ............................................................... 254
Figu e A.2-16 - Manual Mode G a ce – G a ce A – Pa 2 ............................................................... 254
Figu e A.2-17 - Manual Mode G a ce – G a ce A – Pa 3 ............................................................... 255
Figu e A.2-18 - Manual Mode G a ce – G a ce A – Pa 4 ............................................................... 255
Figu e A.2-19 - Manual Mode G a ce – G a ce C ............................................................................ 255
Figu e A.3-1- Highe Hie a chy G a ce - Coo dina ion o Ope a ing Modes ....................................... 256
Figu e A.3-2 - G a ce o Tes Mode - G a ce A – Pa 1 ................................................................... 256
Figu e A.3-3 - G a ce o Tes Mode - G a ce A – Pa 2 ................................................................... 257
Figu e A.3-4 - G a ce o Tes Mode - G a ce A – Pa 3 ................................................................... 258
Figu e A.3- 5 - G a ce o Tes Mode - G a ce A – Pa 4 .................................................................. 258
Figu e A.3- 6 - G a ce o Tes Mode - G a ce A – Pa 3 .................................................................. 258
Figu e A.3- 7 - G a ce o Tes Mode - G a ce C ............................................................................... 259
Figu e A.3- 8 - G a ce o S op Mode – Pa 1 .................................................................................. 259
Figu e A.3- 9 - G a ce o S op Mode – Pa 2 .................................................................................. 260
Figu e A.3- 10 - G a ce o Manual Mode ......................................................................................... 261
Figu e A.4- 1 - P ede ined sc eens o HMI ins alled on he machine body – Pa 1 .......................... 262
Figu e A.4- 2 - P ede ined sc eens o HMI ins alled on he machine body – Pa 2 .......................... 263
Figu e A.4- 3 - P ede ined sc eens o HMI ins alled on he machine body – Pa 3 .......................... 264
Figu e A.4- 4 - P ede ined sc eens o HMI ins alled on he machine body – Pa 4 .......................... 265
Figu e A.4- 5 - P ede ined sc eens o HMI ins alled on he machine body – Pa 5 .......................... 265
Figu e A.5- 1 - LabVIEW in e ace - Local compu e .......................................................................... 266

xix
LIST OF TABLES
Table 1 - A Compa ison o Cha ac e is ics Be ween P e iously P oposed Sys ems [92] ...................... 86
Table 2 - Came a echnical speci ica ions (Adap ed om [92]). ........................................................ 101
Table 3 - P ojec Pe o mance Speci ica ions ................................................................................... 128
Table 4 - Mo phological cha - g een cells: solu ions adop ed in he inal p o o ype [97]. ................. 140
Table 5 - A Compa ison o Cha ac e is ics Be ween P e iously P oposed Sys ems and he Case S udy
[92] ................................................................................................................................................ 168
Table 6 - Resul s ob ained om he USTER TESTER 3 equipmen [92]............................................. 175
Table 7 - Pu ple Co on Ya n Image Analysis Resul s. ...................................................................... 175
Table 8 - Yellow Co on Ya n Image Analysis Resul s. ....................................................................... 176
Table 9 - Image Analysis Resul s o Whi e Polyes e Ya n. ................................................................ 176
Table 10 - Resul s o he ideo analysis o 1km o pu ple co on. .................................................... 180
Table 11 - Classi ica ion o he da ase based on a ious expe imen al condi ions - wi hou using da a
augmen a ion echniques [113]. ..................................................................................................... 186
Table 12 - The de elopmen en i onmen . ....................................................................................... 189
Table 13 - So wa e ools en i onmen . ........................................................................................... 190
Table 14 - Pa ame e s se ings o he aining con igu a ion. ............................................................ 190
Table 15 - Pa ame e s se ings o he aining con igu a ion. ............................................................ 193
Table 16 - Pa ame e s se ings o he lea ning a e aining con igu a ion in Yolo 5s6 – Hai iness Imp o ed
algo i hm. ....................................................................................................................................... 194
Table 17 - Pa ame e s se ings o he aining con igu a ion. ............................................................ 198
Table 18 - Pe o mance me ics inc ease (%) o he op imized YOLO 5s6-Hai iness algo i hm wi h
augmen a ion and he YOLO 5s6 De aul model wi h augmen a ion. ................................................ 198
Table 19 - Me ics compa a ion esul s be ween YOLO 5s6 De aul wi h da a Augmen a ion and imp o ed
YOLO 5s6-Hai iness wi h C2 and Bo -T ans o me Module. ............................................................ 204
Table 20 - Pe o mance compa ison o in oducing he C2 module in he di e en loca ions o he
algo i hm. ....................................................................................................................................... 205
Table 21 - Me ics compa a ion esul s be ween Ac i a ion Func ions wi h da a Augmen a ion. ......... 206
Table 22 - Imp o ed YOLO 5s6 wi h ac i a ion unc ion GeLU. ........................................................ 207
Table 23 - Pe o mance compa ison o di e en head numbe s o MHSA. ........................................ 208
Table 24 - Pe o mance compa ison o di e en High Hype pa ame e s wi h YOLO 5s6 De aul . ...... 209
xx
Table 25 - Pe o mance me ics compa ison be ween p oposed op imized YOLO 5s6-Hai iness and o he
models wi h da a Augmen a ion. ..................................................................................................... 212
Table 26 - k- old C oss Valida ion wi h k = 10 in op imized YOLO 5s6-Hai iness wi h Augmen a ion.. 214
xxi
ABBREVIATIONS
ADAM Adap i e Momen Es ima ion
ANN A i icial Neu al Ne wo k
C2F Coa se o Fine
CV Compu e Vision
DL Deep Lea ning
ELAN E icien Laye Agg ega ion Ne wo k
GELU Gaussian E o Linea Uni
HMI Human Machine In e ace
HSV Hue Sa u a ion Value
IP Image P ocessing
MAE Mean Absolu e E o
PLC P og ammable Logic Con olle
RELU Rec i ied Linea Uni
R-CNN Region Con olu ional Neu al Ne wo k
R-FCN Region Fully Con olu ional Ne wo k
RGB Red G een Blue
SCADA Supe iso y Con ol and Da a Acquisi ion
SGD S ochas ic G adien Descen
SILU Sigmoid Linea Uni s
SPPF Spa ial Py amid Pooling Fas
TEX Linea Mass o Ya n
YOLO You Only Look Once
1. INTRODUCTION
O e iew
This chap e p esen s he amewo k and impo ance o he ya n quali y analysis p oblem unde s udy.
The objec i es o he wo k a e highligh ed, along wi h he de elopmen me hodology and he scien i ic
con ibu ions o he wo k. Finally, he s uc u e o his hesis is p esen ed.
1.1 F amewo k
1.2 Objec i es
1.3 De elopmen me hodology
1.4 Main con ibu ions o he wo k
1.5 Thesis s uc u e
Chap e 1 –
In oduc ion
___________________________________________________________________________
8
To answe he esea ch ques ions, he me hod used was quan i a i e, since eadings,
measu emen s and compa isons o esul s we e ca ied ou ega ding he quali y and pe o mance o he
sys em de eloped h oughou he a ious esea ch s ages [14]. Da a collec ion occu ed h ough he
p oposed sys em de eloped. Thus, he quan i a i e me hod was conside ed he mos app op ia e because
i was in ended o de e mine and e alua e he con inuous imp o emen o he de eloped p oposed
sys em.
Based on an inno a i e esul , whe e he e was a ans e o knowledge o echnology, g ounded
on a eal con ex , an ac ion esea ch s a egy was adop ed [15]. To e alua e he pe o mance o he
sys em de eloped, quan i a i e da a was used, and compa isons we e made wi h exis ing sys ems on he
ma ke and e e ence sys ems in he ex ile indus y.
1.4 Main con ibu ions o he wo k
The scien i ic con ibu ions achie ed while de eloping his wo k a e:
1. De elopmen /op imiza ion o ya n cha ac e iza ion me hods:
• Inno a i e Me hodology: This wo k in oduces an inno a i e me hodology ha o e s
a mo e p ecise me hod o examining/de e mining de ec s in ya n, pa icula ly ocusing
on ya n hai iness.
• Deep Lea ning: Deep lea ning was in eg a ed in o he exis ing algo i hms o he
de ec ion and au oma ic classi ica ion o ya n hai iness ypes. This in eg a ion
subs an ially enhanced he accu acy and classi ica ion capabili ies o he algo i hms.
• Enhancemen s in Algo i hm Design: The de eloped algo i hm, YOLO 5s6 (You Only
Look Once – e sion 5, model s6) - Hai iness [16], in oduced se e al key
enhancemen s, including he in eg a ion o he C2 (coa se o ine) unc ion, a Bo -
T ans o me module, he adop ion o he GeLU (Gaussian E o Linea Uni ) ac i a ion
unc ion, and ine- uning o hype pa ame e s. These imp o emen s aimed o add ess
scale, con ex ual challenges, cap u e spa ial in o ma ion, unde s and complex
ela ionships wi hin images, and e ine he aining da a o be e pa e n cap u e.
• Pe o mance Imp o emen : Expe imen al esul s showcased an inc eased
pe o mance wi h he p oposed YOLO 5s6-Hai iness algo i hm. I ou pe o med he
o iginal YOLO 5, demons a ing a 5-6% imp o emen in he mAP0.5 (mean A e age

Chap e 1 –
In oduc ion
___________________________________________________________________________
9
P ecision a 0.5 In e sec ion o e Union (IoU)) me ic and an 11-12% inc ease in
mAP0.5:0.95 me ic. This pe o mance enhancemen es ablishes i as a mo e e icien
and high-pe o ming neu al ne wo k o ya n hai iness de ec ion, pa icula ly excelling in
de ec ing complex classes like loop ibe s.
2. Ya n image da ase :
• Da ase C ea ion: The da ase c ea ion p ocess was comple e and me iculous,
consis ing o 684 images speci ically cap u ed o ya n hai iness analysis. These images
we e anno a ed and classi ied in o wo ca ego ies: p o uding ibe s and loop ibe s. The
numbe o anno a ions (11037) and images (684) p o ides a ich esou ce o aining
and alida ing deep lea ning algo i hms o ya n de ec classi ica ion.
• Anno a ion and Classi ica ion Me hodology: U iliza ion o LabelME so wa e [17],
[18] in polygonal mode acili a ed p ecise anno a ion o loop ibe s (ma ked in g een)
and p o uding ibe s (ma ked in ed) in he da ase images. This allowed o me iculous
labeling, esul ing in an a e age o 16.1 anno a ions pe image. Addi ionally, he spli o
he da ase in o aining (70%), alida ion (20%), and es (10%) se s ensu es a obus
e alua ion o he model's pe o mance wi hou he use o augmen a ion echniques.
• Visualiza ion and S a is ical Analysis o he Da ase : The da ase cha ac e iza ion
p o ides comp ehensi e insigh s in o he dis ibu ion o anno a ions, bounding box
posi ions, sizes, and s a is ical in o ma ion ega ding loop and p o uding ibe s.
Visualiza ions o hai iness o e a clea unde s anding o he di e si y in bounding box
posi ions and sizes, essen ial o aining models o ecognize objec s wi h a ying
cha ac e is ics and placemen s.
• Imbalance Iden i ica ion and Add essing: The da ase analysis highligh ed an
imbalance be ween loop ibe s and p o uding ibe s. The complexi y o loop ibe s, hei
di e se appea ances, sizes, and occlusions con ibu e o he challenges in de ec ion.
This insigh can guide u u e esea ch o ocus on echniques ha speci ically add ess
he complexi ies associa ed wi h de ec ing loop ibe s.
• Da a Augmen a ion Techniques: To add ess he challenges iden i ied, da a
augmen a ion echniques we e employed using a ious such as lipping, sa u a ion
adjus men , blu , and noise addi ion we e applied o augmen he da ase , esul ing in
an expanded da ase o 1644 images. This augmen ed da ase p o ides a di e se ange
o image a ia ions o obus model aining and e alua ion.
Chap e 1 –
In oduc ion
___________________________________________________________________________
10
• Da ase as a Pla o m o Imp o emen s: This da ase , anno a ed wi h 11037
de ailed ma kings ac oss 684 images, p esen s a aluable esou ce. I no only acili a es
he de elopmen o mo e obus machine lea ning models speci ically designed o ya n
hai iness de ec ion bu also opens pa hs o ad ancing he ield o deep lea ning in ex ile
sciences. By p o iding clea anno a ions, s a is ical insigh s, and di e se augmen a ion
echniques, his da ase s ands as a esou ce o u u e esea ch in ex ile enginee ing
os e ing ad ancemen s in he analysis and unde s anding o ex ile de ec s.
3. De elopmen o algo i hms o image acquisi ion, p ocessing, and analysis:
• Inno a i e Image Cap u e: The algo i hms enable he cap u e o in-mo ion images
o he ya n p oduc ion p ocess, a capabili y no p e iously a ailable in exis ing
comme cial sys ems. This ad ancemen may p o ide a dynamic and con inuous iew o
he ya n p oduc ion p ocess.
• Ad anced Image P e-p ocessing: The algo i hms inco po a e ad anced p e-
p ocessing echniques o enhance he quali y o cap u ed images. This includes
dis o ion co ec ion, noise elimina ion, and con as imp o emen , ensu ing p ecise da a
collec ion o subsequen analysis.
• Segmen a ion and De ec ion: The algo i hms can p ecisely segmen speci ic a eas
o in e es in ya n images, such as he iden i ica ion and di e en ia ion o loop ibe s,
p o uding ibe s, o o he cha ac e is ics, enabling accu a e de ec ion and analysis o
hese a eas.
• Spec al Analysis: Pe o ms spec al analysis based on ya n linea mass a ia ion and
hai iness da a, u ilizing he Fas Fou ie T ans o m (FFT) o iden i y pe iodical
i egula i ies in ya n. Resul s a e a ailable in he use in e ace.
The echnological con ibu ions achie ed while de eloping his wo k a e:
1. De elopmen o a mecha onic de ice:
• De elopmen o an Au oma ic Unwinding and Winding Sys em: This sys em
in oduces an inno a i e app oach by c ea ing an au oma ic unwinding and winding
sys em aimed a e i ying ya n quali y h ough image p ocessing. This non-des uc i e
p o o ype allows o he ya n winding and unwinding wi hou damaging he ya n/bobbins.
I has he abili y o accu a ely mo e he ya n in on o he came a lens, minimizing
Chap e 1 –
In oduc ion
___________________________________________________________________________
11
ib a ions and suppo ing di e en ypes o ya n and bobbins wi hou causing any
des uc i e e ec s o hem.
• Enhanced Ve sa ili y and Po abili y: Unlike exis ing comme cial ya n quali y es ing
machines ha a e la ge, complex, and expensi e, non-des uc i e p o o ype o e s
enhanced e sa ili y and po abili y (weighs less han 30 kg and i s wi hin a compac
box o less han 0.27 m³). I s design acili a es he es ing o a ious ypes o ya ns and
bobbins wi h diame e s up o 160mm (ob ained in p ojec pe o mance speci ica ions),
enabling ex ile companies o conduc on-si e analysis wi hou elying on ex e nal
labo a o ies.
• Moni o ing and Da a S o age: The sys em p o ides an in ui i e local Human-
Machine In e ace (HMI) ha allows use s o con ol ya n quan i y and eloci y. I enables
moni o ing and s o age o da a, ensu ing he access o s a is ical in o ma ion and
assu ing use sa e y and sys em secu i y.
• Image Analysis o Mo e Pa ame e s: Th ough he applica ion o image analysis
echniques, he non-des uc i e p o o ype allows he analysis o mo e pa ame e s
compa ed o exis ing comme cial al e na i es. Addi ionally, he p oposed sys em
cap u es images and ideos in mo ion and collec s pa ame e s like numbe o cables,
wis s ep, wis o ien a ion, and loose ibe s – ea u es which a e no p esen in he
exis ing sys ems.
• Inco po a ion o Mecha onic P inciples: The design and unc ioning o he sys em
in ol es mecha onic p inciples, combining mechanical, au oma ion, and elec onic
enginee ing aspec s o suppo he analysis o ya n quali y. I in eg a es a con ol sys em
ha synch onizes ya n mo emen , main aining s abili y, and ensu ing accu acy du ing
he unwinding and winding p ocesses.
• Imp o ed Sys em Con ol and Use In e ac ion: The con olle a chi ec u e o he
sys em inco po a es a P og ammable Logic Con olle (PLC) o sys em con ol o p o ide
e icien and use - iendly ope a ion. Addi ionally, i includes a Supe iso y Con ol and
Da a Acquisi ion (SCADA) sys em wi h a Local Human Machine In e ace (HMI) o
moni o ing, sys em a iable con ol, and immedia e no i ica ion o signi ican e en s,
such as eme gency, se omechanism anomaly, ya n b eak/absence.
Chap e 1 –
In oduc ion
___________________________________________________________________________
12
2. Enhanced Ya n Analysis Capabili ies:
• Acquisi ion o Va ious Ya n Pa ame e s: Facili a es acquisi ion o a ious ya n
pa ame e s like diame e , linea mass, olume, wis di ec ion, wis s ep, numbe o
cables, hai iness index, loose ibe s, hin places (-50%), hick places (+50%), neps
(+200%), U (%), CV (%), and sH (%) – e en in mo ion, a ea u e absen in exis ing
comme cial sys ems.
• Cus omizable Analysis: Allows use s o ailo analysis pa ame e s (ya n ma e ial,
dimensional ole ance, numbe o ya n sec ions).
• Da a Collec ion and S o age: Collec s and s o es comp ehensi e da a o u u e
s a is ical p ocessing.
• Image and Video Visualiza ion: Allows isualiza ion o bo h images and ideos unde
analysis.
• Use Moni o ing o Image P ocessing Ope a ions: P o ides use moni o ing o
image p ocessing ope a ions.
3. De elopmen o an applica ion o epo gene a ion:
The p oposed sys em includes an applica ion ha p o ides s a is ical da a and gene a es
comp ehensi e epo s. This applica ion plays a undamen al ole bo h in p ac ical use
by he use and o u u e use in ex ile companies, o e ing access o s a is ical
in o ma ion and acili a ing in o med decision-making.
These unc ionali ies enhance he usabili y o he sys em, suppo ing e icien quali y con ol wi h
he po en ial in op imizing he p oduc ion p ocesses wi hin he ex ile indus y.
In summa y, scien i ic con ibu ions ocus on deep lea ning algo i hms (YOLO 5s6-Hai iness) and
he build o a. ya n image da ase . Technological con ibu ions ocus on he de elopmen o a mecha onic
de ice, and p ac ical applica ions ha enable he e ec i e implemen a ion o hese echnologies aiming
o imp o e he quali y o he inal p oduc in he ex ile indus y. Bo h ypes o con ibu ions can ha e a
signi ican impac in he indus y.
Chap e 1 –
In oduc ion
___________________________________________________________________________
13
1.5 Thesis S uc u e
This hesis is s uc u ed in o eigh chap e s. The second chap e , Theo e ical Concep s, p esen s
he ex ile heo e ical concep s and he deep lea ning concep s o unde s and he wo k.
The hi d chap e , Li e a u e Re iew, encompasses a li e a u e e iew in which a su ey o esea ch
p ojec s cen e ed on he analysis o ya n is conduc ed. This chap e is subdi ided in o h ee pa s, he
i s p esen ing a s a e o he a o sys ems ha use image p ocessing, compu e ision and a i icial
in elligence, he second pa p esen ing a compa ison be ween he a ious sys ems and he hi d pa a
c i ical e iew analysis.
The ou h chap e , Image P ocessing and Compu e Vision Sys em o Ya n Analysis, p esen s he
echniques used in he de eloped sys em ega ding image p ocessing and compu e ision o ya n
analysis. Fu he mo e, he chap e del es in o he ha dwa e employed o image acquisi ion, alongside a
desc ip ion o he in e ace so wa e de eloped.
The i h chap e , Deep Lea ning Techniques o Ya n Hai iness Analysis, desc ibes he deep
lea ning echniques used o analyzing and classi ying ya n hai iness. To his end, he imp o ed
a chi ec u e o he YOLO 5s6 neu al ne wo k is p esen ed, including a speci ic discussion o he
composi ion o he modules wi h an emphasis on he modi ied blocks.
The six h chap e , Mecha onic P o o ype De elopmen , p esen s he design o he mecha onic
p o o ype and he de ails o each pa ha cons i u es i , as well as he es s and esul s ob ained.
In he se en h chap e , Expe imen al Resul s and Discussion, he expe imen al e i ica ion is
p esen ed, and he esul s ob ained a e discussed, including a de ailed analysis o he c ea ed da ase
and he e ec s o he changes implemen ed in he p oposed neu al ne wo k.
The eigh h chap e , Conclusions and Fu u e Wo k, p esen s he main conclusions o he s udy, as
well as some p oposals o u u e s eps, wi h he aim o con inuing he wo k, which may allow new lines
o in es iga ion and/o possible echnological de elopmen s.
Summa y o he chap e :
This chap e p o ided an o e iew o he esea ch de eloped, as well as i s ele ance o socie y,
which se ed as he mo i a ion behind his wo k. In his chap e , he me hodology ha aligns wi h he

