Automated garment recognition and assessment system using computer vision and artificial intelligence for blind people

Uni e sidade do Minho
Escola de Engenha ia
Daniel Filipe Coelho Rocha
Au oma ed Ga men Recogni ion and
Assessmen Sys em Using Compu e
Vision and A i icial In elligence o Blind
People
ou ub o de 2024
Au oma ed Ga men Recogni ion and Assessmen Sys em Using Compu e Vision and
A i icial In elligence o Blind People
Daniel Rocha
UMinho | 2024
Daniel Filipe Coelho Rocha
Au oma ed Ga men Recogni ion and
Assessmen Sys em Using Compu e Vision
and A i icial In elligence o Blind People
Tese de Dou o amen o
P og ama Dou o al em Engenha ia Ele ónica e de
Compu ado es
T abalho e e uado sob a o ien ação do
P o esso Dou o Ví o Hugo Mendes da Cos a Ca alho
P o esso a Dou o a Filomena Ma ia da Rocha Menezes de
Oli ei a Soa es
ou ub o de 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
DECLARAÇÃO
Nome: Daniel Filipe Coelho da Rocha
Ende eço ele ónico: [email p o ec ed] Tele one: 917966638 (al e na i o)
Bilhe e de Iden idade/Ca ão do Cidadão: 14147147
Tí ulo da ese: Au oma ed Ga men Recogni ion and Assessmen Sys em Using Compu e Vision and
A i icial In elligence o Blind
O ien ado es: P o esso Dou o Ví o Hugo Mendes da Cos a Ca alho e P o esso a Dou o a Filomena
Ma ia da Rocha Menezes de Oli ei a Soa es
Ano de conclusão: 2024
Designação do Ramo de Conhecimen o do Dou o amen o: Engenha ia Ele ónica e de Compu ado es
DE ACORDO COM A LEGISLAÇÃO EM VIGOR, NÃO É PERMITIDA A REPRODUÇÃO DE QUALQUER PARTE
DESTA TESE/TRABALHO.
Uni e sidade do Minho, 28/10/2024
Assina u a: __________________________________________________

iii
STATEMENT OF INTEGRITY
I he eby decla e ha ing conduc ed my hesis wi h in eg i y. I con i m ha I ha e no used plagia ism o
any o m o alsi ica ion o esul s in he p ocess o he hesis elabo a ion.
I u he decla e ha I ha e ully acknowledge he Code o E hical Conduc o he Uni e si y o Minho.
Uni e si y o Minho, 28 h Oc obe 2024
Full name: Daniel Filipe Coelho Rocha
Signa u e:
____________________________________________________________________
i
ACKNOWLEDGEMENTS
Th oughou his jou ney, I aced nume ous challenges ha made me ques ion and doub mysel .
Howe e , hese obs acles only se ed o s eng hen my esilience and inc ease my de e mina ion o
succeed, su passing mysel each day. Bu none o his would ha e been possible wi hou he igh people
by my side, and o ha , I wan o exp ess my g a i ude o all hose who accompanied me on his jou ney.
To my supe iso , P o esso Ví o Ca alho, I am deeply g a e ul o his guidance, aluable eachings, and
o he excellence, s eng h, and esilience he has always demons a ed. His en husiasm in he pu sui
o knowledge was a cons an sou ce o inspi a ion and di ec ion h oughou his jou ney.
To my co-supe iso , P o esso Filomena Soa es, I ex end my hanks o he ad ice, high s anda ds, and
p o essionalism, as well as o he excellen eedback ha allowed me o imp o e e e y day.
To P o esso Celina P. Leão, I exp ess my g a i ude o he suppo and gene ous sha ing o knowledge
in he ield o s a is ics, always a ailable wi h en husiasm and joy o help.
To all he p o essionals, especially hose a ACAPO B aga, who we e always willing o collabo a e on he
p ojec , my since e hanks. I would like o highligh he con ibu ion o M . Filipe Aze edo, sec e a y o
ACAPO B aga, o his ull a ailabili y, dedica ion, and he en husias ic way he emb aced he p ojec ,
always p oposing imp o emen s o bene i he blind communi y.
To my iend Toninho, I am deeply g a e ul o he suppo and help h oughou his jou ney. Fo he
aluable scien i ic knowledge he sha ed wi h me and o he p ecious hou s he spen wi h me du ing he
oughes momen s, p o ing ha he impossible was, indeed, possible. To you, my deepes hanks o
e e y hing! To my iend João, my hanks o his cons an suppo , his willingness o sha e his as
expe ience and knowledge, and o always being eady o help.
To my amily, his achie emen is also you s. Toge he , we aced days o challenges and o e coming, and
i was you uncondi ional suppo ha ga e me he s eng h o keep pu suing his d eam. None o his
would ha e been possible wi hou he i eless dedica ion and sac i ices o my wi e, who ga e up so much
o allow me o ly eely, always o e ing he bes .
Las ly, I dedica e his achie emen o my daugh e . F om he momen you we e bo n, my jou ney ook
on a new pu pose: o be an example o you and o show you ha any hing is possible when we belie e.
Wa ching you g ow du ing his ime ga e me he esponsibili y o inspi e you, and I hope ha one day,
when you ead hese wo ds, hey will mo i a e you o pu sue you own d eams wi h he same
de e mina ion.
AGRADECIMENTOS
Ao longo des e pe cu so, en en ei inúme os desa ios que me ize am ques iona e du ida de mim
mesmo. No en an o, o am esses obs áculos que o alece am a minha esiliência e aumen a am a minha
on ade de conquis a , supe ando-me a cada dia. Po ém, nada dis o se ia possí el sem as pessoas ce as
ao meu lado, e po isso, que o exp essa a minha g a idão a odos os que me acompanha am nes a
jo nada.
Ao meu o ien ado , P o esso Ví o Ca alho, ag adeço po odo o acompanhamen o, pelos ensinamen os
aliosos, pela excelência, o ça e esiliência que semp e demons ou. O seu en usiasmo na p ocu a de
conhecimen o oi uma on e cons an e de inspi ação e o ien ação ao longo de odo es e pe cu so.
À minha coo ien ado a, P o esso a Filomena Soa es, ag adeço os conselhos, a exigência e o
p o issionalismo, bem como o excelen e eedback que me pe mi i am melho a a cada dia.
À P o esso a Celina P. Leão, exp esso a minha g a idão pelo apoio e pela pa ilha gene osa de
conhecimen o na á ea de es a ís ica, es ando semp e disponí el, com en usiasmo e aleg ia, pa a ajuda .
A odos os p o issionais, especialmen e à ACAPO de B aga, que se mos a am semp e disponí eis pa a
colabo a com o p oje o, o meu since o ag adecimen o. Gos a ia de des aca a colabo ação do S . Filipe
Aze edo, sec e á io da ACAPO de B aga, pela sua o al disponibilidade e dedicação, e pelo en usiasmo
com que ab açou o p oje o, semp e a p opo melho ias que bene iciem a comunidade cega.
Ao meu amigo Toninho, ag adeço p o undamen e pelo supo e e ajuda ao longo des e pe cu so. Pelo
alioso conhecimen o cien í ico que pa ilhou comigo e pelas ho as p eciosas que me dedicou nos
momen os mais di íceis, demons ando que o impossí el e a, a inal, possí el. A i, o meu mui o ob igado
po udo! Ao meu amigo João, o meu ag adecimen o pelo apoio cons an e, pela disponibilidade em
pa ilha a sua as a expe iência e conhecimen o, e po es a semp e p on o a ajuda .
À minha amília, es a conquis a é ambém ossa. Jun os en en ámos dias de desa ios e supe ação, e
oi o osso apoio incondicional que me deu o ças pa a con inua a pe segui es e sonho. Nada disso
e ia sido possí el sem a dedicação incansá el e os sac i ícios da minha esposa, que abdicou de an o
pa a me pe mi i oa li emen e, semp e o e ecendo o seu melho .
Po úl imo, dedico es a conquis a à minha ilha. Desde o momen o em que nasces e, a minha jo nada
ganhou um no o p opósi o: se um exemplo pa a i e mos a - e que udo é possí el quando ac edi amos.
Ve - e c esce du an e es e pe cu so deu-me a esponsabilidade de e inspi a , e espe o que, ao le es
es as pala as um dia, elas e mo i em a pe segui os eus p óp ios sonhos com a mesma de e minação.
i
ABSTRACT
Clo hing managemen is one o he mos signi ican challenges aced by blind and isually impai ed
people. By le e aging ad anced echnologies, such as compu e ision and a i icial in elligence (AI), his
PhD wo k aimed a de eloping and alida ing a mecha onic de ice, he iSigh p o o ype, able o assis in
he iden i ica ion o clo hing ypes, colou s, and condi ions, p ima ily designed o blind and isually
impai ed use s. Valuable insigh s we e ob ained om a na ionwide su ey, conduc ed in collabo a ion
wi h he Associa ion o he Blind and Amblyopes o Po ugal (ACAPO), which e ealed a high demand o
echnological solu ions in ga men managemen , wi h 95.7 % o esponden s exp essing willingness o
adop new echnologies. The iSigh p o o ype in eg a es a sma wa d obe equipped wi h an image
acquisi ion sys em and con olled ligh ing, ensu ing op imal image quali y cap u e o accu a e analysis.
The de ice is con olled ia a use - iendly mobile applica ion, designed o be highly accessible and
in ui i e. The sys em's unc ionali y was igo ously es ed wi h 15 pa icipan s om ACAPO, encompassing
bo h blind and low- ision people. The p o o ype's accu acy in iden i ying clo hing ca ego ies and colou s
was highly acknowledged by use s, wi h 60 % inding i e y p ecise in iden i ying ca ego ies and 80 % in
iden i ying colou s. Addi ionally, 86.7% o pa icipan s a ed he sys em's abili y o de ec s ains and
iden i y Nea Field Communica ion (NFC) ags as highly e ec i e. Finally, he ob ained esul s
demons a ed high le els o use sa is ac ion. Ex ensi e s a is ical analyses con i med signi ican posi i e
co ela ions be ween he iSigh unc ionali ies and in use s' con idence, sel -es eem, well-being, and
independence. These indings highligh he p o o ype's po en ial o signi ican ly enhance he daily li es o
isually impai ed people. Key a eas o u he imp o emen we e also iden i ied by he use s, namely he
educ ion o menu complexi y and addi ion o de ailed ab ic in o ma ion. In conclusion, his PhD wo k
demons a es he easibili y and e ec i eness o in eg a ing compu e ision and AI echnologies in o a
mecha onic de ice (sma wa d obe) o suppo blind and isually impai ed people in ga men
iden i ica ion and managemen . The iSigh p o o ype o e s a obus and use - iendly solu ion ha
signi ican ly imp o es use s' quali y o li e, o e ing oom o u u e e inemen and pa ing he way owa ds
a uly dis up i e p oduc o sma and e ec i e clo hing managemen .
KEYWORDS: A i icial In elligence (AI), Assis i e Technology, Blind People, Clo hing Iden i ica ion and
Modi ica ions, Compu e Vison, Mobile Applica ion, Sma Wa d obe.
xiii
Figu e 4.19: Example o a de ec unde ec ed by he YOLO 5m6 model ha was subsequen ly iden i ied
wi h he aid o da a augmen a ion: a) o iginal image; b) p edic ed image om model YOLOV5m6 wi hou
augmen a ion; c) P edic ed image om model YOLOV5m6 wi h augmen a ion (Rocha, Pin o,
e al
., 2023).
........................................................................................................................................................ 74
Figu e 4.20: Example o a misin e p e a ion o a de ec : a) o iginal image; b) p edic ed image om model
YOLOV5m6; c) p edic ed image om model YOLOV5l6 (Rocha, Pin o,
e al
., 2023). .......................... 75
Figu e 4.21: Examples o p edic ed images om he YOLO 5l6: a) single s ain de ec ion; b) mul iple s ain
de ec ion; c) mul iple hole de ec ion; d) hole de ec ion nea he seam (Rocha, Pin o,
e al
., 2023). .... 76
Figu e 5.1: iSigh Sma Wa d obe P o o ype. .................................................................................... 80
Figu e 5.2 - Schema ic layou o he sma wa d obe p o o ype. ......................................................... 81
Figu e 5.3 Nema 17 s eppe mo o . .................................................................................................. 82
Figu e 5.4: Pic u e o he d i e mo o A4988 used in he de eloped p o o ype. ................................. 83
Figu e 5.5: Pic u e o a Raspbe y Pi 4 Model B used in he de eloped p o o ype. .............................. 84
Figu e 5.6: V3 O icial Raspbe y Pi, 12MP, 120°, came a module, used in he de eloped p o o ype. . 84
Figu e 5.7. ITEAD PN532 NFC eade used in he de eloped p o o ype. ............................................ 86
Figu e 5.8. LED s ip 5M SMD 5050, 300LEDS 60LEDS/M 12V IP20 used in he de eloped p o o ype
(
Fi a LED 5m SMD 5050 300leds 60leds/m 12 IP20 - B anco F io
, 2024). ..................................... 87
Figu e 5.9. Pic u e o he 12V 60W 5A powe supply inco po a ed in he de eloped p o o ype. ........... 88
Figu e 5.10. Chip label NFCTAG213 applied o he de eloped p o o ype. ........................................... 88
Figu e 5.11. Pu pose-buil b anded clo hing hange de eloped o he iSigh p o o ype. ...................... 89
Figu e 5.12. Schema ics o he h ee main phases and communica ion included in he iSigh p o o ype.
........................................................................................................................................................ 90
Figu e 5.13: Flowcha o wa d obe se e . ....................................................................................... 91
Figu e 5.14: Gene al iew o he wa d obe in e io : a) LED s ips applied on he le panel; b) came a
placed in he middle o LED s ips. .................................................................................................... 92
Figu e 5.15: Re lec i e illumina ion in hole de ec . This image shows how ligh e lec ed om a la su ace
is a ec ed by non- la ea u es like holes, which can edi ec ligh ou side he lens's accep ance angle,
c ea ing da k a eas ha e eal su ace de ec s. ................................................................................. 93
Figu e 5.16: In e io o he wa d obe showing he dis ance om he came a o he hange and in e nal
componen s. .................................................................................................................................... 94
Figu e 5.17: Mo o suppo ing he hange . ........................................................................................ 95
Figu e 5.18: NFC ag a ached o he clo hing label. .......................................................................... 96

xi
Figu e 5.19: NFC Reade . ................................................................................................................. 96
Figu e 5.20: Schema ic diag am illus a ing he in eg a ion and connec i i y o sys em componen s wi h
he Raspbe y PI con olle ................................................................................................................ 97
Figu e 5.21: iSigh mobile in e ace block diag am. ........................................................................... 99
Figu e 5.22: iSigh main menu. ...................................................................................................... 100
Figu e 5.23: Read NFC Tag ............................................................................................................ 101
Figu e 5.24: Close Menu. .............................................................................................................. 102
Figu e 5.25: Clo hing lis o uppe clo hes ca ego y. ....................................................................... 103
Figu e 5.26: De ailed in o ma ion co esponding o a selec ed clo hing i em. ................................... 104
Figu e 5.27. O e iew o he “Add new clo hing i em” menu: a) Desc ip ion, Ca ego y, Colou and Type
ields; b) NFC Tag, Size, Pa e n, Washing, S yle and P in ed ields. ................................................. 105
Figu e 5.28: Submenu o he Ca ego ies and Colou s Menu. ........................................................... 106
Figu e 5.29. Example o a e u ned esul om colou analysis. ....................................................... 107
Figu e 5.30. Submenu included in he Modi ica ions menu.............................................................. 108
Figu e 5.31. Example o a e u ned esul om he Modi ica ions Analysis. ...................................... 109
Figu e 5.32. Wa d obe menu de ails. .............................................................................................. 110
Figu e 6.1: Wo k low o he iSigh p o o ype es ing and alida ion p ocess. ...................................... 113
Figu e 6.2: ACAPO membe s in e ac ing wi h he iSigh p o o ype du ing he es ing phase, showcasing
hands-on engagemen : a) Pa icipan using he mecha onic de ice; b) Pa icipan using he iSigh mobile
applica ion. .................................................................................................................................... 116
Figu e 6.3: Dis ibu ion o p o essional backg ound o pa icipan s. ................................................. 119
Figu e 6.4: Age dis ibu ion o he ime ha isual impai men occu ed among pa icipan s. ........... 121
Figu e 6.5: Dis ibu ion o condi ions. .............................................................................................. 122
Figu e 6.6: Visual ep esen a ion o common causes o isual impai men (Bu on
e al
., 2021) ....... 123
Figu e 6.7: Wo d cloud summa izing he isually impai ed use s’ eedback. ..................................... 149
Figu e 6.8 - Wo d cloud summa izing he eedback o blind use s. ................................................... 150
Figu e 6.9 - Wo d cloud summa izing he eedback o low ision use s............................................. 151
x
LIST OF TABLES
Table 2.1: Li e a u e o e iew on ashion image classi ica ion wo ks (Rocha, Soa es,
e al
., 2023a). .. 18
Table 2.2: Summa y o ad anced image segmen a ion echniques o clo hing iden i ica ion .............. 20
Table 2.3: Li e a u e o e iew on ex ile ab ic de ec de ec ion including he au ho , yea , me hod,
da ase , de ec classes and me ics. .................................................................................................. 22
Table 4.1: Summa y o a ailable da ase s o ashion ca ego y classi ica ion (Rocha, Soa es,
e al
.,
2023a). ............................................................................................................................................ 51
Table 4.2: Main cha ac e is ics o p e- ained models (Rocha, Soa es,
e al
., 2023a). ........................ 54
Table 4.3: Hype pa ame e s o model expe imen s (Rocha, Soa es,
e al
., 2023a). ............................ 54
Table 4.4 - Tes pe o mance esul s (Rocha, Soa es,
e al
., 2023a). ................................................. 55
Table 4.5: In e ence ime by ne wo k a chi ec u e (Rocha, Soa es,
e al
., 2023a)............................... 56
Table 4.6: Tes pe o mance esul s wi h augmen ed da a (Rocha, Soa es,
e al
., 2023a). ................. 56
Table 4.7: Classi ica ion epo o he GoogLeNe ne wo k (Rocha, Soa es,
e al
., 2023a). .................. 57
Table 4.8: Hype pa ame e s (Image Size, Op imize , Lea ning Ra e, and Ba ch Size) o model
expe imen s. .................................................................................................................................... 61
Table 4.9 - Main pe o mance esul s o model o clo hing mask segmen a ion (P ecision, Recall and AP
a IoU = 0.50). ................................................................................................................................. 61
Table 4.10: Hype pa ame e s o ine- uned Mask R-CNN (Rocha, Soa es,
e al
., 2023a). .................. 67
Table 4.11: A epo on he e alua ion o Common Objec s in Con ex (COCO) (Rocha, Soa es,
e al
.,
2023a). ............................................................................................................................................ 67
Table 4.12: A summa y o he losses associa ed wi h he model (Rocha, Soa es,
e al
., 2023a). ........ 67
Table 4.13: A dis ibu ion o class de ec s, ocusing on he wo main class de ec s o in e es in his s udy
(Rocha, Pin o,
e al
., 2023). .............................................................................................................. 69
Table 4.14: Con igu a ion o Hype Pa ame e s in Model Tes ing: Image Size, Op imize , Lea ning Ra e,
and Ba ch Size (Rocha, Pin o,
e al
., 2023). ...................................................................................... 72
Table 4.15: Main esul s om he ine- uning o he models wi hou da a augmen a ion o de ec de ec ion
(P ecision, Recall and AP a IoU = 0.50) (Rocha, Pin o,
e al
., 2023). ................................................ 72
Table 4.16: Main pe o mance esul s o he models a e in oducing da a augmen a ion o de ec
de ec ion (P ecision, Recall and AP a IoU = 0.50) (Rocha, Pin o,
e al
., 2023). ................................. 73
Table 4.17: Pe o mance esul s o each model wi h da a augmen a ion and de ec classi ica ion
(P ecision, Recall and AP a IoU = 0.50) (Rocha, Pin o,
e al
., 2023). ................................................ 74
x i
Table 4.18: In e ence ime on es da ase o he di e en YOLO 5 models es ed (Rocha, Pin o,
e al
.,
2023). .............................................................................................................................................. 76
Table 5.1: Nema 17 mo o speci ica ions (
Mo o de Passo Nema 17 p/ Imp esso a 3D - 42BYGH48-23D
,
2024). .............................................................................................................................................. 82
Table 5.2. Desc ip ion o he A4988 speci ica ions (
D i e Pa a Mo o - A4988
, 2024). ...................... 83
Table 5.3: Module came a aspbe y pi 4 speci ica ions (
Raspbe y Pi Came a Module 3
, 2024). ...... 85
Table 5.4: Speci ica ion o lens (
Raspbe y Pi Came a Module 3
, 2024) . .......................................... 85
Table 5.5: Speci ica ions o ITEAD PN532 NFC module (
PN532
, 2024). ............................................ 86
Table 5.6. NFCTAG213 label speci ica ions (
NTAG213
, 2024)........................................................... 89
Table 6.1: Fishe 's Exac Tes esul s o associa ions wi h o e all sa is ac ion. ................................ 134
Table 6.2: Spea man's ank co ela ion esul s. ............................................................................... 135
Table 6.3: Mann-Whi ney es ing esul s o ease o na iga ion and o e all expe ience ac oss di e en
le els o com o wi h echnology. .................................................................................................... 137
Table 6.4: Spea man co ela ion analysis esul s o echnology amilia i y wi h ease o na iga ion and
o e all expe ience. .......................................................................................................................... 138
Table 6.5: Spea man's co ela ion esul s be ween di e en unc ionali ies o he iSigh p o o ype. ... 140
Table 6.6: Spea man's ank co ela ion esul s o equency o echnology use and com o le el. ... 142
Table 6.7: Spea man's ank co ela ion esul s o iSigh unc ionali y and inc eased con idence, sel -
es eem, well-being, and independence. ........................................................................................... 143
Table 6.8: Fiche ´s Exac es esul s o iSigh unc ionali y and inc eased con idence, sel -es eem, well-
being, and independence. ............................................................................................................... 144
Table 6.9: Thema ic analysis o use eedback o he iSigh p o o ype. ............................................ 146
Table 6.10: Dis ibu ion o eedback ca ego ies by isual impai men ype. ...................................... 150
Table 6.11: Usabili y p e e ences by isual impai men ype. ........................................................... 151
Table 6.12: Expec a ions o inno a ions and comme cial adop ion by isual impai men ype. ......... 152
Table 6.13: Sugges ions o imp o emen by isual impai men ype. .............................................. 152
Table 6.14: Impac on daily li e and u u e use by isual impai men ype. ....................................... 153
Table 6.15: Use eedback and co esponding a ec ed componen s in chap e 4 and 5. .................. 156
Table 6.16: Summa y o s a is ical es s, hypo heses, and esul s *[ co esponds o “No
signi ican ” and deno a es none co ela ion/signi icance, while co esponds o “Signi ican ” and
deno es subs an ial impac /co ela ion]. ......................................................................................... 157
x ii
ABBREVIATIONS
AADVDB
Associação De Apoio Aos De icien es Visuais Do Dis i o de B aga
ACAPO
Associação de Ambliopes e Cegos de Po ugal
AI A i icial In elligence
ANN A i icial Neu al Ne wo k
AP A e age P ecision
API Applica ion P og amming In e ace
CAVI Suppo Cen e o Independen Li ing
CEICSH Comissão de É ica pa a a In es igação em Ciências Sociais e Humanas
CMYK Cyan Magen a Yellow Black
CNN Con olu ional Neu al Ne wo k
CPU Cen al P ocessing Uni
DC Di ec Cu en
EU Eu opean Union
GPU G aphics P ocessing Uni
GUI G aphical Use In e ace
Hex Hexadecimal
HSI Hue Sa u a ion Illumina ion
HSL Hue Sa u a ion Luminance
HSV Hue Sa u a ion Value
IDE In eg a ed De elopmen En i onmen
IIC In e -In eg a ed Ci cui
IoT In e ne o Things
IOU In e sec ion O e Union
IP Ing ess P o ec ion
LED Ligh Emi ing Diode
MSCOCO Mic oSo Common Objec s in Con ex
MSVI Mode a e o Se e e Vision Impai men
NFC Nea Field Communica ion
FDENe F equency Domain Enhancemen Ne wo k
QR Quick Response
x iii
RFID Radio Field Iden i ica ion
RGB Red G een Blue
RNID
Regulamen o Nacional de In e ope abilidade Digi al
ROI Region O In e es
RPN Region P oposal Ne wo k
SDG Sus ainable De elopmen Goal
SIFT Scale In a ian Fea u e T ans o m
SGD S ochas ic G adien Descen
SPI Se ial Pe iphe al In e ace
SSD Single Sho Mul iBox De ec o
STA S a is ical Fea u e
SVM Suppo Vec o Machine
UART Uni e sal Asynch onous Recei e /T ansmi e
UID Unique Iden i ie
URL Uni o m Resou ce Loca o
VOC Visual Objec Classes
WHO Wo ld Heal h O ganiza ion

