Authors: Bharti Saxena, Rupali Chaure, Ashish Chourey, Mohit Singh Tomar
Abstract: Here we introduce an empirical exploration of a real-time Hybrid Deep Learning model for Age and Gender Recognition (HDL-AGR) based on facial images collected from multiple unconstrained scenarios. Estimate age and gender from facial images is a classic computer vision problem with applications ranging from human-computer interaction, intelligent surveillance, personalized marketing to healthcare screening. Most existing approaches are limited by low accuracy on far-side age groups, extreme sensitivity to lighting and occlusion, and extreme computational overhead that would preclude real-time deployment. The proposed HDL-AGR framework consists of a backbone (which has been defined as a modified EfficientNet-B4 convolutional base), attention module (Transformer-based), and an output head (dual-branch, trained jointly for age regression and gender classification) to be tuned up to date. The model is trained and evaluated with five benchmark datasets UTKFace, IMDB-WIKI, Adience, CACD and Fair Face containing over the 845K annotated images. Empirical results: HDL-AGR achieves. (i) A new state-of-the-art Mean Absolute Error (MAE) of 3.94 years in age estimation, along with an unprecedented gender classification accuracy of 97.2% and (ii) Operates at an inference speed of 54 frames per second on standard GPU hardware – outperforming all compared peer methods in the process. The contribution of each architectural component is confirmed through ablation studies. Conclusion: Our results identify HDL-AGR as a strong, efficient, and practically deployable approach for online recognition of facial attributes.
DOI: http://doi.org/10.5281/zenodo.20792037
