Age De ec ion Using Deep Lea ning Techniques: A
Compa a i e S udy o CNN, VGG19, and ResNe
Md. Mehedi Hassan, Md. Mah uzu Rahman
Depa men o Compu e Science and Enginee ing,
Da odil In e na ional Uni e si y, Dhaka, Bangladesh
Co esponding Au ho : mehedi15-113[email p o ec ed]
Abs ac
This esea ch ocuses on de eloping an au oma ed sys em o es ima e human age using acial images
based on deep lea ning algo i hms. The s udy compa es he pe o mance o h ee majo a chi ec u es:
Con olu ional Neu al Ne wo k (CNN), VGG19, and ResNe . A da ase consis ing o 13,300 acial
images was collec ed om open-sou ce pla o ms such as Kaggle and Google, di ided in o nine
dis inc age ca ego ies. A e ex ensi e p ep ocessing and aining using op imize s Adam and
RMSP op, he VGG19 model achie ed he highes accu acy o 88.24%, ollowed by CNN (88.14%)
and ResNe (82.63%). These esul s demons a e ha deep con olu ional a chi ec u es a e e ec i e in
acial-based age es ima ion. This s udy con ibu es a compa a i e e alua ion amewo k and
highligh s he impo ance o da a di e si y and model selec ion in deep lea ning-based age p edic ion.
Keywo ds
Deep Lea ning, CNN, VGG19, ResNe , Age De ec ion, Facial Recogni ion
1. In oduc ion
The abili y o es ima e human age om acial ea u es has become a signi ican a ea o esea ch
wi hin compu e ision and a i icial in elligence. Wi h he apid ad ancemen o machine lea ning,
age de ec ion sys ems can now assis in nume ous applica ions including biome ic e i ica ion,
secu i y sys ems, and social media mode a ion. Despi e ex ensi e s udies, challenges emain due o
a ia ions in acial exp essions, ligh ing, and o he en i onmen al ac o s. This esea ch explo es
mul iple deep lea ning a chi ec u es o add ess hese challenges and imp o e he accu acy o acial-
based age es ima ion.
2. Rela ed Wo k
Se e al p e ious s udies ha e explo ed age es ima ion using acial ea u es. Dehshibi and Bas an a d
(2010) p oposed a me hod achie ing 86.64% accu acy on a limi ed da ase . Alonso e al. (2016)
u ilized he FG-NET and Adience da ase s, a aining compa able esul s. Kwon and Lobo (1993)
in oduced ea ly age classi ica ion me hods using w inkle analysis, while Ho ng e al. (2001)
ca ego ized 230 images in o ou age g oups wi h 81.58% accu acy. Recen s udies applying
con olu ional a chi ec u es ha e imp o ed pe o mance, bu limi a ions emain due o da ase
cons ain s and model gene aliza ion. Ou s udy expands upon his by in eg a ing la ge da ase s and
es ing mul iple deep lea ning a chi ec u es.
3. Me hodology
This s udy used a da ase o 13,300 acial images om open-sou ce eposi o ies such as Kaggle and
Google. Images we e di ided in o nine age g oups anging om 0 o 100 yea s. The da ase was spli
in o 60% o aining, 10% o es ing, and 30% o alida ion. Da a p ep ocessing included esizing,
augmen a ion, and no maliza ion. Th ee deep lea ning models we e implemen ed—CNN, VGG19,
and ResNe 50—each ained using Tenso Flow wi h op imize s Adam and RMSP op on Google
Colab GPU. The pe o mance was e alua ed using me ics including p ecision, ecall, F1-sco e, and
o e all accu acy.
4. Resul s and Discussion
Expe imen al analysis showed ha he VGG19 model achie ed he bes accu acy among he es ed
models, eaching 88.24% wi h RMSP op op imize . The CNN model pe o med closely, achie ing
88.14% accu acy wi h RMSP op and 85.47% wi h Adam. ResNe yielded 82.63% accu acy,
indica ing compa a i ely lowe gene aliza ion. The indings sugges ha deepe a chi ec u es wi h
ans e lea ning capabili ies like VGG19 ou pe o m shallowe models o complex acial da ase s.
Howe e , model pe o mance also depended on p ep ocessing quali y and he op imize used du ing
aining.
5. Conclusion and Fu u e Wo k
This esea ch demons a ed ha deep lea ning-based a chi ec u es can achie e high accu acy in age
es ima ion om acial images. Among he compa ed models, VGG19 p o ed mos e ec i e,
highligh ing he s eng h o deep con olu ional ea u e ex ac ion. Fu u e esea ch will ocus on
inc easing da ase di e si y, in eg a ing acial exp ession ecogni ion, and op imizing compu a ional
e iciency. In addi ion, de eloping a eal- ime age es ima ion applica ion could u he enhance
p ac ical usabili y in biome ic and social applica ions.
Re e ences
[1] M.M. Dehshibi and A. Bas an a d, 'A new algo i hm o age ecogni ion om acial images,'
Signal P ocessing, 2010.
[2] J.B. Alonso e al., 'Au oma ic age de ec ion based on acial images,' IEEE CCIS, 2016.
[3] Y.H. Kwon and N.V. Lobo, 'Loca ing acial ea u es o age classi ica ion,' SPIE, 1993.
[4] W.B. Ho ng, C.P. Lee, and C.W. Chen, 'Classi ica ion o age g oups based on acial ea u es,'
Tamkang Jou nal o Science and Enginee ing, 2001.
[5] A.L. Ingole and K.J. Ka ande, 'Au oma ic age es ima ion om ace images using acial ea u es,'
IEEE GCWCN, 2018.
