Academic Edi o : Jian Tang
Recei ed: 21 Feb ua y 2025
Re ised: 5 Ap il 2025
Accep ed: 8 Ap il 2025
Published: 14 Ap il 2025
Ci a ion: Cas o-Bello, M.;
Roman-Padilla, D.B.; Mo ales-Mo ales,
C.; Campos-F ancisco, W.; Ma molejo-
Vega, C.V.; Ma molejo-Dua e, C.;
E angelis a-Alcoce , Y.; Gu ié ez-
Valencia, D.E. Con olu ional Neu al
Ne wo k Models in Municipal Solid
Was e Classi ica ion: Towa ds
Sus ainable Managemen .
Sus ainabili y 2025,17, 3523. h ps://
doi.o g/10.3390/su17083523
Copy igh : © 2025 by he au ho s.
Licensee MDPI, Basel, Swi ze land.
This a icle is an open access a icle
dis ibu ed unde he e ms and
condi ions o he C ea i e Commons
A ibu ion (CC BY) license
(h ps://c ea i ecommons.o g/
licenses/by/4.0/).
A icle
Con olu ional Neu al Ne wo k Models in Municipal Solid
Was e Classi ica ion: Towa ds Sus ainable Managemen
Mi na Cas o-Bello
1
, Dominic B ian Roman-Padilla
1
, Co nelio Mo ales-Mo ales
2,
* , Wil ido Campos-F ancisco
1
,
Ca los Vi gilio Ma molejo-Vega 1, Ca los Ma molejo-Dua e 3, Yane E angelis a-Alcoce 1
and Diego Es eban Gu ié ez-Valencia 1
1Technological Ins i u e o Chilpancingo, Na ional Ins i u e o Technology o Mexico,
Chilpancingo de los B a o 39090, Mexico; [email p o ec ed] (M.C.-B.);
[email p o ec ed] (D.B.R.-P.)
2Technological Ins i u e o San Juan del Río, Na ional Ins i u e o Technology o Mexico,
San Juan del Rio Que e a o 76800, Mexico
3Cen e o Land Policy and Valua ions, Ba celona School o A chi ec u e (ETSAB), Poly echnic Uni e si y o
Ca alonia, 08034 Ba celona, Spain; [email p o ec ed]
*Co espondence: [email p o ec ed]
Abs ac : Municipal Solid Was e (MSW) managemen p esen s a signi ican challenge
o adi ional sepa a ion p ac ices, due o a conside able inc ease in quan i y, di e si y,
complexi y o ypes o solid was e, and a high demand o accu acy in classi ica ion. Image
ecogni ion and classi ica ion o was e using compu e ision echniques allow o op i-
mizing adminis a ion and collec ion p ocesses wi h high p ecision, achie ing in elligen
managemen in sepa a ion and inal disposal, mi iga ing en i onmen al impac , and con-
ibu ing o sus ainable de elopmen objec i es. This esea ch consis ed o e alua ing and
compa ing he e ec i eness o ou Con olu ional Neu al Ne wo k models o MSW de ec-
ion, using a Raspbe y Pi 4 Model B. To his end, he models YOLO 4- iny, YOLO 7- iny,
YOLO 8-nano, and YOLO 9- iny we e ained, and hei pe o mance was compa ed in
e ms o p ecision, in e ence speed, and esou ce usage in an embedded sys em wi h a
cus om da ase o 1883 o ganic and ino ganic was e images, labeled wi h Robo low by de-
limi ing he a ea o in e es o each objec . Image p ep ocessing was applied, wi h esizing
o 640
×
640 pixels and con as au o-adjus men s. T aining conside ed 85% o images and
es ing conside ed 15%. Each aining s age was conduc ed o e 100 epochs, adjus ing
con igu a ion pa ame e s such as lea ning a e, weigh decay, image o a ion, and mosaics.
The p ecision esul s ob ained we e as ollows: YOLO 4- iny, 91.71%; YOLO 7- iny, 91.34%;
YOLO 8-nano, 93%; and YOLO 9- iny, 92%. Each model was applied in an embedded
sys em wi h an HQ came a, achie ing an a e age o 86% CPU usage and an in e ence
ime o 1900 ms. This sugges s ha he models a e easible o applica ion in an in elligen
con aine o classi ying o ganic and ino ganic was e, ensu ing e ec i e managemen and
p omo ing a cul u e o en i onmen al ca e in socie y.
Keywo ds: objec de ec ion; was e; Con olu ional Neu al Ne wo ks; Raspbe y Pi; embedded
sys em; sus ainable de elopmen ; managemen
1. In oduc ion
The managemen o Municipal Solid Was e (MSW) is a majo challenge, due o i s high
gene a ion a e. In 2020, he o al amoun gene a ed wo ldwide was es ima ed a 2.24 bil-
lion ons. This is a ibu ed o apid popula ion g ow h and he ongoing ad ancemen o
Sus ainabili y 2025,17, 3523 h ps://doi.o g/10.3390/su17083523
Sus ainabili y 2025,17, 3523 2 o 18
u baniza ion. P ojec ions o he yea 2050 indica e a 73% inc ease compa ed o he igu es
epo ed in 2020, eaching an es ima ed o al o 3.88 billion ons [
1
]. The ecogni ion and
classi ica ion o Municipal Solid Was e (MSW) images o e s he oppo uni y o iden i y
and ca alog hem using compu e ision echniques, op imizing in elligen and e ec i e
managemen p ocesses o main ain u ban en i onmen quali y, imp o e ci izens’ quali y
o li e, and con ibu e o achie ing sus ainable de elopmen objec i es [
2
]. On he o he
hand, mode n objec iden i ica ion echnology helps o imp o e objec de ec ion; o exam-
ple, ligh weigh deep lea ning algo i hms wi h educed size achie e simila accu acy o
models ha equi e highe cos s and equi emen s. These ha e achie ed no able ad ances,
s anding ou o hei compac size and highe speed wi hou a dec ease in accu acy, and
p o iding heo e ical and echnical suppo ; hey a e pa icula ly use ul in in eg a ed
de ices, allowing he achie emen o supe io le els o classi ica ion e iciency [3–5].
