Human Gender Identification from Facial Images: A Deep Learning Approach

Abstract

These in turn find broad applications in security, human-computer interaction, targeted marketing, and social analytics. In this work, we propose a new hybrid deep architecture that reduces the complexity and improves the accuracy of gender classification. Three different models are proposed and tested, namely: (1) a pre-trained EfficientNetB2 model with a classifier on top performed with an average accuracy of 96.73%, thus proving to have strong capability in feature extraction and classification; (2) a hybrid architecture that combines MobileNetV2 and LSTM to leverage the benefits of sequential learning with an average performance of 80% which may indeed capture the temporal dependencies in facial representations; and (3) CNN-LSTM, which combines convolutional feature extraction with sequential processing, giving an average accuracy of 85%. Our comparative analysis with existing methodologies reveals that the proposed model using EfficientNetB2 outperforms conventional approaches in terms of classification accuracy as well as in terms of computational efficiency, and hence is more suitable for real-world deployment scenarios with resource constraints. Moreover, many experiments has been conducted on benchmark datasets which justifies the robustness, generalizability of our proposed models. This presents more significant insights about the development of efficient and high-performing gender recognition systems based on deep learning from facial images, thereby contributing to the advancement of research in facial analysis.