Chap e 1 –
In oduc ion
___________________________________________________________________________
14
esea ch ques ions was p esen ed, ollowing a posi i is philosophy wi h a deduc i e app oach. Roo ed in
an inno a i e ou come wi h knowledge ans e o echnology wi hin a eal-wo ld con ex , an ac ion
esea ch s a egy is adop ed. This s a egy in ol es an ac i e dynamic, aimed a achie ing he desi ed
p o o ype, and quan i a i e da a was used o assess he pe o mance o he sys em. In his chap e , he
amewo k, objec i es, de elopmen me hodology, p ima y scien i ic/ echnological con ibu ions o he
wo k, and he s uc u e o he hesis we e p esen ed.
2. THEORETICAL CONCEPTS
O e iew
In his chap e , an o e iew o ya n cha ac e iza ion pa ame e s is p esen ed, p o iding a heo e ical
ounda ion o unde s and he de eloped sys em. Addi ionally, a i icial in elligence concep s and he deep
lea ning p inciples necessa y o comp ehending hei applica ion wi hin he wo k a e discussed.
2.1 Ya n Cha ac e iza ion
2.2 S a is ical Pa ame e s o Ya n
2.3 Spec al Analysis
2.4 Image P ocessing & Compu e Vision
2.5 Deep Lea ning
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
16
2.1 Ya n cha ac e iza ion
This subchap e co e s a ious aspec s o ya n p ope ies, including linea mass, diame e , speci ic
olume, hai iness, wis o ien a ion and s ep, de ec s, as well as se e al s a is ical pa ame e s [19], [20],
[21], [22].
2.1.1 Linea mass and diame e
The connec ion be ween he diame e o a ya n and i s linea mass ep esen s a c ucial basis in
ya n analysis, o e ing a simpli ied way o cha ac e iza ion. When he ya n assumes a cylind ical o m, i
is geome ically es ablished ha he ya n diame e is di ec ly ela ed o he squa e oo o i s linea mass
[19], [20], [21], [22], [23]. Figu e 2 p o ides a isual ep esen a ion o a ya n con igu a ion,
demons a ing he obse able co ela ion be ween hese wo pa ame e s.
Figu e 2 - Ya n con igu a ion [19].
The linea mass is exp essed in ex (g/km), and can be de e mined by Equa ion 1 [19], [20], [21],
[22], [23]:
𝑡𝑒𝑥= 𝜙× 𝜌 × 𝜋 (𝑑
2)2×105
↔𝑑= √4 × 𝑡𝑒𝑥/(𝜋× 105×𝜙×𝜌)
(1)
Whe e:
• d is he ya n diame e (cm);
• ex is he linea mass o he ya n (g/km);
• ϕ is he po osi y;
• ρ is he densi y o he ya n ma e ial (exp essed in g/cm3).
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
17
2.1.2 Speci ic olume
The a angemen o he ibe s is e lec ed in he speci ic olume o he ya n, which is de e mined
by he olume occupied by he ibe s and he in e s i ial spaces be ween hem, as depic ed in Equa ion
2 [19], [20], [21], [22], [23]:
𝑉𝑡=𝜋 𝑅2
𝑁 10−5 (2)
whe e:
• R is he ya n adius (mm);
• N is he linea mass o he ya n (g/km);
• V is he speci ic olume (cm3/g).
2.1.3 Hai iness
Hai iness occu s when ce ain ibe s ex end ou wa d om he main body o he ya n, esul ing in
some ibe s o ming loops and/o p o uding om he ya n co e. This p ope y signi ican ly impac s he
appea ance, ex u e, and usabili y o he ex ile, making i a c i ical ac o o he inal p oduc quali y.
De ec ing and measu ing his pa ame e equi es a complex analysis and gene ally canno be de ined by
a single indica o . Typically, i is assessed by conside ing he numbe o ibe s ex ending beyond he ya n
co e pe uni leng h. Fu he mo e, measu ing hai iness allows o he quan i ica ion o a ious s a is ical
pa ame e s essen ial o cha ac e izing ya n quali y and, consequen ly, he ab ics quali y. These
pa ame e s include, bu a e no limi ed o, he s anda d de ia ion o hai iness, hai iness index, and
hai iness mean de ia ion. Figu e 3 p o ides a isual ep esen a ion o hese ibe s along he ya n, which
can be u he ca ego ized in o p o uding ibe s and loop ibe s (hai iness wound a ound he ya n body)
[19], [20], [21], [22].
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
24
Figu e 9 – Example o a complex colo in he RGB scale [32].
The HSV scale is also an addi i e polych oma ic scale whe e pixels a e g ouped in ios. Howe e ,
unlike RGB, each pixel ep esen s one o h ee colo p ope ies: hue, sa u a ion, and b igh ness [33],
[34], [35].
• Hue: alue de e mines he ype o colo being ep esen ed (g een, ed, o ange, pu ple, e c.).
• Sa u a ion: indica es how i id o in ense he colo will be. Fo ins ance, a sa u a ion alue o
255 wi h a ed hue ep esen s a ib an and p onounced ed, while a alue o 50 displays a
mo e mu ed ed.
• Value: de e mines how ligh o da k he ep esen ed colo will be. A high alue, o example,
ep esen s a ligh e shade o ed, while a lowe alue indica es a da ke ed [33], [34], [35].
Figu e 10 p o ides a isual ep esen a ion o his scale.
Figu e 10 – Visual ep esen a ion o he HSV scale [34].

Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
25
2.4.2 Logic Ope a ions
To begin analyzing image p ocessing ope a ions, one o he s a ing poin s can be logical
ope a ions. These ope a ions aim o compa e wo images o equal size and subsequen ly combine hem
in o a single image. The e a e wo main g oups o logical ope a ions: a i hme ic ope a ions and
compa ison ope a ions.
A i hme ic ope a ions aim o c ea e a esul ing image ha is he esul o adding o sub ac ing he
elemen s p esen in wo ini ial images. These ope a ions in ol e pai ing he co esponding pixels in he
wo images ha occupy he same posi ion, and hese pixel pai s a e hen summed o sub ac ed o ob ain
he inal image [36]. Figu e 11 p o ides examples o sub ac ion and addi ion, espec i ely.
Figu e 11 – Examples o image sub ac ion and addi ion [36].
These ope a ions can se e a ious pu poses, bu he mos common use is o enhance o elimina e
speci ic aspec s o an image. Fo ins ance, when iden i ying changes be ween wo images in a ideo, a
commonly used p ocess in ol es sub ac ing hese wo images. This esul s in an image ha highligh s
only he di e ences be ween he pai o images, as exempli ied in Figu e 12.
Figu e 12 – Example o using sub ac ion o highligh changes [36].
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
26
Compa ison ope a ions, on he o he hand, aim o compa e aspec s be ween a pai o images and
gene a e a esul ing image ha combines he wo o iginals in a ious ways. These ope a ions a e mos ly
used wi h bina y pixel alue images, i.e., black (1) and whi e (0). The mos common ypes o compa isons
a e:
• "AND" Compa ison: Compa es he wo images and c ea es an image whe e he only pixels
wi h a alue o 1 a e hose ha ha e his alue in bo h images, assigning a alue o 0 o hose
ha ha e i in only one o he images (Figu e 13) [36].
Figu e 13 – "AND" compa ison demons a ion [36].
• "OR" Compa ison: Compa es he wo images and c ea es an image whe e he pixels wi h a
alue o 1 a e hose ha ha e his same alue in a leas one o he o iginal images (Figu e 14)
[36].
Figu e 14 – "OR" compa ison demons a ion [36].
• "XOR" Compa ison: Pe o ms an ope a ion simila o he "OR" compa ison; howe e , all pixels
wi h a alue o 1 in bo h compa ed images will ha e a alue o 0 in he esul ing image (Figu e
15) [36].
Figu e 15 – "XOR" compa ison demons a ion [36].
The e is also an ope a ion ha , while no belonging o he compa ison g oup, is commonly used
in conjunc ion wi h hem. This ope a ion is in e sion o "NOT," whe e all pixels in an image ake he
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
27
opposi e alue, i.e., all pixels wi h a alue o 1 become 0, and ice e sa, as demons a ed in Figu e 16
[36].
Figu e 16 – Demons a ion o he in e sion ope a ion [36].
2.4.3 “Th esholding” Ope a ion
One o he mos e sa ile g oups o p ocesses in image p ocessing ope a ions is he "Th esholding"
g oup. I s objec i e is o sepa a e a ious de ails in he image o acili a e hei isola ion o enhance hei
p ominence.
The simples p ocesses wi hin his g oup a e bina y h esholding me hods, which in ol e di iding
all he pixels in he image in o wo classes. A pixel alue h eshold is used o c ea e hese classes. Pixels
wi h alues g ea e han he h eshold a e g ouped in o one class, while hose wi h alues below he
h eshold a e g ouped in o ano he [36], [37].
In egula bina y h esholding (Equa ion 10), pixels in he class wi h alues abo e he h eshold a e
assigned he maximum alue o 255, while pixels in he class wi h alues below he h eshold a e
assigned he minimum alue o 0. In he case o in e se bina y h esholding (Equa ion 11), he same
p ocess occu s, bu he classes a e swapped. In o he wo ds, pixels in he class wi h alues abo e he
h eshold a e assigned he minimum alue [36], [37].
𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)= {255 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙 (𝑥,𝑦)>𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑
0 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)≤𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (10)
𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)= {255 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙 (𝑥,𝑦)≤𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑
0 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)>𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (11)
In addi ion o hese p ocesses, he e a e also "Th esholding o ze o" p ocesses and hei in e ses.
These p ocesses unc ion simila ly o bina y h esholding; howe e , only one o he classes is a ec ed.
Fo example, in he egula "Th esholding o ze o" case (Equa ion 12), only he class con aining
alues equal o o less han he h eshold alue is modi ied, assigning all i s pixels he minimum alue o
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
28
0. The opposi e occu s in he case o in e se "Th esholding o ze o" (Equa ion 13), whe e only pixels in
he class abo e he h eshold alue a e a ec ed, assigning hem he maximum alue o 255 [36], [37].
𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)= {𝑉𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦) 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙 (𝑥,𝑦)>𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑
0 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)≤𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (12)
𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)= { 0 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙 (𝑥,𝑦)>𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑
𝑉𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦) 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)≤𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (13)
The las o he bina y h esholding p ocesses is unca ion (Equa ion 14). I ac s in a e y simila
way o in e se "Th esholding o ze o"; howe e , he alues assigned o he pixels in he a ec ed class a e
he h eshold alue i sel , a he han he maximum alue. Figu e 17 p o ides a isualiza ion o he e ec
o each ype o h esholding desc ibed he e [36], [37].
𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)= { 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙 (𝑥,𝑦)>𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑
𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦) 𝑖𝑓 𝑣𝑎𝑙 𝑝𝑖𝑥𝑒𝑙(𝑥,𝑦)≤𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (14)
Figu e 17 – Visual ep esen a ion o all ypes o bina y h esholding [38].
All o hese p ocesses belong o he g oup o bina y h esholding, as hey sepa a e he pixels in he
image in o wo classes, speci ically o g ayscale images. Howe e , he e a e mo e ad anced p ocesses
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
29
ha can isola e pixels in an image in o 3 o mo e classes, which ob iously in ol es a g ea e numbe o
h eshold alues and a highe le el o sepa a ion and ea u e enhancemen in he image. I is also possible
o pe o m hese p ocesses on images in o he colo spec ums, such as RGB o HSV, which can no only
sepa a e classes by colo bu also by b igh ness and sa u a ion [36], [37]. These p ocesses a e
conside ably mo e complex and will no be u he explo ed he e as hey a e beyond he scope o his
wo k.
2.4.4 OTSU me hod
The OTSU me hod is commonly e e ed o as ano he ype o h esholding; howe e , i is mo e
accu a e o s a e ha i is a me hod ha can be associa ed wi h any o he o he h esholding p ocesses.
In o he wo ds, when he OTSU me hod is combined wi h, o example, a bina y h esholding p ocess, i
assis s in de e mining, h ough ma hema ical algo i hm me hodology, he bes h eshold alue choice
[39], [40].
The way he algo i hm ope a es is by es ing all possible alues o he h eshold, seeking he one
ha yields he maximum a iance be ween classes (in his case, classes e e o he image backg ound
and subjec ). This a iance can be calcula ed using equa ion 15 [39], [40]:
𝜎𝑊
2= 𝑊𝑠×𝜎𝑠2+𝑊𝑓×𝜎𝑓2 (15)
Whe e:
• W co esponds o he ela i e weigh o each class, i.e., he p obabili y ha any pixel belongs
o he espec i e class. This means ha he g ea e he numbe o pixels belonging o each
class, he highe hei ela i e weigh .
• s and co espond o he nomencla u e o he subjec and backg ound classes, espec i ely.
Figu e 18 p o ides a isualiza ion o he applica ion o his me hod o an image.

Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
30
Figu e 18 – G aphical ep esen a ion o he OTSU me hod
[39].
2.4.5 Linea Fil e s
Ano he commonly used p ocess in image p ocessing me hodologies is image ans o ma ion using
linea il e s [36], [41], [42]. This p ocess di e s om he p e ious ones in ha he inal alue o a pixel
depends no only on i s ini ial alue bu also on he ini ial alues o i s neighbo s. I is a pa icula ly use ul
p ocess o de ec ing con ou s, poin s, geome ic shapes, and objec s in gene al in an image. Fo
example, i is a p ocess ha can be used o de ec he p esence o a pe son in an image. I can also be
used o apply a ious il e s o he image, such as blu ing (Figu e 19) o adding noise.
Figu e 19 – Example o applying a linea il e o blu ing
[36].
Be o e explaining he p ocess by which he ans o ma ion occu s, i is impo an o cla i y he e m
"linea il e s" in his con ex . A linea il e e e s o a squa e ma ix o o de 3 o highe . Fo easons ha
will be men ioned la e , he o de o a mask ma ix mus always be an odd numbe . Figu e 20 shows an
example o a simple linea il e .
1
9[1 1 1
1 1 1
1 1 1]
Figu e 20 – Example o a linea il e ma ix.
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
31
This ans o ma ion begins by o e laying he il e ma ix on o he pixels o he image, as illus a ed
in Figu e 21.
Figu e 21 – O e laying a il e on image pixels [36].
Nex , a mul iplica ion ope a ion is pe o med be ween he ma ices. Howe e , his ope a ion is no
a ypical ma hema ical ma ix mul iplica ion. The ope a ion in ol es mul iplying each e m in he il e
ma ix by he e m in he same posi ion in he image ma ix. Fo example, ini ially, he e ms a posi ion
(1, 1) in he il e ma ix a e mul iplied by he e m a posi ion (1, 1) in he image ma ix. This p ocess is
epea ed o all e ms in he il e ma ix. A e execu ing his ope a ion, a new ma ix is ob ained. Figu e
22 ep esen s he new ma ix o he example o he p e ious image. I is impo an o no e ha , when
dealing wi h pixel alues, decimal numbe s a e always ounded. In his con ex , only in ege s be ween 0
and 255 a e used.
[0 0 0
0 0 0
0 0 3]
Figu e 22 – The esul ing ma ix om he mul iplica ion o ma ices.
Nex , he sum o all he e ms in he esul ing ma ix is calcula ed. The esul is he inal alue o
he cen al pixel in he image ma ix, as shown in Figu e 23.
Figu e 23 – Replacemen o he inal alue in he ma ix [36].
This p ocess o applica ion o he linea il e , can also be desc ibed by Equa ion 16.
𝑁𝑒𝑤[𝑖,𝑗]= ∑ ∑ 𝑂𝑟𝑖𝑔𝑖𝑛𝑎𝑙[𝑖+𝑘,𝑗+𝑙]×𝐹𝑖𝑙𝑡𝑒𝑟[𝑘,𝑙]
𝑙𝑘 (16)
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
32
This en i e p ocess is epea ed as many imes as he il e i s in o he o iginal image, always wi h
a di e en pixel in he cen al posi ion o he o e lapping ma ices. The ac ha his p ocess equi es he
exis ence o a cen al e m (a e m a he geome ic cen e o he ma ix) in he esul ing ma ix means
ha he linea il e ma ix mus ha e a cen al e m. This equi emen necessi a es ha he linea il e
be a squa e ma ix o odd o de , as e en-o de ma ices do no ha e a single cen al e m. One o he
mos common and e sa ile linea il e s is he Gaussian il e . I s pu pose is o educe noise in an image
by applying a sligh blu o he image. Applying a Gaussian il e is one o he i s s eps in many image
p ocessing me hodologies. Al hough coun e in ui i e, blu ing he image can esul in a clea e inal image
because emo ing ini ial noise leads o a clea e inal image. I is impo an o no e ha he la ge he
o de o he Gaussian il e applied, he g ea e he e ec on he image. Figu e 24 p o ides an example
o he applica ion o a Gaussian il e . The con as o his ype o blu is compa ed wi h ha o Figu e 19.
Figu e 24 – Applica ion o a Gaussian il e [36].
The e m "Gaussian il e " does no e e o a single linea il e bu a he o a ype o il e ha
ollows he same ule. These il e s a e named a e he no mal dis ibu ion in wo dimensions. Figu e 25
gi es an example o a 7x7 o de Gaussian il e ma ix.
Figu e 25 – Example o a 7x7 Gaussian ma ix [36].
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
33
The second ype o il e s o discuss a e line de ec ion il e s. Thei goal is o iden i y he p esence
o lines wi h he desi ed o ien a ion. They a e widely used in applica ions ha seek o iden i y objec s in
an image. The il e ma ix used a emp s o ma ch he shape o he objec o be ound. Figu e 26 p o ides
an example o a e ical line de ec ion il e as well as how i ans o ms he o iginal image.
Figu e 26 – Example o applying a e ical line de ec ion il e [36].
The example shown is a simple applica ion o hese il e s. Mo e commonly, mul iple il e s a e
used o simul aneously de ec e ical, ho izon al, and diagonal lines, e ealing he p esence o mo e
complex shapes. Figu e 27 demons a es he applica ion o h ee il e s o iden i y a complex objec .
Figu e 27 – Using mul iple il e s o de ec a complex shape [36].
Fo he las example, a il e ma ix o o de 3 was used, bu a highe le el o de ail can be achie ed
by inc easing he o de o he il e ma ix. Figu e 28 shows he esul o pe o ming he same ope a ion
bu wi h a 5x5 o de il e .
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
40
A ‘da ase ’, in b oad e ms, e e s o a collec ion o da a ha has been ga he ed and o ganized in
some way. These da a can ep esen in o ma ion abou a a ie y o phenomena o objec s, such as
images, ex , numbe s, o any o he o m o measu able in o ma ion. 'Da ase s' a e essen ial o eeding
machine lea ning algo i hms and enabling hem o acqui e knowledge and make in o med decisions
based on he a ailable da a. Be o e del ing in o he a ious ypes o machine lea ning, i is impo an o
unde s and ha 'da ase s' play a pi o al ole as hey se e as he aw ma e ial upon which algo i hms
lea n and de elop hei capabili ies [46], [47], [48].
F om his poin onwa d, he di e en ypes o machine lea ning a e in oduced, including
supe ised lea ning, unsupe ised lea ning, and ein o cemen lea ning, and p o ide an explana ion o
how hese algo i hms u ilize da ase s in speci ic ways o achie e hei espec i e objec i es [46], [47],
[48].
Figu e 39 – Subse s o a i icial in elligence [46].
• Supe ised lea ning: When he algo i hm has a labeled da ase (da a wi h hei espec i e co ec
ou pu s) and lea ns by compa ing he model ou pu wi h he expec ed ou pu , and i s pa ame e s
a e eadjus ed, i necessa y, un il eaching an accep able h eshold and p ede e mined. Figu e
40 shows he p ocesses in de ail.
Figu e 40 – Supe ised lea ning. S eps: (1) made anno a ions in he da ase ; (2) da ase sepa a ion; (3)

Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
41
algo i hm aining; (4) da ase es o algo i hm e alua ion and pe o mance [48].
Supe ised lea ning can be u he ca ego ized in o wo main ypes: classi ica ion and eg ession.
Classi ica ion in ol es p edic ing a disc e e alue whe e he algo i hm iden i ies inpu da a as belonging
o a speci ic class ( o example, de e mining whe he a pho o con ains a p e-iden i ied animal). The
algo i hm's pe o mance is e alua ed based on how accu a ely i can co ec ly classi y new images
acco ding o he p o ided anno a ions. Reg ession, on he o he hand, deals wi h con inuous da a [46],
[47], [48].
• Unsupe ised lea ning ope a es wi hou p io anno a ions on he da a. In his ca ego y, he
algo i hm is p o ided wi h a da ase lacking anno a ions and seeks o disco e simila i ies among
g oups o da a.
Wi hin unsupe ised lea ning, wo p ima y asks exis : da a clus e ing and in o ma ion ex ac ion.
Da a clus e ing, also known as clus e ing algo i hms (Figu e 41), in ol es g ouping da a based on hei
simila i ies. In o ma ion ex ac ion (Figu e 42), on he o he hand, en ails he algo i hm associa ing new
in o ma ion wi h p e iously ob ained da a, such as p o iding mo ie ecommenda ions on a websi e based
on a use 's pas iewing his o y [46], [47], [48].
Figu e 41 – Clus e ing algo i hm [49].
Figu e 42 – Unsupe ised lea ning–In o ma ion ex ac ion [50].
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
42
• Rein o cemen lea ning in ol es aining machine lea ning models o make a se ies o decisions
based on a ewa d/punishmen amewo k. In essence, his echnique aims o accomplish a
speci ic ask, ini ially employing a ial-and-e o app oach. Subsequen ly, he ou comes o each
a emp , ega dless o success, a e u ilized o aining by means o a ewa d/punishmen
sys em, as depic ed in Figu e 43 [50].
Figu e 43 – Unsupe ised lea ning–In o ma ion ex ac ion [50].
As p e iously e e ed a da ase is a collec ion o da a used as he ounda ion o de eloping a i icial
in elligence algo i hms o a ious o he ypes o da a science p ojec s. Typically, c ea ing a da ase is a
c i ical and ime-in ensi e phase ha signi ican ly impac s he pe o mance o he sys em. I is o en
di ided in o h ee subse s: aining, alida ion, and es da ase s, as illus a ed in Figu e 44 [50].
Figu e 44 - Unsupe ised lea ning—in o ma ion ex ac ion [50].
• T aining da ase : This se consis s o samples used o model de elopmen and con e gence. I
se es as he basis o aining he algo i hm and enabling i o lea n om he p o ided da a.
• Valida ion da ase : In his phase, samples om he alida ion da ase a e employed o objec i ely
assess a model's pe o mance a e he aining p ocess. This s age also allows o ine- uning
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
43
he algo i hm's hype pa ame e s, which can lead o adjus men s and imp o emen s in he
aining phase.
• Tes da ase : The es da ase comp ises samples used o igo ously es ing he algo i hm. I
p o ides a inal e alua ion o he model's pe o mance. I is c ucial ha he samples in he es
da ase a e no used in ei he he aining o alida ion phases o main ain he in eg i y o he
e alua ion esul s.
2.5.2 Neu al Ne wo ks o Objec De ec ion
A signi ican ad ancemen has been made in he ield o compu e ision. Thus a , he eme gence
o algo i hms designed o image classi ica ion has been wi nessed, enabling he disce nmen o he
p esence o objec s such as dogs, ca s, ca s, o bicycles wi hin an image. Howe e , he nex on ie in
his domain is objec de ec ion. Dis inc om classi ica ion algo i hms, objec de ec ion algo i hms aim o
p ecisely loca e objec s o in e es wi hin an image by d awing bounding boxes ( e e ed o as objec
de ec o s) o bounding cu es (ins an segmen a ion) a ound hem [51].
No ably, a single image can con ain mul iple dis inc objec s equi ing de ec ion. One possible
solu ion would in ol e pa i ioning he image in o a ious egions o in e es and employing con olu ional
neu al ne wo ks o classi y he p esence o objec s wi hin hese egions. None heless, his app oach
p esen s challenges, pa icula ly when objec s o in e es exhibi a ying spa ial loca ions and p opo ions
wi hin he image. Consequen ly, selec ing nume ous egions may lead o excessi e compu a ional cos s,
especially in scena ios such as highway su eillance whe e nume ous ehicles need o be de ec ed [51].
To add ess hese issues, a amily o algo i hms was de eloped, including R-CNN (Region-based
Con olu ional Neu al Ne wo k), Fas R-CNN (imp o emen o e R-CNN), Fas e R-CNN (ex ension o Fas
R-CNN), R-FCN (Region-based Fully Con olu ional Ne wo ks), YOLO (You Only Look Once), and o he s
[51].
YOLO
YOLO (You Only Look Once) was in oduced by Joseph Redmon and Ali Fa hadi in 2015 du ing
hei doc o al s udies [52]. Since i s ini ial elease, YOLO has unde gone se e al i e a ions, wi h he la es
e sion being YOLO 8. YOLO ep esen s a depa u e om he egion-based algo i hms discussed ea lie .
I is a con olu ional ne wo k ha p edic s bo h bounding boxes and class p obabili ies o hese boxes.
YOLO di ides he image in o N g ids, each o which has a uni o m SxS dimension. Each g id is
esponsible o de ec ing and pinpoin ing he objec s i encompasses. These g ids p edic he coo dina es
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
44
o bounding boxes ( ela i e o cell coo dina es), as well as he objec label and he p obabili y o i s
p esence wi hin he cell. By ha ing mul iple cells p edic he same objec wi h a ying bounding box
p edic ions, YOLO signi ican ly educes compu a ional equi emen s, as bo h de ec ion and ecogni ion
asks a e e icien ly handled wi hin he image (see Figu e 45).
Figu e 45 - Models he de ec ion as a eg ession p oblem [52].
In Figu e 45, he image is di ided in o an S×S g id, and o each cell o he g id, i p edic s bounding
boxes (B), con idence sco es o hose boxes ( hicke lines indica e highe con idence), and class
p obabili ies (C). These p edic ions a e encoded as a enso o dimensions S×S×(B×5+C) [52], [53].
The YOLO a chi ec u es consis o h ee dis inc componen s, backbone, neck, and head, as
depic ed in Figu e 46.
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
45
Figu e 46 - YOLO a chi ec u e and how i wo ks [53].
Howe e , his p ocess o en esul s in nume ous duplica e p edic ions. To add ess his issue, YOLO
u ilizes he Non-Maximum Supp ession (NMS) echnique [54]. NMS ope a es by elimina ing edundan
bounding boxes wi h lowe p obabili y sco es. YOLO achie es his by e alua ing he p obabili y sco es
associa ed wi h each op ion and selec ing he one wi h he highes sco e as he p ima y choice.
Subsequen ly, i supp esses bounding boxes ha exhibi high In e sec ion o e Union (IoU) wi h he
cu en ly selec ed high p obabili y bounding box (as illus a ed in Figu e 47). This p ocess con inues un il
he e a e no mo e boxes o be emo ed, as demons a ed in Figu e 48.

Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
46
Figu e 47 - In e sec ion o e Union—B1: P edica ion Box and B2: De ec ion Box [55].
Figu e 48 - Non-Maximum Supp ession P ocess example [54].
YOLO s ands ou o i s excep ional speed, achie ing an imp essi e a e o 45 ames pe second,
su passing he pe o mance o o he objec de ec ion algo i hms. Howe e , i is wo h no ing ha YOLO
does ha e a limi a ion in e ms o de ec ing excep ionally small objec s in images, p ima ily s emming
om he inhe en spa ial cons ain s o he algo i hm [54], [55].
Da a Augmen a ion
Da a augmen a ion in YOLO 5 in ol es he p ac ice o gene a ing addi ional aining da a by
applying a ious ans o ma ions o he o iginal images. The p ima y objec i e is o enhance he di e si y
wi hin he aining da ase , he eby imp o ing he model's capaci y o gene alize e ec i ely o new and
unseen da a [16].
Wi hin YOLO 5, da a augmen a ion encompasses a ange o echniques, including andom
c opping, o a ion, lipping, and colo ji e ing, all o which a e employed o c ea e no el aining images.
These ans o ma ions a e andomly applied o he o iginal images du ing he aining p ocess, expanding
he da ase and in oducing a b oade a ie y o images o he model o lea n om [16].
As illus a ed in Figu e 49, you can obse e an example o augmen ed aining images in YOLO 5.
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
47
Figu e 49 - Example o augmen a ion [56].
Th ough da a augmen a ion, he model gains p o iciency in de ec ing objec s unde a ious
condi ions, encompassing di e en o ien a ions, scales, and ligh ing scena ios. This con ibu es o he
enhancemen o he model's accu acy and obus ness [16].
The da a loade in YOLO 5 execu es h ee ca ego ies o augmen a ions [16]:
• Basic Augmen a ions: This ca ego y includes andom c opping and esizing o images, ho izon al
lipping, and andom colo dis o ion. These echniques aim o enhance he model's gene aliza ion
capabili ies;
• Mosaic Augmen a ions: This ca ego y combines ou indi idual images in o a single mosaic image.
This helps he model lea n o de ec objec s ha may appea in p oximi y wi hin a single image;
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
48
• Au o Augmen a ions: This ca ego y employs ein o cemen lea ning o sea ch o he op imal se o
augmen a ions ailo ed o a speci ic da ase . This op imiza ion con ibu es o imp o ed model
pe o mance on ha pa icula da ase .
In addi ion o he abo e, YOLO 5 employs colo space adjus men s as a o m o da a augmen a ion.
This in ol es modi ying he colo balance, b igh ness, con as , and sa u a ion o aining da ase images.
This adap a ion enhances he model's obus ness o a ia ions in eal-wo ld ligh ing condi ions and colo
schemes. The da a loade in YOLO 5 implemen s colo space adjus men s by andomly al e ing he hue,
sa u a ion, and b igh ness o each image du ing aining.
I is impo an o no e ha augmen a ion does no gene a e en i ely new da a; a he , i p esen s
exis ing da a in di e se, andom con igu a ions each ime an image is u ilized. Consequen ly, no wo iews
o an image a e iden ical du ing he aining p ocess.
Se ings o aining augmen a ion a e de ined in a hype pa ame e ile, and hype pa ame e
e olu ion can be employed o op imize hese alues acco ding o speci ic aining equi emen s (as shown
in Figu e 50).
Figu e 50 - T aining augmen a ion de ined in a hype pa ame e ile [56].
Au o Lea ning Bounding Box Ancho s
Wi hin he con ex o YOLO 5, he Au o Lea ning Bounding Box Ancho s (Figu e 51) echnique is
employed o au onomously p oduce ancho boxes me iculously ailo ed o he cha ac e is ics o he
pa icula da ase in use. Th oughou he aining p ocess, he model sc u inizes he dis ibu ion o objec
dimensions and con igu a ions wi hin he aining da ase , subsequen ly ine- uning he ancho box sizes
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
49
and aspec a ios o align wi h hese cha ac e is ics. This adap i e app oach enables he model o mo e
e ec i ely pinpoin objec s p esen ing a ying sizes and shapes, ul ima ely enhancing he p ecision o
objec de ec ion [57].
Figu e 51 - Lea ning Bounding Box Ancho s [57].
To illus a e, conside an ins ance o ancho s ha ha e been de i ed om he aining da a and
inco po a ed in o a YOLO 5 con igu a ion ile (Figu e 52).
Figu e 52 - A ske ch o he objec de ec ion ask.
In his pa icula ins ance, i can be obse ed ha h ee ancho boxes a e de ined, wi h each ancho
box ep esen ed as a lis con aining six alues. These alues co espond o he wid h and heigh o he
ancho boxes a h ee dis inc scales. I is no ewo hy ha hese ancho boxes ha e been deduced om
he aining da a and a e u ilized o gene a ing box p edic ions du ing he in e ence phase.
In he ealm o objec de ec ion (see Figu e 53), he p ima y objec i e is o iden i y and p ecisely
delinea e objec s wi hin an image. This di e s ma kedly om image classi ica ion, whe e he p ima y ask
in ol es ca ego izing he en i e image in o a single class. In objec de ec ion, he challenge lies in he
po en ial exis ence o mul iple ins ances o iden ical o dissimila objec s wi hin he image. The
o e a ching goal is o accu a ely p edic bo h he p esence and he p ecise loca ion o all hese objec s.
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
56
E alua ion Me ics
E alua ion me ics in deep lea ning a e quan i a i e measu es used o assess he pe o mance o
a deep lea ning model in speci ic asks. Common me ics include accu acy, p ecision, ecall, F1-sco e,
loss, and he con usion ma ix. These me ics help in unde s anding how well he model is pe o ming
and compa ing di e en models.
The ini ial e alua ion me ics o a deep lea ning algo i hm like YOLO 5 se e he pu pose o
assessing he pe o mance o he objec de ec ion algo i hm on a speci ic da ase . These me ics,
commonly employed, a e speci ic o YOLO 5 o o he simila algo i hms (Fas e R-CNN - Region-based
Con olu ional Neu al Ne wo k; SSD - Single Sho Mul ibox De ec o ; Re inaNe ; Mask R-CNN - Mask
Region-based Con olu ional Neu al Ne wo k; E icien De ; YOLO 7; YOLO 8) [61], [62]:
1) A e age P ecision (AP): A e age P ecision is a widely used me ic o e alua ing objec de ec ion
pe o mance. I measu es he accu acy o de ec ions a di e en con idence le els ( h esholds) and
calcula es he a ea unde he P ecision-Recall cu e. The highe he AP alue, he be e he model
pe o mance. A e age P ecision is compu ed using he P ecision-Recall (P-R) cu e. The P-R cu e is
ob ained by a ying he con idence h eshold o posi i e de ec ions and calcula ing he P ecision and
Recall o each h eshold. AP is calcula ed by aking he a e age o p ecisions calcula ed a speci ic
poin s on he cu e (e.g., a ecall in e als). AP is as ollows:
AP= 1
n ∑P ( )∆
(17)
Whe e:
• n is he numbe o h eshold poin s on he P-R cu e;
• P( ) is he P ecision a he ecall poin ;
• Δ is he di e ence in ecall be ween wo consecu i e poin s on he P-R cu e.
2) Mean A e age P ecision (mAP): mAP is he a e age o he A e age P ecisions calcula ed o
a ious objec classes. I is a global me ic ha p o ides an o e all iew o he model pe o mance
ac oss all classes. mAP is as ollows:
mAP= 1
C ∑AP
c
=1 (18)
Whe e:
• C ep esen s he numbe o ca ego ies in he da ase . The highe he mAP alue, he be e he
model pe o mance;

Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
57
• [email p o ec ed] o mAP@50 o IoU = 0.50 mean A e age P ecision wi h an In e sec ion o e Union
h eshold o 0.5 o 50%;
• mAP.5:95 means a e age mAP o e di e en IoU h esholds, om 0.5 o 0.95.
3) IoU (In e sec ion o e Union): IoU is a me ic ha measu es he o e lap be ween he bounding
box p edic ed by he model and he g ound u h bounding box o he objec . A high o e lap indica es
an accu a e de ec ion. IoU is as ollows:
IoU=In e sec ion A ea
Union A ea (19)
4) Accu acy: Accu acy measu es he p opo ion o co ec de ec ions ela i e o he o al de ec ions
made by he model. Accu acy is as ollows:
Accu acy=Numbe o Co ec De ec ions
(To al Numbe o De ec ions) (20)
O
Accu acy= TP+TN
TP+TN+FP+FN (21)
Whe e:
• T ue Posi i e (TP): he numbe o samples co ec ly classi ied as posi i e;
• False Posi i e (FP): he numbe o samples inco ec ly classi ied as posi i e;
• False Nega i e (FN): he numbe o samples inco ec ly classi ied as nega i e;
• T ue Nega i e (TN): he numbe o samples co ec ly classi ied as nega i e.
5) P ecision: P ecision (also known as Posi i e P edic i e Value) measu es he p opo ion o co ec
de ec ions among he o al posi i e de ec ions made by he model. P ecision is as ollows:
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛= 𝑇𝑃
(𝑇𝑃+𝐹𝑃) (22)
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
58
6) Recall: Recall (also known as Sensi i i y o T ue Posi i e Ra e) measu es he p opo ion o co ec
de ec ions among he o al ue objec s p esen in he images. Recall is as ollows:
𝑅𝑒𝑐𝑎𝑙𝑙= 𝑇𝑃
(𝑇𝑃+𝐹𝑁) (23)
7) F1-Sco e: The F1-Sco e is he ha monic mean o P ecision and Recall. I is use ul when one desi es
a me ic ha conside s bo h he p ecision and ecall o he model. F1-Sco e is as ollows:
F1− Sco e= 2 x P ecision x Recall
P ecision+Recall (24)
8) Con usion Ma ix: The Con usion Ma ix (Figu e 59) is a able ha shows he coun o co ec and
inco ec de ec ions o each class o objec . I is a use ul ool o e alua ing he model's disc imina ion
abili y ac oss di e en classes.
These e alua ion me ics enable he assessmen o he quali y and pe o mance o he neu al
ne wo k du ing bo h aining and es ing p ocesses.
Figu e 59 –Con usion Ma ix o Bina y Classi ica ion [63].
Typically, in he e alua ion o de ec ion quali y, esul s a e ca ego ized in o ou g oups by
compa ing he model ou pu o he g ound u h. When he model co ec ly de ec s an objec , i is labeled
as a T ue Posi i e (TP). Con e sely, i he model de ec s an objec ha is no p esen in he image, i is
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
59
e e ed o as a False Posi i e (FP). When an objec in he g ound u h goes unde ec ed by he model, i
is e med a False Nega i e (FN). T ue Nega i es (TN), ep esen ing co ec ly unde ec ed objec s o
backg ound, a e no ypically conside ed in he e alua ion o objec de ec ion. These ou ca ego ies o m
he basis o a con usion ma ix.
The P ecision-Recall Cu e and mAP (Mean A e age P ecision)
The P ecision-Recall Cu e and mAP (Mean A e age P ecision) a e impo an e alua ion me ics
in machine lea ning, pa icula ly o asks like objec de ec ion and in o ma ion e ie al [64].
The P ecision-Recall Cu e, depic ed in Figu e 60, p o ides a g aphical ep esen a ion o he
ade-o be ween p ecision, and ecall o a machine lea ning model. P ecision assesses he accu acy o
he model posi i e p edic ions, while ecall (o sensi i i y) e alua es he model abili y o cap u e all he
ac ual posi i e ins ances. This cu e illus a es how p ecision and ecall alues change as he model's
decision h eshold is adjus ed, o e ing insigh s in o how he model pe o mance can be ine- uned o
mee speci ic ask equi emen s. The a ea unde he P ecision-Recall Cu e (AUC-PR) se es as a
quan i a i e measu e o he o e all model pe o mance, wi h a highe AUC-PR alue indica ing supe io
pe o mance [64].
Figu e 60 – An example o he p ecision- ecall cu e [64].
mAP (Mean A e age P ecision), shown in Figu e 61, se es as a comp ehensi e me ic o
assessing a model p edic ion in objec de ec ion and in o ma ion e ie al asks, condensing he o e all
quali y o hose p edic ions in o a single nume ical alue. I is compu ed by aking he a e age o he
a e age p ecision (AP) alues o each class o ca ego y in a mul i-class p oblem. AP ep esen s he a ea
unde he P ecision-Recall Cu e o a single class and quan i ies he model abili y o dis inguish objec s
o ha class om he backg ound. mAP is pa icula ly aluable when e alua ing models in scena ios
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
60
in ol ing mul iple objec classes, as i o e s a uni ied pe o mance me ic ha accoun s o pe o mance
ac oss all classes [65].
Figu e 61 – An example o he p ecision- ecall cu e [65].
In summa y, he P ecision-Recall Cu e illus a es how changes in he classi ica ion h eshold
a ec p ecision and ecall. I helps in unde s anding he model's ade-o be ween making accu a e
posi i e p edic ions and cap u ing all ele an ins ances. mAP, on he o he hand, is a summa izing me ic
ha is pa icula ly use ul o e alua ing models in mul i-class classi ica ion asks, such as objec de ec ion,
by a e aging he pe o mance ac oss di e en classes, p o iding an o e all assessmen o he model
e ec i eness. Bo h me ics a e c ucial o e alua ing he pe o mance o machine lea ning models,
especially in asks whe e p ecision and ecall a e i al, such as objec de ec ion [65].
k- old C oss Valida ion
k- old C oss-Valida ion is a echnique used in machine lea ning and model e alua ion o assess
a model pe o mance and i s gene aliza ion capabili ies. I helps o add ess he p oblem o o e i ing and
o ob ain a mo e eliable es ima e o how well a model is likely o pe o m on unseen da a. k- old C oss-
Valida ion wo ks as ollows [66].
1. Da a Spli ing: The da ase is di ided in o K app oxima ely equal-sized olds o subse s. Fo
example, i k is se o 5, he da ase is di ided in o i e pa s;
2. Model T aining and Tes ing: The aining and e alua ion p ocess is epea ed k imes. In each
i e a ion, one o he k olds is held ou as a alida ion se , while he emaining k-1 olds a e used
o ain he model. The model is hen es ed on he alida ion se ;
3. Pe o mance Me ics: Fo each i e a ion, a pe o mance me ic (e.g., accu acy, mean
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
61
squa ed e o ) is calcula ed based on he model's pe o mance on he alida ion se ;
4. A e age Pe o mance: A e k i e a ions, one has k pe o mance me ic alues (one o each
old). The inal pe o mance me ic is ypically calcula ed as he a e age o hese k alues. This
p o ides a mo e obus es ima e o he model's pe o mance compa ed o a single ain- es spli .
k- old C oss-Valida ion o e s se e al ad an ages in model e alua ion. Fi s ly, i p o ides a mo e
accu a e es ima ion o a model's pe o mance by subjec ing i o es ing on di e en subse s o he da a.
This app oach aids in assessing how well a model gene alizes o unseen da a, as i unde goes es ing on
a ious da a pa i ions, hus enhancing i s eliabili y. Addi ionally, k- old C oss-Valida ion helps mi iga e
he isk o o e i ing, as he model is ained on mul iple da a subse s and es ed on di e en alida ion
se s, ensu ing ha he model's pe o mance is no o e ly ailo ed o a speci ic da ase . This me hod
p o es especially aluable when dealing wi h limi ed da ase s, whe e maximizing da a u ili y is c ucial
[67].
The choice o k (e.g., 5, 10) depends on he speci ic da ase and p oblem. La ge alues o k
educe he a iance in he pe o mance es ima e bu can be compu a ionally expensi e. Smalle alues
may in oduce highe a iance bu equi e less compu a ion. In p ac ice, k- old C oss-Valida ion is a
aluable ool o assessing and compa ing machine lea ning models and selec ing he bes -pe o ming
one o speci ic p oblem [67].
k- old C oss Valida ion s ands as a obus app oach o app aising machine lea ning models. I s
me i s lie in mi iga ing o e i ing and unde i ing issues, ul ima ely deli e ing a mo e p ecise gauge o a
model's e ec i eness on unseen da a. This echnique is ex ensi ely employed in p ac ical scena ios o
model alida ion, compa ison, and he selec ion o an op imal model ailo ed o a speci ic p oblem [68].
A isual ep esen a ion o he me hod's implemen a ion can be obse ed in Figu e 62 [69].

Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
62
Figu e 62 – An example o he p ecision- ecall cu e [69].
The alue k = 10 is equen ly chosen o k- old C oss Valida ion due o se e al compelling
easons [70], [71]:
1. Reduc ion in Resul Va iance: The use o a la ge k alue, such as 10, se es o diminish he
a iance in he esul s. Consequen ly, when k- old C oss Valida ion is conduc ed mul iple imes,
he ou comes a e mo e likely o exhibi consis ency and accu a ely e lec he model's ue
pe o mance;
2. E icien Da a U iliza ion: In he case o k = 10, oughly 90% o he da ase is alloca ed o
aining du ing each i e a ion, wi h only 10% dedica ed o alida ion. This balance is gene ally
conside ed e ec i e o ensu ing ha he model lea ns comp ehensi ely and is igo ously
e alua ed;
3. Equi able Model Compa isons: The selec ion o k = 10 os e s equi able compa isons among
di e en models. This is achie ed by e alua ing all models using he same da ase and adhe ing
o he iden ical c oss- alida ion me hodology.
The p ocedu e o k- old C oss Valida ion wi h k = 10 un olds as ollows [72], [73]:
1. The da ase is pa i ioned in o 10 oughly equi alen segmen s, o olds;
2. The model unde goes 10 ounds o aining and e alua ion. In each i e a ion, one- old is se aside
as he alida ion se , while he emaining 9 olds cons i u e he aining se ;
3. In each i e a ion, pe o mance me ics such as p ecision, ecall, F1-sco e, among o he s, a e
compu ed o assess he model pe o mance;
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
63
4. Upon comple ion o he 10 i e a ions, he pe o mance me ics ob ained in each ound a e
ypically a e aged o yield an agg ega e ep esen a ion o he model's pe o mance. This p o ides
a mo e dependable es ima ion o how he model will pe o m on unseen da a.
2.5.3 YOLO 5 Algo i hm o Objec De ec ion
I is he i h i e a ion in he YOLO se ies o algo i hms, which ha e gained popula i y o hei abili y
o pe o m eal- ime objec de ec ion wi h imp essi e accu acy. Some key ea u es and cha ac e is ics o
YOLO 5 a e as ollows [74]:
• Accu acy and Speed: YOLO 5 is enowned o i s balance be ween accu acy and speed in objec
de ec ion asks. I achie es imp essi e eal- ime pe o mance while main aining high p ecision;
• A chi ec u al Imp o emen s: YOLO 5 in oduces a chi ec u al enhancemen s o e i s
p edecesso s, such as CSPDa kne 53 as he backbone and PANe as he neck, leading o
imp o ed ea u e ex ac ion and agg ega ion;
• Scale Handling: YOLO 5 e ec i ely add esses scale a ia ion, enabling i o de ec objec s o
a ious sizes wi hin an image, hanks o he ea u e py amid and ancho box echniques;
• Adap i e Da a Augmen a ion: YOLO 5 employs da a augmen a ion echniques like andom
c opping, o a ion, and colo ji e ing o enhance i s abili y o gene alize o new and unseen da a;
• Poin -Based Sampling: The PA-Ne (Poin Agg ega ion Ne wo k) Neck in YOLO 5 u ilizes poin -
based sampling o agg ega e ea u es a di e en scales, con ibu ing o i s accu acy;
• Model Va ian s: YOLO 5 comes in di e en a ian s (e.g., YOLO 5s, YOLO 5m, YOLO 5l, and
YOLO 5x) wi h a ying model sizes o sui di e en compu a ional esou ces and deploymen
scena ios;
• S a e-o - he-A Pe o mance: YOLO 5 has achie ed s a e-o - he-a esul s in e ms o accu acy
and e iciency in objec de ec ion asks, making i a popula choice o a ious applica ions,
including au onomous d i ing, su eillance, and mo e;
• Open-Sou ce: YOLO 5 is open-sou ce, allowing esea che s and de elope s o access i s
codebase, cus omize i , and in eg a e i in o hei p ojec s;
• Communi y Suppo : Due o i s popula i y, YOLO 5 has a g owing communi y o use s,
con ibu ing o ongoing de elopmen , imp o emen s, and he a ailabili y o p e- ained models
and esou ces;
• Ve sa ili y: YOLO 5 can be adap ed o a wide ange o objec de ec ion asks, including single-
objec de ec ion, mul i-objec de ec ion, and e en cus om objec de ec ion scena ios.
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
64
In summa y, YOLO 5 is a cu ing-edge objec de ec ion algo i hm known o i s speed, accu acy,
and e sa ili y, making i a signi ican asse in compu e ision applica ions.
Sub- a ian s o YOLO 5
YOLO 5 is a ailable in ou p ima y e sions: small (s), medium (m), la ge (l), and ex a-la ge (x),
wi h each e sion deli e ing inc easingly imp o ed accu acy a es [56]. Addi ionally, he aining ime
equi ed o each a ian a ies (Figu e 63).
Figu e 63 – Sub- a ian s o YOLO 5 [56].
The objec i e o he Figu e 63 is o demons a e he pe o mance o an objec de ec ion model
conce ning i s in e ence ime (X-axis) and o e all e ec i eness (Y-axis). Ini ial indings indica e ha
YOLO 5 excels in achie ing excep ional pe o mance ela i e o o he cu ing-edge echniques.
As illus a ed in he Figu e 63, all YOLO 5 a ian s exhibi as e aining imes when compa ed o
E icien De . E icien De is a ype o objec de ec ion model op imized o e icien and accu a e objec
de ec ion, known o i s abili y o achie e good pe o mance while main aining low compu a ional
demands. No ably, he highes pe o ming YOLO 5 model, YOLO 5x, can p ocess images a a signi ican ly
accele a ed a e while main aining a le el o accu acy simila o ha o he E icien De D4 model [56].
I is impo an o no e ha YOLO 5 pe o mance imp o emen s a e p ima ily a ibu ed o he
aining p ocedu es in PyTo ch, while he undamen al model a chi ec u e emains closely aligned wi h
YOLO 4 [56].
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
65
O e iew o he YOLO 5s6 A chi ec u e
The YOLO 5s6 deep lea ning de ec ion algo i hm has been u ilized o iden i y and cha ac e ize
ya n hai iness, as will be elabo a ed in chap e 5. This choice is mo i a ed by i s op imized deploymen
on esou ce-cons ained edge de ices, ende ing i sui able o manu ac u ing applica ions, as well as i s
un es ic ed accessibili y o esea ch pu poses. Tes s we e conduc ed o compa e di e en YOLO
e sions, and YOLO 5s6 ou pe o med he o he s in e ms o e alua ion me ics, es ablishing i as he
p e e ed choice.
The YOLO 5s6 model comp ises h ee essen ial componen s, as illus a ed in Figu e 64 [75], [76],
[77]:
• Backbone: YOLO 5 employs a con olu ional backbone o ex ac ea u es om inpu images. This
backbone pe o ms ini ial con olu ions and ex ac s low-le el ep esen a ions om he inpu
images;
• Neck: In ce ain a ian s o YOLO 5, such as YOLO 5m, YOLO 5l, and YOLO 5x, a "neck" is
inco po a ed. This neck consis s o a sequence o con olu ional laye s designed o use ea u es
om a ious scales. This capabili y enables he model o cap u e in o ma ion a di e en le els
o esolu ion, enhancing i s abili y o de ec objec s o a ying sizes;
• Head o De ec : The "head" o "De ec " is he inal componen o he model, whe e he objec
de ec ion p ocess occu s. I comp ises se e al con olu ional and de ec ion laye s esponsible o
p edic ing bounding boxes and objec classes wi hin he inpu image.
Figu e 64 illus a es he ne wo k s uc u e o YOLO 5s6, highligh ing he in e play be ween hese
h ee key componen s [76].
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
72
Figu e 70 - Compa ison o ac i a ion unc ions [81].
In esume, hese ac i a ion unc ions play an impo an ole in he aining and pe o mance o
neu al ne wo ks, and he choice o which one o use can depend on he speci ic ask and desi ed esul s.
In YOLO 5s and o he YOLO a ian s, as p e iously e e ed, hese ac i a ion unc ions a e used in he
ne wo k laye s o in oduce non-linea i y and lea ning capabili y.
Hype pa ame e s in YOLO 5s6
Hype pa ame e s in machine lea ning models, including deep lea ning models like YOLO 5, play
a c ucial ole in de ining he aining and da a augmen a ion se ings [82]. When dealing wi h la ge
models, which ypically ha e mo e pa ame e s and complexi y, i is common o adjus ce ain
hype pa ame e s o mi iga e o e i ing and imp o e model gene aliza ion.
Hype pa ame e s o da a augmen a ions a e o en de ined in a sepa a e hype pa ame e ile,
which is p o ided as pa o he aining con igu a ion. In he case o YOLO 5, he hype pa ame e s ha
con ol da a augmen a ions can be ound in he 'hyp.sc a ch.yaml' o 'hyp.sc a ch-low.yaml' ile (o a
simila ile) loca ed wi hin he da a di ec o y.
These augmen a ion hype pa ame e s include se ings o a ious ypes o da a ans o ma ions
such as o a ion, ansla ion, scaling, lipping, and mo e. By con igu ing hese hype pa ame e s, i can be
con olled he ype and in ensi y o augmen a ions applied o he aining da a, which can help imp o e
he model abili y o gene alize o di e en si ua ions.
In Figu e 71, i is explained how i wo ks he YOLO 5 image da a augmen a ion wi h a da ase in
aining phase o a neu al ne wo k wi h de ined hype pa ame e s.

Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
73
Figu e 71 – YOLO 5s6 Image Augmen a ion [82].
I is common p ac ice o de ine hese hype pa ame e s in a sepa a e ile o keep hem o ganized
and easily con igu able, allowing o expe imen a ion wi h di e en augmen a ion s a egies du ing model
aining. Adjus ing hese hype pa ame e s can ha e a signi ican impac on he obus ness and
pe o mance o he ained model.
The hype pa ame e s con igu a ion o aining he YOLO 5 model, a e ollowing explained [82]:
• l 0: The ini ial lea ning a e used by he op imize du ing aining. I is se o 0.01 o SGD and
0.001 o Adam. SGD s ands o "S ochas ic G adien Descen ," which is an op imize used o
adjus he weigh s o a neu al ne wo k du ing aining. "Adam" e e s o ano he op imize called
"Adam" (Adap i e Momen Es ima ion), which is also used o op imize he pa ame e s o a neu al
ne wo k. The alues men ioned (0.01 o SGD and 0.001 o Adam) a e he ini ial lea ning a es
used by hese op imize s o adjus he ne wo k's weigh s du ing aining. Each op imize may
equi e a di e en lea ning a e o wo k e ec i ely based on he p oblem and he neu al ne wo k
a chi ec u e;
• l : The inal lea ning a e o he OneCycleLR a he end o aining, calcula ed as he p oduc o
he ini ial lea ning a e (l 0) and his alue (0.01);
• momen um: The momen um alue used in he SGD op imize o he be a1 alue used in he
Adam op imize ;
• weigh _decay: The weigh decay alue used by he op imize o con ol he penaliza ion o
la ge weigh s in he model. In his case, i is se o 0.0005;
• wa mup_epochs: The numbe o "wa m-up" epochs a he s a o aining du ing which he
lea ning a e is g adually inc eased. I is se o 3.0 epochs;
Chap e 2 –
Theo ical Concep s
___________________________________________________________________________
74
• wa mup_momen um: The ini ial momen um alue du ing he wa m-up pe iod, con igu ed as
0.8;
• wa mup_bias_l : The ini ial lea ning a e o he bias e m du ing he wa m-up pe iod, se o
0.1;
• box, cls, cls_pw, obj, obj_pw: These pa ame e s adjus he weigh s o di e en componen s
o he YOLO 5 loss unc ion, in luencing he impac o di e en pa s o he loss unc ion du ing
aining;
• iou_ : The aining h eshold o In e sec ion o e Union (IoU) used o de e mine whe he a
de ec ion is conside ed a ue posi i e du ing aining. In his case, i is se o 0.20;
• ancho _ : The h eshold ha de e mines whe he an ancho is conside ed ac i e o inac i e;
• l_gamma: The gamma alue used in he ocal loss. In his case, i is se o 0.0.
• hs _h, hs _s, hs _ : These pa ame e s con ol a ia ion in he hue, sa u a ion, and alue o
he HSV colo space du ing da a augmen a ion;
• deg ees, ansla e, scale, shea , pe spec i e: Pa ame e s o adjus ing a ious geome ic
ans o ma ions applied o images du ing da a augmen a ion, such as o a ion, ansla ion, scale,
shea , and pe spec i e;
• lipud, lipl : P obabili ies o applying e ical and ho izon al lipping o images;
• mosaic, mixup, copy_pas e: P obabili ies o applying echniques like mosaic, image mixup,
and segmen copy-pas e du ing da a augmen a ion.
These hype pa ame e s a e c ucial o aining he YOLO 5 model and i s abili y o gene alize o
objec de ec ion. P ope ly uning hese hype pa ame e s is essen ial o ob ain a high-quali y objec
de ec ion model.
Summa y o he chap e :
This chap e p o ided a comp ehensi e ounda ion o heo e ical concep s essen ial o
unde s anding he subsequen de elopmen o a ya n quali y analysis sys em. I encompasses ya n
cha ac e iza ion pa ame e s, s a is ical pa ame e s o ya n, spec al analysis echniques, he p inciples
o image p ocessing and compu e ision, and he undamen als o deep lea ning. These opics
collec i ely se he s age o he wo k objec i es, enabling a deepe unde s anding o he me hodologies
and echnologies in ol ed in his in e disciplina y esea ch endea o . Addi ionally, his chap e in oduces
key concep s ela ed o YOLO 5s, which will be u he explo ed and applied in chap e 5 conside ing
hei ele ance o he wo k goals.
3. LITERATURE REVIEW
O e iew
Technological e olu ion has eached impo an p opo ions in e ms o sys ems ob aining cha ac e is ics
ha measu e ya n quali y and he use o a i icial in elligence o de ec ya n de ec s. The e o e, his
chap e p esen s some s udies ha explo e image p ocessing/compu e ision and a i icial in elligence
app oaches o ob ain ya n cha ac e is ics. I is in ended o analyze wha he scien i ic communi y has
explo ed, he challenges encoun e ed, and he p oblems ha s ill need o be add essed in o de o
unde ake he objec i es o his wo k. Nex , he mos ele an wo ks a e p esen ed. A he end, a
compa a i e able and a c i ical e iew analysis a e p esen ed.
3.1 Use o image p ocessing, compu e ision and a i icial in elligence in ya n analysis
3.2 Compa ison be ween sys ems wi h image p ocessing, compu e ision and a i icial in elligence
echniques
3.3 C i ical e iew analysis
Chap e 3 –
Li e a u e Re iew
___________________________________________________________________________
76
3.1 Use o image p ocessing, compu e ision and a i icial in elligence in
ya n analysis
Technological e olu ion has achie ed signi ican ad ancemen s in sys ems o measu ing ya n
quali y cha ac e is ics. Consequen ly, his subchap e in oduces a ious solu ions and ini ia i es ha
le e age echnology o assess ya n p ope ies. I is impo an o examine he esea ch unde aken by he
scien i ic communi y, he obs acles aced, and he ou s anding issues ha need solu ions o enhancing
esea ch app oaches. Each subchap e highligh s he mos signi ican wo ks om 2020 o 2023,
p o iding an assessmen o he p og ess made in his ield.
Based on he cu en esea ch, limi ed documen ed ins ances o al e na i e ya n analysis sys ems
a e a ailable o compa ison wi h he p oposed de eloped sys em, which will be ex ensi ely discussed in
subsequen chap e s. The analysis pe o med ou line he sho comings and inadequacies ound wi hin
each sys em, he eby p o iding a c i ical e alua ion o hei limi a ions.
3.1.1 Sys em A – S udy by Zhisong Li e al. (2020)
In he compu e ision sys em de eloped by Zhisong Li and colleagues in 2020 [83], hence o h
e e ed o as Sys em A, an e alua ion o ya n quali y was conduc ed. This assessmen ocused on
acqui ing he ollowing pa ame e s: Ya n diame e , De ec s and Hai iness.
These cha ac e is ics we e de e mined using compu e ision and image p ocessing echniques.
In hei p oposed sys em (Figu e 72), hey de ined a diame e image p ocessing uni (DIPU) and selec ed
se e al sampling poin s om he mo ing ya n. The DIPU co esponding o each sampling poin was
segmen ed om he cap u ed ya n images. The a e age DIPU diame e was used o ep esen he ya n
diame e o he es poin s. When ex ac ing ya n images, he DIPU was ca ego ized in o:
• Fo eg ound egion;
• De ined backg ound egion;
• Unknown egion based on g ay le el p ojec ion dis ibu ion cha ac e is ics.
Chap e 3 –
Li e a u e Re iew
___________________________________________________________________________
77
Figu e 72 - Flowcha o he me hod o de ec ing ya n pa ame e s and e alua ing ya n quali y [83].
The unknown egion was u he p ocessed using he Poisson Ma ing me hod [84], which
employed a de ined connec i i y classi ie o sepa a e he ya n image om he backg ound image. A e
he classi ie ex ac ed he ya n co e, hai iness was di e en ia ed using image sub ac ion. Fu he mo e,
ya n de ec s we e analyzed using s a is ical me hods o assess ya n quali y.
The designed sys em employed a closed box as an image cap u e pla o m o shield agains s ay
ligh dis u bance. They used whi e di used glass o elimina e e lec ed ligh and shadows. Fo simula ing
ya n mo emen , a d i e and ol age con olle we e u ilized. The in eg a ed s uc u e o he sys em is
depic ed in Figu e 73. The ya n image cap u e sys em p ima ily consis ed o wo componen s:
Figu e 73 - Sys em de ice o de ec ya n pa ame e s [83].

Chap e 3 –
Li e a u e Re iew
___________________________________________________________________________
78
• Imaging sys em, including a senso , came a lens, and ligh sou ce, wi h a CCD (cha ge-coupled
de ice) image senso o collec ing he sequence o images om he ya n;
• Imaging lens wi h a la ge dep h o ield o ensu e clea analysis o hai iness in di e en posi ions.
Images we e cap u ed using a SONY C Zoom lens wi h a ocal leng h o 12.5 o 75 mm connec ed
o a CCD came a (Basle acA2040-180 km/kc, pixel size 5.5 µm, esolu ion o 2048 × 2048 pixels,
sampling a e o 180 ps) and p ocessed by a pe sonal compu e wi h an image p ocesso .
Ya n hai iness e e s o he numbe o ibe s p ojec ing ou o he ya n co e, was iden i ied as an
impo an indica o a ec ing ya n p ope ies, wea ing e iciency, and ab ic appea ance. In his s udy, he
ya n hai iness index was included as a equi emen o ya n quali y. The au ho s explained ha he ya n
could be di ided in o a ya n co e and hai iness. Using he desc ibed me hod, hey ex ac ed he DIPU
om he ya n co e based on pixel connec i i y o each image, hus enabling he de e mina ion o
hai iness pixels using he image sub ac ion me hod (see Figu e 74).
Figu e 74 - a) New o eg ound image, b) ya n co e, c) hai iness [83].
The au ho s did no p opose speci ic imp o emen s o u u e wo k o his p ojec . They simply
s a ed ha he p oposed me hod and i s co esponding de ec ion sys em acili a e au oma ion and
indus ializa ion in online ya n de ec ion. Se e al o he au ho s ha e also wo ked in he ield o ya n image
p ocessing bu wi hou in oducing signi ican inno a ions compa ed o p e ious esea ch e o s [11],
[85], [86].
Chap e 3 –
Li e a u e Re iew
___________________________________________________________________________
79
3.1.2 Sys em B – S udy by Noman Haleem e al. (2021)
A sys em was also de eloped by Noman Haleem and colleagues in 2021 [87], he eina e e e ed
o as Sys em B. In hei wo k, he au ho s emphasized he po en ial o conduc ing ya n es ing in eal-
ime mode o minimize p ocess la ency, p o ide apid insigh s in o ya n quali y, and enhance he
p oduc ion o high-quali y ya n. They poin ed ou ha while capabili y-based uni o mi y es equipmen
and eme ging image-based uni o mi y es sys ems a e widely used in o line se ings, he e is s ill a lack
o a sui able online sys em o assessing ya n quali y unde ypical indus ial condi ions. As a esponse,
he au ho s o his p ojec p oposed an online uni o mi y es sys em designed speci ically o de ec ing
neps. To achie e his, hey ha nessed image p ocessing echniques and compu e ision.
The sys em hey de eloped (Figu e 75) di ec ly cap u es ya n images on a o a ing ame and
employs he Viola-Jones objec de ec ion algo i hm [88] o eal- ime nep de ec de ec ion. The alida ion
o he nep de ec ion algo i hms, along wi h a compa ison o he new me hod o an exis ing uni o mi y es
equipmen in e ms o nep coun , demons a ed a easonable le el o de ec de ec ion accu acy and
p omising po en ial o applica ions in he ya n spinning indus y. These esul s we e u he alida ed by
compa ing hem wi h USTER exis ing comme cial uni o mi y es ing equipmen (TESTER). The image
collec ion p ocess (see Figu e 75) u ilized a Basle 1440-220 μm digi al came a om Ge many, equipped
wi h a 50 mm lens.
Figu e 75 - An example o a ya n image wi h a nep de ec ob ained using he image acquisi ion sys em [87].
The au ho s le e aged h ee compu e ision models ( e e ed o A, B, and C), all based on he
machine lea ning Viola-Jones algo i hm [88]. These models we e de eloped using he OpenCV compu e
ision lib a y in he Py hon p og amming language. They essen ially unc ioned as image classi ie s, wi h
a ia ions in he inpu da a used du ing hei aining phase. The aining da a comp ised posi i e and
nega i e images in a 2:1 a io, and hese images we e esized by ac o s o 0.2 and 0.1, espec i ely, o
Chap e 3 –
Li e a u e Re iew
___________________________________________________________________________
80
op imize he aining p ocess. Each model unde wen en aining cycles. Each model was ained using
a speci ic ype o image da a: o iginal images o ya ns wi h neps, syn he ic images o neps, o a
combina ion o bo h. This app oach allowed hem o explo e di e en aspec s o nep de ec ion and
op imize he models. A e aining, each model was alida ed using a sepa a e da ase con aining images
o ya ns wi h and wi hou neps. This alida ion p ocess assessed he models' abili y o accu a ely classi y
images as ei he posi i e (wi h neps) o nega i e (wi hou neps)
Figu e 76 showcases a se o conca ena ed images o ya ns exhibi ing he nep de ec , acqui ed
using he p oposed image acquisi ion con igu a ion du ing he ya n spinning p ocess. Addi ionally, he
au ho s no ed ha in hei analysis o a se ies o con inuous ya n shape images, i was c ucial o he
ield o iew o be subs an ial enough o accommoda e he la e al oscilla ions o he ya n du ing he
spinning p ocess.
Figu e 76 - An example o a ya n image wi h a nep de ec ob ained using he image acquisi ion sys em [87].
Figu e 77 p esen s he models con usion ma ices which p o ided insigh s in o he models'
pe o mance in e ms o ue posi i es, ue nega i es, alse posi i es, and alse nega i es.
Chap e 3 –
Li e a u e Re iew
___________________________________________________________________________
81
Figu e 77 - Con usion ma ices ha o e insigh s in o he pe o mance o he model sys em [87].
Fo u u e esea ch di ec ions, he au ho s sugges ed he ollowing [87]:
• Acquisi ion o high-quali y online ya n images h ough an imaging sys em wi h ul a-low
exposu e ime, combined wi h ex e nal ligh ing and an app op ia e op ical se up;
• Expansion o he online ya n uni o mi y measu emen sys em o de ec o he ypes o ya n
de ec s, along wi h necessa y addi ions o he image p ocessing module o c ea e a
comp ehensi e ya n quali y es ing solu ion sui able o indus ial en i onmen s.
• Imp o emen o T aining Models: The au ho s sugges ed adjus ing aining s a egies o
enhance nep de ec ion accu acy. Addi ionally, inc easing he da ase size could imp o e model
pe o mance, while explo ing al e na i e objec de ec ion app oaches could op imize esul s.
• Expansion o he online uni o mi y measu emen sys em: The au ho s belie e ha in he u u e
a sys em o iden i ying addi ional de ec s in he ya n in addi ion o neps could be de eloped.
They also sugges ha image p ocessing modules be modi ied o encompass new de ec
de ec ion equi emen s.
• Finally, he au ho s p opose an in es iga ion in o disc epancies compa ed o he Us e Tes e
3 machine: Conduc u he s udies o unde s and and add ess signi ican di e ences be ween
online and adi ional es ing me hods, imp o ing es ing me hodologies o ensu e alignmen
and accu acy in quali y assessmen .
3.1.3 Sys em C – S udy by Adel El-Geiheini e al. (2020)
Conce ning he applica ion o a i icial in elligence o modeling ya n ensile p ope ies, Adel El-
Geiheini and colleagues in 2020 [89], he eina e e e ed o as Sys em C, emphasized ha he u iliza ion
Chap e 3 –
Li e a u e Re iew
___________________________________________________________________________
88
machine.
10. Analysis Limi a ions: None o he sys ems pe o m spec al analysis on he ya n based on
image p ocessing, and only sys ems B and D use a i icial in elligence o ya n de ec de ec ion.
In he con ex o deep lea ning echniques, a ho ough li e a u e e iew e ealed no ins ances o
sys ems o p o o ypes employing deep lea ning, speci ically he YOLO 5s6 algo i hm, o he
cha ac e iza ion o ya n hai iness.
3.3 C i ical e iew analysis
In he ab ic de ec de ec ion domain, a mul i ude o solu ions has eme ged, capi alizing on
enhanced i e a ions o he YOLO 3, YOLO 4, and YOLO 5 algo i hms. These s udies consis en ly exhibi
imp o ed me ics, unde sco ing he ad ancemen s in his ield [75], [93], [94], [95], [96]. Ne e heless,
a e conduc ing a e iew o he exis ing li e a u e, no ins ances we e iden i ied whe e sys ems o
p o o ypes employed deep lea ning, speci ically he YOLO 5s6 algo i hm, o he cha ac e iza ion o ya n
hai iness [75], [93], [94], [95], [96].
In he ield o ya n analysis, ad ances d i en by image p ocessing and a i icial in elligence ha e
e olu ionized he assessmen o he quali y and cha ac e is ics o ya ns. While p io esea ch has
explo ed he in eg a ion o a i icial in elligence o ab ic analysis, he in ica e applica ion o such
echnology in ya n analysis has emained a ela i ely uncha ed e i o y. Following his pa h, he
subsequen analysis aims o p o ide a c i ical e iew o he a ious ya n analysis sys ems p esen ed in
he p e ious sec ion, highligh ing he dis inc i e ea u es and con ibu ions o each, while also
emphasizing exis ing gaps and wha should be done in his ega d so ha a p oposed sys em can be
c ea ed ha b ing all hese gaps oge he , namely:
• Sys em A [83]:
• Does No Ob ain: Ya n wis o ien a ion, ya n wis s ep, linea mass, mean de ia ion U
(%), coe icien o a ia ion CV (%), hai iness mean de ia ion UH (%), hai iness s anda d
de ia ion sH (%), classi ica ion o hai iness ypes (loop/p o uding ibe s), spec al analysis.
• Limi ed in Ob aining: Numbe o cables, numbe o loose ibe s, olume, non-des uc i e
a ibu es, use o deep lea ning echniques o ya n de ec de ec ion.
• Sys em B [87]:
• Does No Ob ain: Ya n wis o ien a ion, ya n wis s ep, linea mass, mean de ia ion U