Con olu ional Neu al Ne wo ks (CNN) a e deep lea ning models ocused on image
p ocessing and hie a chical ea u e ex ac ion, as well as spa ial locali y modeling. They
consis o mul iple con olu ional and ully connec ed laye s, acili a ing he ex ac ion
o mo e de ailed ea u es and be e image in e p e a ion, and cap u ing, in u n, mo e
global con ex in o ma ion. By unde s anding he gene al s uc u e and seman ics o each
model, i s u ili y can be de e mined. Some examples o deep ne wo k s uc u es a e ResNe ,
Incep ion, E icien Ne , Shu leNe , and YOLO [
2
]. ResNe -50 is a deep CNN a chi ec-
u e ocused on sol ing p oblems such as pe o mance deg ada ion, lea ning di icul y,
and con e gence p oblems wi h inc eased laye s. Applying esidual connec ions o lea n
om mappings and ex ac de ailed ea u es om images, i consis s o con olu ional
laye s, ba ch no maliza ion, pooling, and ully connec ed laye s [
6
]. Shu leNe 2 ocuses
on designing mo e e icien ne wo ks by op imizing memo y usage and ne wo k speed,
main aining he same numbe o inpu and ou pu channels o he con olu ional laye o
minimize memo y access cos , educing he use o g oup con olu ions o sa e esou ces,
simpli ying he ne wo k b anch s uc u e o imp o e pa allel ope a ion, and dec easing
elemen -le el ope a ions o imp o e ne wo k ope a ion [
7
]. You Only Look Once (YOLO)
is an objec de ec ion model designed o eal- ime image p ocessing. By analyzing an
image once while mo ing, i can de ec and classi y objec s h ough p io aining based
on a labeled and ca aloged da ase . I ou pe o ms adi ional CNNs in e ms o e iciency
o eal- ime objec de ec ion, by using a single CNN o p edic bounding boxes and class
p obabili ies simul aneously. In addi ion o ha ing as ame de ec ion, i easons globally,
and pe o ms apid lea ning o iden i y and gene alize objec cha ac e is ics [
8
]. Acco ding
o Ul aly ics, YOLO app oaches objec de ec ion as a single eg ession p oblem by p e-
dic ing bounding boxes and class p obabili ies om ull images in one e alua ion. This
app oach makes YOLO as e han p e ious wo-s age de ec o s, while s ill main aining
high p ecision, and i allows o apid expo , implemen a ion, and execu ion on Raspbe y
Pi de ices. YOLO 4- iny, conside ed a ligh weigh model designed based on YOLO 4, uses
less memo y and ope a es as e , signi ican ly inc easing i s iabili y o implemen a ion
in sma de ices [
9
]. YOLO 7- iny, he ligh weigh e sion o YOLO 7, s ands ou o i s
speed and e iciency, wi h 6.2 M pa ame e s and 13.8 B ope a ions (FLOPs). I is especially
e ec i e in de ec ing objec s o a ious sizes, achie ing pe o mance sco es o 18.8% o
small objec s, 42.4% o medium objec s, and 51.9% o la ge objec s [
10
]. YOLO 8n (nano)
is a ligh weigh model designed o speed and e iciency, and is pa icula ly sui able o
de ices wi h limi ed compu ing esou ces. I conside s only 3.2 M pa ame e s and 8.7 B
FLOPs, is op imized o un on a CPU, and i s smalle a chi ec u e makes i ideal o edge
de ice implemen a ions, while s ill main aining good objec de ec ion pe o mance [
10
].
Finally, YOLO 9 ( iny) is a simpli ied e sion o YOLO 9 designed o esou ce-limi ed
applica ions, ea u ing 2.0 M pa ame e s and 7.7 B FLOPs, wi h an in e ence ime o 2.3 ms
Sus ainabili y 2025,17, 3523 3 o 18
on a T4 GPU wi h Tenso RT10. This e sion is pa icula ly use ul o deploymen on
edge de ices and mobile applica ions whe e compu a ional esou ces a e limi ed, wi hou
signi ican ly sac i icing p ecision [10].
Li e a u e Re iew
In he ield o objec de ec ion models, no able ad ances ha e been epo ed in he
li e a u e. Fo example, he au ho s o [
11
] designed a model called GC-YOLO 5 o
classi ying ba e ies, o ange peels, used pape , cups, and pape bo les. They used 642 im-
ages o aining and 40 o alida ion, achie ing an accu acy o 99.86% a e 500 epochs,
implemen ing hei applica ion wi h a Ja aSc ip in e ace. Simila ly, he au ho s o [
12
]
de eloped he PublicGa bageNe algo i hm wi h a inal accu acy o 96.35%, which classi ies
ou ca ego ies: ki chen was e, ecyclables, haza dous was e, and o he ypes. I is ocused
on mul i asking and is based on CNN a chi ec u e, and was ained wi h a public da ase
o 10,624 images and 10 classes (ki chen was e, plas ic, pape , elec onics, me al, glass,
non- ecyclable pape and plas ics, ex iles, and haza dous was e). I s pe o mance was
imp o ed by op imizing he backbone, da a augmen a ion, hype pa ame e uning, and
label smoo hing. Meanwhile, he au ho s o [
13
] p oposed a was e classi ica ion algo i hm
ha is esis an o he low accu acy caused by ligh and shadow in e e ence, wi h adap i e
capabili y in image ligh ing, using a h eshold eplacemen me hod o educe shadow
noise, along wi h he Canny ope a o o help c op he whi e backg ound in images. The
algo i hm was op imized based on he MLH-CNN model, achie ing an accu acy o 96.77%
wi h he au ho s’ cons uc ed da ase and 93.72% wi h T ashNe . Re e ences [
14
–
16
] de-
eloped a ligh weigh CNN model wi h an e icien algo i hm on an embedded de ice
o de ec ing su ace de ec s in indus ial p oduc s, eaching accu acy le els compa able
o hose o s a e-o - he-a (SOTA) models, while using ewe pa ame e s and equi ing
ewe compu a ions. Addi ionally, he au ho s implemen ed sa e and op imal aul - ole an
con ol, based on neu al ne wo ks, o econs uc sys em dynamics om da a, designing an
adap i e c i ical scheme wi h a non-quad a ic cos unc ion o handle inpu cons ain s and
educe compu a ional load.
Wi h ega d o da ase s, he au ho s o [
2
] de eloped an in eg a ed me hod o was e
image ecogni ion and classi ica ion ha combines ResNe -50, YOLO 5, and weakly su-
pe ised CNN algo i hms, imp o ing he accu acy and e iciency o image ecogni ion.