1 INTRODUCTION
Chap e O e iew
The pu pose o his chap e is o p esen he p oblem s a emen , he mo i a ions, and he scope o he
esea ch. An o e iew o assis i e echnologies o clo hing managemen o blind people is p esen ed
i s . In he ollowing sec ion, he p esen wo k's objec i es and e hical conside a ions a e discussed.
Finally, he esul s o he de eloped scien i ic ac i i y a e p esen ed, as well as he s uc u e o he
disse a ion.
1 In oduc ion
1.1 P oblem S a emen , Mo i a ions, and Scope
1.2 Objec i es
1.3 Con ibu ion o Knowledge and Resul s o he de eloped scien i ic ac i i y
1.4 Resul s o he de eloped scien i ic ac i i y
1.5 Thesis S uc u e
Chap e 1 –
In oduc ion
2
1.1 P oblem S a emen , Mo i a ions, and Scope
Blindness is a condi ion ha a ec s housands o people wo ldwide. An analysis o he Vision A las om
he In e na ional Agency o he P e en ion o Blindness e eals ha wo ldwide, 43 million people a e
blind and 295 million ha e mode a e- o-se e e isual impai men s (
Magni ude and P ojec ions - The
In e na ional Agency o he P e en ion o Blindness
, n.d.). In addi ion, he Wo ld Heal h O ganiza ion
(WHO) epo s ha a leas 2.2 billion indi iduals wo ldwide su e om a nea o dis ance ision
impai men , wi h nea ly hal o hese cases being se e e (
Blindness and Vision Impai men
, n.d.). In
Po ugal, acco ding o he 2011 census (INE, 2012), he e a e app oxima ely 900.000 people wi h ision
di icul ies, o whom app oxima ely 28.000 a e blind. Addi ionally, 23% o popula ion s a e ha ha e
di icul ies in seeing, e en when using glasses o con ac lenses. Such isual impai men ,
e.g.
blindness,
can signi ican ly a ec a pe son's psychological and cogni i e unc ioning. Se e al s udies ha e
demons a ed ha ision impai men is associa ed wi h a a ie y o nega i e heal h ou comes and a
diminished quali y o li e (Chia
e al
., 2006; Langelaan
e al
., 2007).
Blind people can bene i om he use o assis i e echnology by educing he nega i e e ec s o blindness
and imp o ing hei quali y o li e. Sus ainable De elopmen Goal (SDG) 10 (Reduce Inequali ies) is a
undamen al objec i e in he de elopmen o inclusi e echnologies, which aims o educe inequali ies
bo h wi hin and be ween coun ies, p omo ing social, economic, and poli ical inclusion o all people (
Goal
10 | Depa men o Economic and Social A ai s
, 2024). Hence, i is impe a i e o add ess he exis ing
dispa i ies and ba ie s ha a ec his communi y. To mee he unique needs o blind people and enhance
hei o e all quali y o li e, inclusi e echnologies should be de eloped ha ca e o hei speci ic needs
and enhance hei o e all quali y o li e. The ad ancemen o hese echnologies allows people wi h
blindness o b idge he gap and pa icipa e ully in socie y on an equal basis, enhancing hei mobili y,
na iga ion, and o e all independence, he eby p omo ing inclusion.
Despi e he p oli e a ion o sma de ices and he ad ancemen o cu ing-edge echnology o blind
people, mos o he esea ch has ocused on na iga ion, mobili y, and objec ecogni ion, igno ing
aes he ic conside a ions (Bhowmick & Haza ika, 2017; Elmannai & Ellei hy, 2017; Messaoudi
e al
.,
2022). Figu e 1.1 illus a es his end by p o iding an o e iew o esea ch e o s ac oss a ious domains
in e ms o numbe o publica ions be ween 2007-2023. Da a was ob ained om he SCOPUS da abase,
using he ollowing keywo ds sea ch engine: 'Na iga ion AND Blind People', 'Mobili y AND Blind People',
'Objec De ec ion AND Blind People', 'Compu e Vision AND Blind People', and 'Clo hes AND Blind
People'.
Chap e 1 –
In oduc ion
3
Figu e 1.1: Ba cha illus a ing he numbe o publica ions be ween 2007-2023 in a ious esea ch a eas conce ning blind
people. The sea ch s a egies employed include: 'Na iga ion AND Blind People', 'Mobili y AND Blind People', 'Objec
De ec ion AND Blind People', 'Compu e Vision AND Blind People', and 'Clo hes AND Blind People'. Da a was ob ained om
he SCOPUS da abase.
The analysis o he da a p esen ed in Figu e 1.1 indica es ha na iga ion and mobili y- ela ed echnologies
ha e ecei ed signi ican a en ion in ecen yea s, wi h an inc ease in objec de ec ion and compu e
ision esea ch om 2018 onwa ds. I is impo an o no e, howe e , ha he p opo ion o esou ces
alloca ed o aes he ic conside a ions, as e lec ed in he "Clo hes" ca ego y, emains ela i ely low
h oughou he pe iod unde s udy.
Clo hing and s yle p e e ences o di e en occasions o m an in eg al pa o a pe son's iden i y (Johnson
e al
., 2014). This has a signi ican impac on how hey pe cei e hemsel es as well as how hey a e
pe cei ed by o he s (Adam & Galinsky, 2012; Johnson
e al
., 2014). Blindness is o en o e looked when
i comes o aes he ics. E en hough many belie e ha aes he ics is only conce ned wi h isual appeal, i
can hold signi ican ele ance in daily li e. Aes he ics can signi ican ly a ec a pe son's sense o iden i y,
sel -exp ession, well-being and sel -es eem. Ne e heless, blind people may expe ience insecu i y and
s ess when i comes o d essing-up due o he inabili y o ecognize he condi ion o he ga men s. The
inabili y o pe cei e isual cues o e en colou s can make d essing up a daily challenge. Addi ionally,
blind people may be mo e likely o expe ience clo hing s aining and ea ing due o hei inhe en di icul ies
in handling objec s and pe o ming daily asks. Fo ins ance, de ec ing s ains as soon as possible is
essen ial o a oid hem becoming pe manen o di icul o emo e. To add ess his gap, i is essen ial o
de elop echnologies ha ca e o he aes he ic needs o blind people. By in eg a ing inno a i e solu ions
Chap e 1 –
In oduc ion
4
o clo hing, we can enhance he sel -con idence, com o , s yle, and sel -es eem o blind people. This will
no only enhance hei o e all well-being bu will also con ibu e o hei social inclusion and equal
pa icipa ion in socie y.
In spi e o he po en ial o echnological solu ions in he u u e, signi ican challenges emain o be
o e come. Due o hei isual impai men , hese people ace challenges in de ec ing mino i egula i ies
o s ains in clo hing ex u es o ab ics, necessi a ing eliance on o he s o assis ance. This is he p emise
behind he mo i a ion o he p esen wo k - enable blind people o ha e an equal sense o con idence in
wha hey wea , wi hou pe manen ly equi ing human assis ance. Sadly, his pe cep ion is s ill lacking in
he li es o blind people, and clo hing s ill p esen s a daily challenge o hem. To add ess his absence
o suppo , his wo k aimed a de eloping an inno a i e solu ion ha combines s a e-o - he-a esea ch
and echnology wi h con en ional appa a us and u ni u e o common use. This esul ed in he
de elopmen o a wa d obe equipped wi h ideal ligh ing condi ions and image cap u e capabili ies ha
allows de ec ing s ains and de ec s in clo hing i ems, as well as iden i ying hei colou and ca ego y using
A i icial In elligence (AI) algo i hms. By combining his sys em wi h a mobile applica ion, he use can
manage and selec clo hing in elligen ly, which can p opel his o he sel -es eem, con idence, well-being,
and sel -con idence on a daily basis, illing a echnological gap in he aes he ics and image o blind people.
The alida ion o he de eloped wo k was conduc ed h ough a collabo a ion wi h he
Associação de Cegos
e Amblíopes de Po ugal
(ACAPO), he Po uguese Associa ion o he Blind and Amblyopic, and he
Associação de Apoio aos De icien es Visuais do Dis i o de B aga
(AADVDB), he Associa ion o Suppo
o he Visually Impai ed o B aga, which helped o iden i y key a eas o imp o emen .
1.2 Objec i es
This wo k aimed o design and c ea e an inno a i e solu ion ha me ges cu ing-edge esea ch and
echnology wi h e e yday de ices. The esul is a wa d obe designed wi h op imal ligh ing and image
cap u e ea u es ha can de ec s ains and de ec s in clo hing, as well as iden i y colou s and ca ego ies
using A i icial In elligence algo i hms. This endea ou should p omo e he au onomy o blind people in
hei daily ac i i ies, he eby enhancing hei o e all well-being.
To achie e his objec i e, speci ic goals we e se , pa icula ly:
G1. Conduc a na ionwide su ey dis ibu ed by ACAPO o assess use in e es , needs, and ecep i i y o
he s udy;
G2. De elop an algo i hm o he acquisi ion, p ocessing, and analysis o images using AI o iden i y
Chap e 2 –
Li e a u e Re iew
11
2.1 Assis i e Technologies o Blind People
The apid ad ancemen o echnology has signi ican ly enhanced suppo sys ems o isually impai ed
people. This chap e e iews and analyses a ious exis ing solu ions and p ojec s ha align wi h he
echnological amewo k in ended o his esea ch. The objec i e is o p o ide a comp ehensi e
cha ac e iza ion o hese echnologies, while also iden i ying po en ial gaps and a eas o u he
de elopmen and inno a ion. By c i ically e alua ing hese exis ing solu ions, his chap e seeks o
es ablish a ounda ion upon which new, imp o ed me hods can be de eloped o be e suppo he isually
impai ed communi y.
2.1.1 Solu ions o he Visually Impai ed in Clo hing Selec ion
Visually impai ed indi iduals ace signi ican challenges in selec ing app op ia e clo hing, which is
essen ial o pe sonal digni y and con idence. Recen esea ch has ocused on de eloping echnological
solu ions o assis isually impai ed indi iduals wi h clo hing- ela ed asks. Th ough echnological
ad ancemen s, a numbe o solu ions ha e been de eloped o assis he isually impai ed. Among hese
solu ions a e mobile applica ions, In e ne o Things (IoT) sys ems, ac ile labelling sys ems, and inclusi e
ashion design, all o which o e unique ad an ages and ea u es.
To simpli y he ask o online shopping in he ga men sec o , a sys em was p oposed by Yang Huiqiaoand
Peng (2015) ha makes online shopping easie o he isually impai ed. This idea ook shape a e
se e al in e iews wi h isually impai ed s uden s om he Special Educa ion College o Peking Uni e si y.
These in e ac ions highligh ed challenges such as excessi e ad e ising, unin ui i e in e aces, and
insu icien subdi isions o clo hing ca ego ies. Based on his s udy, he au ho s p oposed a web
applica ion o add ess he iden i ied issues and help he isually impai ed make in o med decisions when
pu chasing clo hing. Essen ial in o ma ion abou each i em (name, ype, pho o, p ice, and sales olume)
is displayed immedia ely below he espec i e pho o, simpli ying he sea ch and eading p ocess.
Addi ionally, S angl
e al
. (2018) de eloped he B owseWi hMe sys em, an AI-powe ed online shopping
assis an designed o make online clo hes shopping mo e accessible o isually impai ed people. This
sys em con e s p oduc web pages in o a s uc u ed ep esen a ion, allowing use s o in e ac i ely ask
o speci ic in o ma ion abou p oduc s. The s udy ound ha B owseWi hMe yields accu a e image
desc ip ions and makes he online shopping mo e accessible o isual impai ed. Ga is Filho Síl io José
Viei a and de Assumpção Macedo (2018) in oduced a sys em ha uses NFC ags and Quick Response

Chap e 2 –
Li e a u e Re iew
12
(QR) codes on clo hing o help isually impai ed indi iduals ma ch ou i s. The sys em p o ides audio
eedback on clo hing combina ions h ough an And oid app, enabling use s o selec app op ia e ou i s
independen ly. An audio assis i e echnology ha helps blind indi iduals iden i y clo hing pa e ns and
colou s is p esen ed by N.Swa hi and Jyo hi (2020). Yang
e al
. (2014) de eloped a came a-based
p o o ype sys em ha ecognizes clo hing pa e ns and colou s, p o iding e bal desc ip ions o isually
impai ed use s. Pa icipan s p o ided posi i e eedback abou he sys em, highligh ing i s abili y o
p omo e g ea e independence in hei daily li es. Ano he example, p oposed by (Allam Mahmoud and
ElShaa awy, 2022) uses he Sequen ial Minimal Op imiza ion me hod o clo hing classi ica ion. Simila ly,
he sys em p oposed by Yang
e al
. employs S a is ical Fea u e (STA) and SIFT (Scale In a ian Fea u e
T ans o m) algo i hms o clo hes pa e n ecogni ion. S angl
e al
. (2018) in oduced Vision4All, a deep
lea ning ashion assis ance solu ion ha helps isually impai ed use s o iden i y clo hing a ibu es such
as colou s, ca ego ies, ex u es, ab ics, s yles, g aphics, and ex -based con en on clo hes. A p o o ype
wi h an endoscopic came a a ached o a inge ip, allowing o image cap u e and ecogni ion o colou s
and pa e ns on clo hing su aces, was p oposed by Medei os
e al.
(2017). Rini J.and Thilaga a hi (2015)
p oposed a sys em ha ecognizes ou ypes o pa e ns (checke ed, s iped, i egula , and plain) and
ele en colou s using a came a o image cap u e and a Suppo Vec o Machine (SVM) algo i hm o
classi ica ion. A me hod ha con e s isual in o ma ion ob ained om clo hing in o e bal exposi ion
using Deep Neu al Ne wo ks was p esen ed by Ta eno
e al.
(2020). This sys em cap u es pho os o
clo hing combina ions, which a e hen p e-p ocessed o emo e he pe son's ace and ex ac key
ea u es. A e p ocessing he images, a ma hema ical model is applied o implemen he sys em
holis ically, enabling i o au oma ically ecognize and cha ac e ize he p esen ed clo hing i ems.
Inclusi e ashion design aims o enhance he clo hing selec ion expe ience o isually impai ed
indi iduals by inco po a ing ac ile and unc ional elemen s. Dassole
e al
. (2023) explo ed he
de elopmen o inclusi e ashion using senso y elemen s in clo hing o allow isually impai ed indi iduals
o c ea e ac ile ep esen a ions o shapes. This app oach p omo es indi idual exp ession and au onomy,
cogni i e and c ea i e de elopmen , and a mo e en iching clo hing expe ience.
Despi e he a o emen ioned esea ch, he e a e se e al comme cial solu ions a ailable in he ma ke ,
bo h as s andalone de ices and mobile applica ions. Rega ding comme cial de ices, he Colo ino is a
colou iden i ie ha de ec s app oxima ely 150 di e en colou s and announces hem aloud. I ea u es
h ee olume le els and an op ion o use headphones. Addi ionally, he Colo ino unc ions as a ligh
de ec o , emi ing dis inc acous ic signals o di e en ia e be ween na u al and a i icial ligh sou ces
(
Colo ino – Colo and Ligh De ec o - Ca e ec
, 2024). A simila de ice, he Colo Tes 2000, no only
Chap e 2 –
Li e a u e Re iew
13
iden i ies colou s like he Colo ino bu also i eads da es and imes and de ec s whe he a household
ligh is on o o (
Colo Tes 2000 - Assis i e Technology a Eas e Seals C oss oads
, 2024).
An imp o ed e sion o he Colo ino, he Colo S a P o, o e s a mo e compac and ligh weigh design
o easie use. I ea u es ad anced speech ou pu and enhanced measu emen accu acy. Capable o
de ec ing o e 1,700 colou nuances, pa e ns, and con as s, i uniquely ansla es hese ea u es in o
musical ones. The de ice is also highly e ec i e in iden i ying ligh and i s colou s, making i an in aluable
ool o indi iduals wi h isual impai men s. Addi ionally, he Colo S a P o p o ides colou in o ma ion
in Red G een Blue (RGB) and Hue Sa u a ion Luminance (HSL) o ma s, ecognizes colou pa e ns, and
can de e mine ligh empe a u e, u he inc easing i s unc ionali y (
Colo S a P o – Colo and Ligh
De ec o - Ca e ec
, 2024). The ul ima e e sion, he Colo S a , is a sophis ica ed and po able colou
iden i ica ion de ice ha can dis inguish o e 1,700 unique colou shades, con eyed h ough clea spoken
oice ou pu . I is also adep a measu ing con as , de ec ing Ligh Emi ing Diode (LED) ligh colou s,
gauging ambien ligh in ensi y, and iden i ying pa e ns. This e sa ile de ice p o ides a comp ehensi e
analysis o colou and ligh in a ious en i onmen s (
Colo S a – Talking Colo and Ligh De ec o
, 2024).
In e ms o mobile applica ions, he V7 Aipoly is a mobile applica ion ha enables eal- ime iden i ica ion
o objec s, ex s, and colou s, and communica es wi h he use h ough audio(
V7 Aipoly
, n.d.). Simila ly,
Colou Iden i ie is an iPhone applica ion a ailable on he App S o e ha de ec s he colou o any pixel
in a pho o o image, displaying colou s in RGB, CMYK (Cyan, Magen a, Yellow, Black) and HEX
(Hexadecimal) o ma s along wi h hei names (
Colo Iden i ie : Colo Picke on he App S o e
, 2024).
Mobile applica ions can also play an essen ial ole in human p oximi y, helping o comba social exclusion
and di ec ly add essing SDG 10. Fo example, Be My Eyes is a mobile applica ion designed o assis
isually impai ed people h ough in e ne -based ideo calls. This applica ion acili a es eal- ime
assis ance by connec ing isually impai ed use s wi h olun ee s who a e a ailable o help. I a isually
impai ed indi idual needs o e i y he expi a ion da e on a ca on o milk, hey can u ilize he applica ion
o ini ia e a ideo call. A olun ee ecei es a no i ica ion and, i a ailable, esponds o p o ide he
necessa y assis ance (
Be My Eyes - T azendo Visão Aos Cegos e Pessoas Com Visão Reduzida
, 2024).
This p ocedu e can also be applied o asks such as selec ing clo hing i ems o iden i ying hei colou s.
I is no ewo hy ha he unc ionali y o his applica ion is en i ely dependen on he a ailabili y o
olun ee s.
Muhsin
e al.
(2023) conduc ed a comp ehensi e e iew o subs i u i e assis i e echnologies o people
wi h isual impai men s, ca ego izing hem based on he ype o eedback ( isual, hap ic, o audi o y).
The s udy highligh ed he limi a ions in use expe ience and he challenges o ansla ing nume ous
Chap e 2 –
Li e a u e Re iew
14
esea ch p o o ypes in o p ac ical, widely adop ed solu ions. I emphasized he need o mo e e ec i e
assis i e aids and he impo ance o dis inguishing be ween he needs o indi iduals wi h pa ial ision,
colou blindness, and hose who a e o ally blind. Focusing on speci ic g oups would allow o he
de elopmen o mo e ailo ed and e ec i e assis i e echnologies.
2.1.2 Sma Wa d obe Solu ions
Some esea ch has ocused on de eloping sma wa d obe sys ems o assis blind indi iduals in
managing hei clo hing. These sys ems ypically employ echnologies such as Radio Field Iden i ica ion
(RFID) (Goh
e al.
, 2011), NFC (Alabduljabba , 2022), and IoT (Ka hi a an, 2021) o ga men
iden i ica ion and acking.
Goh
e al.
(2011a) p opose a sma wa d obe sys em using RFID echnology o help use s manage hei
clo hing and make decisions based on p e e ences o colou , s yle, e en s, and emo ions. Ta ge ing busy
en ep eneu s and colou -blind indi iduals, he sys em acks ga men mo emen wi hin he wa d obe
and s o es clo hing da a om RFID ags. This echnology aids use s in making be e wa d obe decisions
and simpli ies clo hing managemen . Alabduljabba (2022) in oduced an IoT-based sma clo hing
sys em which employs NFC echnology o help isually impai ed people manage hei close s
independen ly, allowing o selec and ind app op ia e clo hing wi hin hei close s, h ough he use o
sma phone o scan NFC ags. The sys em pe o med easonably well and me he usabili y equi emen s
o isually impai ed people. A speech ecogni ion so wa e module o a sma close , designed as
assis i e echnology o isually impai ed indi iduals was p esen ed by Ca olina Sauceda, Peña Ing id and
Alejand o Luna Gómez (2021). This sys em allows use s o manage hei clo hing h ough oice
commands, enabling asks such as ga men inse ion, sea ching by desc ip ion, and e ie al om he
wa d obe. The module, de eloped o web-based en i onmen s, shows p omise in enhancing he quali y
o li e o isually impai ed people.
The inc easing demand o solu ions wi hin he concep o sma wa d obes has p omp ed he eme gence
o nume ous solu ions in he ma ke . Each solu ion exhibi s unique capabili ies and can be subca ego ized
in o physical and i ual implemen a ions. Al hough hese solu ions a e no speci ically ocused on blind
indi iduals, hey highligh he g owing in e es and po en ial in de eloping assis i e echnologies ha can
be adap ed o mee he needs o isually impai ed use s. In 2019, he Fashion Applica ion P og amming
In e ace (API) mobile applica ion (
Ou Technology - Fashion Tas e API
, 2024) was de eloped o manage
wa d obes using QR code scanning. This solu ion allows use s o c ea e a i ual wa d obe o o al
Chap e 2 –
Li e a u e Re iew
15
managemen , including adding o emo ing clo hing i ems. The addi ion o i ems is based on scanning a
QR code on he clo hing label using he sma phone came a wi h he ins alled mobile applica ion. Beyond
wa d obe managemen , his sma wa d obe concep p o ides clo hing usage ad ice, sugges s new
pu chases based on use needs, and helps planning ou i s. The echnology is o e ed h ough a mobile
applica ion ea u ing an in ui i e in e ace, dashboa ds, and s a is ical da a showing he mos wo n i ems
and sugges ed ou i s. The Sma Close solu ion (
Sma Close : You Pe sonal S ylis
, 2024), in oduced
in 2018, also uses a mobile applica ion o wa d obe managemen based on clo hing pu chases om
a ious s o es. This applica ion allows use s o ully manage hei wa d obe i ems and ob ain de ailed
s a is ics, such as he mos wo n pieces. Howe e , his solu ion lacks “in elligence”, as i does no use
machine lea ning o compu e ision algo i hms o au onomously iden i y clo hing i ems o p o ide ashion
ad ice. The S ylebook Close App (
S ylebook Close App: A Close and Wa d obe Fashion App o he
IPhone and IPad
, 2024) se es as a comp ehensi e ool o planning and o ganizing wa d obes. Use s
can upload images o hei clo hing i ems and c ea e s ylish ou i s wi h minimal e o . The app includes
a backg ound emo al ea u e o main ain a clean i ual wa d obe. Addi ional ea u es such as a
calenda , packing lis , and inspi a ion boa d allow use s o schedule and ack hei ou i s, p epa e
packing lis s o ips, and analyse hei wa d obe usage. Analogous o S ylebook, Pu eple combines
simplici y wi h unc ionali y. I s in ui i e in e ace pe mi s use s o ca ego ize hei clo hes, plan ou i s,
and pack o ips e icien ly. A dis inc i e ea u e is i s abili y o allow use s o open hei close s o o he
use s on he app o s yling assis ance (
Pu eple - Ou i O ganize App
, 2024). Simila ly, he Chicisimo
app assis s use s in ca aloguing hei clo hes and o e s ou i inspi a ion om ashion en husias s
wo ldwide. Fo ins ance, i a use seeks new ways o s yle a classic ed d ess, he app p o ides images
o a ious s yling op ions used by women globally (
Chicisimo - Build wi h Fashion Shoppe s’ Tas e Da a
,
2024). Cladwell o e s use s daily ou i sugges ions and AI-powe ed s yling ad ice based on he clo hes
al eady p esen in hei wa d obe. This app acili a es he c ea ion o a minimalis close illed wi h capsule
pieces. Unlike o he wa d obe apps, use s do no need o ca alogue hei wa d obe i ems manually.
Ins ead, hey selec om p e-buil capsule wa d obes, and he Cladwell wa d obe is popula ed wi h
gene ic i ems om he chosen capsule. Use s can hen upload pic u es o simila i ems om hei ac ual
wa d obe o eplace he gene ic ones (
Cladwell | Simpli y You Li e Wi h A Capsule Wa d obe
, 2024).
Aclose is a digi al close applica ion ha allows use s o manage all hei ashion i ems in one place, as
well as buy and sell clo hing i ems ha a e no longe needed. I enables he c ea ion o ou i ideas,
p o ides ou i ecommenda ions and s yle analy ics, and acili a es he buying and selling o p e-lo ed
i ems a easonable p ices (Aclose - AI Fashion Assis an on he App S o e, 2024).
Chap e 2 –
Li e a u e Re iew
16
2.2 Clo hing Ca ego y Iden i ica ion
In he las ew yea s, deep lea ning echniques ha e a isen as a g ea me hod o sol e p oblems in
compu e ision, such as image classi ica ion, objec de ec ion, ace ecogni ion and language p ocessing,
whe e con olu ional neu al ne wo ks (CNNs) play an impe a i e ole (Voulodimos
e al
., 2018a). This is
pa icula ly impo an o clo hing ca ego y iden i ica ion, as CNNs can ex ac impo an ea u es om
clo hing images, enabling accu a e iden i ica ion and classi ica ion. The CNNs ha e exhibi ed excellen
esul s and ad ances in image ecogni ion since 2012 (D. Bha
e al
., 2021; Pa el
e al
., 2022), when
AlexNe (K izhe sky
e al
., 2012) was in oduced on he ImageNe La ge Scale Visual Recogni ion
(ILSVRC) (Deng
e al
., 2009;
ImageNe La ge Scale Visual Recogni ion Compe i ion (ILSVRC)
, 2024). The
ImageNe compe i ion consis s o e alua ing se e al algo i hms o la ge-scale objec de ec ion and image
classi ica ion, allowing esea che s o compa e de ec ion ac oss a a ie y o objec classes. In ecen
yea s, se e al CNNs ha e been p esen ed, such as: VGG (Simonyan & Zisse man, 2015), GoogLeNe
(Szegedy
e al
., 2014), SqueezeNe (Iandola
e al
., 2016), Incep ion (Szegedy
e al
., 2015), ResNe (K.
He
e al
., 2015), Shu leNe (Ma
e al
., 2018), MobileNe (Howa d
e al
., 2019; Sandle
e al
., 2018),
E icien Ne (Tan & Le, 2019), RegNe (Radosa o ic
e al
., 2020), among o he s, using hese ne wo ks
o di e en image classi ica ion p oblems.
In line wi h his p emise, some esea che s u ned o he ashion wo ld, making use o he mo e ecen
ad ances in compu e ision o explo e di e se a eas such as ashion de ec ion, ashion analysis and
ashion ecommenda ion, achie ing p omising esul s (W.-H. Cheng
e al
., 2021). Gi en he co e scope
o his hesis, only ashion classi ica ion is co e ed h oughou he wo k. A li e a u e su ey allowed
iden i ying se e al s udies ha a emp ed o handle he classi ica ion o ashion images. Mos o he
au ho s e alua e hei models based on op-k accu acy ega ding clo hing a ibu es ecogni ion, no mally
wi h op-3 and op-5 sco es, which means ha he co ec label is among he op k p edic ed labels.
The esea ch o Chen
e al.
(2015) p esen ed a ne wo k o desc ibing people based in ine-g ained
clo hing a ibu es wi h an accu acy o 48.32%. Simila ly, Liu
e al.
(2016) in oduced he FashionNe
which lea ns clo hing ea u es by join ly p edic ing clo hing a ibu es and landma ks. P edic ed landma ks
a e used o pool o ga e he lea ned ea u es maps. The au ho s epo ed a op-3 classi ica ion accu acy
o 93.01% and a op-5 o 97.01%. Ano he me hod o de ec ashion i ems in a gi en image using deep
con olu ional neu al ne wo ks, was pe o med by Ha a
e al.
(2014), p esen ing a mean A e age P ecision
(mAP) o 31.1%. Likewise, Co biè e
e al.
(2017) p oposed o he me hod based on weakly supe ised
lea ning o classi ying e-comme ce p oduc s, p esen ing a op-3 ca ego y accu acy o 86.30% and a op-