Chap e 3 –
Li e a u e Re iew
___________________________________________________________________________
89
(%), coe icien o a ia ion CV (%), hai iness coe icien H, hai iness mean de ia ion UH (%),
hai iness s anda d de ia ion sH (%), classi ica ion o hai iness ypes (loop/p o uding ibe s),
spec al analysis, hin and hick places de ec s.
• Limi ed in Ob aining: Numbe o cables, numbe o loose ibe s, olume, non-des uc i e
a ibu es.
• Sys em C [89]:
• Does No Ob ain: Ya n wis o ien a ion, ya n wis s ep, mean de ia ion o mass U (%),
hai iness coe icien H, hai iness mean de ia ion UH (%), hai iness s anda d de ia ion sH (%),
classi ica ion o hai iness ypes (loop/p o uding ibe s), Thin, Thick and Nep de ec s,
spec al analysis based in image p ocessing, ya n diame e , linea mass, use o deep lea ning
echniques o ya n de ec de ec ion.
• Limi ed in Ob aining: Numbe o cables, numbe o loose ibe s, olume, non-des uc i e
a ibu es.
• Sys em D [91]:
• Does No Ob ain: Ya n wis o ien a ion, ya n wis s ep, mean de ia ion o mass U (%),
hai iness coe icien H, hai iness mean de ia ion UH (%), hai iness s anda d de ia ion sH (%),
classi ica ion o hai iness ypes (loop/p o uding ibe s), spec al analysis based on image
p ocessing, ya n diame e , olume.
• Limi ed in Ob aining: Numbe o cables, numbe o loose ibe s, olume, non-des uc i e
a ibu es.
The analysis o he cha ac e is ics ob ained by Sys ems A, B, C and D elucida es a c i ical
pe spec i e on he exis ing sys ems and p o o ypes o he ya ns analysis. Each sys em p esen s a mix o
s eng hs and weaknesses, highligh ing he complex na u e o his ield. Al hough Sys em A p esen
ad ances in ce ain pa ame e s, such as mecha onic de elopmen and image analysis o de ec
de ec ion, hei limi a ions in se e al c i ical measu emen s canno be igno ed. Con e sely, Sys em B
employs inno a i e app oaches like a i icial in elligence (AI) and image p ocessing o de ec de ec ion,
ye i also alls sho in essen ial a eas, lacking comp ehensi e pa ame e s necessa y o holis ic ya n
analysis. Simila ly, sys ems C and D in oduce inno a i e aspec s such as AI o de ec de ec ion, bu hey
also ail in c ucial a eas, lacking comp ehensi e pa ame e s essen ial o holis ic ya n analysis.
These gaps collec i ely indica e he inadequacy o any single sys em o comp ehensi ely
add ess all aspec s o ya n analysis, especially he inco po a ion o deep lea ning echniques, image
Chap e 3 –
Li e a u e Re iew
___________________________________________________________________________
90
p ocessing and compu e ision o de ec de ec ion and ob aining a majo i y o ya n pa ame e s, which
a e lacking in mos sys ems.
A c i ical analysis o hese sys ems highligh s a no able disc epancy: none alone p o ide a holis ic
solu ion o ya ns analysis. The absence o a de eloped sys em ha inco po a es all a ibu es leads o
an incomple e unde s anding o ya n quali y, making comp ehensi e assessmen s and quali y con ol in
ex ile p oduc ion p ocesses di icul .
Fu he mo e, he limi a ions o each sys em highligh he gap in he indus y sea ch o a
comp ehensi e solu ion. The absence o spec al analysis based on image p ocessing, he inabili y o
measu e ya n wis o ien a ion, and he lack o ce ain ya n quali y pa ame e s con inue o be un esol ed
challenges in all sys ems. This gap highligh s he need o a comp ehensi e p o o ype ha no only ills
hese gaps bu also add esses he de iciencies p esen in each sys em.
In he sea ch o an e ec i e and comp ehensi e solu ion, i becomes inc easingly e iden ha a
p o o ype capable o assimila ing he s eng hs o Sys ems A, B, C and D and, a he same ime, mi iga ing
hei limi a ions is necessa y. Such a p o o ype should inco po a e mecha onic de elopmen , inco po a e
non-des uc i e a ibu es, encompass a wide ange o ya n pa ame e s o measu emen , and le e age
ad anced echnologies such as deep lea ning o de ec de ec ion. This p o o ype aims o ill he gap in
ya n analysis me hodologies, enabling a deepe unde s anding o ya n quali y and p omo ing
ad ancemen s in ex ile manu ac u ing p ocesses.
In he con ex o deep lea ning echniques, om he au ho s knowledge li e a u e e iew e ealed
no ins ances o sys ems o p o o ypes employing deep lea ning, speci ically he YOLO 5s6 algo i hm, o
he cha ac e iza ion o ya n hai iness.
Following his, he p oposed sys em in his hesis, which esponds o all hese gaps men ioned
abo e and which jus i ied i s c ea ion, is desc ibed in he ollowing chap e s, and se es as a ca alys o
inno a ion in ya n analysis and es ablishing he in oduc ion o deep lea ning o ya n quali y assessmen
wi h he po en ial o allowing p oduc ion op imiza ion.
Summa y o he chap e :
This chap e p o ided an ex ensi e e iew o he li e a u e e e ing o ya n analysis sys ems, wi h a
speci ic ocus on he u iliza ion o image p ocessing, compu e ision, and a i icial in elligence
echniques. Then a compa a i e analysis is p esen ed o p o ide an o e iew o he s eng hs and
limi a ions o he sys ems p esen ed. The chap e concludes wi h a c i ical e iew and analysis o he
discussed li e a u e, o e ing an impa ial assessmen o he ad an ages and limi a ions o di e en ya n
analysis sys ems.
4. IMAGE PROCESSING AND COMPUTER VISION SYSTEM FOR YARN
ANALYSIS
O e iew
This chap e de ails he image p ocessing echniques u ilized in he compu e ision sys em dedica ed o
ya n analysis. I includes segmen a ion algo i hms, ea u e ex ac ion me hods, and pa e n ecogni ion
echniques ailo ed o his pu pose. Addi ionally, he ha dwa e se up o image acquisi ion is discussed,
ou lining he speci ica ions o he imaging de ices used. Fu he mo e, he chap e desc ibes he in e ace
so wa e de eloped o acili a e use in e ac ion wi h he sys em.
4.1 In oduc ion
4.2 Image p e-p ocessing
4.3 Ob aining he ya n co e
4.4 Isola ion o hai iness
4.5 Isola ion o pa icles
4.6 Image cap u e so wa e and ha dwa e
4.7 De eloped in e ace
Chap e 4 –
Image P ocessing and Compu e Vision Sys em o ya n analysis
___________________________________________________________________________
92
4.1 In oduc ion
This chap e ou lines he image p ocessing echniques employed in he de elopmen o
algo i hms ocused on examining ya n cha ac e is ics o de e mine i s quali y. Key pa ame e s include
linea mass, speci ic olume, de ec s in he ya n, and i s le el o hai iness. Addi ionally, he ha dwa e
used o ob ain images is p esen ed, and he in e ace so wa e de eloped is desc ibed.
4.2 Image p e-p ocessing
As i will be desc ibed in Figu e 100 o he lowcha in subchap e 4.7, a e image cap u e he i s s ep
employs he con e sion o he image o g ayscale. This s ep is necessa y e en in he case o using a
came a ha cap u es images only in g ayscale due o he o ma in which he came a na i e so wa e
s o es he images. When hese images a e subsequen ly opened in he "Visual S udio" so wa e, he e is
an impe ec con e sion o pixel alues, which equi es he g ayscale ans o ma ion. A e con e ing he
image o g ayscale (Figu e 84a), he linea iza ion ( h esholding) o he image is ini ia ed. This p ocess is
employed o ans o m he image in o a monoch oma ic o ma wi h he aim o dis inguishing he
backg ound om he ya n. The OTSU linea iza ion me hod was employed, as depic ed in Figu e 84b.
Figu e 84 – a) A ya n in g ayscale (le ); b) Image a e linea iza ion using he OTSU algo i hm ( igh ) [97].
In he nex s ep, a Gaussian il e is applied o he p e ious image (Figu e 85). This il e
in oduces a blu ing e ec o he image, and i s applica ion aims a noise emo al. This app oach will
enable he subsequen image ans o ma ions o be mo e success ul in isola ing o elimina ing desi ed
de ails om he image.
Chap e 4 –
Image P ocessing and Compu e Vision Sys em o ya n analysis
___________________________________________________________________________
93
Figu e 85 - a) G ay scale image (le ); b) Image wi h Gaussian il e ( igh ) [97].
4.3 Ob aining he ya n co e
To acili a e he analysis o he ya n co e exclusi ely, i was necessa y o pe o m mo phological
ope a ions o elimina e he ya n hai iness, po en ial a i ac s in he image, and loose ibe s. The opening
mo phological ope a ion was employed, which in ol es he applica ion o wo consecu i e mo phological
ope a ions using he same s uc u al elemen as a e e ence. Fi s , an e osion ope a ion is applied o he
image in ques ion using a s uc u al elemen , ollowed by a dila ion ope a ion on he same image using
he same s uc u al elemen . These wo consecu i e ope a ions a e e e ed o as an "opening." In e ms
o he s uc u al elemen , also known as he ke nel, a 16 × 16 ma ix was u ilized, whe e all i s elemen s
ha e a alue o 1 [97]. The esul s o his ope a ion a e depic ed in Figu e 86.
Figu e 86 - a) Image p io o he applica ion o he opening mo phological ope a ion (le ); b) Image subsequen
o he implemen a ion o he opening mo phological ope a ion ( igh ) [97].
This p ocedu e allows o begin o disce n he co e o he ya n, which is now clea ly sepa a ed
om all o he de ails in he ini ial image. Howe e , as can be obse ed, he e a e s ill some emnan s o
de ails beyond he co e, pa icula ly he s a ing poin s o he ya n's uzz. Fo his eason, i is necessa y

Chap e 4 –
Image P ocessing and Compu e Vision Sys em o ya n analysis
___________________________________________________________________________
94
o elimina e hese emnan s. To accomplish his, a mo phological closing p ocess is applied. This p ocess
smoo hens he ya n ends, and in ce ain cases whe e aces o ya n uzz a e s ill isible, i emo es any
emaining ya n aces ou side he co e. Figu e 87a) p o ides an example o he ou come.
Figu e 87 - a) Image a e he closing p ocess (le ); b) Isola ed co e image ( igh ) [97].
In his image, he isola ion o he co e can be conside ed comple e. Consequen ly, i becomes
possible o calcula e all he desi ed pa ame e s and cha ac e is ics. Howe e , o u u e ope a ions, as
well as o isual and compu a ional consis ency, a inal in e se "Th esholding" p ocess is ad isable. This
p ocess se es solely o in e he colo s and, consequen ly, he pixel alues. The esul o his
ans o ma ion is illus a ed in Figu e 87b.
Wi h he co e comple ely isola ed, i can begin he ex ac ion o pa ame e s om he ya n. The
i s and mos e iden pa ame e is he ya n diame e , which can be de e mined by coun ing he whi e
pixels in a column o he image. This p ocess is epea ed o all columns, and in he end, he a e age
alue is ob ained o ensu e ha i is no in luenced by he p esence o hick o hin places.
In he case o ideo analysis, he e is an impo an di e en ia ion o highligh . Fo each image,
he ya n diame e is measu ed in only one column o he image. This is done o educe measu emen
edundancy, a oiding measu ing he same sec ion mul iple imes. Addi ionally, i sa es analysis ime and
compu a ional esou ces.
The diame e measu ed in his manne is in pixel uni s, which, on i s own, may no be p ac ical
unless con e ed o a eal-wo ld uni , in his case, millime e s. To con e om pixels o millime e s, he
ac ual leng h o he ya n in he image was measu ed (see Appendix A – Ma hema ical Calcula ion o he
Real Image Size and Lens Magni ica ion Fac o ). Wi h he eal image size known, i is possible o con e
a pixel measu emen o millime e s using Equa ion 28.
𝒅𝒎𝒎 =𝒅𝒑×𝒉
𝟏𝟒𝟒𝟎 (28)
Whe e:
• dmm and dp co espond o he diame e s in millime e s and in pixels, espec i ely;
Chap e 4 –
Image P ocessing and Compu e Vision Sys em o ya n analysis
___________________________________________________________________________
95
• h co esponds o he ac ual heigh o he image, and 1440 co esponds o he numbe o
pixels p esen pe column.
This p ocess can be epea ed o de e mining hick o hin places o he ya n and o de ec ing
neps. Ini ially, he o al numbe o diame e samples is di ided by he desi ed numbe o di isions,
meaning ha he en i e analyzed ya n is di ided in o he sec ions de ined by he use . Nex , he a e age
diame e alue is calcula ed o each o hese sec ions, and his alue is compa ed o he o e all a e age
diame e . The ole ance alue is used as a e e ence poin , om which a sec ion is conside ed ei he
hick o hin. In con as , o neps de ec ion, he dimensional ole ance alue is cons an and se a 200%,
meaning ha sec ions wi h a diame e alue abo e double he a e age a e conside ed a nep.
A highe numbe o di isions esul s in g ea e esolu ion in de ec ing dimensional de ec s in he
ya n. All es s aim o achie e a minimum esolu ion o one sample pe e e y 3 millime e s o ya n, as is
used in USTER machines.
The second and hi d pa ame e s o calcula e a e linea mass and ya n olume, espec i ely.
Re e ing o hei espec i e equa ions, i can be seen ha he calcula ion o linea mass and olume
depends on he leng h o he ya n pe sample, po osi y, ma e ial densi y, and he unknown a iable is he
diame e .
Finally, wi h all he basic ya n pa ame e s calcula ed, i is possible o p oceed o he s a is ical
esul s ela ed o mass, including he mean de ia ion and he coe icien o a ia ion.
4.4 Isola ion o hai iness
To ini ia e he p ocess o isola ing hai iness, i will e u n o he end o he common b anch (Figu e
100), speci ically o he image ea ed wi h he Gaussian il e (Figu e 85b). This image is ini ially
subjec ed o ano he Gaussian il e . The eason o his is ha he con ou de ec ion p ocess ha ollows
is pa icula ly sensi i e o image noise, and o bes de ec he o e all silhoue e o hai iness, i is necessa y
o emo e as much noise as possible.
This p ocess is no pa o he common b anch because, in es s, i was obse ed ha a hea ie
Gaussian il e would comp omise o he esul s, speci ically in pa icle de ec ion (Figu e 86a) and co e
isola ion. A e ob aining an image wi h educed noise, he Canny algo i hm is applied o pe o m a
con ou de ec ion p ocess. The esul ing image will be a silhoue e o he ya n, as shown in Figu e 88.
Chap e 4 –
Image P ocessing and Compu e Vision Sys em o ya n analysis
___________________________________________________________________________
96
.
Figu e 88 - Isola ed ou lines o he ya n.
A e obse a ion, bo h he ya n's silhoue e and ha o i s hai iness a e delinea ed. The goal
om he e is o emo e he whi e pixels in his image, which co espond o he ya n's co e, in o de o
isola e i s hai iness.
To achie e his, he aim is o ob ain an image ha con ains only he co e's pixels, which will be
used o sub ac ion. Fo una ely, one o he p e iously ob ained images can se e his pu pose. Since
he i s b anch o his me hodology was ocused on isola ing he ya n's co e, i s inal image (Figu e 86b)
can be used o his pu pose. Howe e , his image alone canno be used di ec ly o he sub ac ion
p ocess, as image sub ac ion sub ac s he alue o each pixel pai . In he case o sub ac ing wo whi e
pixels (bo h wi h a alue o 255) o wo black pixels (bo h wi h a alue o 0), he esul would be posi i e.
Bu sub ac ing a whi e pixel om a black pixel (0 – 255) would esul in a nega i e alue. Fo his eason,
i s i was needed o ob ain only he pixels ha a e common o bo h he isola ed co e image and he ya n
con ou s image. I was used a logical "AND" ope a ion, which e ains he pixel alues ha a e he same
in bo h images, se ing he o he s o 0, which, in his case, ep esen s a black pixel. An example o he
esul o his ope a ion can be seen in Figu e 89.
Figu e 89 - Con ou s o he isola ed co e.
Chap e 4 –
Image P ocessing and Compu e Vision Sys em o ya n analysis
___________________________________________________________________________
97
Analyzing he igu e, he esul ing image om he logical ope a ion (Figu e 89) isola es he co e's
con ou pixels. Using his image hen is pe o med he sub ac ion ope a ion be ween he ya n con ou s
image (Figu e 88). The esul ing image ob ains he hai iness isola ed (Figu e 90).
Figu e 90 - Con ou s o he isola ed hai iness.
Wi h he hai iness isola ed i is possible o ex ac pa ame e s ela ed o he hai iness. To
de e mine he hai iness coe icien (Equa ion 5), wo pa ame e s om he image a e needed o know: he
ya n's leng h in he image and he leng h o he hai iness. The leng h o he hai iness can be es ima ed
by coun ing he whi e pixels in he image and using he es ablished pixel- o-millime e con e sion.
Howe e , i is impo an o no e ha he images ob ained do no con ain pixels exclusi ely om he ya n's
hai iness, as he ou e co e con ou s a e s ill p esen . To ob ain he ac ual alue o he hai iness leng h
in he image, sub ac ing wice he ya n leng h om he pixel coun is necessa y o elimina e he leng h
measu ed by he co e's con ou pixels.
The second pa ame e o de e mine is he numbe o loose ibe s om he co e, in o he wo ds,
he numbe o loose hai s s ands om he co e. To calcula e his pa ame e , he coun o whi e pixels a
a speci ic dis ance om he co e o he ya n is pe o med in a ow o he image. This coun is done bo h
abo e and below he ya n.
To pe o m his ope a ion, i is necessa y o i s iden i y he uppe and lowe bounda ies o he
ya n co e. Re e ing o Figu e 86b, which shows he isola ed co e, one can analyze column by column o
de e mine he s a and end poin s o he co e. Wi h his in o ma ion i becomes possible o iden i y he
ow ha app oxima ely con ains he co e midpoin . This in o ma ion allows o coun ing he numbe o
loose ibe s loca ed a a ce ain dis ance, bo h abo e and below he co e. I is impo an o no e ha since
he p ocessed image con ains he con ou s o he hai iness, coun ing 2 pixels is equi alen o he p esence
o only one loose end ibe .
Chap e 4 –
Image P ocessing and Compu e Vision Sys em o ya n analysis
___________________________________________________________________________
104
2. Speci ic hai iness da a:
• Hai iness index;
• Numbe o loose ibe s ( his pa ame e is calcula ed only in image analysis, as i becomes
inconsis en in ideo analysis due o he mo emen o he ya n).
3. Faul s da a and s a is ical pa ame e s:
• Thin places;
• Thick places;
• Neps;
• Mean mass de ia ion (U%);
• Coe icien o mass a ia ion (CV%);
• Hai iness s anda d de ia ion (sH%) (calcula ed o ideos/image sequences only, as a
single image does no p o ide enough da a o an exac esul ).
4. Video and ame p ocessing:
• To al ideo ame;
• P ocessed ames;
• F ames wi h e o .
Ha ing p esen ed all he p og am inpu and ou pu pa ame e s, i will now be desc ibed he
me hodology used o ob ain he desi ed esul s. While image and ideo analysis me hodologies di e due
o he o me ocusing on single images and he la e in ol ing mul iple images, he me hodology o
analyzing a single image emains qui e simila . The p ima y dis inc ion lies in he epe i ion o p ocesses
o each image wi hin a ideo du ing ideo analysis, wi h esul s agg ega ed a he end.
Gi en he di e se ange o da a equi ed om a single image, he e is no single sequence o
p ocesses ha p o ides all he necessa y in o ma ion. To ob ain he h ee di e en ypes o images,
namely:
• an image wi h an insula ed ya n co e;
• an image wi h isola ed hai s;
• an image wi h isola ed winding pa icles.