ResNe -50 was used o ex ac ea u es om he images, and weakly supe ised CNNs
we e used o aining and p edic ion. The au ho s e alua ed ou public da ase s: HGI-30,
T ashNe , GINI, and Public Ga bageNe . Wi h HGI-30, he in e ence ime, FLOPs, and
MAPE we e educed by mo e han 48.6%, 46.5%, and 41%, espec i ely, and an accu acy
o 97.02% was achie ed. Meanwhile, he au ho s o [
17
] p esen ed a c i ical analysis o
11 exis ing was e da ase s, collec ing and summa izing p e ious s udies by each au ho and
hei expe imen al esul s. Addi ionally, hey designed wo benchma k da ase s, combining
de ec -was e, classi y-was e, and all he ca ego ies iden i ied in he s udy. Finally, hey
p esen ed a wo-s age de ec o o was e classi ica ion and localiza ion, using E icien De -
D2 o loca e was e and E icien Ne -B2 o classi y i h ough semi-supe ised aining
using unlabeled images, achie ing 70% a e age p ecision in was e localiza ion and 75% in
classi ica ion.
Ano he ac o o conside is he compa a i e pe o mance o models ha can be
applied in embedded sys ems. Fo example, he au ho s o [
18
] p esen ed a model o classi y
ino ganic con aine s in was e p ocessing si es. They compa ed i e CNN a chi ec u es:
Xcep ion, Incep ion V3, ResNe -50, ResNe -50 V2, and DenseNe -201, p e iously ained
wi h a da ase o HDPE plas ic, PET, glass bo les and ja s, cans, ca dboa d, and lexible
plas ic con aine s, collec ed om h ee di e en Kaggle esou ces: a wa e bo le image
Sus ainabili y 2025,17, 3523 4 o 18
da ase , was e classi ica ion, and was e classi ica ion in o 12 classes. Fo aining, hey
conside ed da a augmen a ion, escaling, o a ion, displacemen , il ing, zoom, and lipping,
ob aining a CNN model based on he DenseNe -201 a chi ec u e wi h an accu acy o 0.96%.
Unde his same app oach, he con ibu ions o [
6
] ocused on c ea ing a deep lea ning-
based was e classi ica ion sys em ha enables ecogni ion and ecycling o household
was e h ough a ligh weigh ga bage classi ica ion model, GCNe (Ga bage Classi ica ion
Ne wo k). The Shu leNe 2 model was imp o ed by in eg a ing a pa allel mixed a en ion
mechanism (PMAM), new ac i a ion unc ions, and ans e lea ning, achie ing an a e age
accu acy o 97.9%, and i was applied on a Raspbe y Pi 4B, wi h an in e ence ime o
105 ms. The sys em comple ed he classi ica ion and collec ion o a single objec in 0.88 s.
Simila ly, in [
19
], an in elligen was e classi ica ion sys em based on GNe deep lea ning
was gene a ed h ough Linux and applied on a Raspbe y Pi 4B as he main boa d, in
addi ion o he in eg a ed ha dwa e o a wo-deg ees-o - eedom se o, ouch panel, and
came a. The model was ained wi h he Huawei Ga bage Classi ica ion Challenge Cup
da ase , ob aining an accu acy o 92.62% and an e iciency o 0.63 s, complemen ed by he
de elopmen o a g aphical use in e ace based on Py hon 3.6.8 and QT. Finally, he au ho s
o [
20
] applied a model in Tenso Flow Li e 2.13 on a Raspbe y Pi 4 o eal- ime was e
classi ica ion and ca ego iza ion. The se omo o s we e moun ed on a plas ic boa d ha
sepa a ed he was e acco ding o he model’s classi ica ion and deposi ed i in app op ia e
compa men s depending on i s ype.
The p esen s udy analyzes he e ec i eness o ou CNN models in MSW de ec ion,
using a Raspbe y Pi 4 Model B. YOLO 4- iny, YOLO 7- iny, YOLO 8-nano, and YOLO 9-
iny models we e ained and e alua ed, wi h hei pe o mance compa ed in e ms o
p ecision, in e ence speed, and esou ce usage in an embedded sys em. A da ase was
adjus ed wi h 85% o images o aining and 15% o es ing. In he expe imen s conduc ed,
lea ning a e, weigh decay, epochs, and online aining we e con igu ed, gene a ing image
o a ion and mosaics, and he beha io s o model p ecision and loss we e obse ed. The
compa ison o he models allowed o iden i ica ion o he mos sui able one o eal- ime
MSW de ec ion, conside ing ha dwa e limi a ions and he need o main ain a balance
be ween e iciency and p ecision in a low-cos and low-ene gy-consump ion en i onmen .
The p ecision esul s ob ained we e as ollows: YOLO 4- iny, 91.71%; YOLO 7- iny, 91.34%;
YOLO 8-nano, 93%; and YOLO 9- iny, 92%. Thei applica ion in an embedded sys em
wi h an HQ came a ob ained an a e age o 93% CPU usage and an in e ence ime o
1800 ms.
2. Ma e ials and Me hods
2.1. Ma e ials
Fo his esea ch, a Dell G15 5530 lap op wi h CPU i7 13650HX, GPU RTX 4060 8GB
VRAM, and 32 GB RAM, a Raspbe y Pi 4 model B wi h an HQ came a, Robo low ool [
21
],
YOLO 4 iny [
22
], YOLO 7 iny [
23
], YOLO 8 nano [
24
] and YOLO 9 nano [
25
] we e em-
ployed.
2.2. Me hodology
To e alua e and compa e he e ec i eness o ou con olu ional neu al ne wo k
models o MSW de ec ion, a ou -phase me hodology was de eloped, as shown in Figu e 1.
Sus ainabili y 2025,17, 3523 5 o 18
Sus ainabili y 2025, 17, x FOR PEER REVIEW 5 o 18
Figu e 1. Me hodological diag am.
