Academic Edi o : Jose Na a o
Ped eño
Recei ed: 8 Ma ch 2025
Re ised: 8 Ap il 2025
Accep ed: 10 Ap il 2025
Published: 14 Ap il 2025
Ci a ion: Cas o-Bello, M.;
Rome o-Juá ez, V.M.;
Fuen es-Pacheco, J.; Mo ales-Mo ales,
C.; Ma molejo-Vega, C.V.;
Zagal-Ba e a, S.R.;
Gu ié ez-Valencia, D.E.;
Ma molejo-Dua e, C. R3sNe :
Op imized Residual Neu al Ne wo k
A chi ec u e o he Classi ica ion o
U ban Solid Was e ia Images.
Sus ainabili y 2025,17, 3502. h ps://
doi.o g/10.3390/su17083502
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
R3sNe : Op imized Residual Neu al Ne wo k A chi ec u e o
he Classi ica ion o U ban Solid Was e ia Images
Mi na Cas o-Bello 1, V. M. Rome o-Juá ez 1, J. Fuen es-Pacheco 2,* , Co nelio Mo ales-Mo ales 3,
Ca los V. Ma molejo-Vega 1,*, Se gio R. Zagal-Ba e a 1, D. E. Gu ié ez-Valencia 1
and Ca los Ma molejo-Dua e 4
1Na ional Technological Ins i u e o Mexico, Technological Ins i u e o Chilpancingo,
Chilpancingo de los B a o 39090, México; [email p o ec ed] (M.C.-B.);
[email p o ec ed] (V.M.R.-J.); [email p o ec ed] (S.R.Z.-B.);
[email p o ec ed] (D.E.G.-V.)
2Na ional Technological Ins i u e o Mexico, Na ional Cen e o Resea ch and Technological De elopmen ,
Cue na aca 62490, Mo elos, Mexico
3Na ional Technological Ins i u e o Mexico, Technological Ins i u e o San Juan del Río,
San Juan del Río Que é a o 76800, México; [email p o ec ed]
4Cen 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] (J.F.-P.); [email p o ec ed] (C.V.M.-V.);
Tel.: +52-7772328308 (J.F.-P.); +52-7472555547 (C.V.M.-V.)
Abs ac : Municipal solid was e (MSW) accumula ion is a c i ical global challenge o
socie y and go e nmen s, impac ing en i onmen al and social sus ainabili y. E icien
sepa a ion o MSW is essen ial o esou ce eco e y and ad ancing sus ainable u ban man-
agemen p ac ices. Howe e , manual classi ica ion emains a slow and ine icien p ac ice.
In esponse, ad ances in a i icial in elligence, pa icula ly in machine lea ning, o e mo e
p ecise and e icien al e na i e solu ions o op imize his p ocess. This esea ch p esen s
he de elopmen o a ligh deep neu al ne wo k called R3sNe ( h ee “Rs” o Reduce,
Reuse, and Recycle) wi h esidual modules ained end- o-end o he bina y classi ica ion
o MSW, wi h he capabili y o as e in e ence. The esul s indica e ha he combina ion
o p ocessing echniques, op imized a chi ec u e, and aining s a egies con ibu es o an
accu acy o 87% o o ganic was e and 94% o ino ganic was e. R3sNe ou pe o ms he
p e- ained ResNe 50 model by up o 6% in he classi ica ion o bo h o ganic and ino ganic
MSW, while also educing he numbe o hype pa ame e s by 98.60% and GFLOPS by
65.17% compa ed o ResNe 50. R3sNe con ibu es o sus ainabili y by imp o ing he
was e sepa a ion p ocesses, acili a ing highe ecycling a es, educing land ill dependency,
and p omo ing a ci cula economy. The model’s op imized compu a ional equi emen s
also ansla e in o lowe ene gy consump ion du ing in e ence, making i well-sui ed o
deploymen in esou ce-cons ained de ices in sma u ban en i onmen s. These ad ance-
men s suppo he ollowing Sus ainable De elopmen Goals (SDGs): SDG 11: Sus ainable
Ci ies and Communi ies, SDG 12: Responsible Consump ion and P oduc ion, and SDG 13:
Clima e Ac ion.
Keywo ds: deep lea ning; was e classi ica ion; con olu ion neu onal ne wo k; op imized
a chi ec u e; esidual ne wo ks
1. In oduc ion
In ecen yea s, indus ializa ion and u baniza ion ha e g own conside ably, gene a -
ing la ge amoun s o municipal solid was e (MSW) and making i necessa y o imp o e
Sus ainabili y 2025,17, 3502 h ps://doi.o g/10.3390/su17083502
Sus ainabili y 2025,17, 3502 2 o 20
so ing asks. Fo example, in 2020, app oxima ely 2010 million ons we e gene a ed; he
Wo ld Bank o ecas s ha by 2030, his will inc ease o 2590 million ons and 3400 million
ons by 2050. O hese, app oxima ely 40% a e deposi ed in land ills, 19% a e ecycled, and
11% a e ea ed h ough incine a ion [
1
–
3
]. The e icien sepa a ion o MSW is key o p o-
mo ing en i onmen al sus ainabili y and social well-being. Imp o ing ecycling p ocesses
educes he ecological oo p in , minimizes land ill use, and p omo es he ci cula economy,
aligning wi h global s a egies such as he Sus ainable De elopmen Goals (SDGs). This
esea ch con ibu es o SDGs 11, 12, and 13, p o iding a echnological solu ion o imp o e
MSW managemen and con ibu e o a sus ainable u u e.
Recycling asks gene ally equi e high labo cos s, so he ecogni ion and au oma ic
de ec ion o was e h ough images eplace manual so ing, hanks o ad ances in a i icial
in elligence (AI) [
4
]. Mo eo e , machine lea ning algo i hms aimed a imp o ing p ecision
in au oma ic so ing can add ess complex challenges in ecogni ion and classi ica ion.
Con olu ional neu al ne wo ks (CNNs) comp ise con olu ional laye s esponsible o
ea u e ex ac ion and a ully connec ed laye ha ac s as a classi ie , con aining many
neu ons, making hem a p edominan me hodology o image classi ica ion [4].
Va ious s udies ha e used deep lea ning (DL) models o classi y MSW. Fo example,
in [
5
], a CNN o was e classi ica ion was de eloped using DenseNe 121 and a gene ic
algo i hm o op imize he ully connec ed laye , achie ing 99.60% accu acy wi h a ne wo k
o 8 million pa ame e s. In [
6
], a was e image classi ica ion model based on DenseNe 169
and a ans e lea ning s a egy a e p esen ed, cons uc ing a da ase ea u ing a ied
backg ounds (NWNU-TRASH) and a balanced dis ibu ion. A 70:30 da a spli was used
o aining and es ing, eaching an accu acy o o e 82%.
In [
7
], a hyb id mul ilaye con olu ional neu al ne wo k wi h ew pa ame e s was
de eloped using 64
×
64-pixel images ob ained om T ashNe , achie ing an accu acy
o 92.6%. In [
3
], a ecyclable was e image classi ica ion model is p oposed ha applies
image augmen a ion echniques based on an 18-laye a chi ec u e (ResNe 18). Thei p o-
posal add esses he challenge o iden i ying and classi ying images ha p esen di e se
cha ac e is ics by op imizing he o iginal s uc u e o he esidual ne wo k, e ec i ely
in eg a ing he mos ele an ea u es om he channel maps and comp essing hem wi hin
he spa ial dimension, allowing o a mo e concise and e icien ep esen a ion o isual
in o ma ion wi h an accu acy o 95.87%. In [
8
], a con olu ional neu al ne wo k inspi ed
by VGGNe was p oposed o was e classi ica ion. The MLH-CNN was op imized wi h
ewe pa ame e s and hyb id modules, and T ashNe was used as he da ase , achie ing
an accu acy o 92.6%. In [
9
], he au ho s p oposed an Eco Cycle Classi ying Deep Neu al
Ne wo k (ECCDN-Ne ) model o bina y classi ica ion (o ganic and ino ganic) o was e
images using componen s om Densene 201 and Resne 18. They used 24,705 images,
achie ing a classi ica ion accu acy o 96.10% [
10
]. They compa ed he DenseNe 121, Mo-
bileNe , ResNe 50, and Xcep ion a chi ec u es in he T ashNe da ase , concluding ha
DenseNe 121 pe o ms be e when ine- uning, wi h a p ecision a e o 95%. In [
11
], he
au ho s p esen a VGG16 a chi ec u e o classi y di e en ypes o ga bage, ained wi h
he T ashNe da ase . They achie ed an accu acy o o e 96%. On he o he hand, in [
12
],
a CNN model is p oposed o classi ying was e. Speci ically, he Incep ion- 3 model was
ained wi h a da ase ob ained om di e en sou ces, achie ing a classi ica ion accu acy
o 92.5%.
