Enhanced Face Recognition System Integrating GAN-Augmented CNN Features with ICA and Vision Transformer

Abstract

Facial recognition is a high-technology biometric method that is extensively applied as identity verification, surveillance, and a security system. This paper proposes a hybrid deep learning-based model that can contribute to improved face recognition accuracy and efficiency. This is done by first augmenting the datasets through Generative Adversarial Networks (GANs) that train more synthetic face images to incorporate more and better diversity to the dataset and to reduce overfitting, after which the dataset is fed into a Convolutional Neural Network (CNN) to automatically learn and extract deep spatial features of the augmented images. These extracted features are then narrowed down using Independent Component Analysis (ICA) to select the most important features, removing redundant and irrelevant information. The optimized features are then forwarded to a Vision Transformer (ViT) to be classified by the transformer architecture that takes good consideration of spatial relationships to accurately determine individual faces. Performance evaluation metrics of the proposed system include accuracy of 0.93%, precision of 0.938%, recall of 0.930%, and F1-score of 0.9319%, which show that the proposed system has better recognition performance and strength than the conventional face recognition methods.