Chap e 2 –
Li e a u e Re iew
17
5 92.80% accu acy. La e , Wang
e al.
(2018) p oposed a ashion ne wo k o add ess ashion landma k
de ec ion and ca ego y classi ica ion wi h he in oduc ion o in e media e a en ion laye s o a be e
enhancemen in clo hing ea u es, ca ego y classi ica ion and a ibu e es ima ion. In hei wo k,
accu acies o 90.99% and 95.75% we e epo ed o op-3 and op-5, espec i ely. In ano he s udy by Li
e al.
(2019), he au ho s p esen ed a wo-s eam con olu ional neu al ne wo k wi h one b anch dedica ed
o landma k de ec ion and he o he one o ca ego y and a ibu e classi ica ion, allowing he model o
lea n he co ela ions among mul iple asks and consequen ly imp o emen in he esul s. Accu acies o
93.01% and 97.01% o op-3 and op-5 we e epo ed, espec i ely. The ashion classi ica ion model
p oposed by Cho
e al.
(2019) allowed imp o ing he pe o mance aking in o accoun he hie a chical
dependences be ween class labels eaching accu acies o 91.24% and 95.68% in op-3 and op-5,
espec i ely. A mul i ask deep lea ning a chi ec u e was hen p oposed by Lu
e al.
(2016) ha g oups
simila asks and p omo es he c ea ion o sepa a ed b anches o un ela ed asks, wi h accu acy esul s
o 83.24% and 90.39% o op-3 and op-5, espec i ely. Seo and Shin (2019) p oposed a Hie a chical
Con olu ional Neu al Ne wo k (H-CNN) o ashion appa el classi ica ion. The au ho s demons a ed ha
hie a chical image classi ica ion could minimize he model losses and imp o e i s accu acy, wi h a esul
o 93.3%. The esea ch o Fengzi
e al.
(2020) applied ans e lea ning using pe ained models o
au oma ically label uploaded pho os on ecomme ce indus y. The au ho s epo ed an accu acy o
88.65%. Addi ionally, a condi ion CNN was p oposed by Kolisnik, Hogan, and Zulke nine (2021), based
on b anching con olu ional neu al ne wo ks. The p oposed b anching can p edic hie a chical labels o
an image and he las label p edic in he hie a chy is epo ed wi h an accu acy o 91.0%. A new algo i hm
o clo hing ca ego y classi ica ion o add ess he complexi y o di e se clo hing s yles was p oposed by
X. Liu
e al.
(2023). CloNe imp o es op-1 classi ica ion accu acy by 0.8% and educes model size o
one- hi d compa ed o s a e-o - he-a models. Fu he mo e, Shi
e al.
(2023) in oduce F equency
Domain Enhancemen Ne wo k (FDENe ), a ne wo k o clo hing ca ego y classi ica ion ha cap u es
ad anced ea u es like ex u e and con ou in o ma ion. FDENe uses spec um enhancemen and
dep hwise sepa able con olu ions o imp o e ea u e ex ac ion and educe edundancy. I achie es a
1.3% imp o emen in op-1 accu acy on Deep ashion compa ed o cu en models.
Chap e 2 –
Li e a u e Re iew
18
Table 2.1 summa izes he da a o he a o emen ioned wo ks, including he used da ase s.
Table 2.1: Li e a u e o e iew on ashion image classi ica ion wo ks (Rocha, Soa es,
e al
., 2023a).
Au ho
Da ase
Yea
Accu acy
Ha a
e al.
(2014)
Fashionis a
2014
mAP: 31.1 %
Chen
e al.
(2015)
S ee -da a
2015
Top-1: 48.31 %
Liu
e al.
(2016)
DeepFashion
2016
Top-3: 82.58 %
Lu
e al.
(2016)
DeepFashion
2016
Top-3: 83.24 %
Co biè e
e al.
(2017)
DeepFashion
2017
Top-3: 86.30 %
Wang
e al.
( 2018)
DeepFashion-C
2018
Top-3: 90.99 %
Li
e al.
(2019)
DeepFashion-C
2019
Top-3: 93.01 %
Cho
e al.
(2019)
DeepFashion
2019
Top-3: 91.24 %
Seo and Shin. (2019)
Fashion-MINIST
2019
Top-1: 93.33 %
Fengzi
e al.
(2020)
Fashion P oduc Images
2020
Top-1: 88.65 %1
Kolisnik, Hogan, and Zulke nine(2021)
Fashion P oduc Images
2021
Top-1: 91.0 %1
X. Liu
e al.
(2023)
DeepFashion
2023
Top-1: 68.7 %2
Shi
e al.
(2023)
DeepFashion
2023
Top-1: 71.22 %2
1 Resul s only epo ed o ashion classi ica ion accu acy.
2Resul s only epo ed o DeepFashion da ase .
I becomes clea ha he e has been a g ea e o o build e icien me hods o ashion ca ego y
classi ica ion. Howe e , as he wo ks lis ed in Table 2.1 highligh , he e is a lack o ocus on de eloping
sys ems o aid isual impai ed people, and he e is ye no solu ion capable o co e ing all he di icul ies
s a ed by a blind pe son, namely an au oma ic sys em o clo hing ype iden i ica ion.
2.3 Clo hing Segmen a ion
Image segmen a ion is he p ocess o iden i ying and anno a ing a ious objec s wi hin images o classi y
hem in o dis inc ca ego ies (K. K. Singh & Singh, 2012). This me hod has a wide ange o applica ions,
including d i e assis ance sys ems and disease de ec ion. Wi hin he ealm o segmen a ion, he e a e
wo p ima y ypes: seman ic segmen a ion and ins ance segmen a ion. Gi en ha one o he main goals
o his wo k is o iden i y clo hing i ems o assis isually impai ed people, i is c ucial o pe o m p e-
p ocessing ac ions when cap u ing images o clo hing. This ensu es ha he image p esen ed o he use
is as “clean” as possible. The e o e, i is necessa y o segmen he image and emo e he backg ound
so ha he colou de ec ion algo i hm can achie e he bes esul s by analysing only he pixels ele an
Chap e 2 –
Li e a u e Re iew
19
o he clo hing i em. Below, a b ie e iew o exis ing implemen a ions in clo hing image segmen a ion is
p esen ed.
Yingheng and Yueqi (2020) p esen ed a me hod ha aids ashion analys s and consume s in iden i ying
ypes o clo hing. Segmen a ion esul s a e enhanced using a sys em ha inco po a es Mul i-A en ion
Mask R-CNN, enabling he ex ac ion o de ailed in o ma ion om clo hing i ems. This implemen a ion is
pa icula ly e ec i e in delinea ing bounda ies be ween clo hing i ems, ou pe o ming o he Mask R-CNN
implemen a ions.
Ano he solu ion, desc ibed by Khu ana
e al
. (2018) ocuses on he p oblem o iden i ying and de ec ing
clo hing i ems in images. This me hod di e en ia es be ween a ious ypes o clo hing ha may ha e
simila pa e ns o ex u es. The implemen a ion consis s o wo modules: he i s u ilizes a Fully
Con olu ional Ne wo k o spa ial bounda y segmen a ion, and he second de ec s speci ic ea u es o aid
in ecognizing bounda ies. The esul s demons a e an imp o emen o e exis ing me hods, educing
con usion in images wi h isually simila clo hing i ems made o di e en ma e ials.
DeepFashion2, a me hod designed o ex ac he maximum amoun o da a om pho os o clo hing i ems,
was p oposed by Ge
e al.
(2019). This me hod add esses he limi a ions o he DeepFashion sys em,
aiming o accu a e segmen a ion and iden i ica ion o all clo hing i ems in each pho o. Mo eo e , i
allowed gene a ing a da ase comp ising
ca
. 801,000 anno a ed images wi h iden i ied ypes and s yles.
These images we e segmen ed using masks and bounding boxes. The no el model, Ma ch R-CNN, is
based on Mask R-CNN (K. He, Gkioxa i, Dolla ,
e al
., 2017). The esul s indica e a sligh imp o emen
in A e age P ecision compa ed o he Mask R-CNN and Ma ch R-CNN models.
Complex backg ounds can make image segmen a ion a demanding ask. DeepLabV3+, p oposed by
Wang
e al.
(2021), is a model ha segmen s clo hing i ems wi hin complex backg ounds whe e he
bounda ies migh be ambiguous. A new Neu al Ne wo k a chi ec u e was edesigned o imp o e
segmen a ion pe o mance, showing be e adap abili y o he bounda ies o each clo hing i em. A e
aining he model and es ing i on he da ase , he esul s showed ha accu acy (97.26%), mean
in e sec ion o e union (IoU) (93.23%), and a e age p ecision (AP) (90.56%) alues we e sligh ly highe
han hose ob ained wi h he p e ious DeepLab 3+ model.
Tackling ano he issue, he me hod desc ibed by X. Zhang
e al
. (2020) ecognizes se s o clo hing i ems
and iden i ies he yea hey belong o. He e, a CNN was used o segmen he human body and classi y
he ype o clo hing wi hin he segmen ed a eas wi hou backg ound in e e ence. A da ase o 9,339
images spanning eigh yea s was used o alida e he model, demons a ing good e icacy. Va ious
segmen a ion models we e es ed, wi h SegNe achie ing an IoU alue o app oxima ely 0.951, closely
Chap e 2 –
Li e a u e Re iew
20
ma ching he p oposed me hod IoU alue o 0.951. The p oposed me hod also yielded he bes
classi ica ion esul s, wi h an accu acy o 0.805 compa ed o 0.785 using he ResNe 101 model.
Simila ly, T. Yang
e al.
(2021) p esen ed a me hod o segmen ing images o clo hing i ems cap u ed in
di e en posi ions allows o e ec i ely de ec ing he silhoue es o shi s using ins ance segmen a ion
wi h Mask R-CNN. App oxima ely 9,000 images we e collec ed, segmen ed, and classi ied by a ious
a ibu es. The expe imen al esul s showed ha he model pe o med well in de ec ing sho -slee ed
shi s and colla s, achie ing AP alues o 0.95 and 0.94, espec i ely.
O e all, he conduc ed li e a u e e iew on clo hing segmen a ion unde lines he signi icance o ad anced
segmen a ion echniques in enhancing he accu acy and e ec i eness o algo i hms o iden i ying and
analysing clo hing i ems.
Table 2.2 summa izes he men ioned wo ks on ad anced image segmen a ion echniques o clo hing
iden i ica ion.
Table 2.2: Summa y o ad anced image segmen a ion echniques o clo hing iden i ica ion
Me hod
Au ho s
Key Fea u es
Applica ions
Pe o mance
Fully Con olu ional
Ne wo k
Khu ana
e
al.
(2018)
Spa ial bounda y
segmen a ion, ea u e
de ec ion
Iden i ying and
de ec ing clo hing
i ems
Reduces con usion in
images wi h simila
pa e ns/ ex u es
DeepFashion2
Ge
e al.
(2019)
Ex ac s maximum da a,
masks and bounding
boxes
Accu a e segmen a ion
and iden i ica ion o
clo hing i ems
Sligh imp o emen in AP
compa ed o Mask R-CNN
and Ma ch R-CNN
Mul i-A en ion
Mask R-CNN
Yingheng
and Yueqi
(2020)
Enhances segmen a ion
esul s, de ailed
in o ma ion ex ac ion
Fashion analysis,
consume
iden i ica ion
Ou pe o ms o he Mask R-
CNN implemen a ions in
delinea ing bounda ies
SegNe
X. Zhang
e
al.
(2020)
Human body
segmen a ion, clo hing
ype classi ica ion
Recognizing se s o
clo hing i ems, yea
classi ica ion
IoU: 0.951, Classi ica ion
Accu acy: 0.805
DeepLabV3+
Wang
e al.
(2021)
Segmen s wi hin
complex backg ounds,
edesigned NN
a chi ec u e
Adap able
segmen a ion
pe o mance
Accu acy: 97.26%, Mean
IoU: 93.23%, AP: 90.56%
Ins ance
Segmen a ion wi h
Mask R-CNN
T. Yang
e
al.
(2021)
Segmen s images o
clo hing in di e en
posi ions
De ec ing silhoue es
o shi s
AP: 0.95 (sho -slee ed
shi s), AP: 0.94 (colla s)
3 THEORICAL CONCEPTS
Chap e O e iew
This chap e ou lines he me hodologies used in he p esen wo k. I begins wi h an o e iew o blindness
and con inues wi h a discussion on A i icial In elligence, Machine Lea ning (ML), and Deep Lea ning (DL),
emphasizing hei oles in de eloping “in elligen ” sys ems. The heo e ical concep s o Neu al Ne wo ks
and T ans e Lea ning a e p esen ed, ollowed by he e alua ion me ics used o assess model
pe o mance.
3 Theo ical concep s
3.1 Blindness
3.2 A i icial In elligence, Machine Lea ning and Deep Lea ning
3.3 Neu al Ne wo ks
3.4 T ans e Lea ning
3.5 E alua ion Me ics