Addi ionally, in [
13
], a model o ecyclable was e classi ica ion using deep lea ning
called RWNe was de eloped based on se e al ResNe s uc u es; hey used he T ashNe
da ase wi h image augmen a ion echniques, achie ing an accu acy o 88%. Finally, in [
14
],
a small con olu ional neu al ne wo k was p oposed o was e classi ica ion, ea u ing an
adap i e image b igh ness algo i hm o balance he backg ound b igh ness du ing he
Sus ainabili y 2025,17, 3502 3 o 20
p ep ocessing phase. They also implemen ed a h eshold eplacemen me hod designed
o mi iga e noise gene a ed by shadows. They used he Canny ope a o o c op he whi e
backg ound and educe isual in e e ence, achie ing an accu acy o 96.77% on hei
cus om da ase and 93.72% on he T ashNe da ase . The bina y classi ica ion o MSW
o e s ad an ages ega ding ecycling e iciency, cos educ ion, and p omo ion o a ci cula
economy. I is e ec i e in scena ios wi h limi ed esou ces, low le els o en i onmen al
educa ion, o high olumes o o ganic was e. Con e sely, in si ua ions whe e was e
gene a ion is low, in as uc u e is es ic ed, o managemen objec i es a e p ima ily basic,
i is no necessa y o implemen mul i-class classi ica ion sys ems, as bina y classi ica ion
can be an e icien and e ec i e solu ion.
In pa icula , he bina y classi ica ion o MSW o e s ad an ages ega ding ecycling
e iciency, cos educ ion, and p omo ing a ci cula economy. I is unc ional in scena ios
wi h limi ed esou ces, low le els o en i onmen al educa ion, o high olumes o o ganic
was e. On he o he hand, whe e was e gene a ion is low, he e is limi ed in as uc u e, o
he managemen objec i es a e p ima ily essen ial, i is no necessa y o implemen mul i-
class classi ica ion sys ems because bina y classi ica ion can be an e icien and e ec i e
solu ion. In his sense, bina y classi ica ion ep esen s an e ec i e and obus s a egy,
p o iding a c i ical basis in eal MSW managemen sys ems ha could la e be expanded
o mo e de ailed hie a chical schemes.
Unlike p e ious wo ks ha use basic con olu ional a chi ec u es ha p ocess in o -
ma ion sequen ially, such as hose in [
5
,
6
], his pape p oposes an op imized esidual
a chi ec u e. Residual ne wo ks in oduce skip connec ions ha add he inpu o one laye
o i s ou pu , p ese ing essen ial ea u es o p e ious laye s and hus p e en ing he g adi-
en s om diminishing oo much (g adien anishing p oblem). Using esidual modules
allows o designing deepe neu al ne wo ks, cap u ing mo e complex pa e ns ha a e
use ul o imp o ing hei pe o mance. On he o he hand, while da a om di e en
s udies p esen a iabili y in backg ounds and esolu ions, he da ase used he e seeks o
simpli y isual cha ac e is ics o ocus on lea ning he in insic p ope ies o was e.
This esea ch p esen s he de elopmen o an op imized a chi ec u e based on esidual
ne wo ks ained end- o-end o he bina y classi ica ion o MSW, which can pe o m
as e in e ence.
2. Ma e ials and Me hods
2.1. C ea ion o he Da ase
Da a augmen a ion is an e ec i e s a egy o aining classi ica ion models, pa ic-
ula ly o objec ecogni ion. This is because high-dimensional images p esen a ious
a ia ion ac o s, many o which can be easily eplica ed [15].
Mul iple public da abases we e consul ed o he composi ion o he was e da ase ,
such as T ashNe [
16
], Ga bage_Huawei [
17
], Was e Classi ica ion da a [
18
], Ga bage
Da ase [
19
], Ga bage Classi ica ion [
20
] and Ga bage Pic u es o Classi ica ion [
21
], as
shown in Figu e 1. The images in his da ase con ain he was e i ems agains a con olled
backg ound. This dis inc i e da ase allows he neu al ne wo k o ocus solely on he
inhe en ea u es o he was e i ems, ensu ing no bias is caused by he complex backg ounds
agains which he was e i ems a e placed. A was e classi ica ion sys em ained wi h images
agains con olled backg ounds is pa icula ly bene icial on con eyo bel s wi h cons an
ligh ing and uni o m backg ounds, as well as ecycling machines whe e use s manually
place hei was e i ems.
Sus ainabili y 2025,17, 3502 4 o 20
Sus ainabili y 2025, 17, x FOR PEER REVIEW 4 o 20
Figu e 1. Sample o images om he six da ase s used: (a) T ashNe , (b) Ga bage Huawei, (c) Ga -
bage Classi ica ion Da a, (d) Ga bage Da ase , (e) Ga bage Classi ica ion, and ( ) Ga bage Pic u es
o Classi ica ion.
Each image was inspec ed o elimina e duplica es and ensu e no epe i ions ac oss
he sou ces. Con en -based hash echniques we e used o iden i y and il e simila im-
ages. Addi ionally, he images we e ca e ully labeled in o o ganic and ino ganic ca ego-
ies.
The inal da ase includes 3208 images, di ided in o 1574 images o he o ganic
was e class and 1634 images o he ino ganic was e class, wi h a p opo ional balance
be ween bo h ca ego ies, as shown in Table 1 and Figu e 2.
To ensu e uni o mi y in p ocessing, he images we e con e ed o RGB o ma , e-
ga dless o hei o iginal o ma . This gua an eed ha each image had h ee consis en
channels, which a e essen ial o p ocessing by he neu al ne wo k a chi ec u e. The RGB
colo space was chosen because i p ese es he comple e ch oma ic in o ma ion o im-
ages, unlike g ayscale, which disca ds colo componen s. This p ese a ion o he h ee
channels ( ed, g een, and blue) p o ides he ne wo k wi h addi ional ea u es o classi i-
ca ion, enabling i o iden i y ex u es, con ou s, and colo pa e ns ha a e no pe cep ible
in a monoch oma ic image. Consequen ly, RGB images offe a ia ions in bo h ligh in-
ensi y and hue. Recen li e a u e demons a es ha main aining colo can enhance clas-
si ica ion accu acy; o ins ance, i has been epo ed ha using he RGB scheme p o ides
he mos comple e in o ma ion o he model and esul s in supe io pe o mance com-
pa ed o al e na i e spaces such as HSV, YCbC , o g ayscale images [18].
Figu e 1. Sample o images om he six da ase s used: (a) T ashNe , (b) Ga bage Huawei,
(c) Ga bage Classi ica ion Da a, (d) Ga bage Da ase , (e) Ga bage Classi ica ion, and ( ) Ga bage
Pic u es o Classi ica ion.
Each image was inspec ed o elimina e duplica es and ensu e no epe i ions ac oss
he sou ces. Con en -based hash echniques we e used o iden i y and il e simila images.
Addi ionally, he images we e ca e ully labeled in o o ganic and ino ganic ca ego ies.