Chap e 4 –
Image P ocessing and Compu e Vision Sys em o ya n analysis
___________________________________________________________________________
105
A mul i ace ed app oach in ol ing a ious image p ocessing echniques is employed. The ollowing
lowcha (Figu e 100) illus a es his mul i ace ed app oach used o gene a e each o he h ee image
ypes.
Figu e 100 – Algo i hm Analysis Me hodology Flowcha [92].
Wi h his ool i is possible o p esen in o ma ion in a legible and in ui i e way o he use .
Chap e 4 –
Image P ocessing and Compu e Vision Sys em o ya n analysis
___________________________________________________________________________
106
Summa y o he chap e :
This chap e p o ided a comp ehensi e pe spec i e on he elemen s in ol ed in de eloping image
p ocessing algo i hms. These algo i hms a e designed o e alua e ya n quali y by analyzing i s a ibu es,
such as linea mass, speci ic olume, ya n de ec s, and he le el o hai iness.
5. DEEP LEARNING TECHNIQUES FOR YARN HAIRINESS ANALYSIS
O e iew
The ollowing chap e desc ibes he me hods employed o he de ec ion and classi ica ion o hai iness in
ya n using a new imp o ed YOLO 5s6 con olu ional neu al ne wo k (CNN). In his chap e , a
comp ehensi e b eakdown o he YOLO 5s6 de ec ion me hod is p o ided, elucida ing i s cons i uen
blocks and he enhancemen s ha ha e been in eg a ed. A concise o e iew o he Bo -T ans o me
module, MHSA (Mul i-Head Sel -A en ion) blocks, and he C2 (coa se- o- ine) module, all in eg al
componen s o he op imized YOLO 5s6 Hai iness algo i hm, is p esen ed. Addi ionally, he ac i a ion
unc ions employed in bo h he de aul YOLO 5s6 algo i hm, and he a ian u ilized in he op imized
YOLO 5s6-Hai iness algo i hm a e de ailed, along wi h he a ionale behind hese selec ions.
5.1 YOLO 5s6 Ne wo k – P oposed A chi ec u e
5.2 Bo -T ans o me Module & MHSA Block
5.3 C2 Module
5.4 Hype pa ame e s
5.5 Ac i a ion unc ion GeLU
Chap e 5 –
Deep lea ning echniques o ya n hai iness analysis
___________________________________________________________________________
108
5.1 YOLO 5s6 Algo i hm – P oposed A chi ec u e
Among he a ious e sions o YOLO 5, he YOLO 5s6 a chi ec u e was chosen o he s udy
[75], [76], [77]. Se e al p elimina y es s we e ca ied ou , and i was obse ed ha his model
ou pe o med o he algo i hm a ian s in e ms o o e all esul s.
The decision o op o he YOLO 5s6 e sion ins ead o YOLO 8 (mos ecen e sion) was d i en
by se e al ac o s [102]:
1. Pe o mance supe io i y: Th ough p elimina y es ing, i was e iden ha he YOLO 5s6 model
ou pe o med o he algo i hm a ian s, including YOLO 8, in e ms o o e all esul s.
2. P io i y on p ocessing speed: Gi en he emphasis on speed, pa icula ly c ucial in applica ions
equi ing ins an de ec ion and classi ica ion o ya n ibe s, YOLO 5s6 ema kable ames pe second
(FPS) a e made i he p e e ed choice.
3. U iliza ion o GPU (G aphics P ocessing Uni ) esou ces: Ope a ing in an en i onmen wi h
GPU access, YOLO 5s6 eme ged as he op imal selec ion due o i s op imized pe o mance o GPUs.
Le e aging GPU esou ces signi ican ly accele a ed he p ocess o ibe de ec ion in ya n.
4. Balance be ween accu acy and speed: While YOLO 8 o e s ma ginal highe accu acy,
YOLO 5s6 main ains a compe i i e equilib ium be ween accu acy and speed. This balance aligns
wi h he equi emen s o he de eloped applica ion, necessi a ing bo h accu acy and swi esponse
imes.
5. Use - iendliness: YOLO 5s6 s ands ou as he mo e use - iendly op ion compa ed o YOLO 8.
Buil on he PyTo ch amewo k, i o e s seamless u iliza ion and deploymen o de elope s,
enhancing ease o use du ing applica ion de elopmen and deploymen p ocesses.
Recognizing he hai iness de ec ion pe o mance o he YOLO 5s6 model, an enhanced e sion
named YOLO 5s6-Hai iness was de eloped. This imp o ed model inco po a es a Bo ans o me , a C2
(Coa se- o-Fine) module, and a GeLU (Gaussian E o Linea Uni ) ac i a ion unc ion, as depic ed in
Figu e 101.
Chap e 5 –
Deep lea ning echniques o ya n hai iness analysis
___________________________________________________________________________
109
Figu e 101 - Imp o ed a chi ec u e based on YOLO 5s6 – YOLO 5s6-Hai iness [76].
The imp o ed YOLO 5s6-Hai iness model inco po a es h ee key modi ica ions when compa ed
o he o iginal/de aul YOLO 5s6:
• In he o iginal YOLO 5s6 a chi ec u e, he las C3_1 laye in he Backbone, posi ioned jus be o e
he s anda d SPPF (spa ial py amid pooling as ) module, has been eplaced by he C2 module,
which is a ea u e bo owed om he mo e ad anced YOLO 8 objec de ec ion model [103]. This
change enhances he algo i hm pe o mance and op imiza ion o i s speci ic neu al s uc u e;
• Wi hin he Neck po ion o he a chi ec u e, he p e ious C3_1_F module has been eplaced by
he Bo - ans o me module. This adjus men se es o enhance he model's e iciency in
de ec ing looped ibe s wi hin he cap u ed images;
• The CBS (Con + Ba chNo m + SiLU) module, which plays a undamen al ole in he a chi ec u e
by pe o ming asks like downsampling, dimensionali y augmen a ion and educ ion,
no maliza ion, and applying non-linea i ies o ea u e maps, has unde gone a ans o ma ion. I
has been eplaced by he CBG (Con olu ional, Ba ch No maliza ion, GeLU) module, in oducing

Chap e 5 –
Deep lea ning echniques o ya n hai iness analysis
___________________________________________________________________________
110
a new GeLU ac i a ion unc ion ins ead o he o iginal SiLU (Sigmoid Linea Uni ). This block
consis s o a Con olu ional laye , a Ba ch No maliza ion laye , and he GeLU ac i a ion unc ion
(Figu e 102).
Figu e 102 - The CBG module in he op imized YOLO 5s6-Hai iness [104].
This CBG block is u ilized in mul iple o he blocks (Figu e 103) and in wo dis inc bo leneck blocks
(Figu e 104).
Figu e 103 - The CBG module employed in a ious o he blocks [104].
Figu e 104 - The CBG module is used in wo di e en ypes o bo leneck blocks [104].
The C3 module plays a i al ole in ea u e ex ac ion, comp ising h ee CBG modules and se e al
s acked Bo leneck blocks. The no a ion C3_x indica es he p esence o x s acked Bo leneck blocks. In
Figu e 105, he ea u e map is di ided in o wo b anches a e en e ing he C3 module. One b anch
a e ses he CBG and Bo lenecks, while he o he solely passes h ough he CBG. Subsequen ly, he wo
b anches a e conca ena ed and pass h ough ano he CBG module.
Chap e 5 –
Deep lea ning echniques o ya n hai iness analysis
___________________________________________________________________________
111
Figu e 105 - Two dis inc ypes o CSP blocks (C3) [104].
In each Bo leneck block, wo CBG modules a e inco po a ed. The i s module conduc s a 1x1
con olu ion, educing he channel coun by hal , while he second module pe o ms a 3x3 con olu ion,
doubling he numbe o channels. The ini ial educ ion in dimensionali y aids in enhancing he
unde s anding o ea u e in o ma ion by he con olu ional ke nel, while he subsequen inc ease in
dimensionali y acili a es he ex ac ion o mo e de ailed ea u es. To mi iga e he anishing g adien
p oblem, a esidual s uc u e is employed, ensu ing he addi ion o he inpu and ou pu .
The " anishing g adien p oblem" is a well- ecognized challenge in aining deep neu al ne wo ks,
pa icula ly hose wi h nume ous laye s [76], [77], [105]. This issue a ises du ing backp opaga ion when
g adien s diminish signi ican ly as hey p opaga e backwa d h ough mul iple laye s. Consequen ly, his
esul s in slow o no lea ning in he ea ly laye s, ende ing he aining o deep ne wo ks a duous.
Se e al echniques ha e been de eloped o add ess he anishing g adien p oblem [76], [77], [105]:
1. Ac i a ion Func ions: U ilizing ac i a ion unc ions like ReLU p omo es he low o g adien s,
alle ia ing he anishing g adien p oblem;
2. Ba ch No maliza ion: The use o ba ch no maliza ion no malizes laye inpu s, s abilizing g adien
low du ing aining;
3. Skip Connec ions: The inco po a ion o skip connec ions di ec ly connec s ea lie laye s o la e
laye s, acili a ing smoo he g adien low and mi iga ing he anishing g adien p oblem.
By implemen ing hese echniques, deep lea ning models can e ec i ely ain in deep ne wo ks wi h
mul iple laye s, leading o enhanced con e gence and imp o ed pe o mance.
The p ima y unc ion o he C3 module is o cap u e in ica e pa e ns and de ails in images as he
ne wo k's dep h inc eases. A his s age, 3x3 con olu ions a e widely employed as hey enable he neu al
ne wo k o lea n iche and mo e abs ac ep esen a ions o he objec s p esen ed in he images.
Chap e 5 –
Deep lea ning echniques o ya n hai iness analysis
___________________________________________________________________________
112
When analyzing loop ibe s and p o uding ibe s in ya ns using he YOLO 5 model, each block
assumes a speci ic ole in enhancing he de ec ion o hese hai iness ypes:
1. Backbone:
• Func ion: The Backbone is asked wi h ex ac ing ele an ea u es om ya n images,
encompassing low and mid-le el in o ma ion like edges, ex u es, and pa e ns in he
ya n;
• Impac on de ec ion: An e icien backbone cap u es dis inc i e ea u es o loop ibe s
and p o uding ibe s in ya n. This ea u e ex ac ion p ocess con ibu es o mo e
accu a e subsequen de ec ion.
2. Neck:
• Func ion: The Neck se es as an in e media e componen ha amalgama es in o ma ion
a a ious scales, enabling he model o de ec objec s o di e en sizes and con ex s;
• Impac on de ec ion: Gi en he a ying sizes o loop ibe s and p o uding ibe s in ya n,
he Neck acili a es he cap u e o con ex ual in o ma ion ac oss mul iple scales. This is
pa icula ly ad an ageous o he accu a e de ec ion o smalle o mo e di use hai iness.
3. Head:
• Func ion: The Head is esponsible o objec de ec ion, gene a ing p edic ions ega ding
bounding boxes and objec classes wi hin he ya n ibe s,
• Impac on de ec ion: The Head empowe s he model o pinpoin and classi y loop ibe s
and p o uding ibe s. Bounding box p edic ions speci y he loca ion and size o he
hai iness, while class p edic ions disce n whe he hey a e loop ibe s o p o uding
ibe s.
Each block plays a pi o al ole in he de ec ion o loop ibe s and p o uding ibe s, wi h he
Backbone ex ac ing pe inen ea u es, he Neck consolida ing in o ma ion a di e se scales, and he
Head execu ing he inal de ec ion. This comp ehensi e app oach p o ides he equisi e in o ma ion o
p ecise hai iness iden i ica ion.
Fu he mo e, in he op imized algo i hm, he C2 module was in oduced, a no el elemen de i ed
om YOLO 8, o eplace he C3_1 module in he Backbone sec ion o he a chi ec u e. The C2 module
in eg a es concep s om he C3 module and ELAN (E icien Laye Agg ega ion Ne wo k) [104], [106].
Figu e 106 illus a es he wo dis inc ions be ween he C2 and C3 modules.
Chap e 5 –
Deep lea ning echniques o ya n hai iness analysis
___________________________________________________________________________
113
Figu e 106 - Di e ence be ween C3_x ( op) and C2 module (bo om) [104].
The dispa i y be ween he C3 and C2F blocks can be delinea ed based on hei s uc u e,
objec i e, and applica ion [104], [106]:
1) S uc u e:
C3 Block: The C3 block se es as a s eamlined i e a ion o he CSP block u ilized in YOLO 4 [104],
[106]. I comp ises a sequence o 1x1 and 3x3 con olu ional laye s, ollowed by s acked Bo leneck
modules. A key ea u e o he C3 block is he u iliza ion o a skip connec ion o amalgama e he o iginal
ea u es wi h hose acqui ed wi hin he block.
C2F Block: The C2F block, deno ed as he CSP2 (C oss S age Pa ial Ne wo k 2) block, ex ends he
o iginal CSP block ound in YOLO [104], [106]. I in oduces a no el spli in he inpu , c ea ing wo dis inc
pa hways (main and seconda y). The seconda y pa hway a e ses h ough addi ional con olu ional
Bo leneck modules. Subsequen ly, he ou pu om hese Bo leneck modules is conca ena ed wi h he
ou pu om he main pa hway, leading o he inal con olu ional laye s.
2) Objec i e:
C3 Block: The p ima y aim o he C3 block is o p o ide a ligh e and mo e e icien al e na i e o he
CSP block, op imizing compu a ional pe o mance while p ese ing objec de ec ion accu acy.
C2F Block: In con as , he C2F block is speci ically designed o enhance he capaci y o cap u ing
in o ma ion a a ied esolu ion le els. By pa i ioning he inpu in o wo pa hways and inco po a ing
Chap e 5 –
Deep lea ning echniques o ya n hai iness analysis
___________________________________________________________________________
120
Wi hin he con ex o he p oposed op imized algo i hm YOLO 5s6-Hai iness, he placemen o he
C2F block in he YOLO 5s6 a chi ec u e can exe a signi ican in luence on he de ec ion o loop ibe s
and p o uding ibe s wi hin ya ns. In he YOLO 5s6-Hai iness algo i hm, he C2F block was posi ioned
be o e he SPPF block. The C2F block's unc ion lies in i s capaci y o cap u e and use in o ma ion a
di e se esolu ion le els, while he SPPF block is esponsible o ea u e usion h ough spa ial py amids,
hus augmen ing he ea u e ep esen a ion wi hou he loss o c ucial con ex ual in o ma ion.
Rega ding i s impac on he de ec ion o loop ibe s and p o uding ibe s, in oducing he C2F block
p io o he SPPF block enabled he model o e ec i ely g asp essen ial con ex ual and spa ial in o ma ion
ac oss di e en scales. This capabili y is pa icula ly ad an ageous in he iden i ica ion o in ica e objec s
like loop and p o uding ibe s wi hin ya ns.
5.4 Hype pa ame e s
In he speci ic con ex o de ec ing loop ibe s and p o uding ibe s in ya ns, signi ican changes
we e implemen ed in high-le el hype pa ame e s. The modi ied hype pa ame e s [107] encompassed
"ancho _ " and "scale" wi h he ollowing alues:
ancho _ : 5.0
• The alue ancho _ : 5.0 designa es he h eshold employed o assigning ancho boxes o objec s
in he images, se a 5.0;
• Ancho boxes se e he pu pose o p edic ing he posi ion and size o objec s de ec ed by he
model. Ancho s whose dimensions a e oughly i e imes la ge han he s anda d ancho
dimensions a e exclusi ely conside ed when ancho _ is se o 5.0;
• This adjus men inc eased he model's capabili y o selec la ge and mo e sui able ancho s o
he de ec ion o la ge o unusually sized objec s wi hin he images. This was especially ele an
in he case o objec classes ha signi ican ly exceeded he ypical ancho dimensions o he
YOLO 5s6 model;
• The ele a ion o his pa ame e helped diminish he occu ence o alse posi i e de ec ions, a
aluable cha ac e is ic when he aining da a is p one o noise o complex backg ounds, as is
he case wi h loop ibe s;
• In he pa icula con ex o de ec ing loop ibe s and p o uding ibe s in ya ns, his adap a ion
was ins umen al o accu a ely cap u ing s uc u es ha span ex ensi e a eas and exhibi a ied

Chap e 5 –
Deep lea ning echniques o ya n hai iness analysis
___________________________________________________________________________
121
sizes. Fo ins ance, loop ibe s o p o uding ibe s ha ex end ac oss subs an ial egions o he
image could be mo e e ec i ely de ec ed, gi en ha he images had a esolu ion o 1280x1280
pixels.
scale: 2.0
• A scale o 2.0 was employed, which ansla ed o esizing he images o app oxima ely +/-200%
o 300% beyond o below he s anda d image esolu ion;
• The inc ease in scale esul ed in la ge images, which p o ed o be ad an ageous o he de ec ion
o in ica e de ails in loop ibe s and p o uding ibe s wi hin ya ns;
• I is essen ial o highligh ha esizing images o such an ele a ed scale subs an ially augmen ed
he compu a ional demands, necessi a ing addi ional memo y and p ocessing esou ces;
• These adjus men s played a pi o al ole in achie ing accu a e and comp ehensi e de ec ion.
Gi en ha he images we e o 1280x1280-pixel esolu ion, signi ying a la ge scale, he model's
capaci y o iden i y hese s uc u es wi h g ea e eliabili y was signi ican ly enhanced.
The selec ion o he speci ic alues o "ancho _ " and "scale" was con ingen on he unique
cha ac e is ics o he ya n da ase and he na u e o he loop ibe s and p o uding ibe s ha we e he
a ge o de ec ion.
5.5 Ac i a ion unc ion GeLU
The choice o he GeLU ac i a ion unc ion was pe o med a e a se ies o es s and a ho ough
ma hema ical analysis, aking in o accoun he ollowing conside a ions [108]:
• Cap u ing Complex Non-linea i ies: Loop ibe s and p o uding ibe s exhibi a di e se ange o
shapes and ex u es, ende ing he de ec ion ask in ica e. The GeLU unc ion, dis inguished by i s
heigh ened non-linea i y compa ed o SiLU, is adep a cap u ing in ica e and sub le ela ionships
among he pa e ns inhe en in he ibe s and hei isual cha ac e is ics. This enhancemen
subs an ially augmen s he model's capaci y o acqui e he capabili y o disce n a ious ypes o ya n
hai iness;
• Reduc ion o Vanishing G adien s: Deep neu al ne wo ks can con on he challenge o anishing
g adien s, pa icula ly when employing he SiLU ac i a ion unc ion, cha ac e ized by de i a i es ha
app oach ze o o la ge o nega i e inpu s. GeLU, on he o he hand, boas s smoo h de i a i es ac oss
all poin s, e ec i ely mi iga ing he issue o anishing g adien s and ins illing g ea e s abili y in o he
aining p ocess;
Chap e 5 –
Deep lea ning echniques o ya n hai iness analysis
___________________________________________________________________________
122
• Imp o ed T aining S abili y: GeLU, which app oxima es he cumula i e Gaussian dis ibu ion
unc ion, o e s an addi ional laye o aining s abili y. This s abili y, in u n, u nishes he neu al
ne wo k wi h he capaci y o ine- une i s weigh s mo e p ecisely, ul ima ely enhancing i s abili y o
de ec loop ibe s and p o uding ibe s;
• Supe io Pe o mance Me ics: GeLU consis en ly deli e ed supe io esul s in me ics when
compa ed o o he ac i a ion unc ions, a de ailed exposi ion o which will be p esen ed in chap e 6.
Taking hese ac o s in o accoun , GeLU eme ged as he mos adequa e o he ac i a ion unc ion
in he p oposed op imized YOLO 5 algo i hm.
Summa y o he chap e :
The chap e p esen ed a comp ehensi e explo a ion o c ucial componen s in he enhanced YOLO 5s6
Hai iness algo i hm, ocusing on he Bo -T ans o me Module, MHSA Block, and C2 Module. I explo es
hei in eg a ion and ope a ional mechanisms, showcasing how hese elemen s enhance he neu al
ne wo k's capaci y in in ica e asks such as objec de ec ion and ins ance segmen a ion. Addi ionally, i
p o ides insigh s in o he a ionale behind choosing he GeLU ac i a ion unc ion, highligh ing i s bene i s
in cap u ing complex pa e ns, educing anishing g adien s, ensu ing aining s abili y, and deli e ing
supe io pe o mance me ics. O e all, his comp ehensi e analysis con ibu es signi ican ly o he
imp o ed de ec ion and classi ica ion o hai iness in ya n using he op imized YOLO 5s6 algo i hm.
6. MECHATRONIC PROTOTYPE DEVELOPMENT
O e iew
This chap e p esen s he design o a specialized p o o ype ailo ed o mee indus y demands. The
de elopmen p ocess and key conside a ions a e p esen ed, encompassing mechanical and au oma ion
domains. This p o o ype au oma ically unwinds and winds ya n while allowing i o mo e swi ly pas a
came a lens wi hin p ede ined ole ances and con olled ligh ing condi ions
6.1 Mecha onic sys ems design
6.2 Requi emen s
6.3 Me hodology
6.4 Func ions o a ya n es ing equipmen
6.5 Au oma ic sys em design
6.6 Con olle
6.7 Subsys ems and kinema ic chains
6.8 De elopmen o he au oma ion con ol p og am o he ya n unwinding, conduc ion, and
winding Sys em
6.9 Main enance and emo e access o he mecha onic p o o ype
6.10 Tes s and esul s
6.11 Compa ison be ween he de eloped p o o ype and he o he sys ems ci ed in s a e o a
6.12 C i ical Re iew analysis
Chap e 6 – Mecha onic P o o ype De elopmen
___________________________________________________________________________
124
6.1 Mecha onic sys ems design
The VDI 2206 s anda d is a Ge man echnical s anda d ha p o ides guidelines and me hodologies
o he de elopmen o mecha onic sys ems. I add esses issues ela ed o he de elopmen p ocess,
om iden i ying equi emen s o alida ing and e i ying he inal sys em. The VDI 2206 s anda d is known
o i s sys ema ic app oach and he "V" model ha desc ibes he phases o mecha onic p oduc
de elopmen . I is used by enginee s and indus y p o essionals o ensu e e iciency and quali y in
mecha onic sys em de elopmen [109].
The objec i e o VDI 2206 is o p o ide a suppo me hodology o he de elopmen o mecha onic
sys ems. The main cha ac e is ic o VDI is o be he p ocess, me hod, and ool o he ini ial phase o
de elopmen , ocused on sys em design. The esul o he sys em design is he mecha onic sys em
concep ( he solu ion is es ablished a e e i ica ion and alida ion). This VDI guide is a supplemen o
he VDI 2221 s anda d and VDI 2422 [109].
Acco ding o he VDI 2206 s anda d, he de elopmen o a mecha onic p oduc ollows a "V"
model. In he model depic ed in Figu e 106, cus ome needs a e ans o med in o equi emen s.
Subsequen ly, hese equi emen s a e ans o med in o speci ica ions, as hey de ine he me ic
cha ac e is ics ha he mecha onic p oduc mus mee . A e his, he nex necessa y s ep is Sys em
Design [109].
In he de elopmen o mecha onic sys ems, i is essen ial o ollow an app op ia e me hodology o
ensu e e iciency and use sa is ac ion. The VDI 2206 s anda d [110] o e s an impo an me hodological
amewo k ha encompasses he mul idisciplina y inhe en in such sys ems, in eg a ing elec onic,
mechanical, and in o ma ion echnology componen s.
A he hea o his me hodology is he V-Model, a amewo k ha p oposes a sequence o s eps o
mecha onic sys ems design. Howe e , i is impo an o no e ha no all au oma ion p ojec s s ic ly i
in o a mul idisciplina y app oach.
Following his, he s anda d ela ed o equi emen s, hei ca ego ies, and hei signi icance in
shaping sys ems ha align wi h use and s akeholde needs will be discussed. This explo a ion aims o
es ablish a obus g oundwo k o he e ec i e de elopmen o au oma ion sys ems [110].
The VDI 2206 s anda d p o ides a me hodological amewo k o he mul idisciplina y de elopmen
o mecha onic sys ems, pa icula ly hose ha in ol e elec onic and mechanical componen s in
Chap e 6 – Mecha onic P o o ype De elopmen
___________________________________________________________________________
125
symbiosis wi h in o ma ion echnology. This me hodology e ol es a ound he cen al elemen known as
he V-model (Figu e 109) [110].
Figu e 109 – V-Model o he Design o a Mecha onic Sys em - VDI 2206 [110].
The V-Model p oposes a sequence o s eps o he design o mecha onic sys ems, which a e [110]:
1. Requi emen s - This is he i s s age, consis ing o desc ibing he sys em o be ob ained in he
o m o equi emen s agains which he esul ing sys em will be subsequen ly e alua ed;
2. Concep ual Sys em Design - This s age allows o he c ea ion o a mul idisciplina y solu ion
concep ha leads o he equi ed cha ac e is ics o he sys em o be ob ained. To achie e his,
he main unc ion is di ided in o sub- unc ions, o which indi idual solu ion o ope a ion p inciples
a e ound. Thei pe o mance is e alua ed wi hin he con ex o he sys em ac oss a ious
domains;