Chap e 3 –
Theo ical Concep s
___________________________________________________________________________
28
3.1 Blindness
The human eye can be b oadly compa ed o a pho og aphic came a in e ms o i s unc ioning. Ligh
e lec ed om objec s passes h ough he co nea, pupil, and lens, e en ually eaching he e ina (
The
Ana omy o he ‘Came a’ Eye | UCL Ins i u e o Oph halmology - UCL – Uni e si y College London
, 2024).
He e, specialized cells encode he isual in o ma ion, which is hen ansmi ed o he b ain ia he op ic
ne e, as illus a ed in Figu e 3.1.
Figu e 3.1: Cons i uen s o he human eye, illus a ing he pa hway o ligh om he co nea o he e ina and he op ic ne e,
and compa ing i s unc ioning o a pho og aphic came a(
The Ana omy o he ‘Came a’ Eye | UCL Ins i u e o Oph halmology
- UCL – Uni e si y College London
, 2024) .
Among he i e senses domina ed by mos humans, ision plays a pa icula ly p ominen ole. A disabili y,
by de ini ion, e e s o he pa ial o o al loss o a pa o he human body o i s unc ion, whe he physical
o psychological, which hinde s an indi idual's abili y o pe o m common daily ac i i ies (Enoch
e al
.,
2019; Kho ami-Nejad
e al
., 2016). Visual impai men is one o he signi ican ypes o disabili ies and
can be ca ego ized in o wo main o ms. The i s is low ision, which e e s o isual impai men s ha
canno be ully co ec ed by s anda d means such as glasses o con ac lenses. The second o m is
blindness, which deno es he o al loss o ision (Medei os
e al
., 2017). Visual impai men , whe he
pa ial o o al, signi ican ly a ec s an indi idual's quali y o li e, limi ing hei abili y o engage in daily
ac i i ies independen ly. This limi a ion poin s ou he impo ance o echnological ad ancemen s aimed
a suppo ing isually impai ed people. Inno a i e solu ions such as assis i e de ices and applica ions
ha e been de eloped o enhance hei au onomy and imp o e hei quali y o li e. Technological aids ange
Chap e 3 –
Theo ical Concep s
___________________________________________________________________________
29
om basic ools, like magni ying glasses and b aille eade s, o ad anced digi al applica ions ha le e age
a i icial in elligence and machine lea ning. Fo ins ance, mobile applica ions ha p o ide eal- ime audio
desc ip ions o he use 's su oundings o sma de ices ha assis in iden i ying objec s and na iga ing
spaces a e becoming inc easingly p e alen (Ghazal
e al
., 2019; Najm
e al
., 2022). These echnologies
no only o e p ac ical suppo bu also con ibu e o he social inclusion o isually impai ed people by
enabling hem o pe o m asks ha would o he wise equi e assis ance.
In summa y, he eye unc ions simila ly o a came a by cap u ing and ansmi ing isual in o ma ion o
he b ain. Vision is a c i ical sense ha g ea ly in luences an indi idual's daily unc ioning. Visual
impai men s, whe he in he o m o low ision o blindness, p esen signi ican challenges ha can be
mi iga ed h ough he de elopmen and u iliza ion o a ious assis i e echnologies.
3.2 A i icial In elligence, Machine Lea ning and Deep Lea ning
A i icial in elligence is he de elopmen o compu e sys ems ha can pe o m asks adi ionally handled
by human in elligence,
e.g
, speech ecogni ion, unde s anding o na u al language, and decision-making
(Soo i
e al
., 2023). The ields o machine lea ning (ML) and deep lea ning (DL) a e sub ields o a i icial
in elligence, Figu e 3.2, ha use echniques o au oma ic lea ning o ain compu e models om da a
o e ime, esul ing in imp o ed pe o mance.
Figu e 3.2: In e connec ion be ween he ields/sub ields o AI/ML/DL.
ML in ol es aining compu e models o lea n om da a and make p edic ions (Sa ke , 2021b), whe eas
DL, a subse o ML, employs deep neu al ne wo ks wi h mul iple laye s o s ages h ough which da a is
p ocessed o building a da a-d i en model o disco e da a s uc u es au oma ically (Sa ke , 2021a).
Chap e 3 –
Theo ical Concep s
___________________________________________________________________________
30
3.3 Neu al Ne wo ks
Neu al ne wo ks a e a undamen al componen o deep lea ning, modelled a e he s uc u e and unc ion
o he human b ain. They ypically consis o mul iple laye s, including an inpu laye , one o mo e hidden
laye s, and an ou pu laye as depic ed in Figu e 3.3. The hidden laye s a e c ucial o lea ning complex
pa e ns and ep esen a ions om he da a (K iegesko e & Golan, 2019).
Figu e 3.3: Deep neu al ne wo k wi h wo hidden laye s.
Du ing aining, neu al ne wo ks adjus hei pa ame e s (weigh s and biases) h ough a p ocess called
backp opaga ion. Backp opaga ion is essen ial o he ne wo k o g adually imp o e i s abili y o make
accu a e p edic ions o pe o m speci ic asks (Lecun
e al
., 2015). This lea ning p ocess in ol es
minimizing he di e ence be ween he ne wo k's p edic ions and he ue a ge s, which is ypically
achie ed h ough op imiza ion algo i hms. Op imiza ion algo i hms a e i al in his p ocess, wi h
commonly used me hods including s ochas ic g adien descen (SGD), Adam, and RMSp op (Kingma &
Ba, 2014; Rude , 2016). These algo i hms i e a i ely adjus he lea ning a e and pa ame e s o imp o e
con e gence and ind he op imal se o weigh s and biases ha minimize a p ede ined loss unc ion. The
loss unc ion quan i ies he disc epancy be ween p edic ed and ac ual alues. The deep a chi ec u e o
neu al ne wo ks, cha ac e ized by mul iple hidden laye s, enables hem o cap u e hie a chical ea u es
and in ica e ela ionships wi hin he da a (Lecun
e al
., 2015). Fo example, in image ecogni ion asks,
lowe laye s may de ec basic ea u es such as edges and ex u es, while highe laye s ecognize mo e
complex s uc u es like shapes and objec s. Neu al ne wo ks ha e demons a ed ema kable e sa ili y
and e ec i eness ac oss a wide ange o applica ions, om na u al language p ocessing and speech
ecogni ion o medical diagnosis and au onomous d i ing. This adap abili y s ems om hei abili y o
Chap e 3 –
Theo ical Concep s
___________________________________________________________________________
31
model non-linea ela ionships and ep esen high-dimensional da a comp ehensi ely.
3.3.1 Con olu ional Neu al Ne wo ks
Con olu ional Neu al Ne wo ks a e a class o deep lea ning models speci ically designed o image
ecogni ion and compu e ision asks (K izhe sky
e al
., 2012). They ha e demons a ed ema kable
success in a ious applica ions, including image classi ica ion, objec de ec ion, and image segmen a ion
(Minaee
e al
., 2022; Rawa & Wang, 2017; Voulodimos
e al
., 2018b; Z.-Q. Zhao
e al
., 2019). CNNs a e
composed o mul iple laye s, each asked wi h ex ac ing di e en ypes o in o ma ion om he inpu
da a. The a chi ec u e o CNNs comp ises wo p ima y modules: con olu ion and classi ica ion, as
illus a ed in Figu e 3.4.
Figu e 3.4: The a chi ec u e o a Con olu ional Neu al Ne wo k (CNN) comp ises an inpu laye , se e al al e na ing
con olu ional and max-pooling laye s, ollowed by a ully-connec ed laye and a classi ica ion laye (Alom
e al
., 2019)
Du ing he con olu ion s age, he ne wo k employs il e s o ex ac speci ic elemen s om he images,
abs ac ing and di iding he image in o dis inc pa s. The p ocessed da a is hen passed h ough a pooling
laye , which educes he dimensionali y and compu a ional load by downsampling he ea u e maps. In
he inal s age, he ully connec ed laye classi ies he in o ma ion ob ained om he p e ious laye s. By
in eg a ing hese h ee s ages—con olu ion, pooling, and classi ica ion—CNNs acqui e he abili y o iden i y
and ecognize egions o in e es wi hin each image.
One o he signi ican ad an ages o CNNs is hei educed need o p e-p ocessing compa ed o o he
classi ica ion algo i hms, as hey can op imize il e s h ough lea ning (X. Zhao
e al
., 2024). Each neu on
in a CNN calcula es a alue based on he inpu s and a speci ic ac i a ion unc ion. This alue is adjus ed
using a weigh ec o , which is upda ed du ing he ne wo k's lea ning p ocess. These weigh ec o s
Chap e 3 –
Theo ical Concep s
___________________________________________________________________________
32
ep esen he ea u es ha he ne wo k aims o p io i ize.
3.3.2 Objec De ec ion
Objec de ec ion, which in ol es iden i ying and localizing mul iple objec s wi hin an image, is a
undamen al ask in compu e ision. Se e al CNNs a chi ec u es ha e been de eloped o add ess his
complex ask (S. Aga wal
e al
., 2019). These a chi ec u es can be b oadly ca ego ized in o one-s age and
wo-s age de ec o s based on hei a chi ec u e and p ocessing s eps. One-s age de ec o s, such as YOLO
and Single-Sho Mul iBox De ec o (SSD), pe o m objec de ec ion in a single pass h ough he ne wo k.
They di ide he image in o a g id and p edic bounding boxes and class p obabili ies di ec ly om he ull
image. This app oach is ypically as e and mo e sui able o eal- ime applica ions, hough i may
sac i ice some accu acy compa ed o wo-s age de ec o s(So iany & Ionescu, 2018; Y. Zhang
e al
.,
2021).
Since i s ini ial elease in 2015, YOLO has e ol ed signi ican ly, eaching i s eigh h e sion, YOLO 9 (C.-
Y. Wang
e al
., 2024). Unlike egion-based algo i hms, YOLO di ec ly p edic s bounding box loca ions and
class p obabili ies (Redmon
e al
., 2016). The en i e image is di ided in o a g id, wi h each g id cell
p edic ing he bounding boxes and class p obabili ies o objec s de ec ed wi hin i as depic ed in Figu e
3.5: In he YOLO a chi ec u e, he inpu image is segmen ed in o an S × S g id, wi h each g id cell
p edic ing B bounding boxes cha ac e ized by hei posi ion, size, and con idence sco e. The inal label o
a high-con idence bounding box is de e mined by he class p obabili y map (Redmon
e al
., 2016).

Chap e 3 –
Theo ical Concep s
___________________________________________________________________________
33
Figu e 3.5: In he YOLO a chi ec u e, he inpu image is segmen ed in o an S × S g id, wi h each g id cell p edic ing B bounding
boxes cha ac e ized by hei posi ion, size, and con idence sco e. The inal label o a high-con idence bounding box is
de e mined by he class p obabili y map (Redmon
e al
., 2016).
This app oach allows YOLO o pe o m objec de ec ion in a single pass h ough he ne wo k, enabling
highly e icien and eal- ime objec de ec ion. Each i e a ion o YOLO has b ough imp o emen s in
accu acy, speed, and a chi ec u e, wi h YOLO 9 bene i ing om he la es ad ancemen s in deep lea ning
and compu e ision esea ch (Te en
e al
., 2023; C.-Y. Wang
e al
., 2024).
SSD, in oduced by Liu Wei and Anguelo (2016) p esen s a me hod o de ec ing objec s in images using
a single deep neu al ne wo k. This app oach disc e izes he ou pu space o bounding boxes in o a se o
p io s wi h a ying aspec a ios and scales o each ea u e map loca ion. Du ing p edic ion, he ne wo k
gene a es con idence sco es o each p io , indica ing he p esence o objec s o in e es , and adjus s he
p io s o be e i he objec shapes. Addi ionally, he ne wo k combines p edic ions om mul iple ea u e
maps wi h di e en esolu ions o na u ally handle objec s o a ious sizes. This one-s age de ec o is
enowned o i s balance be ween speed and accu acy. A compa ison o SSD and YOLO a chi ec u es is
depic ed in Figu e 3.6.
Chap e 3 –
Theo ical Concep s
___________________________________________________________________________
34
Figu e 3.6: Compa ison o wo single-sho de ec ion models: SSD and YOLO. The SSD model augmen s he base ne wo k wi h
addi ional ea u e laye s ha p edic adjus men s o de aul boxes o a ying scales and aspec a ios, along wi h co esponding
con idence sco es (Liu Wei and Anguelo , 2016).
Two-s age de ec o s, such as R-CNN, Fas R-CNN, and Fas e R-CNN, ope a e in wo s eps. The i s s age
gene a es egion p oposals ha a e likely o con ain objec s. The second s age p ocesses hese p oposals
o e ine he bounding boxes and classi y he objec s. This app oach gene ally achie es highe accu acy
bu can be slowe and mo e compu a ionally in ensi e.
The R-CNN (Gi shick
e al
., 2014), Figu e 3.7, was one o he i s success ul CNN-based me hods o
objec de ec ion. I in oduced a mul i-s ep pipeline consis ing o Region P oposal, Fea u e Ex ac ion, and
Objec Classi ica ion and Localiza ion. While R-CNN made signi ican con ibu ions o he ield, i s
compu a ional and memo y ine iciencies limi ed i s p ac ical use in eal- ime and la ge-scale applica ions.
Figu e 3.7: The mul i-s ep pipeline o R-CNN, including egion p oposal, ea u e ex ac ion, and objec classi ica ion and
localiza ion (Gi shick
e al
., 2014) .
Chap e 3 –
Theo ical Concep s
___________________________________________________________________________
35
To add ess he ine iciencies o R-CNN, Fas R-CNN was de eloped. This a chi ec u e op imizes he objec
de ec ion pipeline, esul ing in as e and mo e e icien p ocessing (Gi shick, 2015). Key imp o emen s
include ea u e sha ing, a uni ied a chi ec u e, and end- o-end aining. Fas e R-CNN, Figu e 3.8, u he
ex ends and imp o es upon he o iginal R-CNN a chi ec u e, by in oducing a no el Region P oposal
Ne wo k (RPN) ha sha es con olu ional ea u es wi h he objec de ec ion ne wo k (Ren
e al
., 2015).
Figu e 3.8: Fas e R-CNN is a uni ied ne wo k o objec de ec ion, whe e he RPN module unc ions as he 'a en ion'
mechanism wi hin he ne wo k (Ren
e al
., 2015).
This inno a ion add esses he compu a ional and memo y ine iciencies o R-CNN. Fas e R-CNN also
se es as he ounda ion o Mask R-CNN, which adds ins ance segmen a ion capabili ies.
Re inaNe , p oposed by Lin
e al.
(2018) in oduces he Focal Loss, a loss unc ion ha add esses he
class imbalance p oblem by ocusing mo e on ha d- o-classi y samples. This a chi ec u e s ikes a balance
be ween he speed o one-s age de ec o s and he accu acy o wo-s age de ec o s. Objec de ec ion
con inues o e ol e wi h new me hods and imp o emen s, d i en by ongoing ad ancemen s in deep
lea ning and compu e ision. The choice o he ideal a chi ec u e depends on he speci ic applica ion
equi emen s, such as accu acy, speed, and he capabili y o ope a e in eal- ime.
3.3.3 Image Segmen a ion
Image segmen a ion, a closely ela ed bu dis inc ask om objec de ec ion, in ol es pa i ioning an
image in o mul iple segmen s o egions o loca e objec s and bounda ies accu a ely. Unlike objec
de ec ion, which p o ides bounding boxes a ound de ec ed objec s, image segmen a ion iden i ies he
Chap e 3 –
Theo ical Concep s
___________________________________________________________________________
36
exac pixels belonging o each objec , p o iding a mo e de ailed unde s anding o he scene. Ins ance
segmen a ion, a sub ype o image segmen a ion, no only de ec s objec s bu also segmen s each objec 's
pixels o p o ide a pixel-le el mask. This app oach is pa icula ly use ul in applica ions whe e p ecise
objec bounda ies a e equi ed (Sha ma
e al
., 2022) .
Se e al popula deep lea ning models a e used o image segmen a ion. U-Ne (Ronnebe ge
e al
., 2015)
employs a U-shaped a chi ec u e o he e icien segmen a ion o medical images, demons a ing high
e icacy wi h small da ase s and deli e ing p ecise segmen a ion esul s. Fully Con olu ional Ne wo ks
(FCNs) a e capable o p ocessing images o any size and gene a ing spa ial maps by subs i u ing ully
connec ed laye s in con en ional CNNs wi h con olu ional laye s, enabling pixel-by-pixel segmen a ion o
en i e images. SegNe (Bad ina ayanan
e al
., 2016) ea u es an encode -decode ne wo k s uc u e
u ilized in asks such as scene unde s anding and objec ecogni ion. The encode cap u es con ex ual
in o ma ion om he image, while he decode le e ages his con ex o p ecise localiza ion and
segmen a ion o objec s. DeepLab (L.-C. Chen
e al
., 2017) is dis inguished by i s use o a ous
con olu ions, which cap u e mul i-scale con ex h ough mul iple pa allel il e s.
Mask R-CNN, an ex ension o Fas e R-CNN, is a p ominen a chi ec u e o ins ance segmen a ion. Mask
R-CNN enhances Fas e R-CNN's objec de ec ion by p edic ing pixel masks o de ec ed objec s. I begins
wi h a backbone ha ex ac s ea u e maps om he inpu image, cap u ing hie a chical and con ex ual
in o ma ion. The RPN gene a es egion p oposals (bounding box candida es) o po en ial objec s. These
p oposals a e p ocessed h ough he Region o In e es (ROI) Align laye o ex ac ixed-size ea u e maps,
which a e hen passed h ough wo b anches: he egion o in e es classi ie and he bounding box
eg esso (K. He, Gkioxa i, Dolla ,
e al
., 2017). The classi ie p edic s he p obabili ies o objec classes,
and he eg esso e ines he bounding box coo dina es. An addi ional b anch, he mask p edic ion head,
p oduces a bina y mask o each objec ins ance, ep esen ing pixel-le el segmen a ion and dis inguishing
he objec om he backg ound as illus a ed in Figu e 3.9
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
43
cha ac e iza ion o he pa icipan , such as age, gende , academic quali ica ions and cu en occupa ion.
The second pa o he ques ionnai e,
Iden i ica ion o he ype o disabili y
, enabled he analysis o
condi ions ha caused blindness in he pa icipan s, as well as he age g oup in which hese condi ions
occu ed.
The hi d pa ,
Clo hing
, includes 6 ques ions o iden i y he impo ance and conce ns ha blind people
ha e ega ding he ac ions hey need o do o ge d essed, o choose clo hes and how hey do shopping.
In his subjec , some ques ions we e p esen ed as s a emen s. The answe s we e based on a 5-poin
Like
scale o ag eemen , impo ance o equency.
The ou h pa ,
Technology
, includes 3 ques ions ha aim o assess he knowledge, usabili y and
sa is ac ion ha blind communi y has in ela ion o he a ailable suppo i e echnologies. Howe e , i is
also possible o assess he p oximi y ha his communi y has wi hin his echnology.
The i h and las pa ,
Resea ch and de elopmen
, includes 3 main ques ions and o he 2
complemen a y. In his sec ion, i is in ended o assess he a ailabili y and he in e es in de eloping a
echnology ha suppo s he blind communi y iden i ying and combining clo hes. Mo eo e , his sec ion
aims o assess i blind people wan his echnology o be a ailable. The sugges ions and opinions o he
pa icipan s we e aken in o conside a ion in he inal e sion.
4.1.4 Implemen a ion
The da a collec ed using google o ms om July o Sep embe 2020 and he pa icipa ion was olun a y
and anonymous. The da a was ga he ed in a able and analysed using he s a is ical so wa e SPSS
(S a is ical Package o he Social Sciences) e sion 22. A his ini ial s age o analysis o he collec ed
da a, a desc ip i e s a is ic in which da a was ansla ed in o pe cen age alues was conside ed.
4.1.5 Pa icipan s Cha ac e iza ion
The sample o was essen ially composed by blind associa es o ACAPO, who we e asked o ill a pu pose-
buil su ey, deli e ed by email. A o al o 26 pa icipan s eplied o he su ey, o which 23 ull esponses
ul illed all he se condi ions, and wi h only 3 being excluded due o no ul illing he a o emen ioned
condi ions. In he p esen s udy, all he analyses we e made conside ing he ull esponses.
To ob ain he cha ac e iza ion o he pa icipan s in his su ey, 4 a iables we e de ined: age, gende ,
academic quali ica ions and p o ession.

Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
44
I was obse ed ha he e was no equal dis ibu ion be ween age g oups, wi h indi iduals be ween 45
and 54 yea s old p esen ing 52.2% o he answe s, he younges pa icipan being 18 yea s old, and he
oldes being 65 yea s old. Due o he wide ange o pa icipan ’s age, du ing he analysis o he su ey,
he a iable age was g ouped in o wo main g oups, one o pa icipan s being 44 o less yea s old, and
he second o pa icipan s being 45 yea s old o mo e. Rega ding gende and educa ion, mos o he
pa icipan s we e emale (73.9%), and had highe educa ion (43.5%), wi h only 26.1% ha ing, a leas ,
nine yea s o basic educa ion. In he analysis o he occupa ions, i was ound ha mos o hem we e
unemployed (30.4%), which highligh s he di icul y o his popula ion o ge in eg a ed in he labou
ma ke . F&B ( ood & be e age) and adminis a i e se ices we e he occupa ion a eas wi h g ea e
ep esen a i eness, accoun ing o 13% each. O he ca ego ies include all e i ed people and s uden s,
ep esen ing 17%.
4.1.6 Iden i ica ion o he Type o Disabili y
The second pa o he ques ionnai e con ains wo ques ions enabling each pa icipan o iden i y he
causes and ypes o condi ions ha caused he pa icipan s' loss o ision, as ollows:
QII.1: Since wha age a e you isually impai ed?
QII.2: Wha caused you disabili y?
On he basis o he ecei ed esponses, he majo i y (69.6%) became blind du ing he i s yea o li e.
This shows ha hese pa icipan s do no ha e basic isual no ions abou he “wo ld” a ound hem. An
iden ical esul was ob ained ega ding he eason o he loss o ision, ha is, 65.2% had los hei ision
due o congeni al disease, and only 8.7% as a esul o an acciden . The emaining 26.1% e e s o
acqui ed diseases o condi ions. Congeni al disease occu s du ing p egnancy, which explains he high
numbe o pa icipan s wi h ision loss be o e eaching one yea old. Hence, he condi ions ha o igina ed
blindness o ision loss we e g ouped in wo main ca ego ies, congeni al disease and acqui ed disease,
wi h glaucoma and e inopa hy being he main causes o ision loss, espec i ely.
4.1.7 Aes he ics and Clo hing
The way we d ess has a big impac on ou daily li es and i is c ucial o ou well-being. This pe cep ion
anges om he impo ance ha socie y gi es o he way we d ess o he impo ance o choosing clo hing
o he wo kplace, o help con ey he igh and app op ia e image o he con ex in which we a e inse ed
in. So, in he hi d pa o he ques ionnai e, h ee s a emen s we e conside ed:
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
45
QIII.1: Vision is one o he senses ha domina es he li e o human beings. I allows hem o know and
pe cei e he wo ld a ound hem, while gi ing meaning o objec s, concep s, ideas and as es.
QIII.2: How o d ess and he s yle we p e e o di e en occasions is pa o someone's iden i y. Blind
people do no ha e his sense, and d essing can be a di icul and s ess ul ask. Wi h he ad ancemen
o echnology, i is impo an o minimize all he limi a ions o a blind pe son whils o ganizing hei own
clo hing.
QIII. 3: The lack o knowledge abou he colou s, he ype o pa e n o he condi ion o he ga men s
makes his a daily challenge, in which he cu en esou ces a e no he bes .
And he ollowing h ee ques ions:
QIII.4: A e you wo ied abou ma ching ga men s?
QIII.5: How o en do you need help buying clo hing?
QIII.6: Whe e do you buy you clo hes?
Due o he ela i ely small size o he sample, o all s a emen s, he scale was ecoded so ha only wo
g oups we e possible: "ag ee" and "disag ee", showing a posi i e o nega i e esponse o each
s a emen . Fig.1 shows he dis ibu ion o esul s o he di e en s a emen s.
Figu e 4.1: Posi i e and nega i e esponse dis ibu ion o each s a emen (Rocha
e al
., 2022).
The majo i y o pa icipan s ecognize ha he sense o ision is c ucial o he a ibu ion o meanings and
ecogni ion o e e y hing ha su ounds us. The impo ance o aes he ics as a de ini ion o ou iden i y is
also in ag eemen among he pa icipan s, as well as he daily di icul ies expe ienced by pa icipan s due
o he lack o esou ces. As expec ed, all pa icipan s ha e ein o ced he impo ance o combine clo hing
o hem, he majo i y o which (73.9%) saying ha hey equen ly seek help o shop hei clo hes, mainly
om amily membe s, suppo cen e p o essionals and CAVI (Suppo Cen e o Independen Li ing).
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
46
When asked whe e hey buy hei clo hes, all pa icipan s e e ed physical s o es, wi h 39.1% also
e e ing o do shopping online.
4.1.8 Technology
Technology has e ol ed conside ably in his ield; howe e , i is necessa y o unde s and whe he he e is
enough echnology o minimize he obs acles aced by people wi h blind condi ion o ision loss, in hei
daily li es. So, in he ou h pa o he ques ionnai e, a e he iden i ica ion o he de ice and sys em
used, wo ques ions we e conside ed:
QIV.1: A e you awa e o any echnology o clo hing iden i ica ion?
QIV.2: How sa is ied a e you wi h he echnology a ailable on his opic?
Mos o pa icipan s (65.2%) demons a ed o be awa e o exis ing clo hing iden i ica ion echnologies.
Sma phone apps we e poin ed as main sou ce o aid o de ec colou s. As an al e na i e, he same use s
op o Colo ino (
Colo ino – Colo and Ligh De ec o - Ca e ec
, 2024). Howe e , only 26.1% e e ed o
be sa is ied wi h he exis ing echnology, which ein o ces he impo ance o he opic add essed on his
wo k con ibu es o he mo i a ion behind i . O e all,
ca.
89% o pa icipan s e e ed hey used
sma phones, 72% used a lap op, 33% used mobile phones and 11% able de ices. Mos sma phone
use s e e ed hey had he IOS ope a ing sys em.
4.1.9 Resea ch and De elopmen
Aes he ics has an impac on he image and appea ance o ou body, di ec ly in luencing ou well-being,
sel -es eem and op imism in “day- o-day” li e. The esea ch and de elopmen conduc ed in his wo k
sough o include hese cha ac e is ics. The e o e, in he i h pa o he ques ionnai e, h ee main
ques ions we e conside ed:
QV.1: Would i be impo an o de elop some hing ha would make i possible o iden i y and combine
clo hing?
QV.2: How likely is i o use a echnology ha includes he ac o iden i ying and combining clo hing?
QV.3: Did you need help in comple ing his su ey?
Acco ding o Figu e 4.2, mos pa icipan s (91.3%) ag eed ha a amewo k should be de eloped o help
iden i ying and ma ching ga men s. In line wi h he p e ious answe , almos all pa icipan s (95.7%) we e
willing o use echnology o help hem iden i ying and combining clo hes. When choosing he pla o m o
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
47
use he echnology, he sma phone and he lap op we e he p e e ed choices, accoun ing o 43.2% and
35.1%, espec i ely. Las ly, only 17.4% o pa icipan s s a ed ha hey sough some help o ill ou he
ques ionnai e.
Figu e 4.2: Resul s dis ibu ion o QV.1 and QV.2 (Rocha
e al
., 2022).
4.1.10 Conclusions o he su ey
Based on he collec ed and analysed in o ma ion, i becomes impo an o de elop ools ha allow he
iden i ica ion and combina ion o ga men s. Mo eo e ,
ca
. 95.5% o he esponden s a e a ailable o use
echnology o help hem in his ask. Al hough he sample dimension used in he analysis could be
conside ed small (n=23) o be ep esen a i e o he blind popula ion a na ional le el, he au ho s belie e
ha he collec ed esul s ein o ce he cen al impo ance o aes he ics on clo hing o blind people, who
a e becoming inc easingly inse ed in he labou ma ke , ollowing global and Eu opean guidelines o
inclusion (
Reduzi as Desigualdades • ODS - BCSD Po ugal
, n.d.).
Th ough he su ey unde aken by he pa icipan s, i is possible o conclude ha no knowing colou s,
pa e ns, and he o e all condi ion o clo hes is a majo challenge ha blind people ace e e y day. This
comes o emphasize he need o iden i y and combine clo hing. Despi e o all he di icul ies, he exis ing
o e o di e en echnologies and suppo is e y limi ed. So a , mobile phone applica ions and Colo ino
emain he leading sou ces o suppo used by blind people. Analysing he commen s, pa icipan s ha e
demons a ed hei en husiasm and willingness o collabo a e in his esea ch.
4.2 Sys em Requi emen s
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
48
A e analysing he esul s o he conduc ed su ey and summa izing he key aspec s o a sys em ha
could help blind people choose and cha ac e ize hei clo hes, especially wi h ega d o hei well-being,
he sys em a chi ec u e and equi emen s we e designed. The design phase in ol ed iden i ying c i ical
unc ionali ies and componen s ha would e ec i ely add ess he needs highligh ed by he su ey
esponden s. The p oposed sys em(Rocha
e al
., 2021) aims o de elop a comp ehensi e solu ion ha
includes he ollowing key equi emen s:
• Clo hing Classi ica ion: The sys em should accu a ely iden i y and ca ego ize clo hing i ems based
on ype (e.g., shi s, d esses).
• Colou De ec ion: I should p ecisely de e mine he colou s o clo hing i ems o assis use s in
making app op ia e wa d obe choices.
• De ec De ec ion: The sys em mus iden i y de ec s such as s ains and holes o ensu e ha use s
a e awa e o he condi ion o hei clo hing.
• Use -F iendly Mobile In e ace: The mobile applica ion should be highly accessible and in ui i e,
allowing blind and low- ision use s o na iga e and use he sys em wi h ease.
• Sma Wa d obe In eg a ion: The sys em should include a sma wa d obe equipped wi h an
image acquisi ion sys em and con olled ligh ing o cap u e high-quali y images o clo hing i ems.
• NFC Tagging: The sys em should suppo NFC agging o acili a e he o ganiza ion and e ie al
o clo hing in o ma ion.
• Clo hing Managemen : The applica ion should enable use s o manage hei wa d obe, including
adding new i ems, upda ing in o ma ion, and e i ying i em de ails.
Once he sys em equi emen s we e de ined, i was possible o ou line he me hodology o i s concep ion,
as explained in he ollowing sec ion.
4.3 Sys em O e iew
The de elopmen o he iSigh p o o ype was s uc u ed a ound h ee main componen s: (i) he collec ion
o da a and he applica ion o deep lea ning algo i hms o ga men iden i ica ion and modi ica ion
de ec ion, (ii) he design and implemen a ion o he mecha onic de ice, speci ically he au oma ed
wa d obe sys em, and (iii) he c ea ion o a use - iendly mobile applica ion. The wo k low illus a ing he
main s eps o his me hodology is p esen ed in Figu e 4.3.

Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
49
Figu e 4.3: Schema ic ep esen a ion o he wo k low comp ised in he iSigh p o o ype de elopmen .
The ollowing subsec ions p o ide a de ailed accoun o each componen . Ini ially, he da a collec ion
p ocess and he deploymen o ad anced deep lea ning algo i hms a e discussed, ocusing on hei ole
in accu a ely iden i ying ga men s and de ec ing modi ica ions. Subsequen ly, he design, cons uc ion,
and in eg a ion o he au oma ed wa d obe sys em is elabo a ed (Chap e 5), highligh ing i s impo ance
as he co e mecha onic elemen o he iSigh p o o ype. Finally, he de elopmen o he mobile applica ion
is explo ed, wi h an emphasis on i s use - iendly in e ace and seamless in eg a ion wi h he au oma ed
wa d obe, ensu ing accessibili y and ease o use o isually impai ed use s.
4.4 Deep Lea ning App oaches
This subsec ion de ails h ee expe imen s using deep lea ning echniques. The i s expe imen add esses
he classi ica ion o clo hing ca ego ies using CNNs. The second expe imen expands his app oach by
inco po a ing bo h segmen a ion and classi ica ion o clo hing ca ego ies, in addi ion o ex ac ing colou
in o ma ion. Las ly, he hi d expe imen ocuses on iden i ying modi ica ions in clo hing i ems.
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
50
4.4.1 Fi s Expe imen - Clo hing Ca ego y Classi ica ion using CNN
In he i s expe imen , s a e-o - he-a deep lea ning echniques we e le e aged h ough ans e lea ning
o pe o m ca ego y classi ica ion using CNNs (Rocha
e al
., 2019). The pipeline o applying deep lea ning
o clo hing ca ego y classi ica ion in ol es se e al key s eps. The p ocess begins wi h he selec ion,
p epa a ion, and p e-p ocessing o a da ase . In pa allel, se e al da a augmen a ion echniques a e
employed o examine hei e ec s on he esul s wi h and wi hou i s use. Finally, ans e lea ning is
implemen ed, using s a e-o - he-a deep lea ning models o pe o m ca ego y classi ica ion using CNNs.
Figu e 4.4 illus a es his wo k low.
Figu e 4.4: Wo k low me hodology o he classi ica ion o clo hing ca ego ies (Rocha, Soa es,
e al
., 2023a).
4.4.1.1 Da ase p epa a ion
To apply deep lea ning o clo hing ca ego y classi ica ion, a da ase was needed be o e being ed in o he
neu al ne wo ks. As a i s s ep, i was necessa y o iden i y a da ase ha could mee his wo k’s
equi emen s. Conside ing ha all images aken by he blind people a e collec ed in a con olled
en i onmen and wi h one i em o clo hing a a ime (Rocha
e al
., 2020), and based on he au ho s'
p e ious esea ch (Rocha
e al
., 2021), he au ho s we e able o conclude ha a easonable quan i y o
da a was equi ed o aining and ob aining mo e accu a e esul s. Fo his eason, a esea ch su ey
was conduc ed o iden i y a ailable da ase s - Table 4.1.
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
51
Table 4.1: Summa y o a ailable da ase s o ashion ca ego y classi ica ion (Rocha, Soa es,
e al
., 2023a).
Da ase
Yea
# o
pho os
Desc ip ion
DeepFashion-C (Z. Liu, Luo,
e al
.,
2016)
2016
289,222
Anno a ed wi h clo hing bounding box, pose
a ia ion ype, landma k isibili y, clo hing ype,
ca ego y, and a ibu es.
Fashion Landma k Da ase (Z. Liu,
Yan,
e al
., 2016)
2016
123,016
Anno a ed wi h clo hing ype, pose a ia ion
ype, landma k isibili y, clo hing bounding box,
and human body join .
FashionMinis (Xiao
e al
., 2017)
2019
70,000
G ayscale image da ase associa ed wi h a label
om 10 classes.
DeepFashion2 (Ge
e al
., 2019)
2019
491,000
A e sa ile benchma k o ou asks including
clo hes de ec ion, pose es ima ion,
segmen a ion, and e ie al.
Fashion P oduc Images (Fashion
P oduc Images Da ase | Kaggle,
n.d.)
2019
44,400
Anno a ed wi h gende , mas e ca ego y,
subca ego y, a icle ype, base colou , season,
yea , usage and p oduc desc ip ion.
Based on he cha ac e is ics o each da ase , i was decided o use he Fashion P oduc Images Da ase ,
which is a smalle esolu ion e sion o he da ase . This da ase p o ides a a ie y o a ibu es ha a e
ele an o he cu en and u u e needs o he p ojec , such as ca ego y, colou , season, among o he s.
Howe e , despi e he la ge numbe o anno a ed ea u es included in he da ase , i s ill p esen s
unbalanced da a be ween a icle ypes, which we e e e ed o as “ca ego ies” in his wo k. Figu e 4.5
shows he clo hing ca ego ies o which a leas 500 images we e conside ed in hei dis ibu ion,
i.e.
each en y on he
Xx
axis ep esen s a speci ic clo hing ca ego y.
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
52
Tshi s
Shi s
Casual Shoes
Wa ches
Spo s Shoes
Ku as
Tops
Handbags
Heels
Sunglasses
Walle s
Flip Flops
Sandals
B ie s
Bel s
Backpacks
Socks
Fo mal Shoes
Pe ume and Body Mis
Jeans
Sho s
T ouse s
Fla s
0
1000
2000
3000
4000
5000
6000
7000
8000
Reco ds
Ca ego y
Figu e 4.5: Dis ibu ion o he numbe o eco ds o each a icle ype/ca ego y (Fashion P oduc Images da ase ) (Rocha,
Soa es,
e al
., 2023a).
Fo a comp ehensi e and unbiased compa ison be ween di e en ca ego ies, he da ase was ca e ully
cu a ed by selec ing exac ly 500 eco ds om he ini ial comple e ange o each ca ego y as shown in
Figu e 4.5. This app oach a oids imbalanced da a, ensu ing ha each ca ego y in he da ase is
adequa ely ep esen ed. A ep esen a i e i em om each ca ego y is illus a ed in Figu e 4.6, allowing o
a be e unde s anding o he da ase composi ion and di e si y.
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
59
4.4.1.4 Conclusions o Fi s Expe imen
This s udy p esen s an analysis o clo hing ype iden i ica ion and s ain de ec ion o blind indi iduals using
deep lea ning models. Th ough comp ehensi e benchma king, i was demons a ed ha a deep lea ning
model, pa icula ly GoogleLeNe , can achie e up o 91% F1-Sco e in iden i ying clo hing ype ca ego ies,
ep esen ing a signi ican imp o emen o e p e ious li e a u e. The s udy highligh ed he e ec i eness o
ans e lea ning, whe e ine- uning all ne wo k laye s and le e aging p e- ained weigh s yielded supe io
esul s compa ed o adi ional me hods ha add new head laye s. Augmen ed da a echniques, such as
andom ho izon al lipping, we e shown o u he enhance model pe o mance, while andom o a ions
nega i ely impac ed esul s. This unde sco es he impo ance o ca e ully selec ing augmen a ion
s a egies o op imize model accu acy. I was also obse ed ha he dep h o he a chi ec u e (
e.g.,
ResNe -18 s. ResNe -50) did no signi ican ly a ec alida ion accu acy, hough i did in luence he
compu a ional e iciency. ResNe -18, o ins ance, p o ided he as es in e ence imes wi hou
comp omising accu acy. Despi e he p omising esul s, he e a e limi a ions o his s udy. One no able
cons ain is he eliance on images o clo hing wo n by models, which may in oduce edundancy and
lead o misclassi ica ions be ween simila clo hing a icles, such as casual shoes and spo shoes, o -
shi s and ops. To build on hese indings, u he es s a e necessa y o de e mine he mos e ec i e
ypes o da a augmen a ion o he models used. Addi ionally, add essing he limi a ions ela ed o image
sou ces could enhance he model's obus ness and gene alizabili y.
In summa y, he ine- uned GoogleLeNe model, enhanced wi h app op ia e da a augmen a ion
echniques, ou pe o ms exis ing me hods and o e s a iable solu ion o clo hing ype iden i ica ion o
blind indi iduals. This s udy’s esul s align wi h he accu acy le els epo ed Kolisnik, Hogan, and
Zulke nine (2021), despi e di e ences in class speci ica ions, demons a ing he model’s po en ial o
p ac ical applica ions.
4.4.2 Second Expe imen - Clo hing Ca ego y Segmen a ion and Classi ica ion and
Colou Ex ac ion
The second expe imen ex ended he i s by in oducing clo hing segmen a ion. While he i s expe imen
ocused on image classi ica ion, his one e alua es he p oblem as ins ance segmen a ion, e ining
p ede ined ca ego ies o accommoda e di e en backg ounds and enabling he analysis o pixel colou s
as a esul o segmen a ion. Hence, his p ocess can be used o classi y clo hing ca ego ies and

Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
60
cha ac e ize indi idual pixels based on hei colou s. The wo k low is depic ed in Figu e 4.8.
Figu e 4.8 – Wo k low o clo hing segmen a ion and colou ex ac ion.
4.4.2.1 Da ase collec ion
The de eloped da ase consis s o 2,000 images ha a e equally dis ibu ed ac oss eigh ca ego ies o
ga men s: d esses, jacke s, pan s, polos, shi s, shoes, sho s, and -shi s. The p esen da ase is an
enhancemen and expansion o he p e ious one, inco po a ing a e ined selec ion o he mos signi ican
ca ego ies.
In addi ion o e ining he exis ing ca ego ies, new images we e sou ced om pe sonal wa d obes and
included di e se backg ounds in o de o inc ease he da ase 's obus ness. As a esul , each ga men
ca ego y is ully ep esen ed unde a a ie y o condi ions, imp o ing i s applicabili y in eal-li e scena ios.
A sample o he images added o he da ase is depic ed in Figu e 4.9.
Figu e 4.9 – Samples o images added o he da ase om pe sonal wa d obes: a) jacke , b) jeans, c) d ess, and d) sho s.
A spli o 70/20/10 was used o di ide he da ase in o aining, alida ion, and es se s, ensu ing an
equal dis ibu ion o da a ac oss all ca ego ies. As he objec i e o he s udy was no only o classi y
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
61
ga men s, bu also o segmen hem o ex ac pixel-le el colou in o ma ion, he s udy was app oached
as a p oblem o segmen a ion. Fo his pu pose, he YOLO 8s model was employed, and he ne wo k was
ine- uned. While aining, a a ie y o da a augmen a ion echniques we e applied, such as changes in
hue, alue, and sa u a ion, ho izon al lips, and scaling. As a esul o hese augmen a ion s a egies, he
model was enhanced in e ms o obus ness and gene alizabili y.
Based on es ing and expe imen a ion h oughou he aining p ocess, Table 4.8 p esen s he
hype pa ame e s ha yielded he bes esul s.
Table 4.8: Hype pa ame e s (Image Size, Op imize , Lea ning Ra e, and Ba ch Size) o model expe imen s.
Pa ame e s
Value
Image Size
640x640 pixels
Op imize
SGD
Lea ning Ra e
0.01
Ba ch Size
16
The esul s o his s udy, including pe o mance me ics o each ca ego y, a e p esen ed in Table 4.9.
P ecision indica es a low pe cen age o alse posi i es ela i ely o he o al numbe o pixels co ec ly
iden i ied as pa o an objec mask. Recall e e s o he pe cen age o pixels co ec ly iden i ied as
belonging o an objec mask compa ed wi h he o al numbe o pixels ac ually belonging o ha mask - a
low pe cen age indica es ew alse nega i es. A mAP a IoU = 50 measu es he a e age p ecision o he
model's p edic ions when he in e sec ion o e union (IoU) is, a leas , 50%. Table 4.9 summa izes he
pe o mance esul s.
Table 4.9 - Main pe o mance esul s o model o clo hing mask segmen a ion (P ecision, Recall and AP a IoU = 0.50).
Ca ego y
P ecision
Recall
AP a IoU = 0.50
AP
all
0.941
0.896
0.965
0.949
D ess
0.849
0.917
0.941
0.906
Jacke
1
0.911
0.99
0.955
Pan s
0.989
0.875
0.97
0.96
Polo
0.953
0.837
0.96
0.952
Shi
0.863
0.875
0.925
0.916
Shoes
0.996
1
0.995
0.995
Sho s
0.833
0.833
0.964
0.955
T-shi
0.917
0.917
0.971
0.954
Table 4.9 p esen s he pe o mance esul s o he YOLO 8s model in he ask o segmen ing clo hing
ac oss a ious ca ego ies, as well as de ailed me ics o each ype o ga men , ocusing on segmen a ion
masks. Fo he o e all pe o mance o he model ac oss all ca ego ies ("all"), he model achie ed a
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
62
p ecision o 0.941, a ecall o 0.896, and an AP o an IoU o 0.5 o 0.965. These esul s indica e a high
o e all pe o mance o he model when i comes o segmen ing masks. In he d ess ca ego y, he model
had a p ecision o 0.849, a ecall o 0.917, and a mAP o 0.941. E en hough hese esul s indica e good
pe o mance, he p ecision is sligh ly lowe han in o he ca ego ies. The jacke ca ego y demons a ed
an excellen pe o mance wi h a p ecision o 1.0 and a ecall o 0.91, esul ing in a mAP o 0.99,
demons a ing an excellen abili y o he model in iden i ying and segmen ing jacke s. In he pan s
ca ego y, a p ecision o 0.989, a ecall o 0.875 and a mAP o 0.97 we e ob ained, indica ing almos a
pe ec pe o mance in he mask segmen a ion p ocess. The polo ca ego y had a p ecision o 0.953,
ecall o 0.837, and mAP o 0.96. Despi e he high p ecision, he lowe ecall indica es ha he ga men
may no be ully de ec ed. Shi s ("shi ") achie ed a p ecision o 0.863, a ecall o 0.875, and a mAP o
0.925, showing balanced pe o mance, bu wi h oom o imp o emen . I is no ewo hy ha shoes had
an almos pe ec p ecision o 0.996, a ecall o 1.0, and a mAP o 0.995, indica ing excellen pe o mance
in his ca ego y. In he case o sho s, he p ecision was 0.992, he ecall was 0.833, and he mAP was
0.964. The low ecall indica es ha some ins ances may no ha e been de ec ed despi e he high
p ecision. Las ly, he -shi s ca ego y demons a ed good o e all pe o mance wi h a easonable balance
be ween p ecision and ecall, being 0.887 and 0.917 espec i ely, also achie ing a mAP o 0.971.
Acco ding o Figu e 4.10, he con usion ma ix o he YOLO 8s model p o ides de ailed insigh in o i s
pe o mance in he segmen a ion o clo hing ac oss a ious ca ego ies.
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
63
Figu e 4.10 – Illus a ion o he ela ionship be ween p edic ed and ue alues in he con usion ma ix o each clo hing
ca ego y.
This ma ix – Figu e 4.10 - p o ides an e alua ion o whe e he model pe o ms well and whe e i ails.
The lowe alue o sho s may be a ibu ed o misclassi ica ion wi h d esses and shi s, likely due o he
simila poses o he models wea ing hese ga men s. Fu he mo e, -shi s and polos a e some imes
con used p obably due o hei simila shapes. Figu e 4.11 illus a es some esul s om segmen a ion.
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
64
Figu e 4.11: Example o in e ence images wi h ca ego y and espec i e segmen a ion mask: (a) O iginal image o pan s; (b)
Segmen ed image o pan s; (c) O iginal image o a shi ; (d) Segmen ed image o he shi .
A e segmen a ion, he ex ac ed mask was used o de e mine he colou o he clo hing i em. A pixel-
le el analysis o he esul ing mask was conduc ed since he model gene ally pe o ms well in
classi ica ion. This in ol es ex ac ing he segmen a ion mask o he clo hing and con e ing i o he HSV
colou space o de e mine he exac colou o he ga men . Figu e 4.12 illus a es he en i e wo k low.

Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
65
Figu e 4.12: Wo k low o colou ex ac ion.
Using he black-and-whi e mask o he segmen ed egion, a bi wise AND ope a ion was pe o med on he
o iginal image. This p ocess p oduces an image con aining only he egion o in e es , namely segmen ed
clo hing, in he RGB colou space. This image is con e ed in o a HSV colou space in o de o ex ac he
colou componen s. Due o he sepa a ion o he ch ominance componen (Hue) om he sa u a ion and
alue componen s, he HSV colou space o e s ad an ages when selec ing a desi ed hue as compa ed
o he RGB colou space. The hue o a pa icula colou is de ined as an angula alue be ween 0 and
360 deg ees, wi h 0 deg ees ep esen ing pu e ed, 120 deg ees ep esen ing g een, and 240 deg ees
ep esen ing blue. Since he con ex o he p oblem does no equi e dis inguishing be ween ligh and
da k shades o speci ic colou s, he p ima y colou s o be de ec ed we e de ined. The e o e, speci ic
colou in e als we e used o de ine he Hue componen analysis, including black, whi e, ed, g een,
yellow, blue, o ange, b own and pink. The colou ex ac ion p ocess, al hough no quan i ied by speci ic
me ics such as accu acy o p ecision, con ibu es signi ican ly o he o e all unc ionali y o he sys em
by accu a ely iden i ying and dis inguishing a ious clo hing colou s. By con e ing he segmen ed egions
in o he HSV colou space, i is possible o isola e he hue componen , which simpli ies he iden i ica ion
o he p ima y colou s ega dless o a ia ions in ligh ness o sa u a ion. Despi e he absence o
quan i a i e me ics, he quali a i e assessmen con i ms ha he de ined hue in e als e ec i ely co e
he in ended p ima y colou s. This in e nal alida ion suppo s he sys em's capabili y o p o ide
accessible and p ecise clo hing iden i ica ion.
4.4.2.2 Conclusions o Second Expe imen
Following ine- uning wi h da a augmen a ion echniques, he YOLO 8s model was e ec i e a segmen ing
a ious ca ego ies o clo hing. Almos all ca ego ies showed high pe o mance me ics o segmen a ion
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
66
masks, wi h jacke s and shoes achie ing nea -pe ec esul s. E en so, he e is s ill oom o imp o emen
in ca ego ies such as d esses and sho s, whe e ecall could be imp o ed. Following he segmen a ion
p ocess, he mask o he segmen ed egion was used o analyse he colou s p esen in he clo hing i em.
Due o he ac ha i is no necessa y o di e en ia e be ween de ailed colou nuances in his con ex ,
nine essen ial colou s we e de ined. The e o e, he HSV colou space was used, in which he Hue
componen was de ined o encompass he de ined colou s.
4.4.3 Thi d Expe imen - Iden i ica ion o Modi ica ions on Clo hing
The hi d expe imen aimed a iden i ying de ec s in clo hing,
i.e.
, holes and s ains, by using compu e
ision ad ances, namely deep lea ning algo i hms. This expe imen was di ided in o wo s ages, he i s
one aiming a he c ea ion o a small da ase o s ain de ec ion using a neu al ne wo k, and he second
one a expanding he da ase and in oducing o he ypes o de ec s, as well as using a di e en neu al
ne wo k o de ec de ec ion and classi ica ion.
4.4.3.1 Fi s S age
Du ing he ini ial phase o he i s s age, images we e cap u ed in o de o es ablish a da ase . No e ha ,
a he ime o he s udy, he e was no publicly a ailable da abase on clo hing wi h s ains. Then, a neu al
ne wo k was ained using ans e lea ning echniques o de ec he p esence o s ains, as shown in
Figu e 4.13.
Figu e 4.13: Wo k low o he me hodology used o de ec ing s ains (Rocha
e al
., 2023a).
As a esul , a da ase wi h app oxima ely 104 images was compiled. The ga men s in his da ase may
display mul iple s ains ha a e dis ibu ed ac oss a ious sec ions, esul ing in a o al o app oxima ely
300 s ains. These images we e aken om pe sonal wa d obes whe e co ee and wine s ains we e
in en ionally applied o clo hing, as i is shown in Figu e 4.14.
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
67
Figu e 4.14: S ains dis ibu ion on clo hing om he da ase , a) co ee s ain; b) wine s ain; c) mul iple s ains; d) backside s ain
(Rocha, Soa es,
e al
., 2023a).
Despi e ga men s ha ing wo ypes o s ains, he anno a ion p ocess was ca ied ou conside ing only one
class, namely, "s ain". Due o he anno a ions in he da ase and he limi ed amoun o da a, he ans e
lea ning me hod was employed. By applying ans e lea ning, mask_ cnn_R_50_FPN_3x wi h ResNe 50
was u ilized as he backbone. A Mask R-CNN o s ain de ec ion was implemen ed using he De ec on2
lib a y,(Wu
e al
., 2019) he hype pa ame e s a e p esen ed in Table 4.10.
Table 4.10: Hype pa ame e s o ine- uned Mask R-CNN (Rocha, Soa es,
e al
., 2023a).
Pa ame e s
Value
I e a ions
360
Leaning a e
0.001
Ba ch size
1
As an ex ension o Fas e R-CNN, Mask R-CNN inco po a es ins ance segmen a ion as an enhancemen .
By adding pixel-le el segmen a ion, his ne wo k has an ad an age o e Fas e R-CNN, which labels e e y
pixel ha belongs o a de ec ed objec . The main esul s o he ne wo k pe o mance and model losses
om Mask R-CNN a e p esen ed in Table 4.11 and Table 4.12, espec i ely.
Table 4.11: A epo on he e alua ion o Common Objec s in Con ex (COCO) (Rocha, Soa es,
e al
., 2023a).
Me hod
AP
AP a IoU = 0.50
AP a IoU = 0.75
Bounding Box
0.549
0.857
0.672
Segmen a ion
0.540
0.858
0.674
Table 4.12: A summa y o he losses associa ed wi h he model (Rocha, Soa es,
e al
., 2023a).
To al Loss
Loss Classi ica ion
Loss Box Reg ession
Loss Mask
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
68
0.480
0.053
0.154
0.240
Acco ding o he COCO e alua o , he esul s p esen ed in Table 4.11 a e p omising, pa icula ly o he
AP a IoU = 0.50 o 0.857, despi e he small quan i y o da a. I is no ewo hy ha he highes e i ied
loss was 0.240 (Table 4.12), indica ing ha he segmen a ion o he s ain p esen ed a mo e challenging
ask. The e alua ion o he model allowed o conclude ha he misclassi ica ions we e mainly ela ed o
he de ec ion o b and logos in he clo hing and he low con as be ween he s ain and he clo hing colou ,
as shown in Figu e 4.15.
Figu e 4.15: Example o an i em o clo hing ha has been misclassi ied (Rocha, Soa es,
e al
., 2023a).
This i s s age p esen s an inno a i e me hod o de ec ing s ains om a clo hing image. The p esen ed
me hod, using a da ase o clo hing wi h s ains caused by wine and co ee, demons a es he abili y o a
deep lea ning algo i hm o accu a ely iden i y and loca e s ains on clo hing. Despi e he ob ained esul s
look p omising, i is expec ed ha hey could be u he imp o ed when a la ge da ase is used.
4.4.3.2 Second S age
A e e alua ing he esul s o he ini ial s age, which included only a limi ed numbe o images depic ing
only one ype o de ec ,
i.e.
s ain, i became appa en ha a b oade app oach could imp o e he esul s.
As pa o he s udy, i was decided o explo e he inclusion o a new de ec ype, holes, along wi h a
subs an ial expansion o he da ase . This expansion was in ended o no only di e si y he ypes o
conside ed de ec s, bu also en ich he da ase o p o ide a mo e comp ehensi e desc ip ion o he
p oblem. Mo eo e , i employed a di e en neu al ne wo k a chi ec u e ha add essed he p oblem
h ough objec de ec ion a he han segmen a ion. In addi ion, echniques o da a augmen a ion we e
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
75
YOLO 5m6 model. Howe e , conside ing he con ex o his applica ion, p io i izing ecall o e p ecision
may be mo e bene icial. In o he wo ds, i is p e e able o he model o ha e ewe alse nega i es han
alse posi i es. Figu e 4.20 illus a es an example o a alse posi i e, whe e he bu onholes a e
misin e p e ed as a de ec . This highligh s he impo ance o ha ing ep esen a i e images ha include
such scena ios.
Figu e 4.20: Example o a misin e p e a ion o a de ec : a) o iginal image; b) p edic ed image om model YOLOV5m6; c)
p edic ed image om model YOLOV5l6 (Rocha, Pin o,
e al
., 2023).
Based on he pe o mance o he models ac oss all expe imen s, he model YOLO 5l6 exhibi s he bes
gene aliza ion o unseen da a,
i.e.
, es da ase , when compa ed o o he models. Figu e 4.21 displays
p edic ed images ha encompass a ious scena ios, including a ia ions in illumina ion, backg ounds,
mul iple de ec s, as well as challenging a eas.

Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
76
Figu e 4.21: Examples o p edic ed images om he YOLO 5l6: a) single s ain de ec ion; b) mul iple s ain de ec ion; c) mul iple
hole de ec ion; d) hole de ec ion nea he seam (Rocha, Pin o,
e al
., 2023).
The main cons ain o using his model in a p ac ical con ex is he compu a ional cos . Such impac was
e alua ed h ough he calcula ion o he in e ence ime o he es wi h he da ase . Table 4.18 exhibi s
he esul s o he in e ence ime on he es da ase .
Table 4.18: In e ence ime on es da ase o he di e en YOLO 5 models es ed (Rocha, Pin o,
e al
., 2023).
Model
In e ence Time (s)
YOLOV5s6
0.0092
YOLOV5m6
0.0112
YOLOV5l6
0.0157
These esul s sugges ha , despi e incu ing compu a ional cos s, all models a e belie ed accep able due
o he negligible equi ed ime. Thus, he indings indica e ha implemen ing objec de ec ion echnology
wi h augmen ed da a may be a success ul s a egy o iden i ying de ec s in clo hing. This s udy s ands
ou om p e ious esea ch wo k (L. Cheng
e al
., 2023), as i e alua es de ec de ec ion on clo hing as a
whole, ins ead o ocusing on zoomed-ou images o de ec s on s e ched ex iles and wi hou backg ound.
Chap e 4 –
Sys em De elopmen
___________________________________________________________________________
77
Upon compa ing he new da ase c ea ed in his wo k wi h hose om ex ile indus y i becomes clea
how challenging he ask p esen ed in his wo k is, mainly due o he p esence o ce ain ea u es such
as bu onholes, which could po en ially be in e p e ed as de ec s. Fu he mo e, his app oach was p o ed
o be e ec i e in high demanding con ex s, namely wi h w inkled ex iles, a ious backg ounds, di e en
illumina ion, and di e se pa ne s.
4.4.3.8 Conclusions o Thi d Expe imen and u u e wo k
The de ec ion and classi ica ion o clo hing de ec s was success ully ca ied ou wi h a deep lea ning
app oach. An enhanced da ase was cons uc ed wi h new ypes o s ains and wi h holes. Th ough he
ine- uning o h ee models om he YOLO 5 objec de ec o , a o al o h ee expe imen s was ca ied ou .
Da a augmen a ion was demons a ed o be essen ial o a be e gene aliza ion o he model, allowing o
achie e highe p ecision esul s. S ill, ecall alues demons a e ha he model can be imp o ed o
minimize alse nega i es. Maximum p ecision, ecall, and a e age p ecision (AP) alues o 0.915, 0.543,
and 0.747, espec i ely, we e achie ed wi h he YOLO 5l6 model o de ec de ec ion and classi ica ion.
The de ec ion o holes was ound o be mo e challenging han he de ec ion o s ains, which emphasizes
he impo ance o in eg a ing he indings o his s udy in an au oma ic wa d obe ha could ake mul iple
came as cap u e om di e en clo hing pe spec i es.
The da ase buil in his wo k demons a es ha objec de ec ion echnology can be used o accu a ely
de ec and classi y de ec s on clo hing au onomously. Mo eo e , i ep esen s he i s s ep o he c ea ion
o a mobile applica ion ha can e ec i ely de ec mul iple de ec s on clo hing, based on he in eg a ion o
hese indings in an au oma ed close sys em as a u u e s ep.
5 SMART WARDROBE PROTOTYPE
Chap e O e iew
This chap e p esen s a de ailed desc ip ion o he sma wa d obe p o o ype and i s key componen s,
ollowed by an o e iew o he mobile applica ion de eloped o con ol he sys em. The aim o his chap e
is o ou line he design and essen ial elemen s necessa y o he de elopmen o a mecha onic
p o o ype — a sma wa d obe sys em capable o iden i ying clo hing cha ac e is ics such as ype, colou ,
and al e a ions wi hin a con olled en i onmen .
5 Sma Wa d obe P o o ype
5.1 P o o ype Sys em Design
5.2 Main Componen s o he De eloped P o o ype
5.3 Mobile Applica ion In e ace
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
79
5.1 P o o ype Sys em Design
The p ima y objec i e is o simula e a p o o ype closely esembling eal-wo ld scena ios, om which he
en i e ha dwa e in as uc u e can be seamlessly ans e ed o indi idual wa d obes. The e o e, i is
essen ial o ca e ully selec adap able componen s o ensu e he p o o ype's e ec i eness. Hence, he
p o o ype mus mee a numbe o speci ica ions ha a e essen ial o he in eg a ion and unc ionali y o
he p oduc , namely:
• Iden i y he ype, colou , and modi ica ions o he ga men ;
• Implemen a NFC eade o ex ac each ga men 's unique iden i ie ;
• Main ain a con olled ligh ing en i onmen o ensu e accu a e assessmen ;
• Au oma e he o a ion o ga men s o e alua e bo h sides;
• Include a came a o acqui ing images;
• Es ablish TCP/IP communica ion o con ol all ha dwa e componen s emo ely.
To p o e he concep (Rocha
e al
., 2020) and de ise a eadily deployable solu ion, a compac IKEA
wa d obe measu ing 80 cm by 50 cm by 30 cm (heigh x wid h x dep h) was ini ially selec ed o he
e alua ion o app op ia ely sized ga men s (Sil a
e al
., 2023). The sys em, howe e , aced some
limi a ions. The cons ained dimensions o he wa d obe es ic ed i s use o ga men s o diminu i e
p opo ions, p ima ily ailo ed o child en. Fu he , he imaging p ocess was hampe ed by pa ial cap u e,
which necessi a es he o a ion o he came a and he subsequen applica ion o a s i ching algo i hm o
inco po a e di e en segmen s o he image. Also, as he e ec i eness o he NFC eade depends on he
o ien a ion o he ga men , a a ia ion in he accu acy o he ead occu ed. Las ly, he wa d obe's
i emo able suppo in as uc u e complica ed he p ocess o ga men eplacemen , making a quick
ga men change a di icul ask. As a esul , hese challenges led o he design o a new p o o ype
a chi ec u e, as shown in Figu e 5.1.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
80
Figu e 5.1: iSigh Sma Wa d obe P o o ype.
The dimensions o he sma wa d obe p o o ype depic ed in Figu e 5.1, 123 cm by 120 cm by 63 cm
(heigh x wid h x dep h), enabled i s e ec i eness o demons a e he p oposed solu ion and i s
anspo a ion o es ing pu poses. In he nex subsec ion a de ailed desc ip ion o i s componen s is
p o ided.
5.2 Main Componen s o he De eloped P o o ype
Be o e de ailing he indi idual componen s o he sma wa d obe p o o ype, an o e iew is p o ided in
Figu e 5.2, which illus a es he schema ic layou o he p o o ype and highligh s he in e connec ions
be ween key ha dwa e componen s, including he Raspbe y PI, DC mo o , came a, ligh ing sys em, and
NFC eade .

Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
81
Figu e 5.2 - Schema ic layou o he sma wa d obe p o o ype.
This schema ic, Figu e 5.2, p o ides a ep esen a ion o he Sma Wa d obe P o o ype, highligh ing he
key componen s and hei oles wi hin he sys em. These componen s a e ou lined and desc ibed in de ail
below:
• Raspbe y PI: Ac s as he cen al uni , managing he ope a ion and coo dina ion o all o he
ha dwa e elemen s.
• Came a: Cap u es images o he ga men s o acili a e isual analysis and iden i ica ion.
• Ligh s: P o ide con olled illumina ion, ensu ing clea and consis en image cap u e.
• DC Mo o : Ro a es he ga men s, allowing he came a o assess hem om mul iple angles.
• NFC Reade : Uniquely iden i ies each ga men using Nea Field Communica ion (NFC)
echnology.
The sys em is ul ima ely accessed and con olled h ough a TCP/IP communica ion in e ace, enabling
emo e ope a ion and in eg a ion wi h ex e nal de ices.
In he ollowing subsec ions, a de ailed analysis o each componen is p o ided.
5.2.1 DC Mo o
Nema 17 s eppe mo o , model 42BYGH48-23D (
Mo o de Passo Nema 17 p/ Imp esso a 3D -
42BYGH48-23D
, 2024) - Figu e 5.3 - is a ype o elec ic mo o widely used in applica ions equi ing
p ecise posi ioning, such as o a ing a sha a a speci ic angle. Due o i s excep ional cos -e ec i eness,
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
82
his mo o has a high o que and a limi ed s ep angle o 1.8 deg ees, which gua an ees accu a e and
eliable pe o mance.
Figu e 5.3 Nema 17 s eppe mo o .
Table 5.1 p o ides he speci ica ions o he DC mo o used in he p o o ype.
Table 5.1: Nema 17 mo o speci ica ions (
Mo o de Passo Nema 17 p/ Imp esso a 3D - 42BYGH48-23D
, 2024).
Speci ica ions
Desc ip ion
To que
55 N.cm
S ep
1.8 deg ees
Numbe o S eps
200
Phase Cu en
1.7A
Numbe o Phases
2
Numbe o Wi es
4
Sha Diame e
5mm
Mo o Leng h
48mm
5.2.2 D i e Mo o
The d i e mo o A4988(
D i e Pa a Mo o - A4988
, 2024) - Figu e 5.4- is a widely used s eppe mo o
d i e enowned o i s e iciency in con olling bipola s eppe mo o s. I p o ides an e ec i e means o
con olling he cu en and pola i y o he mo o coils, enabling p ecise mo ion con ol.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
83
Figu e 5.4: Pic u e o he d i e mo o A4988 used in he de eloped p o o ype.
Equipped wi h cu en limi a ion and adjus able se ings, along wi h o e hea ing p o ec ion and i e
di e en mic os ep esolu ions (up o 1/16-s ep), his mo o is designed o mee he demands o a ious
mo ion con ol applica ions. Table 5.2 desc ibes he main speci ica ions o he A4988 d i e mo o .
Table 5.2. Desc ip ion o he A4988 speci ica ions (
D i e Pa a Mo o - A4988
, 2024).
Speci ica ions
Desc ip ion
Con ol Me hod
S ep and di ec ion con ol
Logic Vol age
3-5.5V
Mo o Ou pu
Vol age
8-35V
Mic os ep
Resolu ions
Full-s ep, Hal -s ep, 1/4-s ep, 1/8-s ep, 1/16-s ep
Cu en Con ol
Adjus able, allowing se ing o maximum ou pu cu en using a po en iome e , enabling u iliza ion o ol ages
abo e nominal mo o ol age o highe s ep a es
Vol age Regula o
Buil -in
P o ec ion Fea u es
Cu en o e load and sho -ci cui p o ec ion
The speci ica ions p esen ed abo e make he A4988 s eppe mo o d i e used in a wide ange o mo ion
con ol sys ems, o e ing e sa ili y, eliabili y, and p ecise con ol o e s eppe mo o ope a ions.
5.2.3 Raspbe y PI
The Raspbe y Pi 4 Model B (
Raspbe y Pi 4 Model B
, 2024) - Figu e 5.5 - is a single-boa d minicompu e .
This cen al p ocessing uni is equipped wi h a 1.5GHz quad-co e A m® Co ex®-A72 CPU and 8GB o
RAM. Because o i s high-pe o mance CPU and ample memo y, he Raspbe y Pi 4 Model B is ideally
sui ed o p ocessing-in ensi e applica ions. The ad anced g aphics capabili y and high- esolu ion ideo
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
84
ou pu capabili ies allow o obus mul imedia p ocessing. Mo eo e , bo h USB 3.0 and USB 2.0 po s,
ue Gigabi E he ne , and dual HDMI ou pu s p o ide e sa ile connec i i y and as da a ans e s.
Figu e 5.5: Pic u e o a Raspbe y Pi 4 Model B used in he de eloped p o o ype.
5.2.4 Came a
The O icial Raspbe y Pi Came a Module V3 (
Raspbe y Pi Came a Module 3
, 2024) - Figu e 5.6 -
equipped wi h a 12MP senso , is designed o o e high- esolu ion imaging capabili ies sui able o a
a ie y o applica ions. This module u ilizes he Sony IMX708 senso (
Raspbe y Pi Came a Module 3
,
2024), which boas s ad anced ea u es and speci ica ions ailo ed o enhanced pe o mance.
Figu e 5.6: V3 O icial Raspbe y Pi, 12MP, 120°, came a module, used in he de eloped p o o ype.
The came a module's high esolu ion o 12MP ensu es de ailed image cap u e, while he senso 's
compac size (7.4mm diagonal) and pixel dimensions (1.4μm × 1.4μm) con ibu e o high sensi i i y and
image quali y. Table 5.3 de ails he came a speci ica ions.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
91
Figu e 5.13: Flowcha o wa d obe se e .
Figu e 5.14 illus a es he ini ial s a e o he wa d obe when i is powe ed on. The wa d obe’s in e io is
en i ely co e ed wi h ch oma key ab ic o p o ide a uni o m colou o image acquisi ion, enhancing
image consis ency. The LED s ips loca ed on he le panel a e powe ed by a 12V powe supply, as shown
in Figu e 5.14a. Posi ioned cen ally is he o icial Raspbe y Pi Came a Module V3, as depic ed in Figu e
5.14b.

Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
92
Figu e 5.14: Gene al iew o he wa d obe in e io : a) LED s ips applied on he le panel; b) came a placed in he middle o
LED s ips.
The ligh sou ce is essen ial o ensu ing he quali y o images cap u ed by he came a, p o iding su icien
illumina ion o he clea iden i ica ion o clo hing i ems ega dless o en i onmen al condi ions. In his
echnique, bo h he senso (came a) and he ligh sou ce a e posi ioned on he same side o he objec
being cap u ed. This me hod is ypically employed o ob ain de ailed in o ma ion abou he ex u e and
o he su ace cha ac e is ics o he objec . This ype o illumina ion can be ca ego ized in o wo sub ypes:
one ha p ima ily le e ages specula e lec ions and ano he ha p edominan ly u ilizes e lec ed ays
cons i u ing di use e lec ion. In scena ios whe e ligh e lec ed by a la su ace is cap u ed by he op ical
sys em, non- la cha ac e is ics such as holes can e lec he ligh ou side he lens's maximum accep ance
angle, he eby e ealing da k ea u es, as schema ized in Figu e 5.15.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
93
Figu e 5.15: Re lec i e illumina ion in hole de ec . This image shows how ligh e lec ed om a la su ace is a ec ed by non-
la ea u es like holes, which can edi ec ligh ou side he lens's accep ance angle, c ea ing da k a eas ha e eal su ace
de ec s.
A he same ime, he se e ,
i.e.
he Raspbe y Pi moun ed on he ex e nal le panel (
Raspbe y Pi 4
Model B
, 2024), is ini ialized and begins awai ing HTTP eques s om he g aphical use in e ace (GUI),
i.e.
mobile in e ace.
I he eques is o cap u e an image, he came a is ini ialized o execu e he cap u e. As explained be o e,
his speci ic came a was selec ed based on i s 120° lens wide diagonal ield o iew, which allows o a
minimal dis ance o e ec i ely cap u e he en i e y o he ga men . This wide-angle lens ensu es
comp ehensi e co e age acili a ing de ailed imaging o he clo hing i em, ensu ing no pa o he ga men
is omi ed in he cap u e p ocess.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
94
Figu e 5.16: In e io o he wa d obe showing he dis ance om he came a o he hange and in e nal componen s.
Upon ecei ing he o a ion eques , he mo o si ua ed on he op igh side (Figu e 5.16 and Figu e
5.17), which suppo s he hange (Figu e 5.16 and Figu e 5.17), is ac i a ed o o a e 180 deg ees. This
ea u e enables he sys em o cap u e ga men s om bo h sides, enhancing ope a ional e iciency and
lexibili y. The abili y o o a e he hange mo o by 180 deg ees is c ucial o p ocesses ha equi e access
o bo h sides o he ga men , such as inspec ion o modi ica ions.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
95
Figu e 5.17: Mo o suppo ing he hange .
Ano he po en ial eques ha can be handled by he sys em in ol es eading he NFC ag a ached o
each ga men . The placemen o hese ags on ga men s has been designed o be minimally in asi e.
Thus, he ag can be ound ei he on he o iginal clo hing label, as illus a ed in Figu e 5.18.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
96
Figu e 5.18: NFC ag a ached o he clo hing label.
To acili a e his, an NFC eade has been in eg a ed in o he wa d obe sys em (Figu e 5.19). This eade
is asked wi h e ie ing he unique iden i ie (UID) code om he NFC ags, as pe he use 's eques .
Speci ically, an NFC eade is s a egically posi ioned on he ex e io le side o he wa d obe o ensu e
seamless eading o he NFC ags a ixed o he ga men s, as demons a ed in Figu e 5.19.
Figu e 5.19: NFC Reade .
The in eg a ion and connec i i y o all sys em componen s wi h he Raspbe y Pi con olle a e
comp ehensi ely illus a ed in he schema ic diag am p o ided in Figu e 5.20.

Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
97
Figu e 5.20: Schema ic diag am illus a ing he in eg a ion and connec i i y o sys em componen s wi h he Raspbe y PI
con olle
This diag am (Figu e 4.41) o e s a de ailed o e iew o he se up, highligh ing he in e acing o he NFC
eade , came a module, and mo o o a ion wi h he Raspbe y PI, he eby elucida ing he sys em's
ope a ional amewo k. The da a cap u ed om image p ocessing and NFC ag eading is ansmi ed
om he Raspbe y PI o he GUI. Addi ionally, he mobile in e ace p ocesses he ou pu om he
wa d obe se e and, when necessa y, communica es wi h AI models o in e ence.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
98
5.3 Mobile Applica ion In e ace
The pu pose o his sec ion is o explo e he de elopmen o a mobile in e ace, emphasizing compliance
wi h accessibili y s anda ds.
5.3.1 Accessibili y
In he Eu opean Union (EU), he s anda d ha encompasses accessibili y equi emen s o he web and
mobile applica ions is he EN301549 s anda d (
In odução à No ma Eu opeia EN 301 549 -
Acessibilidade.Go .P
, 2024). This s anda d is essen ially a copy o he WCAG 2.1 'AA' compliance. In
Po ugal, hese equi emen s a e included in he
Regulamen o Nacional de In e ope abilidade Digi al
(RNID) as dec eed by D.L. No. 83/2018(<i>DL n.o 83/2018 - Acessibilidade Dos Sí ios Web e Das
Aplicações Mó eis - Acessibilidade.Go .P </i>, 2024). Acco ding o his dec ee, "Accessibili y" e e s o
he p inciples and echniques o be obse ed in he design, cons uc ion, main enance, and upda ing o
websi es and mobile applica ions o make hei con en mo e accessible o use s, especially people wi h
disabili ies. In compliance wi h A icle 5, his accessibili y en ails se e al equi emen s ega ding
in o ma ion and na iga ion which his s udy mus also adhe e o in o de o be accessible o he blind.
The e o e, in he con ex o in o ma ion, he ollowing aspec s a e conside ed:
1. P esen a ion o in o ma ion and in e ace componen s in a manne pe cei able by use s.
2. Ensu ing he unc ionali y o he use in e ace, making su e ha he componen s and na iga ion
a e ac ionable.
3. In o ma ion and he ope a ion o he use in e ace mus be easily unde s ood.
4. In e p e a ion by assis i e echnologies o be eliably in e p e ed by a wide ange o use agen s.
Adhe ing o hese p inciples ensu es ha he iSigh mobile applica ion aligns wi h bo h EU and
Po uguese accessibili y s anda ds, he eby acili a ing an inclusi e use expe ience o isually impai ed
indi iduals.
5.3.2 Mobile In e ace Fea u es
Gi en he in ended use base o Po uguese-speaking people wi h isual impai men s, he applica ion has
been me iculously designed in Po uguese and de eloped using a c oss-pla o m language. This s a egic
choice acili a es comp ehensi e es ing and deploymen ac oss bo h iOS and And oid de ices, ensu ing
b oad accessibili y and unc ionali y. To add ess he ini ial p ojec equi emen s, a de ailed block diag am
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
99
ep esen ing he en i e layou o he applica ion was de eloped and i is depic ed in Figu e 5.21.
Figu e 5.21: iSigh mobile in e ace block diag am.
The applica ion is exclusi ely accessible o people wi h disabili ies, le e aging sc een eade s o B aille
displays as g aphical in e aces be ween he blind use and he compu e . These in e aces in e p e he
in o ma ion displayed on he sc een, enabling seamless in e ac ion and na iga ion. Al hough de eloped
using a c oss-pla o m language, he applica ion was p ima ily designed and es ed o iOS de ices o
ensu e obus pe o mance and accessibili y. Upon examining Figu e 5.21, i becomes appa en ha
ollowing he login p ocess, six p ima y elemen s eme ge: Tag, Clo hes, Ca ego ies, Colou s,
Modi ica ions, and Wa d obe.
To na iga e he applica ion, use s mus enable VoiceO e , a ges u e-based sc een eade ha allows
iPhone use wi hou isual inpu . This ea u e p o ides audi o y desc ip ions o on-sc een con en , anging
om ba e y le els o calle iden i ica ion and he cu en applica ion in use. When he use ouches o
swipes he sc een, VoiceO e announces he name o he elemen unde he inge , including icons and
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
100
ex , and enables in e ac ion wi h bu ons and na iga ion o o he sc eens using speci ic VoiceO e
ges u es. The iSigh main menu is p esen ed below in Figu e 5.22.
Figu e 5.22: iSigh main menu.
The i s menu (Figu e 4.43), NFC Tag, is used o ead ags associa ed wi h clo hing i ems. Upon selec ion,
he use is in o med o he need o ensu e he ag is co ec ly placed on he clo hing i em, as illus a ed
in Figu e 5.23. Once con i med, he use can p ess he bu on o ead he ag.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
107
Figu e 5.29. Example o a e u ned esul om colou analysis.
Despi e in ol ing wo HTTP eques s - one o image cap u e and ano he o in e ence - he pipeline is
swi , and he use is no i ied o he success ul ope a ion h ough a hap ic ale on hei de ice. Fo
ca ego y eques s, he pipeline emains he same, di e ing only in he in e ence esponse. A e ecei ing
he esul , he use can always add he clo hing i em, being edi ec ed o he o m na iga ion panel as
p e iously desc ibed.
The Modi ica ions menu is simila o he Ca ego y and Colou menus, bu he la e allows he analysis o
clo hing i ems om bo h sides using a bu on o o a e he i em - Figu e 5.30.

Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
108
Figu e 5.30. Submenu included in he Modi ica ions menu.
The pipeline o his p ocess di e s in he model u ilized, employing a speci ic model designed o
modi ica ions in e ence. This model is ailo ed o de ec and analyse de ec s such as s ains o holes in
he clo hing i em. Upon comple ion o he in e ence p ocess, he esul s a e p esen ed o he use ,
p o iding de ailed in o ma ion abou he numbe o de ec ed s ains o holes on he ga men , as illus a ed
in Figu e 5.31. This de ailed eedback enables he use o make in o med decisions ega ding he
condi ion and main enance o hei clo hing i ems, ensu ing hey can add ess any issues p omp ly. This
unc ionali y is pa icula ly use ul o main aining he quali y and appea ance o hei wa d obe, u he
enhancing he o e all use expe ience.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
109
Figu e 5.31. Example o a e u ned esul om he Modi ica ions Analysis.
The inal menu pe ains o he wa d obe se ings - Figu e 5.32. In his menu, use s can con igu e IP
add esses and po s o communica ion wi h he wa d obe se e and he AI model se e h ough he
Se ings submenu. The NFC Tag submenu allows o eading NFC ags, simila o he i s menu, bu
u ilizing he NFC eade p esen in he wa d obe.
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
110
Figu e 5.32. Wa d obe menu de ails.
Al hough his ea u e may seem edundan , i is use ul when he mobile de ice's NFC eade is
una ailable.
In summa y, designing an applica ion ha speci ically add esses he needs o blind people and
inco po a ing ad anced echnologies such as NFC, AI models, his s udy ep esen s a signi ican
ad ancemen in accessibili y and usabili y o blind use s. The in eg a ion o hese echnologies ensu es
ha use s can independen ly manage hei wa d obe, make in o med choices, and e icien ly o ganize
hei clo hing i ems. This comp ehensi e app oach no only enhances he use expe ience bu also
empowe s blind people by p o iding hem wi h he necessa y ools o g ea e au onomy and sel - eliance
in hei daily li es.
Chap e Summa y
This chap e o e s a comp ehensi e o e iew o he de elopmen and componen s o he sma wa d obe
p o o ype, de ailing i s design, co e elemen s, and he associa ed mobile applica ion in e ace aimed a
Chap e 5 –
Sma Wa d obe P o o ype
___________________________________________________________________________
111
imp o ing wa d obe managemen o isually impai ed indi iduals.
The p o o ype sys em was enginee ed o iden i y clo hing cha ac e is ics such as ype, colou , and
modi ica ions h ough a mecha onic se up, which is emo ely con olled ia a mobile applica ion. The
ini ial p o o ype encoun e ed limi a ions ela ed o spa ial cons ain s and image cap u e di icul ies.
Consequen ly, he design was e ised o a la ge model. The e ised p o o ype ea u es a ch oma key
ab ic lining o acili a e uni o m image cap u e and inco po a es a DC mo o (Nema 17) o o a ing
ga men s o enable iewing o bo h sides. Key componen s o he de eloped p o o ype include he Nema
17 s eppe mo o , an A4988 mo o d i e , a Raspbe y Pi 4 Model B, a high- esolu ion 12MP came a, an
NFC eade u ilizing he PN532 chip, a con olled LED s ip o ligh ing, and a 12V powe supply. NFC
NTAG213 ags we e employed o label clo hing i ems, allowing o he eading o unique iden i ie s. The
sys em ope a es h ough a Raspbe y Pi, which se es as a local se e handling eques s om a mobile
in e ace. These eques s encompass cap u ing images, eading NFC ags, and con olling he mo o o
ga men o a ion. A wide-angle came a aids in cap u ing comp ehensi e images, while he mo o o a es
he hange 180 deg ees o p o ide a ull iew o bo h sides o he clo hing. The NFC eade is s a egically
posi ioned o ensu e e icien ag eading.
An accessible and use - iendly mobile applica ion was de eloped o enable blind use s o independen ly
o ganize and iden i y clo hing i ems. Ensu ing accessibili y was a p ima y objec i e; he e o e, he
applica ion adhe es o s anda ds such as EN301549 in he EU and WCAG 2.1 AA equi emen s. In
Po ugal, hese s anda ds a e ein o ced by he
Regulamen o Nacional de In e ope abilidade Digi al
(RNID)
unde DL No. 83/2018, ensu ing ha he applica ion is pe cei able, ope able, unde s andable, and
compa ible wi h assis i e echnologies such as sc een eade s and B aille displays. The mobile in e ace
was designed wi h a ocus on a Po uguese-speaking use base wi h isual impai men s, employing a
c oss-pla o m de elopmen app oach o compa ibili y wi h bo h iOS and And oid de ices. Fea u es
include VoiceO e o iOS, acili a ing in e ac ion h ough audi o y cues and ges u es. The applica ion
suppo s NFC ag eading, close managemen , and clo hing ca ego iza ion ia AI analysis based on colou
and ype. Use s in e ac wi h he sys em h ough he mobile applica ion, which communica es wi h he
Raspbe y Pi ia HTTP eques s. This includes image cap u e, NFC ag eading, and mo o con ol. The
applica ion p ocesses he da a and, when necessa y, in e aces wi h AI models o p o ide de ailed
eedback on clo hing i ems, including iden i ica ion and ecommenda ions based on use p e e ences.
In summa y, he iSigh p o o ype is designed o o e an accessible and p ac ical solu ion o wa d obe
managemen , wi h he po en ial o signi ican ly enhance independence and imp o e he daily li es o
isually impai ed use s.
6 EVALUATION OF PROTOTYPE
Chap e O e iew
The ollowing chap e desc ibes he esul s o a su ey conduc ed o alida e he iSigh p o o ype.
Membe s o ACAPO (
Associação dos Cegos e Amblíopes de Po ugal
) pa icipa ed in a s uc u ed es ing
p o ocol ollowed by a de ailed ques ionnai e. The goal was o ga he eedback on he usabili y,
accessibili y, and o e all impo ance o he iSigh p o o ype o he isually impai ed communi y.
6 E alua ion o P o o ype
6.1 E hical Conside a ions
6.2 Tes ing P o ocol
6.3 Ques ionnai e Me hodology
6.4 Resul s and Analysis
6.5 S a is ical analysis and ele an indings

Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
113
Chap e 6 p esen s he e alua ion and alida ion o he iSigh p o o ype, designed o assis blind
indi iduals in managing hei clo hing. The chap e s a s wi h e hical conside a ions, ensu ing
pa icipan s' sa e y and well-being h ough e hical app o als and in o med consen . The es ing p o ocol
in ol ed a s uc u ed p ocess whe e pa icipan s we e in oduced o he p o o ype, engaged in hands-on
es ing, and guided h ough scena ios o e alua e speci ic unc ionali ies like colou and ca ego y
e alua ion, de ec ion o modi ica ions, adding new clo hing i ems, and da abase e i ica ion. A e he
es ing, pa icipan s p o ided immedia e eedback and comple ed a comp ehensi e ques ionnai e. The
ques ionnai e me hodology in ol ed de eloping a de ailed se o ques ions, ini ially es ed wi h ACAPO
boa d membe s, o ga he in-dep h eedback on he p o o ype. The esul s and analysis sec ion p esen s
he indings om he ques ionnai e, co e ing demog aphics, ypes o isual impai men , echnology use,
accessibili y o he mobile applica ion, usabili y o he p o o ype, and he pe cei ed impo ance o he
iSigh sys em. S a is ical analysis was conduc ed o explo e ela ionships such as he co ela ion be ween
ease o na iga ion and use sa is ac ion, he impac o echnology amilia i y on ease o na iga ion and
o e all expe ience, and he e ec i eness o he p o o ype in iden i ying clo hing ea u es. The analysis also
examined use sugges ions o imp o emen s and compa ed eedback be ween blind use s and hose
wi h low ision.
The chap e concludes wi h a summa y o he s a is ical es s, hypo heses, esul s, and signi icance le els,
highligh ing key indings and a eas o u he de elopmen . This comp ehensi e e alua ion demons a es
he iSigh p o o ype's po en ial o enhance he independence and quali y o li e o isually impai ed use s.
Figu e 6.1 summa izes he main s eps and i e a ions in Chap e 5:
Figu e 6.1: Wo k low o he iSigh p o o ype es ing and alida ion p ocess.
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
114
This wo k low illus a es he s uc u ed p ocess ollowed in e alua ing and alida ing he iSigh
p o o ype, om e hical conside a ions and es ing p o ocols o de ailed s a is ical analyses and use
eedback, p o iding a clea o e iew o he chap e 's con en and me hodology.
6.1 E hical Conside a ions
To ensu e he sa e y and well-being o he pa icipan s, e hical conside a ions we e pa amoun . A consen
o m is included in Appendices 2 and 3, which documen s ha he es ing p o ocol and he ques ionnai e
we e submi ed o he E hics Commi ee o Resea ch in Social and Human Sciences (Comissão de É ica
pa a a In es igação em Ciências Sociais e Humanas - CEICSH) o he Uni e si y o Minho and ecei ed
app o al unde code CEICSH 185/2023. The co esponding e hical app o al documen is p o ided in
Appendix 4. P io o pa icipa ing in he s udy, each pa icipan signed an in o med consen o m
acknowledging hei unde s anding o he s udy's pu pose, p ocedu es, po en ial isks, and hei igh o
wi hd aw a any ime wi hou any consequence. As a esul o his consen , pa icipan s we e made awa e
o hei in ol emen and he na u e o he s udy, adhe ing o e hical s anda ds and p o ec ing hei p i acy
and igh s.
6.2 Tes ing P o ocol
To ensu e a comp ehensi e e alua ion o he iSigh p o o ype, a s uc u ed es ing p o ocol was ollowed.
Pa icipan s we e gi en de ailed ins uc ions on how o use he p o o ype, which included he ollowing
s eps:
1. In oduc ion o he P o o ype: each pa icipan was p o ided wi h an o e iew o he iSigh
p o o ype, including i s pu pose and i s ea u es.
2. Hands-on Tes ing: pa icipan s we e gi en a se pe iod o in e ac wi h he p o o ype
independen ly, allowing hem o explo e i s unc ionali ies and in e ace.
3. Guided Scena io: To e alua e he iSigh p o o ype, pa icipan s we e guided h ough a se ies o
asks designed o es i s unc ionali y in he ollowing eal-wo ld scena ios:
a. Ca ego iza ion and Colou E alua ion - Pa icipan s we e asked o eques an assis an o
p o ide a piece o clo hing o colou e alua ion. They we e ins uc ed o place he
clo hing i em in he wa d obe, ensu ing i was pa allel o he sides. Pa icipan s hen
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
115
na iga ed o he "Colou s" menu and cap u ed an image o he clo hing i em, e i ying
he colou e alua ion esul s wi h he assis an . Op ionally, pa icipan s could epea he
p ocess in he "Ca ego y" menu o de e mine he ca ego y o he clo hing i em and e i y
hese esul s wi h he assis an as well.
b. E alua ion o Modi ica ions on a Clo hing I em: pa icipan s eques ed an assis an o
p o ide a clo hing i em wi h modi ica ions, such as s ains o holes. They placed he i em
in he wa d obe, ensu ing i was pa allel o he sides. Pa icipan s accessed he
"Modi ica ions" menu and cap u ed an image o he clo hing i em, subsequen ly
con i ming he modi ica ion esul s wi h he assis an .
c. Adding a Clo hing I em: pa icipan s eques ed an assis an o p o ide a piece o clo hing.
They placed he clo hing i em in he wa d obe, ensu ing i was pa allel o he sides.
Pa icipan s opened he "Clo hing" sec ion and selec ed "Add New Clo hing I em." A e
cap u ing he image, hey e i ied he au oma ic illing o cha ac e is ics such as ca ego y
and colou , comple ing any addi ional in o ma ion as desi ed. Pa icipan s also had he
op ion o add he NFC ag au oma ically using he p o ided bu on.
d. Ve i ica ion in he Da abase: pa icipan s eques ed an assis an o p o ide a p e iously
added clo hing i em. F om he main menu o he applica ion, hey accessed he "NFC
Tag" unc ion, b ough he ag o he clo hing i em close o he NFC eade , and con i med
ha he i em was egis e ed in he da abase. Al e na i ely, hey could use he wa d obe's
NFC eade , loca ed in he uppe le co ne , accessed ia he "Wa d obe" menu ollowed
by "NFC Tag”.
4. Feedback Session: a e comple ing he asks, pa icipan s discussed hei expe iences wi h a
acili a o , p o iding immedia e eedback on any issues o sugges ions o imp o emen .
Upon comple ing he es ing p o ocol, pa icipan s we e asked o comple e a comp ehensi e ques ionnai e
o documen hei expe iences and e alua ions. Figu e 6.2 shows pa icipan s engaging wi h he iSigh
p o o ype du ing he hands-on es ing phase.
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
116
Figu e 6.2: ACAPO membe s in e ac ing wi h he iSigh p o o ype du ing he es ing phase, showcasing hands-on
engagemen : a) Pa icipan using he mecha onic de ice; b) Pa icipan using he iSigh mobile applica ion.
6.3 Ques ionnai e Me hodology
To e alua e he e ec i eness and usabili y o he iSigh p o o ype, a comp ehensi e ques ionnai e was
de eloped. The de ailed ques ionnai e is p o ided in Appendix A5. Be o e dis ibu ion o he pa icipan s,
a pilo es was conduc ed wi h wo membe s o he ACAPO boa d. This pilo es ing allowed o
imp o emen s and co ec ions in he seman ic o mula ion o he ques ionnai e, ensu ing cla i y and
e ec i eness in cap u ing he necessa y da a.
The ques ionnai e was designed o ga he de ailed eedback on a ious aspec s o he iSigh p o o ype. A
o al o 15 pa icipan s olun a ily ook pa in he s udy. The esponses we e analysed using SPSS 22,
and a
Like
scale was employed o measu e le els o ag eemen , impo ance, o equency ac oss a ious
ques ions. In he
Like
scale used, a lowe sco e indica ed a nega i e opinion, while a highe sco e
indica ed a posi i e opinion. The ques ionnai e comp ised 36 ques ions di ided in o six main sec ions,
desc ibed as ollows:
I. Pe sonal Iden i ica ion: his sec ion ga he ed demog aphic in o ma ion abou he pa icipan s,
including age, gende , academic quali ica ions, and cu en occupa ion.
a) b)
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
123
Figu e 6.6: Visual ep esen a ion o common causes o isual impai men (Bu on
e al
., 2021)
III. Technology
The esponses in his hi d sec ion highligh ed he cu en s a e o echnology adop ion among he blind
communi y. I p o ides insigh s on hei amilia i y wi h suppo i e echnologies, hei sa is ac ion le els,
and he equency o use.
QIII.1: Do you cu en ly use any applica ion dedica ed o clo hing?
The esponses indica e ha he majo i y o pa icipan s, 67 %, do no cu en ly use any applica ions
dedica ed o managing clo hing, while 33 % do use such applica ions. This sugges s ha he e is a
signi ican oppo uni y o he iSigh p o o ype o ill a gap in echnology usage among isually impai ed
people in he con ex o clo hing managemen . Fo hose who indica ed usage o clo hing- ela ed
applica ions, u he de ails we e sough on he speci ic applica ions used. One pa icipan epo ed using
colou de ec o s on an And oid mobile de ice, and ou pa icipan s epo ed using he Colo ino de ice.
This addi ional in o ma ion highligh s ha among hose using clo hing- ela ed echnology, he Colo ino
de ice is he mos commonly used. Colo ino is a well-known colou iden i ie o isually impai ed people,

Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
124
which helps use s dis inguish colou s in hei clo hing. The use o colou de ec ion applica ions and
de ices emphasizes he impo ance o colou iden i ica ion in clo hing managemen o he isually
impai ed communi y. In e es ingly, among he use s o hese echnologies, only one use had a low
li e acy le el. This sugges s ha he adop ion o such de ices is no signi ican ly hinde ed by li e acy
ba ie s, indica ing hei use - iendly na u e and accessibili y. Addi ionally, he majo i y o hese people
a e s ill engaged in p o essional ac i i ies and ha e no ye e i ed. This de ail highligh s he con inued
need o p ac ical and e icien solu ions o suppo isually impai ed people in hei ac i e p o essional
li es, ensu ing hey can manage hei clo hing choices e ec i ely and main ain hei p o essional
appea ance.
QIII.2: To wha ex en do you eel com o able using echnological de ices such as
sma phones and compu e s?
The esponses a ied, wi h he majo i y indica ing a mode a e le el o com o . Speci ically, 40 % o he
pa icipan s epo ed eeling e y com o able, 53.3 % el somewha com o able, and 6.7 % el
somewha uncom o able. These da a sugges ha mos pa icipan s eel, a leas , somewha com o able
using echnological de ices, wi h a subs an ial po ion eeling e y com o able. This o e all com o wi h
echnology is highly mo i a ing o he adop ion o he iSigh p o o ype, as i indica es ha mos use s a e
likely o be ecep i e o using a new echnological solu ion o clo hing managemen . Unde s anding hese
com o le els helps in designing use in e aces and ea u es ha ca e o he a ying deg ees o
echnological p o iciency among he pa icipan s.
QIII.3: How o en do you use echnological de ices in you daily li e?
The equency o echnological de ice usage was o e whelmingly in a ou o daily use. Fou een
pa icipan s indica ed daily use (
ca
. 93.3 %), while only one pa icipan ep esen ing
ca.
6.7 % epo ed
occasional use. These da a indica e ha mos pa icipan s use echnological de ices daily. This high
equency o daily use u he suppo s he po en ial accep ance and in eg a ion o he iSigh p o o ype
in o hei daily ou ines. The p e alence o daily echnology use among pa icipan s highligh s hei
amilia i y and com o wi h digi al ools, making hem well-sui ed o adop ing new echnological solu ions
like he iSigh p o o ype.
QIII4.: How would you a e you abili y o use so wa e and applica ions speci ically
designed o blind people, such as sc een eade s?
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
125
The esponses showed ha while a signi ican numbe o pa icipan s (46.7 %) a e hei skills as a e age,
he e is also a conside able p opo ion wi h e y high (26.6 %) and high (13.3 %) p o iciency le els. Only
13.3 % a ed hei skills as low. This ange o abili ies unde sco es he impo ance o designing he iSigh
p o o ype o be accessible and use - iendly o use s wi h a ying le els o echnical p o iciency. By
ensu ing ha he p o o ype is in ui i e and easy o na iga e, i can ca e o he needs o bo h highly skilled
use s and hose wi h less expe ience using specialized so wa e o he isually impai ed.
QIII.5: To wha ex en do you belie e ha echnologies imp o e you quali y o li e and
independence?
Ten pa icipan s, ep esen ing ca. 66.7 %, el ha echnologies imp o e hei quali y o li e and
independence e y signi ican ly, while i e pa icipan s, 33.3 %, belie ed ha echnologies imp o e hei
li es signi ican ly, wi h only one pa icipan , 6.7 %, elling ha he imp o emen is mode a e. This high
le el o pe cei ed bene i om echnology unde sco es he c i ical ole ha echnological solu ions like
he iSigh p o o ype can play in enhancing he daily li es and au onomy o isually impai ed people.
Designing he iSigh p o o ype o mee hese expec a ions can u he ein o ce i s posi i e impac and
ensu e i s widesp ead adop ion and use.
QIII.6: How o en do you seek o lea n and explo e new echnologies adap ed o you needs?
The in e es in lea ning and explo ing new echnologies adap ed o isually impai ed people’s needs was
high. The ga he ed da a e eal ha a signi ican majo i y o pa icipan s, 73.3 %, a e always looking o
lea n and explo e new echnologies, while 20 % do so egula ly, and only one pa icipan , ep esen ing
6.7 %, a ely seeks new echnologies. This high le el o engagemen wi h adap i e echnologies sugges s
ha he iSigh p o o ype has a ecep i e audience eage o in eg a e new solu ions in o he daily li es o
he pa icipan s. This en husiasm o explo ing new echnologies highligh s he po en ial o posi i e
ecep ion and sus ained use o he iSigh p o o ype among he isually impai ed communi y.
QIII.7: How do you e alua e he suppo and accessibili y o echnologies o blind people?
The suppo and accessibili y o echnologies o isually impai ed people ecei ed mos ly posi i e
eedback. Speci ically, 53.3 % a ed he suppo as sa is ac o y, while 40 % a ed i as e y sa is ac o y,
and only 6.7 % had no opinion. This eedback unde sco es he impo ance o con inuing o de elop and
e ine assis i e echnologies o mee he needs and expec a ions o isually impai ed use s. Ensu ing ha
he iSigh p o o ype o e s obus suppo and high accessibili y will be c ucial o i s success and adop ion
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
126
among his g oup o use s.
QIII.8: To wha ex en do you eel ha echnologies acili a e g ea e social and
p o essional inclusion in you li e?
The ole o echnologies in acili a ing g ea e social and p o essional inclusion was o e whelmingly
posi i e. The ob ained da a e eal ha 93.3 % o he pa icipan s belie e ha echnologies con ibu e e y
signi ican ly o hei social and p o essional inclusion, while he emaining 6.7 % conside he impac o
be jus signi ican . This highligh s he c i ical ole ha echnology plays in enabling isually impai ed people
o pa icipa e mo e ac i ely in social and p o essional se ings. The s ong endo semen o echnology's
ole in inclusion suppo s he con inued de elopmen and implemen a ion o inno a i e solu ions like he
iSigh p o o ype o u he enhance hese oppo uni ies o isually impai ed use s.
QIII.9: Do you conside echnologies o be a use ul ool o you educa ion o aining?
The use ulness o echnologies as ools o educa ion o aining was highly endo sed. The collec ed da a
indica e ha all pa icipan s ind echnologies o be bene icial o hei educa ion o aining, wi h 46.7 %
a ing hem as ex emely use ul and 53.3 % as e y use ul. This s ong endo semen highligh s he
impo ance o in eg a ing educa ional and aining unc ionali ies in o he iSigh p o o ype. By ensu ing
ha he p o o ype suppo s educa ional and aining needs, i can p o ide subs an ial bene i s o isually
impai ed people, enhancing hei lea ning expe iences and p o essional de elopmen oppo uni ies.
IV. Accessibili y o he iSigh Mobile Applica ion
Feedback on he accessibili y and usabili y o he iSigh mobile applica ion was ga he ed h ough de ailed
ques ions in he ou h sec ion. Pa icipan s p o ided aluable sugges ions and opinions, which we e
c ucial o e ining he applica ion's use in e ace and ensu ing i me he high s anda ds o accessibili y
equi ed by he isually impai ed communi y.
QIV.1: How do you e alua e he ease o na iga ion in he mobile applica ion?
The esponses we e gene ally posi i e, indica ing a a ou able use expe ience. A o al o 66.7 % o
pa icipan s a ed he na iga ion as easy, while 33.3 % a ed i as e y easy. This eedback sugges s ha
he mobile applica ion's in e ace is use - iendly and accessible o isually impai ed use s. The high
a ings o ease o na iga ion indica e ha he design and layou o he applica ion e ec i ely me he
needs o he a ge use g oup. Ensu ing ease o na iga ion is c ucial o he success ul adop ion and
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
127
con inued use o he applica ion, as i di ec ly impac s he o e all use expe ience. The posi i e e alua ion
o na iga ion wi hin he mobile applica ion highligh s he impo ance o main aining simplici y and
in ui i eness in he use in e ace design. This will help use s e icien ly pe o m asks and access ea u es
wi hou unnecessa y complexi y o us a ion. The eedback also p o ides aluable insigh s o u he
e inemen and imp o emen , ensu ing ha he applica ion emains accessible and use - iendly o all
use s.
QIV.2: Does he mobile applica ion p o ide image desc ip ions o ex ual al e na i es o
isual elemen s adequa ely?
The esponses indica ed a s ong o e all sa is ac ion wi h he applica ion's p o ision o image desc ip ions
and ex ual al e na i es. Speci ically, 53.3 % o pa icipan s a ed he desc ip ions as excellen , 26.7 % as
good, and 20 % as adequa e. No pa icipan s a ed he desc ip ions as poo o e y poo . The posi i e
eedback on he applica ion's image desc ip ions and ex ual al e na i es emphasizes he e ec i eness o
hese ea u es in enhancing accessibili y. P o iding clea and comp ehensi e desc ip ions allows isually
impai ed use s o ully engage wi h isual con en , making he applica ion mo e inclusi e and usable. The
high a ings o image desc ip ions and ex ual al e na i es sugges ha he applica ion mee s o exceeds
he expec a ions o mos use s in his ega d. This aspec o he applica ion's accessibili y is c ucial, as i
ensu es ha all use s, ega dless o isual abili y, can access and unde s and he con en p esen ed
wi hin he app. The eedback also highligh s he impo ance o con inuing o p io i ize and e ine hese
ea u es o main ain high s anda ds o accessibili y and use sa is ac ion.
QIV.3: How do you e alua e he cla i y and o ganiza ion o he ex ual con en in he mobile
applica ion?
The da a indica e ha he majo i y o pa icipan s ound he ex ual con en o be ei he e y clea o clea .
Speci ically, 53.3 % o pa icipan s a ed he con en as e y clea , while 46.6 % a ed i as clea . The e
we e no esponses indica ing ha he con en was unclea . The high a ings o cla i y and o ganiza ion
o ex ual con en sugges ha he applica ion e ec i ely communica es in o ma ion in a way ha i is
easily unde s andable o isually impai ed use s. This cla i y is essen ial o ensu ing ha use s can
na iga e he applica ion and access he in o ma ion hey need wi hou con usion o di icul y. The posi i e
eedback ega ding he ex ual con en highligh s he impo ance o main aining high s anda ds o cla i y
and o ganiza ion. Clea and well-o ganized con en helps use s e icien ly inding and comp ehending he
in o ma ion hey need, enhancing hei o e all expe ience wi h he applica ion. This eedback is aluable
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
128
o ongoing imp o emen s, ensu ing ha he applica ion con inues o mee he needs o i s use s
e ec i ely.
QIV.4: How do you e alua e he o e all use expe ience o he mobile applica ion?
The esponses show a high le el o o e all sa is ac ion wi h he mobile applica ion. Speci ically, 66. 7% o
pa icipan s a ed hei expe ience as e y sa is ac o y, while 33.3 % a ed i as sa is ac o y. No
pa icipan s a ed hei expe ience as unsa is ac o y o e y unsa is ac o y. The high a ings o o e all
use expe ience sugges ha he mobile applica ion success ully mee s he needs and expec a ions o
isually impai ed use s. The posi i e e alua ions indica e ha he applica ion p o ides a sa is ac o y and
use - iendly expe ience, which is c ucial o ensu ing con inued use and accep ance. The eedback on
he o e all use expe ience emphasizes he impo ance o main aining a high le el o use sa is ac ion
h ough con inuous imp o emen s and upda es. By add essing any po en ial issues and inco po a ing
use eedback, he applica ion can u he enhance i s usabili y and e ec i eness, ensu ing i emains a
aluable ool o isually impai ed people.
V. Usabili y o he iSigh P o o ype
The usabili y o he iSigh p o o ype was assessed h ough ques ions ocused on he ease o use and u ili y
o he de ice. Pa icipan s' eedback indica ed how in ui i e hey ound he p o o ype and i s e ec i eness
in mee ing hei needs. The i h sec ion also explo es he p e e ed me hods o echnology deli e y and
a ailabili y o he blind communi y.
QV.1: E alua e in he ease o use o he iSigh p o o ype, which includes bo h he mobile
applica ion and he sma wa d obe.
The esponses we e o e whelmingly posi i e, indica ing a high le el o use - iendliness and accessibili y.
A o al o 53.3 % o pa icipan s ound he iSigh p o o ype o be e y easy o use, while 46.7 % ound i
jus easy o use. No pa icipan s epo ed di icul ies o a ed he ease o use as less han easy. This
eedback sugges s ha he iSigh p o o ype success ully mee s he usabili y equi emen s o i s a ge
use g oup, ensu ing ha he combina ion o he mobile applica ion and sma wa d obe is in ui i e and
s aigh o wa d o isually impai ed use s. The high a ings o ease o use indica e ha he design and
unc ionali y o bo h he mobile applica ion and he sma wa d obe a e well-in eg a ed and accessible.
This is c ucial o encou aging he adop ion and consis en use o he iSigh p o o ype, as use s a e mo e
likely o engage wi h echnology ha hey ind easy o na iga e and ope a e.

Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
129
QV.2: E alua e he accu acy o he iSigh p o o ype in iden i ying he ca ego y and colou s
o clo hing i ems.
Fo he accu acy in iden i ying he ca ego y o clo hing i ems, 60 % o pa icipan s a ed i as " e y
p ecise," while 40 % a ed i jus as "p ecise." Fo he accu acy in iden i ying he colou s o clo hing i ems,
80% o pa icipan s a ed i as " e y p ecise," while 20 % a ed i as "p ecise."
Mos pa icipan s ound he iSigh p o o ype o be highly accu a e in iden i ying bo h he ca ego y and
colou s o clo hing i ems. Howe e , a small numbe o pa icipan s (20 %) epo ed a di e gence in hei
esponses, a ing he accu acy as "p ecise" o ca ego y iden i ica ion and " e y p ecise" o colou
iden i ica ion. The emaining pa icipan s p o ided consis en a ings o bo h ca ego y and colou
accu acy. This sugges s ha while he iSigh p o o ype gene ally pe o ms well in bo h a eas, he e is a
pe cei ed highe accu acy in colou iden i ica ion compa ed o ca ego y iden i ica ion among some use s.
QV.3: E alua ed he iSigh p o o ype's abili y o de ec s ains and iden i y NFC ags.
The esponses o de ec ing s ains we e o e whelmingly posi i e, wi h he majo i y, 86.7 %, a ing i as
"highly e ec i e." Simila ly, he iden i ica ion o NFC ags also ecei ed high a ings, wi h mos
pa icipan s, 86.7 %, inding i "highly e ec i e." The e was a s ong co ela ion be ween he pa icipan s'
a ings o hese wo unc ionali ies, indica ing ha use s gene ally ound bo h ea u es o be highly
e ec i e. Speci ically, pa icipan s who a ed he iSigh p o o ype as " e y p ecise" o bo h ca ego y and
colou iden i ica ion consis en ly a ed i s abili y o de ec s ains and iden i y NFC ags as "highly e ec i e."
Pa icipan s who a ed ca ego y iden i ica ion as "p ecise" bu colou iden i ica ion as " e y p ecise" also
showed high a ings o de ec ing s ains and NFC ags, hough wi h some mino a ia ions. Those who
a ed bo h ca ego y and colou iden i ica ion as "p ecise" ended o a e s ain and NFC ag de ec ion as
"e ec i e," hough some a ed hese capabili ies highe . This consis ency ac oss esponses sugges s ha
he iSigh p o o ype pe o ms eliably ac oss di e en unc ionali ies, p o iding use s wi h an e ec i e ool
o managing hei clo hing. The s ong pe o mance in colou iden i ica ion and NFC ag de ec ion is
pa icula ly no ewo hy, as i highligh s he p o o ype's abili y o handle complex asks wi h high p ecision.
In summa y, he usabili y and accu acy o he iSigh p o o ype a e highly a ed by pa icipan s, indica ing
ha he p o o ype mee s he needs and expec a ions o isually impai ed use s. The posi i e eedback on
ease o use and accu acy unde sco es he e ec i eness o he design and unc ionali y o he iSigh
p o o ype. Con inuing o e ine hese aspec s based on use eedback will be impo an o u he
enhancing he use expe ience and ensu ing he success ul adop ion o he echnology. The s ong
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
130
co ela ion be ween esponses ac oss di e en unc ionali ies ein o ces he p o o ype's o e all eliabili y
and e ec i eness in assis ing isually impai ed people in managing hei wa d obe. The addi ional
commen s and sugges ions p o ide aluable insigh s o u u e imp o emen s, ensu ing he sys em
emains use -cen ic and esponsi e o he needs o i s use s.
VI. Assessmen o he Impo ance o iSigh
This assessmen aims o e alua e he pe cei ed impo ance o he iSigh p o o ype among use s,
speci ically examining i s e ec i eness in acili a ing he clo hing selec ion p ocess, iden i ying speci ic
cha ac e is ics and modi ica ions in ga men s, and enhancing wa d obe o ganiza ion. By analysing use
eedback and sa is ac ion, he s udy seeks o elucida e he p o o ype's impac and pinpoin a eas o
po en ial enhancemen .
QVI.1: E alua e he impo ance o he iSigh p o o ype h ough wo s a emen s conce ning
i s unc ionali y. S1: "iSigh acili a es he p ocess o selec ing clo hes, while allowing
iden i ica ion o cha ac e is ics and modi ica ions o he o iginal s a e." S2: "iSigh
o ganizes my clo hes e ec i ely."
The esponses we e compellingly posi i e. Fo he s a emen S1, 66.7 % o pa icipan s s ongly ag eed,
and 33.3 % jus ag eed. Simila ly, o he s a emen S2, 66.7 % o pa icipan s s ongly ag eed, while
33.3 % jus ag eed. These esul s indica e a high le el o ag eemen among pa icipan s ega ding he
impo ance and e ec i eness o he iSigh p o o ype in bo h acili a ing he selec ion o clo hes and
o ganizing hem e ec i ely. The majo i y o pa icipan s s ongly ag ee ha he p o o ype signi ican ly aids
in hese aspec s, sugges ing ha he sys em mee s c i ical needs o isually impai ed use s in managing
hei wa d obe. The consis en posi i e esponses ac oss bo h s a emen s e lec he pe cei ed alue o
he iSigh p o o ype in imp o ing daily ou ines and enhancing he independence o isually impai ed
people. This eedback unde sco es he sys em's po en ial o make a meaning ul impac on he li es o i s
use s by p o iding p ac ical and eliable solu ions o clo hing managemen .
QVI.2: To assess whe he he use o he iSigh p o o ype inc eased hei eelings o
con idence, sel -es eem, well-being, and independence h ough ou s a emen s. S1: “The
use o iSigh inc eased my sense o con idence”. S2: “The use o iSigh inc eased my sense
o sel -es eem”. S3: “The use o iSigh inc eased my sense o well-being”. S4: “The use o
iSigh inc eased my sense o independence”.
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
131
The esponses indica e a e y posi i e impac o he iSigh p o o ype on hese aspec s. Fo S1, 66.7 % o
pa icipan s s ongly ag eed, whe eas 33.3 % jus ag eed. Fo S2, 66.7 % o pa icipan s s ongly ag eed,
while 33.3% ag eed. Fo S3, 66.7 % o pa icipan s s ongly ag eed, and 33.3 % ag eed. Finally, o S4,
66.7 % o pa icipan s s ongly ag eed, and 33.3 % ag eed. The e is a high le el o consis ency in he
esponses ac oss all ou s a emen s, wi h all pa icipan s ei he s ongly ag eeing o ag eeing ha he
iSigh p o o ype posi i ely impac ed hei con idence, sel -es eem, well-being, and independence. This
sugges s a obus and consis en posi i e pe cep ion o he p o o ype's bene i s. All pa icipan s who
s ongly ag eed on one s a emen also s ongly ag eed on he o he s a emen s. Simila ly, pa icipan s
who ag eed on one s a emen also ag eed on he o he s a emen s. This uni o mi y in esponses indica es
ha pa icipan s consis en ly expe ienced an inc ease in con idence, sel -es eem, well-being, and
independence h ough he use o he iSigh p o o ype. The lack o di e gen esponses u he suppo s
he conclusion ha he iSigh p o o ype has a comp ehensi e posi i e e ec on use s' li es. The uni o m
posi i e eedback ac oss all ou s a emen s unde sco es he signi ican impac o he iSigh p o o ype on
enhancing he quali y o li e o isually impai ed use s. The inc ease in con idence, sel -es eem, well-
being, and independence highligh s he p o o ype's e ec i eness in add essing he e e yday challenges
aced by hese use s.
QVI.3: To assess he impo ance o he iSigh p o o ype in iden i ying clo hes, colou s, and
s ains/di .
Pa icipan s uni o mly ecognized he iSigh p o o ype as a c ucial ool in hei daily li es. Fo he
impo ance o iSigh in iden i ying clo hes, 66.7 % o pa icipan s ound i ex emely impo an , and 33.3 %
ound i e y impo an . Simila ly, o iden i ying colou s, 66.7 % o pa icipan s ound i ex emely
impo an , and 33.3 % ound i e y impo an . Rega ding he iden i ica ion o s ains/di , 66.7 % o
pa icipan s ound i ex emely impo an , and 33.3 % ound i e y impo an . These esponses indica e
ha pa icipan s uni o mly ound he iSigh p o o ype o be an ex emely impo an ool in iden i ying
clo hes, colou s, and s ains/di . The high pe cen age o pa icipan s a ing i as "Ex emely Impo an "
e lec s he p o o ype's signi ican ole in aiding hei daily ou ines and enhancing hei independence.
This consis ency in high a ings ac oss all h ee a eas highligh s he comp ehensi e u ili y o he iSigh
p o o ype in assis ing isually impai ed people in managing hei clo hing e ec i ely.
QVI.4: Whe he hey belie e ha he iSigh p o o ype would imp o e hei quali y o li e in
ela ion o clo hing managemen .
Chap e 6 –
E alua ion o P o o ype
___________________________________________________________________________
132
The esponses indica e ha he majo i y o pa icipan s belie e ha iSigh would signi ican ly enhance
hei quali y o li e. Fo he s a emen , "iSigh imp o es my quali y o li e in ela ion o clo hing
managemen ," 80.0 % o pa icipan s indica ed ha i g ea ly imp o es hei quali y o li e, while 20.0 %
indica ed ha i signi ican ly imp o es hei quali y o li e. These esponses e lec a s ong consensus
among pa icipan s ega ding he bene icial impac o he iSigh p o o ype on hei daily li es. The majo i y
also belie es ha he iSigh p o o ype g ea ly imp o es hei quali y o li e, while he emaining pa icipan s
s ill iew i as a signi ican imp o emen . This indica es ha he iSigh p o o ype is pe cei ed as a aluable
ool ha can subs an ially aid isually impai ed people in managing hei clo hing mo e e ec i ely,
ul ima ely enhancing hei o e all quali y o li e.
QVI.5: Abou he speci ic aspec s o hei daily li es whe e he iSigh sys em has a posi i e
impac .
The op ions included acili a ing he choice o clo hes o speci ic occasions, helping o iden i y he colou s
o clo hes mo e e ec i ely, enabling he iden i ica ion o s ains/di on clo hes, con ibu ing o g ea e
au onomy in managing hei wa d obe, and o he unspeci ied bene i s. The esponses we e highly posi i e,
wi h e e y pa icipan selec ing all he op ions, excep o one pa icipan who did no selec he i s op ion.
QVI.6 Do you ha e any speci ic sugges ions o imp o emen s you would like o see
implemen ed in he sys em o make i e en mo e use ul in you daily li e?
To p o ide a comp ehensi e o e iew and a oid edundancy, eedback om sec ion IV (Ques ion: "Please
sha e any addi ional commen s o speci ic sugges ions o imp o e he mobile applica ion based on you
expe ience.") and sec ion V (Ques ion: "Please sha e any addi ional commen s o speci ic sugges ions o
imp o e he sys em based on you expe ience.") we e combined. The agg ega ed esponses highligh he
inno a i e na u e o he iSigh p o o ype and i s po en ial o signi ican ly enhance he daily li e o isually
impai ed people. Use s app ecia e he applica ion’s abili y o aid in clo hing selec ion, iden i ica ion o
cha ac e is ics, and acili a ing o ganiza ion, no ing ha i could help hem d ess independen ly and
e icien ly. Many use s sugges ed making he applica ion as e by including ewe menus o s eamline
he p ocess. The u ili y o he applica ion is e iden , wi h pa icipan s emphasizing i s use ulness in daily
ou ines and exp essing eage ness o ha e such a sys em a home. Se e al use s sugges ed
enhancemen s o he applica ion's ea u es, such as including de ailed in o ma ion abou clo hing ab ics,
eading label cha ac e is ics, and iden i ying s ains mo e accu a ely. The applica ion’s abili y o dis inguish
colou s and simila i ems, like shoes, was also highligh ed as a c i ical a ea o imp o emen . Despi e