2.2.1. Da ase Design
The design o he da ase o 1883 images in ol ed labeling based on a compa a i e
analysis e alua ing he ypes o was e, as well as image cha ac e is ics such as esolu ion,
whi e backg ounds wi hou i egula edges, uncon olled backg ounds in g een a eas,
and was e combina ions. In addi ion, se e al diffe en da ase s, including Ga bage Clas-
si ica ion and T ashne , we e used, which we e classi ied and balanced in o ino ganic,
wi h 997 images (53%), and o ganic, wi h 886 images (47%). Fo labeling, he Robo low
ool was used, manually delimi ing he a eas o in e es o he objec s o be de ec ed.
Subsequen ly, image p ep ocessing me hods we e applied o esize he images o 640 ×
640 pixels and pe o m au oma ic con as adjus men s. Addi ionally, da a augmen a ion
echniques we e implemen ed, o a ing he images a angles o +15° and −15°, and apply-
ing ho izon al and e ical lips o expand he di e si y o he da ase , (Table 1, Figu e 2).
Table 1. Collec ed da ase s.
Name No. o Ca ego ies No. o Im-
ages Anno a ion Image Type Au ho
UAVVas e 1 (ga bage was e in g een a eas) 772 Segmen a ion Was e cap u e in
open ields [25]
T ashne 6 (glass, pape , ca dboa d, plas ic, me al,
gene al was e) 2527 Classi ica ion
Clean backg ound
o whi e back-
g ound
[26]
Was e Classi i-
ca ion da a 2 (o ganic and ecyclable objec s) ~25,000 Classi ica ion Gene a ed by
Google sea ches [27]
Ga bage Classi-
ica ion
12 (ba e y, biological, b own glass, ca d-
b
oa d, clo hes, g een glass, me al, pape ,
plas ic, shoes, ash, whi e glass)
15,150 Classi ica-
ion/De ec ion
Combining he
“clo hing da ase ”
and he web sc ap-
ping ool
[28]
T ashBox 7 (medical was e, elec onic was e, plas-
ics, pape , me al, glass, ca dboa d) 17,785 Classi ica-
ion/De ec ion
Gene a ed by he
web [29]
Figu e 1. Me hodological diag am.
2.2.1. Da ase Design
The design o he da ase o 1883 images in ol ed labeling based on a compa a i e
analysis e alua ing he ypes o was e, as well as image cha ac e is ics such as esolu ion,
whi e backg ounds wi hou i egula edges, uncon olled backg ounds in g een a eas, and
was e combina ions. In addi ion, se e al di e en da ase s, including Ga bage Classi i-
ca ion and T ashne , we e used, which we e classi ied and balanced in o ino ganic, wi h
997 images (53%), and o ganic, wi h 886 images (47%). Fo labeling, he Robo low ool was
used, manually delimi ing he a eas o in e es o he objec s o be de ec ed. Subsequen ly,
image p ep ocessing me hods we e applied o esize he images o 640
×
640 pixels and
pe o m au oma ic con as adjus men s. Addi ionally, da a augmen a ion echniques we e
implemen ed, o a ing he images a angles o +15
◦
and
−
15
◦
, and applying ho izon al and
e ical lips o expand he di e si y o he da ase , (Table 1, Figu e 2).
Table 1. Collec ed da ase s.
Name No. o Ca ego ies No. o Images Anno a ion Image Type Au ho
UAVVas e 1 (ga bage was e in g een
a eas) 772 Segmen a ion Was e cap u e in
open ields [26]
T ashne
6 (glass, pape , ca dboa d,
plas ic, me al, gene al
was e)
2527 Classi ica ion
Clean backg ound
o whi e
backg ound
[27]
Was e
Classi ica ion
da a
2 (o ganic and ecyclable
objec s) ~25,000 Classi ica ion Gene a ed by
Google sea ches [28]
Ga bage
Classi ica ion
12 (ba e y, biological,
b own glass, ca dboa d,
clo hes, g een glass, me al,
pape , plas ic, shoes, ash,
whi e glass)
15,150 Classi ica ion/
De ec ion
Combining he
“clo hing da ase ”
and he web
sc apping ool
[29]
T ashBox
7 (medical was e, elec onic
was e, plas ics, pape ,
me al, glass, ca dboa d)
17,785 Classi ica ion/
De ec ion
Gene a ed by he
web [30]
Sus ainabili y 2025,17, 3523 6 o 18
Sus ainabili y 2025, 17, x FOR PEER REVIEW 6 o 18
Figu e 2. Da ase ep esen a ion buil .
2.2.2. YOLO Model T aining
Fo aining he models, he hype pa ame e s and he cus om da ase o 1883 images
we e conside ed, using 85% o aining and 15% o es ing, along wi h online da a aug-
men a ion echniques such as lipud: 0.5, lipl : 0.5, and mosaic: 1.0 (Table 2).
Table 2. T ained models.
Model Se ings: YOLO 4 Tiny Model Se ings: YOLO 7 Tiny Model Se ings: YOLO 8 Nano Model Se ings: YOLO 9 Tiny
Inpu 640 × 640 Inpu 640 × 640 Inpu 640 × 640 Inpu 640 × 640
Lea ning a e 0.002 Lea ning a e 0.001 Lea ning a e 0.001 Lea ning a e 0.001
Weigh decay 0.0002 Op imize Adam Op imize Adam Op imize Adam
Op imize Adam Momen um 0.937 Momen um 0.937 Momen um 0.937
Momen um 0.937 Ba chsize 32 Ba chsize 32 Ba chsize 32
Ba chsize 32 Subdi isions 8 Subdi isions 8 Subdi isions 8
Subdi isions 8 To al epoch 100 To al epoch 100 To al epoch 100
To al i e a ions 6000
Fo aining YOLO, he inpu image is di ided in o a g id o size, whe e each cell
gene a es a egion o in e es and p edic s he numbe o bounding boxes, along wi h i s
con idence le el in each o hem, and he numbe o model classes wi h hei p obabili ies.
The p edic ions we e encoded in a enso o size S × S (B × 5 + C), as shown in Figu e 3
[30], whe e, S is he numbe o cells in he image g id; B is he numbe o bounding boxes;
C is he numbe o model classes; and 5 is he in o ma ion associa ed wi h each bounding
box (x, y, w, h, con idence).
Figu e 2. Da ase ep esen a ion buil .
2.2.2. YOLO Model T aining
Fo aining he models, he hype pa ame e s and he cus om da ase o 1883 images
we e conside ed, using 85% o aining and 15% o es ing, along wi h online da a
augmen a ion echniques such as lipud: 0.5, lipl : 0.5, and mosaic: 1.0 (Table 2).