Chap e 6 – Mecha onic P o o ype De elopmen
___________________________________________________________________________
126
3. Speci ic Sys em Design - This s ep in ol es de eloping sec ions o he global solu ion ela ed o
a ious a eas in ol ed (Mechanical Enginee ing, Elec onics, Compu e Science, among o he
domains);
4. In eg a ion - The esul s o a ious componen s o he global de eloped solu ion wi hin each
domain a e in eg a ed o ob ain an o e all solu ion o he sys em;
5. P ope y Ve i ica ion - This s age in ol es compa ing he p ope ies o he de eloping sys em wi h
hose de ined o i and is associa ed wi h he concep s o e i ica ion and alida ion. Ve i ica ion
in ol es o mally analyzing whe he he sys em mee s all p oposed equi emen s. Valida ion
in ol es in o mally checking i he sys em is sui able, ypically done in an in o mal manne . The
p ope y e i ica ion p ocess con inues h oughou he p ojec , al hough i is o mally loca ed
du ing sys em in eg a ion. This can be done i ually, physically, o in a mixed app oach;
6. Modeling and Model Analysis - This s age uns pa allel o all s eps excep equi emen s and
in ol es assessing he sys em's p ope ies using models o compu e -based simula ion ools.
Ano he me hodology ollowed o his mecha onic p o o ype, is based on he a ional
me hodology p esen ed by Nigel C oss, which is simila o bu wi h some imp o emen s o e he
me hodology ou lined in he Ge man enginee ing s anda d VDI 2221 [110].
The gene al me hodology p esen ed in VDI 2221 consis s o se en phases:
1. Cla i ica ion o objec i es;
2. Func ion s uc u e;
3. Solu ion p inciples and hei combina ions;
4. Di ision in o modules;
5. De elopmen o he main modules;
6. De elopmen o all modules;
7. P oduc ion p epa a ion.
Howe e , he "Requi emen s" s age emains o undamen al impo ance, as i se es as he
ounda ion o he p ojec . Requi emen s a e he building blocks upon which he au oma ion sys em will
Chap e 6 – Mecha onic P o o ype De elopmen
___________________________________________________________________________
127
be cons uc ed, and hei p ope de ini ion is essen ial o ensu e ha he sys em mee s he needs and
expec a ions o he use s. This is whe e a jou ney o unde s and equi emen s begins, a opic o u mos
signi icance in he de elopmen o he au oma ion sys em o winding and unwinding ya n [111].
The "Requi emen s" se es as a means o ensu e ha he sys em mee s he p oposed needs
and expec a ions, cons i u ing he p ima y inpu o he p ojec . In his ega d, he ollowing concep s
ela ed o his heme a e p esen ed [111].
6.2 Requi emen s
In he con ex o de eloping he mecha onic p o o ype, his subchap e plays a cen al ole. The
objec i es se o his esea ch in ol e he design o a mecha onic de ice capable o e icien ly winding
and unwinding he ya n, enabling in-dep h analyses o a ious aw ma e ials. To achie e his challenge,
i is c ucial o es ablish a comp ehensi e unde s anding o he equi emen s unde lying his mecha onic
sys em. The "Requi emen s" ho oughly explo es he p ope ies ha he de ice mus possess o mee he
needs and cons ain s o use s, highligh ing he dis inc ion be ween explici and implici , unc ional, and
non- unc ional equi emen s. These equi emen s o m he ounda ion upon which he mecha onic
p o o ype will be designed and cons uc ed, ensu ing i ul ills i s pu pose o e icien ly winding and
unwinding he ya n in pe ec alignmen wi h he hesis objec i es.
Requi emen s a e he p ope ies ha a pa icula sys em mus possess o mee he needs and
cons ain s imposed by he use . Speci ically applied o he ield o au oma ion/so wa e, hey can be
di ided in o wo ca ego ies: explici and implici , and unc ional and non- unc ional.
Explici equi emen s a e he elemen s di ec ly eques ed by clien s, while implici equi emen s
a e hose included by he de elopmen eam based on hei echnical knowledge, e en i hey ha e no
been men ioned by he clien .
Func ional equi emen s a e elemen s associa ed wi h he sys em's unc ionali y, while non-
unc ional equi emen s ela e o aspec s o i s pe o mance. Func ional equi emen s should be cohe en
and co e all he sys em's needs, cha ac e izing i s esponse o ce ain s imuli wi hou being bound o
design and implemen a ion aspec s in o de o maximize he ange o solu ions o explo e.
Non- unc ional equi emen s, by hei na u e, do no in luence he sys em's unc ionali ies o be
de eloped and can be classi ied in o eigh ypes: appea ance, usabili y, pe o mance, ope abili y
Chap e 6 – Mecha onic P o o ype De elopmen
___________________________________________________________________________
128
(cha ac e is ics ha enable he sys em o ope a e p ope ly in i s en i onmen ), main enance and suppo ,
secu i y, cul u al and poli ical aspec s, and legal aspec s.
6.3 Me hodology
Following he goal ee me hod, he main objec i es conside ed o he p ojec we e six [97]:
1. Sa e y;
2. Simplici y;
3. Reliabili y;
4. Po abili y;
5. Ve sa ili y;
6. Lowe cos .
F om he objec i es, i was hen necessa y o quan i y and de ine some pa ame e s ha he
machine mus mee o achie e he goals. The bounda ies se , acco ding o he me hod o pe o mance
speci ica ions, we e as ollows in Table 3.
Table 3 - P ojec Pe o mance Speci ica ions
Objec i e
Values
Reason o choosing his op ion
Analyze a ious
ype o ya n
ma e ials
The machine should be compa ible
wi h co on and polyes e ;
B eaking Fo ce ≥ 1 N;
Compa ible wi h a ia ions in
b eaking ension ≤ |10|%.
Ve sa ili y: Analyzing a ious ypes o ya n ma e ials,
diame e s, and bobbin dimensions allows he
p o o ype o be e sa ile, accommoda ing di e en
ma e ials and sizes commonly used in ex ile
indus ies. This e sa ili y enables b oade
applica ions and compa ibili y wi h di e se ya n
ypes, ca e ing o a wide ange o p oduc ion needs.
Analyze a ious
ype o ya n
diame e s
Compa ible wi h diame e s ≤ 2 mm.
Compa ible wi h a ia ions in
diame e ≤ |10|%.
Analyze a ious
bobbin
dimensions
Compa ible wi h bobbins wi h an
ou e base diame e ≤ 160 mm and
heigh ≤ 200 mm.
Adhe ence o
Speed ≤ 4.2 m/s wi h a maximum
Accu acy and P ecision: Adhe ence o speed and
Chap e 6 – Mecha onic P o o ype De elopmen
___________________________________________________________________________
129
speed
ole ances
a ia ion o 0.5%.
posi ion ole ances ensu es ha he machine
ope a es wi h high accu acy and p ecision. This is
c ucial in ya n analysis, as p ecise measu emen s
a e necessa y o quali y con ol and consis ency in
p oduc ion.
Adhe ence o
posi ion
ole ance
De ia ion in posi ion ≤ 1 mm.
Machine and
use sa e y
In case o an eme gency, he
machine should s op wi hin ≤ 1
second.
Sa e y: P io i izing machine and use sa e y is
undamen al. Ensu ing ha he machine s ops
p omp ly in case o eme gencies, as well as ha ing
simple and in ui i e se up p ocedu es, con ibu es o
a sa e wo king en i onmen and educes he isk o
acciden s o mishaps du ing ope a ion.
Simple, in ui i e,
and quick se up
Se up ime ≤ 2 minu es.
Du abili y and Main enance: Reducing wea o
mo ing componen s is essen ial o he du abili y
and longe i y o he machine. By minimizing wea
and ea , main enance equi emen s a e educed,
leading o cos sa ings, and ensu ing consis en
pe o mance o e ime.
Reduced wea o
mo ing
componen s
Op imal ope a ing condi ions ≥ 10
km.
Quali y Con ol: A oiding ya n b eakage is c i ical o
main aining he in eg i y o he ya n being analyzed.
I ensu es ha he p ocess emains unin e up ed
and p e en s ma e ial was e, allowing o accu a e
analysis wi hou comp omising quali y.
A oid ya n
b eakage
A oid ≤ 5 ya n b eaks pe 10 km.
Reduced Mass
Mass ≤ 30 kg.
Po abili y and P ac icali y: Reducing he mass and
olume o he machine makes i mo e po able and
space e icien . This aspec is aluable o indus ies
whe e space is limi ed o whe e mobili y and
lexibili y a e necessa y, enabling easie
anspo a ion and ins alla ion in di e en se ings.
Reduced
Volume
Volume ≤ 0.275 m3.
Lowe Cos
1.300,00€ (only ha dwa e and
so wa e)
This cos ad an age allows o mo e accessible ini ial
in es men , making he p o o ype inancially mo e
easible o es ing o de elopmen pu poses
compa ed o he signi ican ly highe cos o he
USTER es e 3.
232
[37] P. K. Sahoo, S. Sol ani, and A. K. C. Wong, “A su ey o h esholding echniques,”
Compu e
Vision, G aphics and Image P ocessing
, ol. 41, no. 2. 1988. doi: 10.1016/0734-
189X(88)90022-9.
[38] “OpenCV: Miscellaneous Image T ans o ma ions.” Accessed: Oc . 16, 2023. [Online]. A ailable:
h ps://docs.openc .o g/4.x/d7/d1b/g oup__imgp oc__misc.h ml
[39] Mu huk ishnan, “O su’s me hod o image h esholding explained and implemen ed,”
Mu huk ishnan
. 2020.
[40] L. To ok, “Mé odo de O su,” Ins i u o de Compu ação – Uni e sidade Fede al Fluminense (UFF).
[41] Emanuele. T ucco and Alessand o. Ve i, “In oduc o y echniques o 3-D compu e ision,” p.
343, 1998, Accessed: Oc . 16, 2023. [Online]. A ailable:
h ps://www. esea chga e.ne /publica ion/220692096_In oduc o y_ echniques_ o _3-
D_compu e _ ision
[42] A. C. Bo ik and S. T. Ac on, “Basic Linea Fil e ing wi h Applica ion o Image Enhancemen ,” in
The Essen ial Guide o Image P ocessing
, 2009. doi: 10.1016/B978-0-12-374457-9.00010-X.
[43] L. Ding and A. Gosh asby, “On he canny edge de ec o ,”
Pa e n Recogni
, ol. 34, no. 3, 2001,
doi: 10.1016/S0031-3203(00)00023-6.
[44] S. Anand, T. Thi ya, and S. Jee a, “Edge De ec ion using Di ec ional Fil e Bank,”
In J Appl In
Sys
, ol. 1, no. 4, pp. 21–27, May 2012, doi: 10.5120/ijais12-450162.
[45] “OpenCV: Mo phological T ans o ma ions.” Accessed: Oc . 17, 2023. [Online]. A ailable:
h ps://docs.openc .o g/4.x/d9/d61/ u o ial_py_mo phological_ops.h ml
[46] Ma lab, “ ein o cemen Lea ning o Enginee s, Pa 1: Wha Is ein o cemen Lea ning? -
ouTube,” Ma lab ouTube Channel.
[47] J. de A. B. A aujo-Filho, A. N. Assunção Júnio , M. A. Gu ie ez, and C. . Nomu a, “A i icial
In elligence and Ca diac Imaging: We need o alk abou his,”
ARQUIVOS BRASILEIROS DE
CARDIOLOGIA - IMAGEM CARDIOVASCULAR
, ol. 32, no. 3, 2019, doi: 10.5935/2318-
8219.20190034.
[48] P. Manickam
e al.
, “A i icial In elligence (AI) and In e ne o Medical Things (IoMT) Assis ed
Biomedical Sys ems o In elligen eal hca e,”
Biosenso s
, ol. 12, no. 8. 2022. doi:
10.3390/bios12080562.
[49] “Algo i mos de clus e ing - Visão ge al.” Accessed: Oc . 17, 2023. [Online]. A ailable:
h ps://isolu ion.p o/p / /machine-lea ning-wi h-py hon/clus e ing-algo i hms-
o e iew/algo i mos-de-clus e ing- isao-ge al