The inal da ase includes 3208 images, di ided in o 1574 images o he o ganic was e
class and 1634 images o he ino ganic was e class, wi h a p opo ional balance be ween
bo h ca ego ies, as shown in Table 1and Figu e 2.
Sus ainabili y 2025,17, 3502 5 o 20
Table 1. Sou ce in o ma ion and numbe o images o each da ase .
Da ase Name Link Numbe o Images
T ashNe h ps://gi hub.com/ga y hung/ ashne , accessed on
21 Decembe 2024 993
Ga bage_HuaWei h ps://www.kaggle.com/da ase s/xiaohoua/
ga bage-huawei, accessed on 10 Decembe 2024 548
Was e Classi ica ion da a h ps://www.kaggle.com/da ase s/ echsash/was e-
classi ica ion-da a, accessed on 21 Decembe 2024 731
Ga bage Da ase h ps://www.kaggle.com/da ase s/sumn2u/ga bage-
classi ica ion- 2, accessed on 10 Decembe 2024 175
Ga bage Classi ica ion (12 classes) h ps://www.kaggle.com/da ase s/mos a aabla/
ga bage-classi ica ion, accessed on 21 Decembe 2024 275
Ga bage Pic u es o Classi ica ion
h ps://www.kaggle.com/da ase s/masco 9183/
ga bage-pic u es- o -classi ica ion, accessed on 10
Decembe 2024
339
Sus ainabili y 2025, 17, x FOR PEER REVIEW 5 o 20
Table 1. Sou ce in o ma ion and numbe o images o each da ase .
Da ase Name Link Numbe
o Images
T ashNe h ps://gi hub.com/ga y hung/ ashne , accessed on 21 Decembe 2024 993
Ga bage_HuaWei h ps://www.kaggle.com/da ase s/xiaohoua/ga bage-huawei, accessed on 10 De-
cembe 2024 548
Was e Classi ica ion da a h ps://www.kaggle.com/da ase s/ echsash/was e-classi ica ion-da a, accessed on
21 Decembe 2024 731
Ga bage Da ase h ps://www.kaggle.com/da ase s/sumn2u/ga bage-classi ica ion- 2
,
accessed on
10 Decembe 2024 175
Ga bage Classi ica ion
(12 classes)
h ps://www.kaggle.com/da ase s/mos a aabla/ga bage-classi ica ion
,
accessed on
21 Decembe 2024 275
Ga bage Pic u es o
Classi ica ion
h ps://www.kaggle.com/da ase s/masco 9183/ga bage-pic u es- o -classi ica ion
,
accessed on 10 Decembe 2024 339
Figu e 2. Samples om he da ase o e alua e he p oposed ne wo k: (a) o ganic was e and (b)
ino ganic was e.
2.2. Da a Augmen a ion
Image augmen a ion was pe o med online du ing ne wo k aining using he Im-
ageDa aGene a o (h ps://www. enso low.o g/api_docs/py hon/ /ke as/p ep o-
cessing/image/ImageDa aGene a o (accessed on 8 Ma ch 2025)) class om he deep
lea ning neu al ne wo k lib a y Ke as [22]. The da a augmen a ion echniques applied
included escaling, o a ion, zoom, wid h shi , heigh shi , shea , ho izon al lip, e ical
lip, and ill mode; he anges o hese pa ame e s a e de ailed in Table 2. The da a dis i-
bu ion was 85% o aining and 15% o alida ion, wi h a sepa a e se o images used o
es ing. The es se comp ised 569 images (bo h o ganic and ino ganic), which we e e i-
ied o ensu e ha none o hese images we e epea ed o p e iously included in he ain-
ing and alida ion se s. This ca e ul sepa a ion gua an ees ha he pe o mance epo ed
o he R3sNe model is objec i e and genuinely ep esen a i e o i s ac ual gene aliza ion
capaci y, as he e alua ion is ca ied ou exclusi ely on da a en i ely new o he model.
Figu e 2. Samples om he da ase o e alua e he p oposed ne wo k: (a) o ganic was e and
(b) ino ganic was e.
To ensu e uni o mi y in p ocessing, he images we e con e ed o RGB o ma , ega d-
less o hei o iginal o ma . This gua an eed ha each image had h ee consis en channels,
which a e essen ial o p ocessing by he neu al ne wo k a chi ec u e. The RGB colo space
was chosen because i p ese es he comple e ch oma ic in o ma ion o images, unlike
g ayscale, which disca ds colo componen s. This p ese a ion o he h ee channels ( ed,
g een, and blue) p o ides he ne wo k wi h addi ional ea u es o classi ica ion, enabling
i o iden i y ex u es, con ou s, and colo pa e ns ha a e no pe cep ible in a monoch o-
ma ic image. Consequen ly, RGB images o e a ia ions in bo h ligh in ensi y and hue.
Recen li e a u e demons a es ha main aining colo can enhance classi ica ion accu acy;
o ins ance, i has been epo ed ha using he RGB scheme p o ides he mos comple e
in o ma ion o he model and esul s in supe io pe o mance compa ed o al e na i e
spaces such as HSV, YCbC , o g ayscale images [18].
Sus ainabili y 2025,17, 3502 6 o 20
2.2. Da a Augmen a ion
Image augmen a ion was pe o med online du ing ne wo k aining using he Image-
Da aGene a o (h ps://www. enso low.o g/api_docs/py hon/ /ke as/p ep ocessing/
image/ImageDa aGene a o (accessed on 8 Ma ch 2025)) class om he deep lea ning
neu al ne wo k lib a y Ke as [
22
]. The da a augmen a ion echniques applied included
escaling, o a ion, zoom, wid h shi , heigh shi , shea , ho izon al lip, e ical lip, and
ill mode; he anges o hese pa ame e s a e de ailed in Table 2. The da a dis ibu ion was
85% o aining and 15% o alida ion, wi h a sepa a e se o images used o es ing.
The es se comp ised 569 images (bo h o ganic and ino ganic), which we e e i ied o
ensu e ha none o hese images we e epea ed o p e iously included in he aining and
alida ion se s. This ca e ul sepa a ion gua an ees ha he pe o mance epo ed o he
R3sNe model is objec i e and genuinely ep esen a i e o i s ac ual gene aliza ion capaci y,
as he e alua ion is ca ied ou exclusi ely on da a en i ely new o he model.
Table 2. Values used o da a augmen a ion echniques in he aining phase.
ImageDa aGene a o
Pa ame e Value
Rescale 1./255
Ro a ion [0, 5]
Zoom [0.0, 0.10]
Wid h shi [0.0, 0.1]
Heigh shi [0.0, 0.1]
Shea ange [0.0, 0.5]
Ho izon al lip T ue
Ve ical lip T ue
Fill mode nea es
To eed he model wi h he p ocessed images, da a gene a o s ( low_ om_da a ame)
we e used, allowing images o be loaded di ec ly om hei pa hs. These gene a o s
we e se up wi h a ba ch size o 16 images, a a ge size o 224
×
224 pixels, and bina y
classi ica ion mode o he labels. Addi ionally, he aining gene a o used image andom-
iza ion
(shu le = T ue)
, while he alida ion and es gene a o s did no employ his op ion
(shu le = False), ensu ing consis ency in he model’s e alua ion.
This p ocess ensu ed ha he da a used o ain he model we e balanced and e icien ly
p epa ed, imp o ing o e all pe o mance and he sys em’s gene aliza ion capaci y.
The ba ch size hype pa ame e was de e mined empi ically by examining he max-
imum numbe o images, along wi h hei weigh s and biases, ha could i in o he
a ailable GPU VRAM memo y o aining. By conside ing mo e examples pe ba ch, he
g adien becomes mo e ep esen a i e o he en i e da ase , acili a ing smoo he con e -
gence and educing he numbe o i e a ions pe epoch. In a small ba ch, he g adien is
compu ed wi h ewe examples pe i e a ion, esul ing in a noisy g adien ha can hinde
aining con e gence.
