JoWUA

Abstract

Facial recognition technology plays a crucial role in modern biometric authentication, surveillance systems, and secure access control applications. However, real-world deployment remains challenging due to variations in illumination, background noise, facial expressions, and other environmental distortions that often reduce recognition accuracy. Most of the prevailing methods perceive face recognition as a pixel-based issue and make extensive use of traditional convolutional neural networks or handcrafted descriptors, which often are not able to realize the intricate interactions between various regions of the faces. This paper will propose a deep learning architecture to overcome these constraints by combining the method of multivariate features extraction with a stacked attention-based stacked Long Short-Term Memory (LSTM) architecture. The implementations of the proposed approach would involve a normalization of Z scores to equalize the pixel intensity of the facial images and then Independent Component Analysis (ICA) to identify statistically independent and discriminating facial features, which include edges, contours and textures. The latter are then handled with a stacked attention LSTM model, which is learned to operate at the sequential level and consequently selectively attends and concentrates on important parts of the face such as the eyes, nose and mouth and attenuates background noise. The framework was tested on two evaluation datasets of CelebA and CASIA-WebFace that consist of large-scale facial images with various variations. Experimental results demonstrate that the proposed system achieves recognition rates of 96% on CelebA and 92% on CASIA-WebFace, with an RMSE of 34.82, indicating improved robustness and generalisation compared with conventional deep learning models. These findings confirm the effectiveness of the proposed approach for reliable and scalable facial recognition applications.