233
[50] Isha Salian, “Di e ence Be ween Supe ised, Unsupe ised, & ein o cemen Lea ning |
NVIDIA Blog,” nVidia.
[51] A. M. Obeso, J. Benois-Pineau, M. S. Ga cía Vázquez, and A. Á. . Acos a, “Visual s in e nal
a en ion mechanisms in deep neu al ne wo ks o image classi ica ion and ob ec de ec ion,”
Pa e n Recogni
, ol. 123, 2022, doi: 10.1016/j.pa cog.2021.108411.
[52] J. edmon, S. Di ala, . Gi shick, and A. Fa hadi, “ ou only look once: Uni ied, eal- ime objec
de ec ion,” in
P oceedings o he IEEE Compu e Socie y Con e ence on Compu e Vision and
Pa e n Recogni ion
, 2016. doi: 10.1109/CVPR.2016.91.
[53] “C_5. olo V4 explained in ull de ail-EN - Deep Lea ning Bible - 4. Objec De ec ion - Eng.”
Accessed: Oc . 17, 2023. [Online]. A ailable: h ps://wikidocs.ne /167833
[54] “Non-maximum Supp ession (NMS). A Technique o emo e duplica es and… | by
Sambasi a ao. K | Towa ds Da a Science.” Accessed: Oc . 17, 2023. [Online]. A ailable:
h ps:// owa dsda ascience.com/non-maximum-supp ession-nms-93ce178e177c
[55] “Ob ec T acking wi h DeepSO T and OLO-NAS: A P ac i ione ’s Guide | Deci.” Accessed: Oc .
17, 2023. [Online]. A ailable: h ps://deci.ai/blog/objec - acking-wi h-deepso -and-yolo-nas-
p ac i ione s-guide/
[56] “Gi ub - ul aly ics/yolo 5 a blog. obo low.com.” Accessed: Oc . 17, 2023. [Online]. A ailable:
h ps://gi hub.com/ul aly ics/yolo 5? e =blog. obo low.com
[57] “Wha a e Ancho Boxes in Ob ec De ec ion?” Accessed: Oc . 17, 2023. [Online]. A ailable:
h ps://blog. obo low.com/wha -is-an-ancho -box/
[58] “Wha is OLO 5? A Guide o Beginne s.” Accessed: Oc . 17, 2023. [Online]. A ailable:
h ps://blog. obo low.com/yolo 5-imp o emen s-and-e alua ion/
[59] “A 5-laye dense block wi h a g ow h a e o k = 4. Each laye akes all... | Download Scien i ic
Diag am.” Accessed: Oc . 17, 2023. [Online]. A ailable: h ps://www. esea chga e.ne / igu e/A-
5-laye -dense-block-wi h-a-g ow h- a e-o -k-4-Each-laye - akes-all-p eceding_ ig2_306885833
[60] “E icien De : Scalable and E icien Ob ec De ec ion”, Accessed: Oc . 17, 2023. [Online].
A ailable: h ps://gi hub.com/google/au oml/ ee/
[61] L. Zhu, J. Zhang, and C. Jia, “An Imp o ed OLO 5-based Me hod o Su ace De ec De ec ion
o S eel Pla e,” in
P oceedings - 2022 Chinese Au oma ion Cong ess, CAC 2022
, 2022. doi:
10.1109/CAC57257.2022.10055055.
234
[62] J. Zhang, J. Zhang, K. Zhou, . Zhang, . Chen, and X. an, “An Imp o ed OLO 5-Based
Unde wa e Objec -De ec ion F amewo k,”
Senso s
, ol. 23, no. 7, 2023, doi:
10.3390/s23073693.
[63] “Con usion Ma ix o ou Mul i-Class Machine Lea ning Model | by Joydwip Mohajon |
Towa ds Da a Science.” Accessed: Oc . 17, 2023. [Online]. A ailable:
h ps:// owa dsda ascience.com/con usion-ma ix- o -you -mul i-class-machine-lea ning-model-
9aa3b 7826
[64] “Mean A e age P ecision (mAP): De ini ion, Me ics, and Applica ions in Compu e Vision |
Label ou Da a.” Accessed: Oc . 17, 2023. [Online]. A ailable:
h ps://labelyou da a.com/a icles/mean-a e age-p ecision-map
[65] “Mean A e age P ecision (mAP) Explained: E e y hing ou Need o Know.” Accessed: Oc . 17,
2023. [Online]. A ailable: h ps://www. 7labs.com/blog/mean-a e age-p ecision
[66] O. C onie, M. Mo adi, and C. A. N. Biscio, “A c oss- alida ion-based s a is ical heo y o poin
p ocesses,”
Biome ika
, 2023, doi: 10.1093/biome /asad041.
[67] N. Gha ami
e al.
, “Au oma ic segmen a ion o p os a e M I using con olu ional neu al
ne wo ks: In es iga ing he impac o ne wo k a chi ec u e on he accu acy o olume
measu emen and MRI-ul asound egis a ion,”
Med Image Anal
, ol. 58, 2019, doi:
10.1016/j.media.2019.101558.
[68] D. Niu, . Liang, . Wang, M. Wang, and W. C. ong, “Icing o ecas ing o ansmission lines
wi h a modi ied Back P opaga ion Neu al Ne wo k-Suppo Vec o Machine-Ex eme Lea ning
Machine wi h Ke nel (BPNN-SVM-KELM) based on he a iance-co a iance weigh de e mina ion
me hod,”
Ene gies (Basel)
, ol. 10, no. 8, 2017, doi: 10.3390/en10081196.
[69] A. M. And ew, “An In oduc ion o Suppo Vec o Machines and O he Ke nel-based Lea ning
Me hods,”
Kybe ne es
, ol. 30, no. 1. 2001. doi: 10.1108/k.2001.30.1.103.6.
[70] V. . Phung and E. J. hee, “A igh-accu acy model a e age ensemble o con olu ional neu al
ne wo ks o classi ica ion o cloud image pa ches on small da ase s,”
Applied Sciences
(Swi ze land)
, ol. 9, no. 21, 2019, doi: 10.3390/app9214500.
[71] Q. en, M. Li, and S. an, “Tec onic disc imina ion o oli ine in basal using da a mining
echniques based on ma o elemen s: a compa a i e s udy om mul iple pe spec i es,”
Big
Ea h Da a
, ol. 3, no. 1, 2019, doi: 10.1080/20964471.2019.1572452.
[72] X. Zhang and C. A. Liu, “Model a e aging p edic ion by K- old c oss- alida ion,”
J Econom
, ol.
235, no. 1, 2023, doi: 10.1016/j.jeconom.2022.04.007.
235
[73] E. Se inç, “An empowe ed AdaBoos algo i hm implemen a ion: A COVID-19 da ase s udy,”
Compu Ind Eng
, ol. 165, 2022, doi: 10.1016/j.cie.2021.107912.
[74] “ OLO Algo i hm o Ob ec De ec ion Explained [ Examples].” Accessed: Oc . 17, 2023.
[Online]. A ailable: h ps://www. 7labs.com/blog/yolo-objec -de ec ion
[75] Z. Guo, . uang, Z. Shan, J. uang, Z. ou, and W. Li, “A digi al implan a ion sys em o Z-
di ec ion ya n o h ee-dimensional p e o m based on lexible o ien ed wo en p ocess,”
Eng Appl
A i In ell
, ol. 116, 2022, doi: 10.1016/j.engappai.2022.105385.
[76] A. Z. Sellam
e al.
, “Deep Lea ning Solu ion o Quan i ica ion o Fluo escence Pa icles on a
Memb ane,”
Senso s
, ol. 23, no. 4, 2023, doi: 10.3390/s23041794.
[77] . u and Z. Zhu, “Sim-YOLO 5s: A me hod o de ec ing de ec s on he end ace o li hium
ba e y s eel shells,”
Ad anced Enginee ing In o ma ics
, ol. 55, 2023, doi:
10.1016/j.aei.2022.101824.
[78] . K. Jung and G. S. Choi, “Imp o ed OLO 5: E icien Ob ec De ec ion Using D one Images
unde Va ious Condi ions,”
Applied Sciences (Swi ze land)
, ol. 12, no. 14, 2022, doi:
10.3390/app12147255.
[79] Z. ang, “Ac i a ion Func ion: Cell ecogni ion Based on oLo 5s/m,”
Jou nal o Compu e and
Communica ions
, ol. 09, no. 12, 2021, doi: 10.4236/jcc.2021.912001.
[80] S. El wing, E. Uchibe, and K. Doya, “Sigmoid-weigh ed linea uni s o neu al ne wo k unc ion
app oxima ion in ein o cemen lea ning,”
Neu al Ne wo ks
, ol. 107, 2018, doi:
10.1016/j.neune .2017.12.012.
[81] J. Dohe y, B. Ga dine , E. Ke , N. Siddique, and S. S. Man i, “Compa a i e S udy o Ac i a ion
Func ions and Thei Impac on he OLO 5 Ob ec De ec ion Model,” in
Lec u e No es in
Compu e Science (including subse ies Lec u e No es in A i icial In elligence and Lec u e No es
in Bioin o ma ics)
, 2022. doi: 10.1007/978-3-031-09282-4_4.
[82] “Abou hype pa ame e s · ul aly ics/yolo 5 · Discussion #7319.” Accessed: Oc . 22, 2023.
[Online]. A ailable: h ps://gi hub.com/ul aly ics/yolo 5/discussions/7319
[83] Z. Li, P. Zhong, X. Tang, . Chen, S. Su, and T. Zhai, “A New Me hod o E alua e a n
Appea ance Quali ies Based on Machine Vision and Image P ocessing,”
IEEE Access
, ol. 8,
2020, doi: 10.1109/ACCESS.2020.2972967.
[84] J. Sun, J. Jia, C. K. Tang, and . . Shum, “Poisson ma ing,” in
ACM SIGGRAPH 2004 Pape s,
SIGGRAPH 2004
, 2004. doi: 10.1145/1186562.1015721.
236
[85] . Sun, Z. Li, . Pan, J. Zhou, and W. Gao, “Measu emen o long ya n hai based on hai iness
segmen a ion and hai iness acking,”
Jou nal o he Tex ile Ins i u e
, ol. 108, no. 7, 2017, doi:
10.1080/00405000.2016.1240144.
[86] J. Jing, M. uang, P. Li, and X. Ning, “Au oma ic measu emen o ya n hai iness based on he
imp o ed M M F segmen a ion algo i hm,”
Jou nal o he Tex ile Ins i u e
, ol. 109, no. 6,
2018, doi: 10.1080/00405000.2017.1368106.
[87] N. aleem, M. Bus eo, and A. Del Bue, “A compu e ision based online quali y con ol sys em
o ex ile ya ns,”
Compu Ind
, ol. 133, 2021, doi: 10.1016/j.compind.2021.103550.
[88] A. Soleimanipou and G. . Chegini, “A ision-based hyb id app oach o iden i ica ion o
An hu ium lowe cul i a s,”
Compu Elec on Ag ic
, ol. 174, 2020, doi:
10.1016/j.compag.2020.105460.
[89] A. El-Geiheini, S. ElKa eb, and M. R. Abd-Elhamied, “ a n Tensile P ope ies Modeling Using
A i icial In elligence,”
Alexand ia Enginee ing Jou nal
, ol. 59, no. 6, 2020, doi:
10.1016/j.aej.2020.07.049.
[90] B. Yang
e al.
, “Pa ame e iden i ica ion o p o on exchange memb ane uel cell ia Le enbe g-
Ma qua d backp opaga ion algo i hm,”
In J Hyd ogen Ene gy
, ol. 46, no. 44, 2021, doi:
10.1016/j.ijhydene.2021.04.130.
[91] M. R. Abd-Elhamied, W. A. Hashima, S. ElKa eb, I. Elhawa y, and A. El-Geiheini, “P edic ion o
Co on a n’s Cha ac e is ics by Image P ocessing and ANN,”
Alexand ia Enginee ing Jou nal
,
ol. 61, no. 4, 2022, doi: 10.1016/j.aej.2021.08.057.
[92] F. Pe ei a
e al.
, “In elligen Compu e Vision Sys em o Analysis and Cha ac e iza ion o a n
Quali y,”
Elec onics (Swi ze land)
, ol. 12, no. 1, 2023, doi: 10.3390/elec onics12010236.
[93] X. Luo, Z. Cheng, Q. Ni, . Tao, and . Shi, “De ec de ec ion algo i hm o ab ic based on
de o mable con olu ional ne wo k,”
Tex ile Resea ch Jou nal
, ol. 93, no. 9–10, 2023, doi:
10.1177/00405175221143742.
[94] . Jin and Q. Niu, “Au oma ic Fab ic De ec De ec ion Based on an Imp o ed OLO 5,”
Ma h
P obl Eng
, ol. 2021, 2021, doi: 10.1155/2021/7321394.
[95] G. Lin, K. Liu, X. Xia, and . an, “An E icien and In elligen De ec ion Me hod o Fab ic
De ec s based on Imp o ed OLO 5,”
Senso s
, ol. 23, no. 1, 2023, doi: 10.3390/s23010097.
[96] . Zuo and . Du, “De ec ion me hod o he e o opic ibe based on imp o ed OLO 5,”
MATEC
Web o Con e ences
, ol. 358, 2022, doi: 10.1051/ma eccon /202235801008.
237
[97] P. Caldas, F. Sousa, F. Pe ei a, . Lopes, and J. Machado, “Au oma ic sys em o ya n quali y
analysis by image p ocessing,”
Jou nal o he B azilian Socie y o Mechanical Sciences and
Enginee ing
, ol. 44, no. 11, 2022, doi: 10.1007/s40430-022-03875-3.
[98] “Au omação Indús ial | OM ON, Po ugal.” Accessed: No . 26, 2023. [Online]. A ailable:
h ps://indus ial.om on.p /p /home
[99] “STC Se ies (USB3.0 Se ies) USB3.0 Se ies/Fea u es | OM ON Indus ial Au oma ion India.”
Accessed: No . 26, 2023. [Online]. A ailable: h ps://www.om on-
ap.co.in/p oduc s/ amily/3874/
[100] “Visual S udio: IDE and Code Edi o o So wa e De elope s and Teams.” Accessed: No . 27,
2023. [Online]. A ailable: h ps:// isuals udio.mic oso .com/
[101] “OpenCV - Open Compu e Vision Lib a y.” Accessed: No . 27, 2023. [Online]. A ailable:
h ps://openc .o g/
[102] E. Casas, L. Ramos, E. Bendek, and F. Ri as-Eche e ia, “Assessing he E ec i eness o OLO
A chi ec u es o Smoke and Wild i e De ec ion,”
IEEE Access
, ol. 11, 2023, doi:
10.1109/ACCESS.2023.3312217.
[103] Z. Zhang, “D one-YOLO: An E icien Neu al Ne wo k Me hod o Ta ge De ec ion in D one
Images,”
D ones
, ol. 7, no. 8, 2023, doi: 10.3390/d ones7080526.
[104] G. u and X. Zhou, “An Imp o ed OLO 5 C ack De ec ion Me hod Combined wi h a Bo leneck
T ans o me ,”
Ma hema ics
, ol. 11, no. 10, 2023, doi: 10.3390/ma h11102377.
[105] A. M. oy and J. Bhadu i, “DenseSP -YOLO 5: An au oma ed damage de ec ion model based
on DenseNe and Swin-T ans o me p edic ion head-enabled YOLO 5 wi h a en ion
mechanism,”
Ad anced Enginee ing In o ma ics
, ol. 56, 2023, doi:
10.1016/j.aei.2023.102007.
[106] A. S ini as, T. . Lin, N. Pa ma , J. Shlens, P. Abbeel, and A. Vaswani, “Bo leneck ans o me s
o isual ecogni ion,” in
P oceedings o he IEEE Compu e Socie y Con e ence on Compu e
Vision and Pa e n Recogni ion
, 2021. doi: 10.1109/CVPR46437.2021.01625.
[107] H.-P.-T. Van and V.-D. oang, “Insula o De ec ion in In elligen Moni o ing Based on olo Family
and Cus omizing ype pa ame e s,”
Jou nal o Technical Educa ion Science
, no. 75A, 2023,
doi: 10.54644/j e.75a.2023.1308.
[108] B. Singh, S. Pa el, A. Vi ay a giya, and . Kuma , “Analyzing he impac o ac i a ion unc ions
on he pe o mance o he da a-d i en gai model,”
Resul s in Enginee ing
, ol. 18, 2023, doi:
10.1016/j. ineng.2023.101029.

238
[109] I. G aessle and J. en ze, “The new V-Model o VDI 2206 and i s alida ion das Neue V-Modell
de VDI 2206 und seine Validie ung,”
A -Au oma isie ungs echnik
, ol. 68, no. 5, 2020, doi:
10.1515/au o-2020-0015.
[110] J. Gausemeie and S. Moeh inge , “VDI 2206- A New Guideline o he Design o Mecha onic
Sys ems,”
IFAC P oceedings Volumes
, ol. 35, no. 2, pp. 785–790, Dec. 2002, doi:
10.1016/S1474-6670(17)34035-1.
[111] C. C. Insau alde and A. Zoi l, “Sys em equi emen s in indus ial au oma ion: A e iew o
Modeling Me hodologies o Con ol So wa e A chi ec u es,” in
IEEE In e na ional Con e ence on
Indus ial In o ma ics (INDIN)
, 2013. doi: 10.1109/INDIN.2013.6622947.
[112] “(PDF) indin2012 JMachado ESeab a FINAL mai 2012.” Accessed: Oc . 29, 2023. [Online].
A ailable:
h ps://www. esea chga e.ne /publica ion/258764397_indin2012_JMachado_ESeab a_FINAL
_mai_2012
[113] G. Ba bie i and D. A. Gu ie ez, “A GEMMA-GRAFCET Me hodology o enable Digi al Twin based
on Real-Time Coupling,” in
P ocedia Compu e Science
, 2021. doi:
10.1016/j.p ocs.2021.01.122.
[114] E. S. J. Machado, “T a amen o da Pa agem de Eme gência – Caso de Aplicação,” Guima ães,
2007.
[115] . amana han, “The IEC 61131-3 p og amming languages ea u es o indus ial con ol
sys ems,” in
Wo ld Au oma ion Cong ess P oceedings
, 2014. doi:
10.1109/WAC.2014.6936062.
[116] “USE ’S MANUAL S SMAC CP Se ies CP2E CPU Uni So wa e”.
[117] F. Pe ei a, L. Magalhães, A. A. San os, A. F. da Sil a, K. An osz, and J. Machado, “De elopmen
o an Au oma ed Wooden andle Packaging Sys em wi h In eg a ed Coun ing Technology,”
Machines 2024, Vol. 12, Page 122
, ol. 12, no. 2, p. 122, Feb. 2024, doi:
10.3390/MACHINES12020122.
[118] F. Pe ei a and C. Felguei as, “Lea ning Au oma ion om emo e Labs in ighe Educa ion,” in
ACM In e na ional Con e ence P oceeding Se ies
, 2020. doi: 10.1145/3434780.3436689.
[119] “Ewon Flexy | IIoT Da a Ga eway | Ewon.” Accessed: No . 27, 2023. [Online]. A ailable:
h ps://www.ewon.biz/p oduc s/ewon- lexy
[120] F. Pe ei a and J. Machado,
Sis emas de Acesso Remo o a Máquinas e P ocessos Indus iais –
Pa e I
, 2023 d ed., ol. 22. Po o: Publindús ia, 2023.
239
[121] J. G. Pin o, J. Mon ei o, . Vasconcelos, and F. O. Soa es, “A new sys em o di ec
measu emen o ya n mass wi h 1mm accu acy,” in
P oceedings o he IEEE In e na ional
Con e ence on Indus ial Technology
, 2002. doi: 10.1109/ICIT.2002.1189336.
[122] A. Spa a igna, E. B oglia, and S. Lugli, “Beyond capaci i e sys ems wi h op ical measu emen s
o ya n e enness e alua ion,”
Mecha onics
, ol. 14, no. 10, 2004, doi:
10.1016/j.mecha onics.2004.07.002.
[123] F. Pe ei a, L. Pin o, J. Machado, F. Soa es, . Vasconcelos, and V. Ca alho, “ a n ai iness -
Loop & P o uding Fibe s Da ase ,” ol. 1, 2023, doi: 10.17632/DKV6J6FW6C.1.
[124] “ obo low: Gi e you so wa e he powe o see ob ec s in images and ideo.” Accessed: Oc .
24, 2023. [Online]. A ailable: h ps:// obo low.com/
[125] D. Mah o and S. C. ada , “Emo ion p edic ion o ex ual da a using GloVe based eBi-
CuDNNLSTM model,”
Mul imed Tools Appl
, 2023, doi: 10.1007/s11042-023-16062-w.
[126] I. S. Isa, M. S. A. osli, U. K. uso , M. I. F. Ma uzuki, and S. N. Sulaiman, “Op imizing he
ype pa ame e Tuning o OLO 5 o Unde wa e De ec ion,”
IEEE Access
, ol. 10, 2022, doi:
10.1109/ACCESS.2022.3174583.
[127] “Tips o Bes T aining esul s - Ul aly ics OLO 8 Docs.” Accessed: No . 27, 2023. [Online].
A ailable: h ps://docs.ul aly ics.com/yolo 5/ u o ials/ ips_ o _bes _ aining_ esul s/#model-
selec ion
[128] “En enda dis ância ocal, ângulo de isão e cí culo de imagem – apenas imagens.” Accessed:
Oc . 29, 2023. [Online]. A ailable: h ps://apenasimagens.com/p /dis ancia- ocal-ci culo-
imagem/
APPENDIXES
A.1 Ma hema ical Calcula ion o he Real Image Size and Lens Magni ica ion Fac o
A.2 Full sys em con olle speci ica ion
A.3 Highe Hie a chy G a ce - Coo dina ion o Ope a ing Modes
A.4 P ede ined sc eens o HMI ins alled on he machine body
A.5 LabVIEW in e ace - Local compu e
A.6 De eloped Wo k Dissemina ion
Appendixes
___________________________________________________________________________
241
A.1 Ma hema ical Calcula ion o he Real Image Size and Lens
Magni ica ion Fac o
Be o e any measu emen s can be made on a cap u ed image, i is necessa y o p ecisely know
he dimensions o a e e ence objec om which all o he desi ed dimensions can be ex apola ed. The
bes e e ence objec is he pixel, as i is he elemen al objec o he image and has a simple geome ic
shape (a squa e).
To ob ain he eal dimensions o a pixel, he aim is o de e mine he dimensions o he en i e
image. Then, i is simply a ma e o di iding he wid h o he image by he numbe o pixels in a ow o
he image.
To de e mine he dimensions o he image a he heigh o he ya n, he eal dimensions o he
image backg ound we e i s calcula ed. Knowing he dis ance be ween he image backg ound and he
came a, as well as he dis ance be ween he came a and he ya n, allows o a compa a i e calcula ion
o ob ain he eal leng h o he ya n in he image. This is possible because he a ea o he image sec ion
inc eases p opo ionally wi h he dis ance be ween he sec ion and he came a lens, as demons a ed in
Figu e A.1-1.
Figu e A.1-1 - Field o View Visualiza ion [128]
To ob ain he ac ual size o he image backg ound, an image was cap u ed wi h a backg ound
con aining a known 2-millime e squa e. Then, a simple ule o h ee was applied o ind he o al leng h
o he image (Y), which is 4.23 millime e s. Knowing ha he image dimension is 1440x1080 pixels, he
heigh o he image can be di ec ly calcula ed as 3.17 millime e s.
Wi h he size o he image backg ound calcula ed, he dis ance be ween he image backg ound
and he came a (Z) and he dis ance be ween he came a and he ya n (B) we e measu ed, esul ing in
41.6 millime e s and 12.8 millime e s, espec i ely.
Wi h all hese dimensions, a simple ule o h ee can be applied o calcula e he ac ual leng h o
Appendixes
___________________________________________________________________________
248
G a ce C
Figu e A.2-9 - G a ce C - Tes Mode

Appendixes
___________________________________________________________________________
249
G a ce o Tes Te mina ion Mode
Figu e A.2-10 - G a ce o Tes Te mina ion Mode – Pa 1
Appendixes
___________________________________________________________________________
250
Figu e A.2-11 - G a ce o Tes Te mina ion Mode – Pa 2
Appendixes
___________________________________________________________________________
251
Figu e A.2-12 - G a ce o Tes Te mina ion Mode – Pa 3
Appendixes
___________________________________________________________________________
252
G a ce o S op Mode
Figu e A.2-13 - G a ce o S op Mode – Pa 1
Appendixes
___________________________________________________________________________
253
Figu e A.2-14 - G a ce o S op Mode – Pa 2

Appendixes
___________________________________________________________________________
254
Manual Mode G a ce
G a ce A
Figu e A.2-15 - Manual Mode G a ce – G a ce A – Pa 1
Figu e A.2-16 - Manual Mode G a ce – G a ce A – Pa 2
Appendixes
___________________________________________________________________________
255
Figu e A.2-17 - Manual Mode G a ce – G a ce A – Pa 3
Figu e A.2-18 - Manual Mode G a ce – G a ce A – Pa 4
G a ce C
Figu e A.2-19 - Manual Mode G a ce – G a ce C
Appendixes
___________________________________________________________________________
256
A.3 Highe Hie a chy G a ce - Coo dina ion o Ope a ing Modes
Figu e A.3-1- Highe Hie a chy G a ce - Coo dina ion o Ope a ing Modes
G a ce o Tes Mode
G a ce A
Figu e A.3-2 - G a ce o Tes Mode - G a ce A – Pa 1
Appendixes
___________________________________________________________________________
257
Figu e A.3-3 - G a ce o Tes Mode - G a ce A – Pa 2
Appendixes
___________________________________________________________________________
264
Figu e A.4- 3 - P ede ined sc eens o HMI ins alled on he machine body – Pa 3

Appendixes
___________________________________________________________________________
265
Figu e A.4- 4 - P ede ined sc eens o HMI ins alled on he machine body – Pa 4
Figu e A.4- 5 - P ede ined sc eens o HMI ins alled on he machine body – Pa 5
Appendixes
___________________________________________________________________________
266
A.5 LabVIEW in e ace - Local compu e
Figu e A.5- 1 - LabVIEW in e ace - Local compu e
Appendixes
___________________________________________________________________________
267
A.6 De eloped Wo k Dissemina ion
All con ibu ions esul ing om he de eloped wo k a e desc ibed below:
F.1. Jou nal publica ions
[1] Filipe Pe ei a, Leand o Pin o, Filomena Soa es, Rosa Vasconcelos, José Machado and Ví o
Ca alho. A No el Deep Lea ning-Based Ya n Hai iness Cha ac e iza ion Me hodology Using an Imp o ed
YOLO 5 Algo i hm.
IEEE Access 2024. Submi ed
[2] Filipe Pe ei a, Leand o Pin o, Filomena Soa es, Rosa Vasconcelos, José Machado and Vi o
Ca alho, ONLINE YARN HAIRINESS - LOOP & PROTRUDING FIBERS DATASET, Da a in B ie , 2024,
110355, ISSN 2352-3409, h ps://doi.o g/10.1016/j.dib.2024.110355.
[3] Filipe Pe ei a, Alexand e Macedo, Leand o Pin o, Filomena Soa es, Rosa Vasconcelos, José
Machado and Vi o Ca alho, In elligen Compu e Vision Sys em o Analysis and Cha ac e iza ion o Ya n
Quali y, Elec onics, 2023,
12
, 236. h ps://doi.o g/10.3390/elec onics12010236.
F.2. Con e ence publica ions
[1] Filipe Pe ei a, José Machado, Filomena Soa es, Rosa Vasconcelos and Vi o Ca alho, Design o a
Ya n Analysis Mecha onic P o o ype o Tex ile Indus y, DESIGNCOMMIT 2024, B aga, Po ugal, 20-22
May 2024 (In p ess).
[2] Rica do Pin o, Filipe Pe ei a, Vi o Ca alho, Filomena Soa es and Rosa Vasconcelos, Ya n linea
mass de e mina ion using image p ocessing: i s insigh s, IECON 2019 - 45 h Annual Con e ence o he
IEEE Indus ial Elec onics Socie y, Lisbon, Po ugal, 2019, pp. 198-203,
doi:10.1109/IECON.2019.8926650.
[3] Filipe Pe ei a, Vi o Ca alho, Rosa Vasconcelos, Filomena Soa es, A Re iew in he use o A i icial
In elligence in Tex ile Indus y, Inno a ions in Mecha onics Enginee ing – ICIENG, 2021, Lec u e No es
in Mechanical Enginee ing, Sp inge , h ps://doi.o g/10.1007/978-3-030-79168-1_34.
Appendixes
___________________________________________________________________________
268
F.3. Book publica ions
[1] Filipe Pe ei a, Vi o Ca alho, Filomena Soa es, Rosa Vasconcelos, José Machado, 6 - Compu e
ision echniques o de ec ing ya n de ec s, Tex ile Ins i u e Book Se ies, Applica ions o Compu e Vision
in Fashion and Tex iles, Woodhead Publishing, 2018, Pages 123-145, ISBN 9780081012178,
h ps://doi.o g/10.1016/B978-0-08-101217-8.00006-3.
F.4. Da ase publica ion
[1] Filipe Pe ei a, Leand o Pin o, José Machado, Filomena Soa es, Rosa Vasconcelos and Vi o
Ca alho, Ya n Hai iness - Loop & P o uding Fibe s Da ase , Mendeley Da a, 2023, V1,
h ps://da a.mendeley.com/da ase s/dk 6j6 w6c/1