2.3. R3sNe Ne wo k P oposal
ResNe was in oduced in 2016 by [
22
], achie ing i s place in he ImageNe image
classi ica ion challenge. Howe e , models o his ype equi e a lo o ha dwa e esou ces,
mainly because hey a e designed o sol e a mo e complex p oblem. This leads o model
size and in e ence ime sho comings when applied o speci ic asks, such as classi ying
was e images [
23
]. These a chi ec u es’ complexi y and high numbe o pa ame e s can be
a disad an age in en i onmen s wi h limi ed compu a ional esou ces o when low-cos
and e icien deploymen is equi ed.
Sus ainabili y 2025,17, 3502 7 o 20
The c ea ion o R3sNe a chi ec u e is based on i s abili y o build deep neu al ne wo ks
wi hou acing adi ional p oblems like anishing g adien s and o e i ing (see Figu e 3).
“R3” e e s o was e managemen ’s h ee “Rs”: Reduce, Reuse, and Recycle. Likewise,
he allusion o ResNe , he base a chi ec u e wi h esidual modules, is main ained. The
ResNe in oduced by [
22
] has p o en highly e ec i e in aining deep ne wo ks, hanks o
hei skip connec ions ha acili a e g adien low du ing backp opaga ion. R3sNe is a
modi ied and ligh e e sion o he well-known ResNe 18 a chi ec u e, explici ly designed
o mee he needs o his s udy. By educing he dep h and adjus ing he s uc u e o he
esidual blocks, R3sNe main ains he compu a ional e iciency and gene aliza ion capaci y
o ResNe 18 while being op imized o bina y classi ica ion asks.
Sus ainabili y 2025, 17, x FOR PEER REVIEW 7 o 20
pe o mance in classi ying o ganic and ino ganic was e. These adap a ions a e c ucial o
deploying he model in esou ce-cons ained en i onmen s and accele a ing aining and
in e ence imes. Addi ionally, by modi ying he esidual blocks o handle ea u e dimen-
sions efficien ly, R3sNe acili a es efficien in o ma ion low h ough he ne wo k, ensu -
ing ha disc imina i e ea u es a e p ese ed and effec i ely u ilized o he inal classi i-
ca ion. This op imiza ion is essen ial o asks whe e diffe ences be ween classes can be
sub le and equi e p ecise ea u e ex ac ion, as in he case o o ganic was e, whe e he
ex ac ed ea u es include a mo e signi ican a iabili y in ex u es and colo s, e lec ing
he di e si y o ma e ials such as ood emnan s, lea es, and o he biological componen s.
On he o he hand, ino ganic was e usually exhibi s smoo h su aces wi h mo e uni o m
pa e ns and e lec ions cha ac e is ic o ma e ials like plas ics, glass, and me als.
Al hough R3sNe is inspi ed by well-known a chi ec u es such as ResNe 18, Res-
NeX , and MobileNe V2, i inco po a es speci ic modi ica ions ha op imize i s compu a-
ional efficiency and pe o mance in pa icula asks o MSW bina y classi ica ion. Unlike
ResNe 18, ou ne wo k signi ican ly educes he numbe o esidual laye s o minimize
he numbe o pa ame e s ( om 11.7 M o 0.3 M). Unlike ResNeX , nei he g ouped con-
olu ions no pa allel pa hs we e applied, bu a he simple esidual connec ions wi h
ReLU ac i a ions and ba ch no maliza ion we e p io i ized, u he simpli ying he s uc-
u e. Finally, conce ning MobileNe V2, no in e ed con olu ional blocks wi h sepa able
con olu ions we e used ins ead o s anda d esidual blocks wi h classic 3 × 3 con olu ions.
These modi ica ions we e made wi h he explici in en ion o ob aining an efficien , ligh ,
and as model ha could be deployed in de ices wi h limi ed esou ces while main ain-
ing high pe o mance in accu acy and gene aliza ion capaci y.
Below a e he main cha ac e is ics o he a chi ec u e used (see Figu e 3), including
he ke nel size 𝑘, he numbe o il e s 𝑓, he ke nel egula ize 𝑘𝑟_𝑙2, he pool size 𝑝𝑠,
and he s ide 𝑠. Table 3 shows he numbe o pa ame e s o R3sNe and he ResNe 50
ne wo ks. R3sNe has only 0.3 M pa ame e s, so i needs much less memo y and s o age
space, acili a ing i s deploymen in de ices wi h limi ed ha dwa e esou ces. Al hough
R3sNe has mo e GFLOPS han ResNe 18, i is s ill much ligh e han ResNe 50. R3sNe
equi es almos 1.5 imes mo e ope a ion han ResNe 18, bu he pa ame e s a e educed
by 97%. R3sNe main ains a manageable le el o GFLOPS, allowing a balance be ween
efficiency and compu ing capaci y.
Figu e 3. Diag am o R3sNe a chi ec u e p oposed.
Figu e 3. Diag am o R3sNe a chi ec u e p oposed.
Inspi ed by a ian s such as ResNeX [
24
] and ligh weigh a chi ec u es like Mo-
bileNe V2 [
25
], R3sNe inco po a es modi ica ions o educe he numbe o pa ame e s
and lowe compu a ional esou ce consump ion wi hou signi ican ly comp omising pe -
o mance in classi ying o ganic and ino ganic was e. These adap a ions a e c ucial o
deploying he model in esou ce-cons ained en i onmen s and accele a ing aining and
in e ence imes. Addi ionally, by modi ying he esidual blocks o handle ea u e dimen-
sions e icien ly, R3sNe acili a es e icien in o ma ion low h ough he ne wo k, ensu ing
ha disc imina i e ea u es a e p ese ed and e ec i ely u ilized o he inal classi ica ion.
This op imiza ion is essen ial o asks whe e di e ences be ween classes can be sub le
and equi e p ecise ea u e ex ac ion, as in he case o o ganic was e, whe e he ex ac ed
ea u es include a mo e signi ican a iabili y in ex u es and colo s, e lec ing he di e si y
o ma e ials such as ood emnan s, lea es, and o he biological componen s. On he o he
hand, ino ganic was e usually exhibi s smoo h su aces wi h mo e uni o m pa e ns and
e lec ions cha ac e is ic o ma e ials like plas ics, glass, and me als.
Al hough R3sNe is inspi ed by well-known a chi ec u es such as ResNe 18, ResNeX ,
and MobileNe V2, i inco po a es speci ic modi ica ions ha op imize i s compu a ional e -
iciency and pe o mance in pa icula asks o MSW bina y classi ica ion. Unlike ResNe 18,
ou ne wo k signi ican ly educes he numbe o esidual laye s o minimize he numbe o
pa ame e s ( om 11.7 M o 0.3 M). Unlike ResNeX , nei he g ouped con olu ions no pa -
allel pa hs we e applied, bu a he simple esidual connec ions wi h ReLU ac i a ions and
ba ch no maliza ion we e p io i ized, u he simpli ying he s uc u e. Finally, conce ning
Sus ainabili y 2025,17, 3502 8 o 20
MobileNe V2, no in e ed con olu ional blocks wi h sepa able con olu ions we e used
ins ead o s anda d esidual blocks wi h classic 3
×
3 con olu ions. These modi ica ions
we e made wi h he explici in en ion o ob aining an e icien , ligh , and as model ha
could be deployed in de ices wi h limi ed esou ces while main aining high pe o mance
in accu acy and gene aliza ion capaci y.
Below a e he main cha ac e is ics o he a chi ec u e used (see Figu e 3), including
he ke nel size
k
, he numbe o il e s
, he ke nel egula ize
k _l
2, he pool size
ps
,
and he s ide
s
. Table 3shows he numbe o pa ame e s o R3sNe and he ResNe 50
ne wo ks. R3sNe has only 0.3 M pa ame e s, so i needs much less memo y and s o age
space, acili a ing i s deploymen in de ices wi h limi ed ha dwa e esou ces. Al hough
R3sNe has mo e GFLOPS han ResNe 18, i is s ill much ligh e han ResNe 50. R3sNe
equi es almos 1.5 imes mo e ope a ion han ResNe 18, bu he pa ame e s a e educed
by 97%. R3sNe main ains a manageable le el o GFLOPS, allowing a balance be ween
e iciency and compu ing capaci y.
