Abstract:
Automated gender and age estimation from facial images are important for many realworld
applications. Although, several studies have been proposed in the past, most of them are
proposed as individual models and a considerable performance gap is noticed. Moreover, deep
learning based approaches treated their model as a black box classifier and hence their model’s
knowledge representation is not understandable and difficult to further improve. In this
manuscript, we have proposed a simple and efficient CNN model architecture by considering
gender and age estimation as a multi-label classification problem. The proposed model is trained
and then evaluated on the publicly available Adience benchmark dataset. Experimental results
demonstrated that the proposed model showed better performance than the similar approaches
with an accuracy of 84.20 % on gender estimation and an accuracy of 57.60 % on age estimation.
In addition, we have proposed a visualization technique to explain the classification results and
then the gender-specific and age group-specific landmark facial regions are identified.