Gender recognition is a significant subfield in artificial intelligence and can be used in many applications, for example: security systems, suggestions, and biometric identification. This work evaluates the performance of MobileNet and ViT architectures for gender classification using real world and synthetic datasets. From the experimental results, we find that the overall performance of MobileNet is outstanding in real-life applications with the highest accuracy, recall, and F1 score due to feature extraction with its lightweight architecture. On the other hand, ViT is powerful, but it needs a lot of training data to fine-tune its performance. Various Explainable AI (XAI) techniques, including Grad CAM and LIME, were utilized to improve model interpretability by shedding light on the rationale behind the decision-making process and highlighting the most salient features that contribute to the classification process. These results indicate that MobileNet is suitable for real-time applications, while ViT shows potential for optimization through fine tuning and additional data. Future studies should also investigate hybrid models, strategies for mitigating bias, and improved explainability for fairness and reliability in gender classification systems.
Saini et al. (Sat,) studied this question.