Table 3. The numbe o pa ame e s o R3sNe , ResNe 18, and ResNe 50 wi h an inpu image size o
224 ×224 pixels.
Model Pa ame e s (M) GFLOPS
R3sNe 0.3 2.70
ResNe 18 11.7 1.81
ResNe 50 25.6 7.75
•Inpu Laye
The inpu da a consis o 224
×
224-pixel images wi h h ee colo channels (RGB).
These images a e p ocessed by he ne wo k h ough a se ies o con olu ional laye s ha
ex ac ele an ea u es o he subsequen was e classi ica ion in o o ganic and ino -
ganic ca ego ies.
•Ini ial Con olu ional Laye
The ne wo k begins wi h a con olu ional laye ha uses a 7
×
7 il e , a s ide o 1,
and 16 il e s. This laye aims o ex ac ini ial spa ial ea u es om he inpu images. The
wo-dimensional con olu ion ope a ion can be ma hema ically ep esen ed as ollows:
Y[i,j]=
M−1
∑
m=0
N−1
∑
n=0
X[i+m,j+n]·W[m,n]+b(1)
whe e Y[i,j] is he ou pu alue a posi ion
(i,j)
a e con olu ion,
X[i,j]
is he inpu alue
a posi ion
(i,j)
,
W[m,n]
is he il e alue a posi ion
(m,n)
,
b
is he bias, and
M
and
N
a e
he il e ’s dimensions.
Fo he i s con olu ional laye , he ResNe app oach o using 7
×
7 il e sizes, which
cap u e la ge and mo e global pa e ns, was e ained. In he case o he images wi h
esidues used, il e s wi h a la ge ecep i e ield in he i s con olu ional laye a e inc ed-
ibly bene icial since he inpu images a e 224
×
224 ( ela i ely la ge) and con ain deb is
a di e en scales, cap u ing ex u es and edges o a ying deb is sizes. The 7
×
7 il e s
quickly educe he spa ial dimensions o he inpu , which dec eases he compu a ional cos
o he ollowing con olu ional laye s. Also, each il e has a o al o 49 weigh s, allowing
o a b oade a ie y o ini ial pa e ns o be lea ned. Then, o achie e deepe laye s wi h
mo e channels, 3
×
3 il e s a e selec ed o he deepe con olu ional laye s. These il e s
acili a e he ex ac ion o ine de ails o e ine he classi ica ion and enable a con olled
inc ease in he numbe o pa ame e s.
Sus ainabili y 2025,17, 3502 9 o 20
•Ba ch No maliza ion
A e con olu ion, ba ch no maliza ion is applied be o e he ac i a ion unc ions o a
laye o s abilize lea ning and accele a e model con e gence. This ope a ion adjus s he
ac i a ions o each ba ch so ha hey ha e a mean nea ze o and a a iance close o one,
helping o mi iga e he co a ia e shi p oblem. The ope a ion is de ined as ollows:
ˆ
xi=xi−µB
qσ2
B+ϵ
(2)
µB=1
m
m
∑
i=1
xi(3)
σ2
B=1
m
m
∑
i=1
(xi−µB)2(4)
whe e
xi
is he ac i a ion alue o inpu
i
,
µB
and
σ2
B
a e he mean and a iance o he
ba ch, espec i ely,
ϵ
is a small cons an alue o a oid di ision by ze o,
m
is he ba ch size,
and
ˆ
xi
is he no malized alue. The no maliza ion p ocess du ing aining is as ollows:
Fi s , he ba ch mean and a iance a e calcula ed o each channel o he ea u e map using
Fo mulas (3) and (4)
. Then, Fo mula (2) is used o no malize he ac i a ions, ensu ing hey
ha e a mean o 0 and a iance o 1. An addi ional ans o ma ion is applied o escale
and shi he no malized alues wi h lea nable pa ame e s
γ
and
β
o ob ain
yi
using
Fo mula (5). Du ing in e ence, only cumula i e s a is ics a e used.
yi=γˆ
xi+β(5)
•ReLU Ac i a ion Func ion
Finally, a ReLU ac i a ion unc ion in oduces non-linea i y in o he model, enabling
he ne wo k o lea n mo e complex da a ep esen a ions. I is de ined as ollows:
ReLU(x)=max(0, x)(6)
This indica es ha o any inpu
x
, he unc ion e u ns o
x
i i is g ea e han 0, and
0 i xis less han o equal o 0. Ma hema ically, his can be exp essed as ollows:
ReLU(x)=(x i x >0
0i x ≤0(7)
This unc ion is widely used in deep lea ning neu al ne wo ks because i add esses he
anishing g adien p oblem, making he model easie o ain and o en achie ing be e
pe o mance. This sequence o ope a ions (con olu ion, ba ch no maliza ion, and ReLU
ac i a ion) allows he ne wo k o ex ac and enhance signi ican ea u es om he inpu im-
ages, acili a ing he subsequen was e classi ica ion in o o ganic and ino ganic ca ego ies.
•Residual Blocks
A e he ini ial block, he model comp ises ou esidual blocks. The i s block
con ains 16 il e s, wo epe i ions, and one s ide; he second block consis s o 32 il e s,
wo epe i ions, and wo s ides; he hi d block consis s o 64 il e s, one epe i ion, and wo
s ides; and he ou h block consis s o 128 il e s, one epe i ion, and wo s ides. These
blocks enable he c ea ion o deep ne wo ks wi hou he issues associa ed wi h anishing
Sus ainabili y 2025,17, 3502 16 o 20
de ia ions (be ween 0.01 and 0.02) indica e pe o mance s abili y ac oss di e en uns,
sugges ing ha he esul s a e ep oducible and consis en . I s alida ion accu acy shows
beha io like aining, al hough i s a s om a sligh ly highe alue (0.73 a epoch 0). The
model expe iences a mode a e inc ease in he ea ly epochs, eaching 0.77–0.79 be ween
epochs 5 and 10. Subsequen ly, his me ic con e ges o alues o 0.82–0.83 in he la e
i e a ions, e lec ing a good abili y o gene alize he pa e ns lea ned du ing aining. The
closeness be ween alida ion and aining accu acy sugges s ha he model does no o e i ,
e en a e 30 epochs.
Rega ding consis ency, he s anda d de ia ions in alida ion dec ease om 0.02 o
0.03 du ing he ea ly epochs and o 0.01 owa d he end, demons a ing ha i becomes
mo e p edic able as i con e ges. The es accu acy begins a 0.76 and imp o es g adually
in he ea ly epochs, like alida ion. In he inal i e a ions, his me ic eaches alues closely
aligned o alida ion alues. The di e ence be ween he wo me ics is less han 1%,
con i ming ha he model gene alizes e ec i ely and consis en ly.
The minimum obse ed loss eaches 0.4, indica ing ha ResNe 50 s ill p esen s a sig-
ni ican ma gin o e o . This loss alue sugges s ha he model lacks comple e con idence
in i s p edic ions, as he e is a conside able p obabili y o inco ec classi ica ion. Compa ed
o R3sNe , which has a loss o less han 0.1, R3sNe pe o ms be e in accu acy and elia-
bili y by le e aging image ea u es mo e e ec i ely and achie ing g ea e gene aliza ion,
esul ing in a educed loss and highe con idence in he p edic ions.
In his s udy, 10 epe i ions we e ca ied ou , and o ensu e he obus ness and e-
liabili y o he esul s, he images in he aining and es se s we e selec ed andomly.