Table 2. T ained models.
Model Se ings:
YOLO 4 Tiny
Model Se ings:
YOLO 7 Tiny
Model Se ings:
YOLO 8 Nano
Model Se ings:
YOLO 9 Tiny
Inpu 640 ×640 Inpu 640 ×640 Inpu 640 ×640 Inpu 640 ×640
Lea ning a e 0.002 Lea ning a e 0.001 Lea ning a e 0.001 Lea ning a e 0.001
Weigh decay 0.0002 Op imize Adam Op imize Adam Op imize Adam
Op imize Adam Momen um 0.937 Momen um 0.937 Momen um 0.937
Momen um 0.937 Ba chsize 32 Ba chsize 32 Ba chsize 32
Ba chsize 32 Subdi isions 8Subdi isions 8Subdi isions 8
Subdi isions 8To al epoch 100 To al epoch 100 To al epoch 100
To al i e a ions 6000
Fo aining YOLO, he inpu image is di ided in o a g id o size, whe e each cell
gene a es a egion o in e es and p edic s he numbe o bounding boxes, along wi h i s
con idence le el in each o hem, and he numbe o model classes wi h hei p obabili ies.
The p edic ions we e encoded in a enso o size S
×
S (B
×
5 + C), as shown in Figu e 3[
31
],
whe e, S is he numbe o cells in he image g id; B is he numbe o bounding boxes; C is
he numbe o model classes; and 5 is he in o ma ion associa ed wi h each bounding box
(x, y, w, h, con idence).
The models we e adjus ed wi h hei own hype pa ame e s, and hei p ecision, ecall,
and F1-Sco e we e calcula ed (Table 3, Figu e 4).
Sus ainabili y 2025,17, 3523 7 o 18
Sus ainabili y 2025, 17, x FOR PEER REVIEW 7 o 18
Figu e 3. Rep esen a ion o YOLO model aining, adap ed om [30].
The models we e adjus ed wi h hei own hype pa ame e s, and hei p ecision, e-
call, and F1-Sco e we e calcula ed (Table 3, Figu e 4).
Figu e 4. Con usion ma ices gene a ed.
Figu e 3. Rep esen a ion o YOLO model aining, adap ed om [31].
Table 3. YOLO model calcula ions.
Model Clase Valo es P ecision Recall F1-Sco e IoU
YOLO 9 iny
Ino ganic TP = 0.97
FN = 0.19
FP = 0.04
0.97
0.97 +0.04
0.96
0.97
0.97 +0.19
0.836
2×0.96 ×0.836
0.96 +0.836
1.605
1.796
0.893
0.97
0.97 +0.19 +0.04
0.97
1.2
0.808
O ganic TP = 0.93
FN = 0.04
FP = 0.06
0.93
0.93 +0.06
0.939
0.93
0.93 +0.04
0.958
2×0.939 ×0.958
0.939 +0.958
1.799
1.897
0.948
0.93
0.93 +0.04 +0.06
0.93
1.03
0.903
Valida ion Me ics
To e alua e he pe o mance o he classi ica ion o de ec ion model, he numbe s
o T ue Posi i es (TPs), False Posi i es (FPs), T ue Nega i es (TNs), and False Nega i es
(FNs) ob ained om he con usion ma ices gene a ed in Figu e 4we e conside ed, and
calcula ed in ela ion o he ollowing:
The In e sec ion o e Union (IoU)
Each inpu image con ains bounding boxes (g ound- u h); du ing he lea ning p ocess,
new p edic ed bounding boxes a e gene a ed. When he e is an objec wi hin he a ea o
in e es , he gene a ed bounding box o e laps wi h he e e enced one, calcula ing he
union a ea wi h he ob ained con idence, ob aining a ela ionship be ween he in e sec ion
zone and he union a ea; see Equa ion (1) and Tables 3and 4[32].
IoU =TP
TP +FN +FP (1)
Sus ainabili y 2025,17, 3523 8 o 18
Sus ainabili y 2025, 17, x FOR PEER REVIEW 7 o 18
Figu e 3. Rep esen a ion o YOLO model aining, adap ed om [30].
The models we e adjus ed wi h hei own hype pa ame e s, and hei p ecision, e-
call, and F1-Sco e we e calcula ed (Table 3, Figu e 4).
Figu e 4. Con usion ma ices gene a ed.
Figu e 4. Con usion ma ices gene a ed.
Table 4. Me ics ob ained om YOLO models.
Model P ecision Recall F1-Sco e IoU
YOLO 4 iny 0.91 0.94 0.85 0.60
YOLO 7 iny 0.91 0.93 0.91 0.83
YOLO 8 nano 0.94 0.97 0.85 0.64
YOLO 9 iny 0.92 0.94 0.92 0.85
P ecision
P ecision measu es he p opo ion o co ec de ec ions among all de ec ions made. A
high p ecision alue indica es ha he model is making ew alse de ec ions; see Equa ion (2)
and Tables 3and 4[20].
P ecision =TP
TP +FP (2)
Recall
Recall measu es he p opo ion o eal objec s ha a e co ec ly de ec ed. A high ecall
alue indica es ha he model is inding mos o he objec s p esen in he images; see
Equa ion (3) and Tables 3and 4[33].
Recall =TP
TP +FN (3)
F1-sco e
Acco ding o Ul aly ics, his me ic allows o he obse a ion o he balance be ween
p ecision (P) and ecall (R), e alua ing model pe o mance in a mo e balanced way, iden i y-
Sus ainabili y 2025,17, 3523 9 o 18
ing whe he a model wo ks well ac oss all classes o whe he i is biased owa ds a speci ic
class; see Equa ion (4) and Tables 3and 4.
F1−sco e =2×P×R
P+R(4)
Ano he impo an aspec o conside in model aining is he g aphical ep esen a ion
o me ics o isualize he model’s pe o mance du ing i s aining (Figu e 5).
Sus ainabili y 2025, 17, x FOR PEER REVIEW 9 o 18
Table 4. Me ics ob ained om YOLO models.
