Research Information System
INTERNATIONAL GRADUATE RESEARCH SYMPOSIUM - IGRS’23, İstanbul, Turkey, 2 - 05 May 2023, pp.1
A Study on Gender and Age Recognition Using VGGFace
Murat
Sekmen1, Kemal
Bıçakcı1
1 Istanbul Technical University, Institute of Informatics
sekmenmu20@itu.edu.tr, kemalbicakci@itu.edu.tr
ABSTRACT
Among cybersecurity access and authorisation solutions, the use of
"password" alone is insufficient. Among the supporting solutions,
facial recognition stands out for its personalisation and practicality for the
user. However, facial recognition technology also has different application
areas. The aim of this study is to test whether VGGFace derived models for
gender and age recognition can achieve the same success on a dataset prepared
by us.
In our study, we created a face dataset specific to Turkey, as detailed
age information in Bogazici Face Database (the only open source Turkish face
database) was not available. Our dataset is divided into two parts, studio
(constrained) and free (unconstrained) environment, with approximately 5
selfies taken by 183 and 849 participants respectively, resulting in a
male-female selfie distribution of 744-385 and 3795-1236. The photos were taken
with the same mobile phone and their dimensions, lighting, distance and detail
parameters were varied.
The code shared by Serengil used gender and age models, which are
derivatives of the VGGFace model. The photos were first subjected to face
detection and alignment processes using RetinaFace; the detected faces were
framed and parts outside the frame were discarded to reduce noise in the photo.
The frame dimensions were reduced to 224x224, which is the size accepted by
VGGFace. In the resizing process, the frame is reduced proportionally with the
long edge, and in the next step, the short edge is filled with black dots.
The correct prediction rate for males and females in the studio
environment was 100% and 69% respectively, while in the free environment it was
99.82% and 89.16%. As the model had a false-negative-rate for females of 30.9%
and 10.84% in the studio and outdoor environments respectively, it was
concluded that the model did not have a tendency to classify as
"male".
When examined independently of environment, the precision and
sensitivity of female predictions were recorded as 84.39% and 92.22%
respectively, while these rates were 94.71% and 99.85% for male predictions.
The mean age of the dataset is 22 years. The average absolute error in
age estimation was 8.87 overall, while it was 7.59 and 9.17 in the studio and
outdoor environments, respectively. It was noted that the results of our tests
were lower than the general average of 4.65 in Serengil's study.
In the scenario where we directly shortened the edges of the frame,
regardless of the proportions, performance increased by an average of 1 point.
Therefore, it is possible to improve performance on our dataset with additional
work. On the other hand, we can say that we have obtained some evidence that a
model may not give the same result in every dataset.
The local data collected as part of the project also includes digitally
stored biometric photos taken from our participants' ID cards. In the next
stage of our study, we plan to compare these biometric photos with selfie
photos to improve our results.
This study was funded by the Scientific Research Projects Coordination
Unit of Istanbul Technical University (İTÜ BAP). Project No 43647.
Keywords: Age recognition; Gender recognition; Cyber
security; Database