This app oach allowed he model’s pe o mance o be e alua ed unde di e en da a
dis ibu ions and a oided possible biases associa ed wi h a speci ic pa i ion. The bes
case among he 10 epe i ions was selec ed o he esul s, co esponding o he ins ance
ha main ained he lowes loss alue (ac oss he aining, alida ion, and es se s). This
app oach ensu es ha he epo ed esul s e lec he maximum pe o mance achie ed by
R3sNe . Addi ionally, he con usion ma ices p esen ed in Figu e 6co espond o his bes
case o R3sNe and p e ained ResNe 50, p o iding a clea and ep esen a i e iew o he
model’s abili y o classi y he images co ec ly.
Sus ainabili y 2025, 17, x FOR PEER REVIEW 16 o 20
R3sNe and p e ained ResNe 50, p o iding a clea and ep esen a i e iew o he
model’s abili y o classi y he images co ec ly.
Figu e 6. Con usion ma ix o he models (a) R3sNe and (b) p e ained ResNe 50.
4. Discussion
The classi ica ion o MSW has e ol ed signi ican ly due o ad ances in con olu ional
neu al ne wo ks (CNNs). This s udy implemen ed an op imized e sion o ResNe called
R3sNe , speci ically designed o add ess bina y was e classi ica ion wi h compu a ional
efficiency and high accu acy. Se e al ele an aspec s s and ou when compa ing he e-
sul s ob ained wi h ela ed esea ch. In e ms o pe o mance, R3sNe achie ed an accu-
acy o 87% o o ganic was e classi ica ion and 94% o ino ganic was e, su passing he
esul s ob ained by con en ional CNN models in [23], which a e aged 80.88% accu acy
in bo h classi ica ions using a ligh weigh i e-laye a chi ec u e wi h lowe esolu ion
images. On he o he hand, he au ho s o [14] implemen ed a small CNN ha achie ed
96.77% on hei cus om da ase and 93.72% on he T ashNe se ; hey employed ad anced
p ep ocessing echniques such as adap i e b igh ness algo i hms and Canny ope a o s
o backg ound c opping, which may ha e con ibu ed o imp o ed accu acy. In compa -
ison, R3sNe , by in eg a ing esidual blocks wi h sho cu connec ions, ReLU ac i a ion
unc ions, and ba ch no maliza ion, p o es o be an efficien solu ion ha mi iga es com-
mon issues such as anishing g adien s and o e i ing. Mo eo e , he Bina y Focal
C ossen opy loss unc ion effec i ely handled class imbalance.
In e ms o compu a ional efficiency, R3sNe seeks o balance pe o mance. In con-
as o models like DenseNe 121 ained o MSW [5,6] wi h many pa ame e s (o e eigh
million), R3sNe signi ican ly educes compu a ional equi emen s while main aining
compe i i e accu acy. This is especially use ul o implemen a ions in esou ce-limi ed
en i onmen s like IoT de ices. Re . [23] used a da ase aken om [18] ha con ained
25,077 images, which was one o he da ase s also used in his s udy; howe e , duplica e
images we e iden i ied and emo ed du ing he da ase cu a ion phase. This s ep is c ucial
o a oid aining biases and ensu e he model pe o mance calcula ion is ep esen a i e
and eliable. Fu he mo e, a consis en e alua ion using a speci ic da a combina ion al-
lowed o mo e obus and gene alizable esul s.
Al hough he olume o da a used is ela i ely limi ed, i was conside ed sufficien
o effec i ely ain he R3sNe model because o he s a egies implemen ed o maximize
sample di e si y. Online da a augmen a ion echniques and duplica e emo al enhanced
Figu e 6. Con usion ma ix o he models (a) R3sNe and (b) p e ained ResNe 50.
Sus ainabili y 2025,17, 3502 17 o 20
4. Discussion
The classi ica ion o MSW has e ol ed signi ican ly due o ad ances in con olu ional
neu al ne wo ks (CNNs). This s udy implemen ed an op imized e sion o ResNe called
R3sNe , speci ically designed o add ess bina y was e classi ica ion wi h compu a ional
e iciency and high accu acy. Se e al ele an aspec s s and ou when compa ing he esul s
ob ained wi h ela ed esea ch. In e ms o pe o mance, R3sNe achie ed an accu acy o
87% o o ganic was e classi ica ion and 94% o ino ganic was e, su passing he esul s
ob ained by con en ional CNN models in [
23
], which a e aged 80.88% accu acy in bo h
classi ica ions using a ligh weigh i e-laye a chi ec u e wi h lowe esolu ion images. On
he o he hand, he au ho s o [
14
] implemen ed a small CNN ha achie ed 96.77% on hei
cus om da ase and 93.72% on he T ashNe se ; hey employed ad anced p ep ocessing
echniques such as adap i e b igh ness algo i hms and Canny ope a o s o backg ound
c opping, which may ha e con ibu ed o imp o ed accu acy. In compa ison, R3sNe , by
in eg a ing esidual blocks wi h sho cu connec ions, ReLU ac i a ion unc ions, and ba ch
no maliza ion, p o es o be an e icien solu ion ha mi iga es common issues such as
anishing g adien s and o e i ing. Mo eo e , he Bina y Focal C ossen opy loss unc ion
e ec i ely handled class imbalance.
In e ms o compu a ional e iciency, R3sNe seeks o balance pe o mance. In con as
o models like DenseNe 121 ained o MSW [
5
,
6
] wi h many pa ame e s (o e eigh
million), R3sNe signi ican ly educes compu a ional equi emen s while main aining
compe i i e accu acy. This is especially use ul o implemen a ions in esou ce-limi ed
en i onmen s like IoT de ices. Re . [
23
] used a da ase aken om [
18
] ha con ained
25,077 images, which was one o he da ase s also used in his s udy; howe e , duplica e
images we e iden i ied and emo ed du ing he da ase cu a ion phase. This s ep is c ucial
o a oid aining biases and ensu e he model pe o mance calcula ion is ep esen a i e and
eliable. Fu he mo e, a consis en e alua ion using a speci ic da a combina ion allowed
o mo e obus and gene alizable esul s.
Al hough he olume o da a used is ela i ely limi ed, i was conside ed su icien
o e ec i ely ain he R3sNe model because o he s a egies implemen ed o maximize
sample di e si y. Online da a augmen a ion echniques and duplica e emo al enhanced
he a iabili y o he aining se wi hou necessi a ing addi ional da a collec ion. These
s a egies inc eased he compelling di e si y o examples submi ed o he model, mi -
iga ing he limi a ions a ising om ewe samples and educing he isk o o e i ing.
Online da a augmen a ion applies new ans o ma ions o he o iginal images, gene a ing
a ia ions a each aining epoch and ensu ing di e en images a each i e a ion. Thus,
R3sNe cap u ed a mo e obus se o o e all ea u es. A mo e ex ensi e da ase could
u he imp o e he pe o mance and gene alizabili y o he model, as a mo e signi ican
numbe o examples ends o co e he a iabili y o he esiduals in a mo e complex
manne . Howe e , he esul s indica e ha he model achie ed compe i i e pe o mance
despi e he limi ed size. This sugges s ha he amoun o a ailable da a en iched by da a
augmen a ion and duplica e cleaning was adequa e o ain R3sNe e ec i ely, achie ing a
sa is ac o y balance be ween model complexi y and he in o ma ion p o ided.
T adi ional con olu ional a chi ec u es wi h many pa ame e s o en o e i when
access o limi ed da ase s is a ailable. Public da ase s o MSW classi ica ion a e equen ly
ela i ely small; o ins ance, a ound 25,000 images a e usually a ailable o bina y asks
like di e en ia ing be ween o ganic and ino ganic. This s ands in s a k con as o ex en-
si e da ase s like ImageNe . T aining ne wo ks wi h millions o pa ame e s om sc a ch
in hese si ua ions can esul in models memo izing he aining da a a he han lea n-
ing gene alizable pa e ns. While con en ional models such as VGG-16, ResNe 50, o
Dense-Ne -121 ha e exhibi ed high accu acy in o he asks, hey comp ise many laye s
Sus ainabili y 2025,17, 3502 18 o 20
and pa ame e s (wi h VGG-16 con aining app oxima ely 138 million pa ame e s) and e-
qui e subs an ial compu ing powe . Consequen ly, hey a e unsui able o applica ions in
esou ce-cons ained en i onmen s such as embedded de ices. E en a emp s o educe o
eeze laye s in p e- ained models o en yield bulky models ha ad e sely a ec s o age
and ope a ional e iciency.