Model
P ecision
Recall
F1-sco e
IoU
YOLO 4 iny 0.91 0.94 0.85 0.60
YOLO 7 iny
0.91
0.93
0.91
0.83
YOLO 8 nano
0.94
0.97
0.85
0.64
YOLO 9 iny
0.92
0.94
0.92
0.85
Ano he impo an aspec o conside in model aining is he g aphical ep esen a-
ion o me ics o isualize he model’s pe o mance du ing i s aining (Figu e 5).
Figu e 5. G aphs o models’ pe o mance in aining.
2.2.3. Embedded Sys em Implemen a ion
Du ing implemen a ion o he embedded sys em, he ained weigh s o each model
we e mig a ed o he Raspbe y Pi 4 Model B de ice, whe e each one was execu ed in a
Py hon 3.13.0 language en i onmen in Thonny IDE o s a model de ec ion. Expe i-
men al es s we e conduc ed wi h fi e o ganic (gene al ood was e) and ino ganic (plas-
ics, pape , glass, aluminum) was e i ems. This allowed o he iden ifica ion o hei be-
ha io in a eal en i onmen .
YOLO 4 iny: This de ec ed o ganic and ino ganic esidues wi h difficul y, due o
ligh ing and backg ound p oblems and ins abili y acco ding o he objec ’s posi ion, and
i had a Cen al P ocessing Uni (CPU) usage o a ound 80% (Figu e 6).
Figu e 5. G aphs o models’ pe o mance in aining.
2.2.3. Embedded Sys em Implemen a ion
Du ing implemen a ion o he embedded sys em, he ained weigh s o each model
we e mig a ed o he Raspbe y Pi 4 Model B de ice, whe e each one was execu ed in a
Py hon 3.13.0 language en i onmen in Thonny IDE o s a model de ec ion. Expe imen al
es s we e conduc ed wi h i e o ganic (gene al ood was e) and ino ganic (plas ics, pape ,
glass, aluminum) was e i ems. This allowed o he iden i ica ion o hei beha io in a eal
en i onmen .
YOLO 4 iny: This de ec ed o ganic and ino ganic esidues wi h di icul y, due o
ligh ing and backg ound p oblems and ins abili y acco ding o he objec ’s posi ion, and i
had a Cen al P ocessing Uni (CPU) usage o a ound 80% (Figu e 6).
Sus ainabili y 2025,17, 3523 16 o 18
included in he aining se . This di e s om ou s udy, whe e we used eal- ime images
wi h ou YOLO models in hei mos compac e sions. Ano he impo an aspec is
he e alua ion using [email p o ec ed], whe e [
38
] ob ained a maximum a e age o 0.613 in hei
YOLO 5m model, which is lowe han he esul s ob ained in his esea ch: YOLO 7- iny
= 0.949, YOLO 8-nano = 0.976, and YOLO 9- iny = 0.968. Addi ionally, e . [
38
] did no
conduc es s on an embedded sys em. Howe e , e s. [
14
–
16
] de eloped a ligh weigh
CNN model wi h an e icien algo i hm on an embedded de ice o de ec ing su ace de ec s
in indus ial p oduc s, achie ing accu acy compa able o s a e-o - he-a models; his aligns
wi h ou wo k on he implemen a ion o ligh weigh YOLO models in embedded sys ems.
Rega ding he da ase , i ag ees wi h [
2
], in which T ash-ne and Public Ga bageNe we e
also used o ain he MSW de ec ion model. I is wo h men ioning ha [
2
] e alua ed a
single model wi h ou da ase s, while in his esea ch, ou models we e e alua ed wi h
a combined da ase o T ash-ne and Public Ga bageNe . The use o a Raspbe y Pi 4B
de ice as he main boa d o objec de ec ion in embedded sys ems [
6
] is consis en wi h
his wo k. Rega ding in e ence imes, in [
6
], GCNe , MobileNe 2, and ResNe 50 we e
e alua ed, ob aining imes o 200 ms, 200 ms, and 600 ms, espec i ely, which a e di e en
om he 1800 ms ob ained in his expe imen a ion; howe e , he models in his s udy
de ec ed mo e han one ype o was e simul aneously, while hose in [
6
] only de ec ed a
single ype o was e.
5. Conclusions
YOLO e sions o objec de ec ion models equi e high-spec de ices, wi hou consid-
e ing esou ce consump ion in hei implemen a ion; hey a e commonly no applied in
embedded sys ems, such as he Raspbe y Pi 4 Model B wi h he 12.3 MP HQ came a, due
o hei limi ed CPU and GPU capabili ies. Rega ding he compa ison o he models in
hei di e en e sions, he pe o mance o objec de ec ion models has been in es iga ed.
This esea ch p esen s a compa a i e analysis o ou YOLO models es ed unde he same
en i onmen al condi ions and wi h he same da ase s, he YOLO 4- iny, YOLO 7- iny,
YOLO 8-nano, and YOLO 9- iny e sions, o de e mine each one’s pe o mance in e ms
o CPU usage, in e ence ime, FPS, p ecision, ecall, [email p o ec ed], F1-sco e, and model size.
The esul s ob ained show ha YOLO 9- iny achie ed a e y solid pe o mance in he
expe imen al es s, wi h an a e age p ecision o 92.11%, a ecall o 94.97%, an F1-sco e o
0.729, a compac size o 4.43 MB, and an absence o e o s du ing es ing, posi ioning i as
he bes model in e ms o s abili y and o e all pe o mance. Meanwhile, YOLO 7 demon-
s a ed high p ecision in ecognizing was e, bu also de ec ed non-exis en objec s, exhibi ed
high CPU usage (97%), showed dependence on he objec ’s loca ion and ligh ing, showed
weakness in de e mining dis ances wi h o ganic was e, and had long in e ence imes o
objec ype iden i ica ion. This esea ch will allow us o e alua e he beha io o ligh weigh
YOLO models applied on a Raspbe y Pi 4B de ice, wi hou elying on emo e sys ems,
o implemen a ion, o example, in an in elligen con aine o classi ying o ganic and
ino ganic was e, which would ensu e p ope managemen and con ibu e o he Sus ainable
De elopmen Goals, speci ically Goal 13: Clima e Ac ion.