One limi a ion o he cu en s udy is ha he model was es ed on only wo classes
(o ganic and ino ganic), which migh no en i ely e lec i s po en ial in a mul i-class
scena io. The e o e, u u e wo k is planned o ex end he analysis o a mul iclass p oblem,
which may equi e adjus men s in he model a chi ec u e and hype pa ame e uning o
op imize i s pe o mance in mo e complex asks.
R3sNe exhibi s excep ional accu acy, ea u ing swi ini ial lea ning, s abiliza ion
du ing he inal epochs, and s ong gene aliza ion capabili ies. The op imized a chi ec u e
and app op ia e aining con igu a ion enhance hese esul s, making i a p ac ical choice
o MSW classi ica ion asks.
The esul s demons a e R3sNe ’s po en ial o e icien and accu a e MSW classi ica-
ion. Howe e , u u e esea ch could ocus on alida ing he model wi h mo e ex ensi e
and di e se da ase s and inco po a ing ad anced p ep ocessing echniques o imp o e i s
pe o mance, e.g., analyzing esidues in images wi h complex backg ounds and a iable
ligh ing condi ions by applying ans e lea ning using ou p e- ained ne wo k.
R3sNe has excellen po en ial o eal-wo ld was e so ing. I s ligh weigh design
allows i o be in eg a ed in o au oma ic so ing sys ems (con eyo bel s o ecycling
machines), educing cos s and imp o ing was e sepa a ion e iciency. This could educe he
amoun o was e in land ills and suppo ci cula economy objec i es, especially in egions
wi h limi ed esou ces. Howe e , challenges such as up on ha dwa e cos s (came as
and o he senso s) and en i onmen al a ia ions could a ec i s pe o mance, equi ing
op imiza ions such as da a collec ion unde di e se condi ions.
Compa ed o ecen s udies, R3sNe aligns wi h he end owa d mo e e icien models.
Fo example, he wo k by [
27
] used MobileNe wi h ans e lea ning and achie ed 93.35%
accu acy on 2400 images. Ou model shows lowe accu acy, possibly due o he lack o
ans e lea ning. Re . [
27
] highligh s he e ec i eness o eusing p e- ained weigh s,
sugges ing ha R3sNe could bene i om his echnique o imp o e i s pe o mance,
especially wi h small da ase s.
Ou model pe o mance is sligh ly lowe han ha o [
28
], who achie ed an accu acy
o 93.28% in classi ying o ganic and ecyclable was e using an imp o ed con olu ional
neu al ne wo k a chi ec u e. This di e ence could be due o he di e en op imiza ion and
egula iza ion s a egies employed. While [
28
] used deep con olu ional laye s, LeakyReLU,
and d opou , ou app oach does no inco po a e hese imp o emen s, which limi s ea u e
ex ac ion. Fu he mo e, he g ea e quan i y and quali y o da a used by [
28
] (25,077 im-
ages, 70% aining, 30% es ) sugges s ha a la ge o mo e balanced da ase could imp o e
he pe o mance o ou a chi ec u e. Inco po a ing modi ica ions such as LeakyReLU o
adjus ed d opou a es could bene i ou p oposal, especially in scena ios wi h high a i-
abili y o was e. Al hough ou accu acy is sligh ly lowe , R3sNe is op imized o e iciency
wi h signi ican ly ewe pa ame e s, making i sui able o en i onmen s wi h limi ed
compu a ional esou ces. In conclusion, R3sNe o e s an e icien and scalable solu ion
o MSW bina y classi ica ion, ou pe o ming hea ie models such as ResNe 50 and being
compe i i e agains ecen app oaches; while he imp o ed DCNN o [
28
] p io i izes accu-
acy in mo e demanding con ex s o eal-wo ld applica ions, R3sNe is ideal in low-cos
en i onmen s and DCNN could be p e e ed whe e e o educ ion is a p io i y. Fu u e
esea ch could combine he e iciency o R3sNe wi h he accu acy o models such as DCNN,
op imizing i s implemen a ion in was e managemen sys ems unde a iable condi ions.
Sus ainabili y 2025,17, 3502 19 o 20
5. Conclusions
This s udy p esen s an op imized a chi ec u e based on ResNe , called R3sNe , which
is explici ly designed o bina y MSW classi ica ion. The esul s highligh he p oposed
model’s high pe o mance, achie ing an accu acy o 87% in o ganic was e classi ica ion
and 94% in ino ganic was e. These me ics su pass hose o ResNe 50 and demons a e ha
a ela i ely small a chi ec u e ea u ing esidual blocks, ReLU ac i a ion unc ions, ba ch
no maliza ion, and he Bina y Focal C ossen opy loss unc ion signi ican ly con ibu es o
he model’s pe o mance and s abili y.
The R3sNe a chi ec u e o e s compe i i e accu acy and op imizes he use o compu a-
ional esou ces, making i sui able o deploymen in ha dwa e-cons ained en i onmen s,
such as IoT de ices. Addi ionally, i s design acili a es gene aliza ion and minimizes com-
mon issues in deep ne wo ks, such as o e i ing and anishing g adien s. Compa ed o
ResNe 50, R3sNe educes hype pa ame e s by 98.60% and GFLOPS by 65.17%.
I is impo an o emphasize ha he e icien design o R3sNe con ibu es o sus-
ainabili y in se e al ways. Fi s , i educes he bu den on land ills, p omo es a ci cula
economy, and enables mo e e ec i e was e so ing. Second, he educed compu a ional
demand is e lec ed in ene gy consump ion, making i sui able o implemen a ion in en i-
onmen s wi h ha dwa e limi a ions, such as IoT de ices used in sma ci y applica ions.
These con ibu ions indica e ha he esea ch con ibu es o SDGs 11, 12, and 13, ein o cing
he model’s po en ial o sus ainable u ban de elopmen .
In conclusion, R3sNe is a signi ican ad ance in he in elligen managemen o MSW.
I o e s a solu ion o he challenges ha a ise, ensu ing ha echnological de elopmen
goes hand in hand wi h sus ainable de elopmen .
Au ho Con ibu ions: Concep ualiza ion, M.C.-B., J.F.-P. and V.M.R.-J.; me hodology, J.F.-P., M.C.-B.
and V.M.R.-J.; so wa e, D.E.G.-V. and S.R.Z.-B.; alida ion, M.C.-B., J.F.-P., V.M.R.-J. and C.M.-D.;
o mal analysis, J.F.-P., C.M.-M. and C.M.-D.; in es iga ion, V.M.R.-J. and J.F.-P.; esou ces, V.M.R.-J.,
J.F.-P. and C.M.-M.; da a cu a ion, V.M.R.-J. and J.F.-P.; w i ing—o iginal d a p epa a ion, V.M.R.-J.,
M.C.-B. and J.F.-P.; w i ing— e iew and edi ing, J.F.-P., V.M.R.-J. and M.C.-B.; isualiza ion, C.M.-D.
and C.V.M.-V.; supe ision, C.V.M.-V.; p ojec adminis a ion, M.C.-B., J.F.-P. and V.M.R.-J. All au ho s
ha e ead and ag eed o he published e sion o he manusc ip .
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 ase can be ob ained om he ollowing link: h ps://gi hub.
com/ ic o MRJDe /da ase R3sNe , accessed on 9 Feb ua y 2025.
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.