Au ho Con ibu ions: Concep ualiza ion, C.M.-M., W.C.-F., C.M.-D., Y.E.-A. and D.E.G.-V.; Me hod-
ology, M.C.-B. and C.M.-D.; So wa e, D.B.R.-P., W.C.-F. and Y.E.-A.; Valida ion, M.C.-B., D.B.R.-P. and
C.M.-D.; Fo mal analysis, M.C.-B., D.B.R.-P. and D.E.G.-V.; In es iga ion, M.C.-B., D.B.R.-P., C.M.-M.
and C.V.M.-V.; Da a cu a ion, D.B.R.-P.; W i ing—o iginal d a , M.C.-B. and D.B.R.-P.; W i ing—
e iew & edi ing, M.C.-B., D.B.R.-P., C.V.M.-V. and C.M.-D.; Visualiza ion, D.E.G.-V.; Supe ision,
M.C.-B. and C.M.-M.; P ojec adminis a ion, C.M.-M. All au ho s ha e ead and ag eed o he
published e sion o he manusc ip .
Sus ainabili y 2025,17, 3523 17 o 18
Funding: This esea ch ecei ed no ex e nal unding.
Ins i u ional Re iew Boa d S a emen : No applicable.
In o med Consen S a emen : No applicable.
Da a A ailabili y S a emen : The da a associa ed wi h he s udy will be a ailable in he eposi o y
o he Na ional Technological Ins i u e o Mexico/Technological Ins i u e o Chilpancingo, h ps:
// inacional. ecnm.mx/handle/TecNM/149 (accessed on 1 Feb ua y 2025). The da a p esen ed in
his s udy a e a ailable upon eques om he co esponding au ho .
Con lic s o In e es : The au ho s decla e no con lic s o in e es .
Re e ences
1.
Jin, B.; Li, W. Spa ial E ec s and D i ing Fac o s o Consump ion Upg ades on Municipal Solid Was e Eco-E iciency, Conside ing
Emission Ou pu s. Sus ainabili y 2025,17, 2356. [C ossRe ]
2.
Zhou, Y.; Wang, Z.; Zheng, S.; Zhou, L.; Dai, L.; Luo, H.; Zhang, Z.; Sui, M. Op imiza ion o Au oma ed Ga -bage Recogni ion
Model Based on ResNe -50 and Weakly Supe ised CNN o Sus ainable U ban De elopmen . Alex. Eng. J. 2024,108, 415–427.
[C ossRe ]
3.
De Ca olis, B.; Ladogana, F.; Macchia ulo, N. YOLO T ashNe : Ga bage De ec ion in Video S eams. In P oceedings o he
2020 IEEE Con e ence on E ol ing and Adap i e In elligen Sys ems (EAIS), Ba i, I aly, 27–29 May 2020; pp. 1–7.
4.
Falasche i, L.; Manoni, L.; Palma, L.; Pie leoni, P.; Tu che i, C. Embedded Real-Time Vehicle and Pedes ian De ec ion Using a
Comp essed Tiny YOLO 3 A chi ec u e. IEEE T ans. In ell. T ansp. Sys . 2024,25, 19399–19414. [C ossRe ]
5.
He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Lea ning o Image Recogni ion. In P oceedings o he 2016 IEEE Con e ence on
Compu e Vision and Pa e n Recogni ion (CVPR), Las Vegas, NV, USA, 26 June–1 July 2016; pp. 770–778.
6.
Wang, L.; Ji, W.; Wang, G.; Feng, Y.; Du, M. In elligen Design and Op imiza ion o Exe cise Equipmen Based on Fusion Algo i hm
o YOLO 5-ResNe 50. Alex. Eng. J. 2024,104, 710–722. [C ossRe ]
7.
Chen, Z.; Yang, J.; Chen, L.; Jiao, H. Ga bage Classi ica ion Sys em Based on Imp o ed Shu leNe V2. Resou . Conse . Recycl.
2022,178, 106090. [C ossRe ]
8.
Sachin, C.; Manasa, N.; Sha ma, V.; AA, N.K. Vege able Classi ica ion Using You Only Look Once Algo i hm. In P oceedings
o he 2019 In e na ional Con e ence on Cu ing-edge Technologies in Enginee ing (ICon-CuTE), U a P adesh, India, 14–16
No embe 2019; pp. 101–107.
9.
Liu, S.; Jin, Y.; Ruan, Z.; Ma, Z.; Gao, R.; Su, Z. Real-Time De ec ion o Seedling Maize Weeds in Sus ainable Ag icul u e.
Sus ainabili y 2022,14, 15088. [C ossRe ]
10. Ul aly ics Ul aly ics Home. A ailable online: h ps://docs.ul aly ics.com/es#how-can-i- ain-a-cus om-yolo-model-on-my-
da ase (accessed on 22 Ma ch 2025).
11.
Wu, Z.; Zhang, D.; Shao, Y.; Zhang, X.; Zhang, X.; Feng, Y.; Cui, P. Using YOLO 5 o Ga bage Classi ica ion. In P oceedings o
he 2021 4 h In e na ional Con e ence on Pa e n Recogni ion and A i icial In elligence (PRAI), Yibin, China, 20–22 Augus 2021;
pp. 35–38.
12.
Zeng, M.; Lu, X.; Xu, W.; Zhou, T.; Liu, Y. PublicGa bageNe : A Deep Lea ning F amewo k o Public Ga bage Classi ica ion. In
P oceedings o he 2020 39 h Chinese Con ol Con e ence (CCC), Shenyang, China, 27–29 July 2020; pp. 7200–7205.
13.
Yang, Z.; Xia, Z.; Yang, G.; L , Y. A Ga bage Classi ica ion Me hod Based on a Small Con olu ion Neu al Ne wo k. Sus ainabili y
2022,14, 14735. [C ossRe ]
14.
Zhang, D.; Hao, X.; Wang, D.; Qin, C.; Zhao, B.; Liang, L.; Liu, W. An E icien Ligh weigh Con olu ional Neu al Ne wo k o
Indus ial Su ace De ec De ec ion. A i . In ell. Re . 2023,56, 10651–10677. [C ossRe ]
15.
Zhang, D.; Wang, Y.; Meng, L.; Yan, J.; Qin, C. Adap i e C i ic Design o Sa e y-Op imal FTC o Unknown Nonlinea Sys ems
wi h Asymme ic Cons ained-Inpu . ISA T ans. 2024,155, 309–318. [C ossRe ]
16.
Zhang, D.; Hao, X.; Liang, L.; Liu, W.; Qin, C. A No el Deep Con olu ional Neu al Ne wo k Algo i hm o Su ace De ec
De ec ion. J. Compu . Des. Eng. 2022,9, 1616–1632. [C ossRe ]
17.
Majch owska, S.; Mikołajczyk, A.; Fe lin, M.; Klawikowska, Z.; Plan ykow, M.A.; Kwasig och, A.; Majek, K. Deep Lea ning-Based
Was e De ec ion in Na u al and U ban En i onmen s. Was e Manag. 2022,138, 274–284. [C ossRe ] [PubMed]
18.
Laksono, P.W.; Anisa, A.; P iyanda i, Y. Deep Lea ning Implemen a ion Using Con olu ional Neu al Ne wo k in Ino ganic
Packaging Was e So ing. F ankl. Open 2024,8, 100146. [C ossRe ]
19.
Fu, B.; Li, S.; Wei, J.; Li, Q.; Wang, Q.; Tu, J. A No el In elligen Ga bage Classi ica ion Sys em Based on Deep Lea ning and an
Embedded Linux Sys em. IEEE Access 2021,9, 131134–131146. [C ossRe ]
Sus ainabili y 2025,17, 3523 18 o 18
20.
Gude, D.K.; Banda i, H.; Challa, A.K.R.; Tasneem, S.; Tasneem, Z.; Bha acha jee, S.B.; Lali , M.; Flo es, M.A.L.; Goyal, N.
T ans o ming U ban Sani a ion: Enhancing Sus ainabili y h ough Machine Lea ning-D i en Was e P ocessing. Sus ainabili y
2024,16, 7626. [C ossRe ]
21. Robo low. A ailable online: h ps://app. obo low.com/login (accessed on 24 Oc obe 2024).
22. YOLO 4. A ailable online: h ps://gi hub.com/AlexeyAB/da kne (accessed on 22 No embe 2024).
23. YOLO 7. A ailable online: h ps://gi hub.com/WongKinYiu/yolo 7 (accessed on 22 No embe 2024).
24. YOLO 8. A ailable online: h ps://gi hub.com/ul aly ics/ul aly ics (accessed on 24 No embe 2024).
25. YOLO 9. A ailable online: h ps://gi hub.com/WongKinYiu/yolo 9 (accessed on 24 No embe 2024).
26.
K a , M.; Piechocki, M.; P ak, B.; Walas, K. Au onomous, Onboa d Vision-Based T ash and Li e De ec ion in Low Al i ude
Ae ial Images Collec ed by an Unmanned Ae ial Vehicle. Remo e Sens. 2021,13, 965. [C ossRe ]
27. Thung, G.; Yang, M. Classi ica ion o T ash o Recyclabili y S a us; CS229; S an o d Uni e si y: S an o d, CA, USA, 2016.
28.
Sashaank Seka Was e Classi ica ion Da a. A ailable online: h ps://www.kaggle.com/da ase s/ echsash/was e-classi ica ion-
da a/da a (accessed on 9 Decembe 2024).
29.
Mos a a Mohamed Ga bage Classi ica ion (12 Classes). A ailable online: h ps://www.kaggle.com/da ase s/mos a aabla/
ga bage-classi ica ion (accessed on 9 Decembe 2024).
30.
Kumse y, N.V.; Bha Nekka e, A.; Kama h, S.S.; Kuma , M.A. T ashBox: T ash De ec ion and Classi ica ion Using Quan um
T ans e Lea ning. In P oceedings o he 2022 31s Con e ence o Open Inno a ions Associa ion (FRUCT), Helsinki, Finland,
27–29 Ap il 2022; pp. 125–130.
31.
Redmon, J.; Di ala, S.K.; Gi shick, R.B.; Fa hadi, A. You Only Look Once: Uni ied, Real-Time Objec De ec ion. In P oceedings
o he IEEE Con e ence on Compu e Vision and Pa e n Recogni ion (CVPR), Nash ille, TN, USA, 11–15 June 2015. [C ossRe ]
32.
Kim, K.; Kim, K.; Jeong, S. Applica ion o YOLO 5 and 8 o Recogni ion o Sa e y Risk Fac o s a Cons uc ion Si es.
Sus ainabili y 2023,15, 15179. [C ossRe ]
33.
Gao, X.; Wang, G.; Qi, J.; Wang, Q.; Xiang, M.; Song, K.; Zhou, Z. Imp o ed YOLO 7 o Sus ainable Ag icul u e Signi ican ly
Imp o es P ecision Ra e o Chinese Cabbage (B assica Pekinensis Rup .) Seedling Bel (CCSB) De ec ion. Sus ainabili y 2024,
16, 4759. [C ossRe ]
34.
Sallang, N.C.A.; Islam, M.T.; Islam, M.S.; A shad, H. A CNN-Based Sma Was e Managemen Sys em Using Tenso Flow Li e and
LoRa-GPS Shield in In e ne o Things En i onmen . IEEE Access 2021,9, 153560–153574. [C ossRe ]
35.
Fu, C.-Y.; Liu, W.; Ranga, A.; Tyagi, A.; Be g, A.C. DSSD: Decon olu ional Single Sho De ec o . a Xi 2017, a Xi :1701.06659.
[C ossRe ]
36. Tan, M.; Pang, R.; Le, Q. V E icien De : Scalable and E icien Objec De ec ion. a Xi 2019, a Xi :1911.09070. [C ossRe ]
37.
Lin, T.-Y.; Goyal, P.; Gi shick, R.; He, K.; Dollá , P. Focal Loss o Dense Objec De ec ion. a Xi 2017, a Xi :1708.02002. [C ossRe ]
38.
Pa el, D.; Pa el, F.; Pa el, S.; Pa el, N.; Shah, D.; Pa el, V. Ga bage De ec ion Using Ad anced Objec De ec ion Techniques.
In P oceedings o he 2021 In e na ional Con e ence on A i icial In elligence and Sma Sys ems (ICAIS), Coimba o e, India,
25–27 Ma ch 2021; pp. 526–531.
Disclaime /Publishe ’s No e: The s a emen s, opinions and da a con ained in all publica ions a e solely hose o he indi idual
au ho (s) and con ibu o (s) and no o MDPI and/o he edi o (s). MDPI and/o he edi o (s) disclaim esponsibili y o any inju y o
people o p ope y esul ing om any ideas, me hods, ins uc ions o p oduc s e e ed o in he con en .