Kaza, S.; Yao, L.C.; Bhada-Ta a, P.; Van Woe den, F. Wha a Was e 2.0: A Global Snapsho o Solid Was e Managemen o 2050; Wo ld
Bank: Washing on, DC, USA, 2018. [C ossRe ]
2.
Lin, K.; Zhao, Y.; Kuo, J.H.; Deng, H.; Cui, F.; Zhang, Z.; Zhang, M.; Zhao, C.; Gao, X.; Zhou, T.; e al. Towa d Sma e Managemen
and Reco e y o Municipal Solid Was e: A C i ical Re iew on Deep Lea ning App oaches. J. Clean. P od. 2022,346, 130943.
[C ossRe ]
3.
Zhang, Q.; Zhang, X.; Mu, X.; Wang, Z.; Tian, R.; Wang, X.; Liu, X. Recyclable Was e Image Recogni ion Based on Deep Lea ning.
Resou . Conse . Recycl. 2021,171, 105636. [C ossRe ]
4.
Liu, X.; Zhou, Q.; Zhao, J.; Shen, H.; Xiong, X. Faul Diagnosis o Ro a ing Machine y unde Noisy En i onmen Condi ions
Based on a 1-D Con olu ional Au oencode and 1-D Con olu ional Neu al Ne wo k. Senso s 2019,19, 972. [C ossRe ] [PubMed]
5.
Mao, W.L.; Chen, W.C.; Wang, C.T.; Lin, Y.H. Recycling Was e Classi ica ion Using Op imized Con olu ional Neu al Ne wo k.
Resou . Conse . Recycl. 2021,164, 105132. [C ossRe ]
Sus ainabili y 2025,17, 3502 20 o 20
6.
Zhang, Q.; Yang, Q.; Zhang, X.; Bao, Q.; Su, J.; Liu, X. Was e Image Classi ica ion Based on T ans e Lea ning and Con olu ional
Neu al Ne wo k. Was e Manag. 2021,135, 150–157. [C ossRe ] [PubMed]
7.
Malik, M.; Sha ma, S.; Uddin, M.; Chen, C.-L.; Wu, C.-M.; Soni, P.; Chaudha y, S. Was e Classi ica ion o Sus ainable De elopmen
Using Image Recogni ion wi h Deep Lea ning Neu al Ne wo k Models. Sus ainabili y 2022,14, 7222. [C ossRe ]
8.
Shi, C.; Tan, C.; Wang, T.; Wang, L. A Was e Classi ica ion Me hod Based on a Mul ilaye Hyb id Con olu ion Neu al Ne wo k.
Appl. Sci. 2021,11, 8572. [C ossRe ]
9.
Bin Islam, S.; Sumon, S.I.; Majid, M.E.; Kashem, S.B.A.; Nashba , M.; Ash a , A.; Khandaka , A.; Kunju, A.K.A.; Hasan-Zia, M.;
Chowdhu y, M.E. ECCDN-Ne : A Deep Lea ning-Based Technique o E icien O ganic and Recyclable Was e Classi ica ion.
Was e Manag. 2025,193, 363–375. [C ossRe ] [PubMed]
10.
Keskin, E.F.; ˙
I¸sleyen, O.; Demi han, H.; Keskin, ¸S.R. T ash Classi ica ion Using Deep Lea ning Models. In P oceedings o he 2023
Cong ess in Compu e Science, Compu e Enginee ing, & Applied Compu ing (CSCE), Las Vegas, NV, USA, 24–27 July 2023;
pp. 318–323.
11.
Abdu, H.; Noo , M.H.M. Domes ic T ash Classi ica ion wi h T ans e Lea ning Using VGG16. In P oceedings o he 2022 IEEE
12 h In e na ional Con e ence on Con ol Sys em, Compu ing and Enginee ing (ICCSCE), Penang, Malaysia, 21–22 Oc obe 2022;
pp. 137–141.
12.
Azis, F.A.; Suhaimi, H.; Abas, E. Was e Classi ica ion Using Con olu ional Neu al Ne wo k. In P oceedings o he 2020 2nd
In e na ional Con e ence on In o ma ion Technology and Compu e Communica ions, Kuala Lumpu , Malaysia, 12–14 Augus
2020; Associa ion o Compu ing Machine y: New Yo k, NY, USA, 2020; pp. 9–13.
13.
Lin, K.; Zhao, Y.; Gao, X.; Zhang, M.; Zhao, C.; Peng, L.; Zhang, Q.; Zhou, T. Applying a Deep Residual Ne wo k Coupling wi h
T ans e Lea ning o Recyclable Was e So ing. En i on. Sci. Pollu . Res. 2022,29, 91081–91095. [C ossRe ] [PubMed]
14.
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 ]
15. Good ellow, I.; Bengio, Y.; Cou ille, A. Deep Lea ning; MIT P ess: Camb idge, MA, USA, 2016.
16.
Thung, G. Ga y hung/T ashne : Da ase o Images o T ash; To ch-Based CNN o Ga bage Image Classi ica ion. A ailable
online: h ps://gi hub.com/ga y hung/ ashne ? ab= eadme-o - ile (accessed on 20 Decembe 2024).
17.
Xiaohoua. Ga bage_HuaWei. A ailable online: h ps://www.kaggle.com/da ase s/xiaohoua/ga bage-huawei (accessed
on 9 Decembe 2024).
18.
Seka , S. 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
(accessed on 20 Decembe 2024).
19. Kunwa , S. Ga bage Da ase . A ailable online: h ps://www.kaggle.com/ds /10182596 (accessed on 9 Decembe 2024).
20.
Mohamed, M. 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 20 Decembe 2024).
21.
Ga bage Pic u es o Classi ica ion. A ailable online: h ps://www.kaggle.com/da ase s/masco 9183/ga bage-pic u es- o -
classi ica ion?selec =Ga bage+classi ica ion (accessed on 9 Decembe 2024).
22.
He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Lea ning o Image Recogni ion. In P oceedings o he 2016 IEEE Compu e
Socie y Con e ence on Compu e Vision and Pa e n Recogni ion, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778. [C ossRe ]
23.
Nnamoko, N.; Ba owclough, J.; P oc e , J. Solid Was e Image Classi ica ion Using Deep Con olu ional Neu al Ne wo k.
In as uc u es 2022,7, 47. [C ossRe ]
24.
Xie, S.; Gi shick, R.; Dolla , P.; Tu, Z.; He, K. Agg ega ed Residual T ans o ma ions o Deep Neu al Ne wo ks. In P oceedings o
he IEEE Con e ence on Compu e Vision and Pa e n Recogni ion (CVPR), Honolulu, HI, USA, 21–26 July 2017.
25.
Sandle , M.; Howa d, A.; Zhu, M.; Zhmogino , A.; Chen, L.-C. MobileNe V2: In e ed Residuals and Linea Bo lenecks. In P o-
ceedings o he IEEE Con e ence on Compu e Vision and Pa e n Recogni ion (CVPR), Sal Lake Ci y, UT, USA,
18–23 June 2018
.
26.
Lin, T.-Y.; Goyal, P.; Gi shick, R.; He, K.; Dollá , P. Focal Loss o Dense Objec De ec ion. In P oceedings o he 2017 IEEE
In e na ional Con e ence on Compu e Vision (ICCV), Venice, I aly, 22–29 Oc obe 2017; pp. 2999–3007.
27.
Suganda Gi sang, A.; P a ama, H.; San o Agus inus, L.P. Classi ica ion O ganic and Ino ganic Was e wi h Con olu ional Neu al
Ne wo k Using Deep Lea ning. In . J. In ell. Sys . Appl. Eng. 2023,11, 343–348.
28.
Chhab a, M.; Sha an, B.; Elba achi, M.; Kuma , M. In elligen Was e Classi ica ion App oach Based on Imp o ed Mul i-Laye ed
Con olu ional Neu al Ne wo k. Mul imed. Tools Appl. 2024,83, 84095–84120. [C ossRe ]
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 .
