Architectural Innovations in CNNs for Robust Face Recognition Across Varied Lighting and Poses: A Comparative Performance Analysis

Abstract

Face recognition (FR) is a fundamental task in computer vision with applications in security, healthcare, and human–computer interaction. Although convolutional neural networks (CNNs) have significantly advanced FR performance, existing systems remain highly sensitive to variations in illumination, pose, and image quality. Moreover, reliance on benchmark datasets alone often limits generalizability to real-world conditions. In this work, a customized lightweight CNN architecture was designed to enhance recognition accuracy under diverse lighting and pose variations. The approach integrates both the Labeled Faces in the Wild (LFW) dataset and a locally collected dataset, ensuring evaluation under benchmark and real-world conditions. A robust preprocessing pipeline—including cropping, normalization, and augmentation—further strengthens the model’s generalization. To avoid undertraining, model optimization was guided by validation loss and early stopping rather than fixed epoch counts. Experimental results show that the proposed model achieves 99.67% accuracy on the local dataset and 93.33% accuracy on LFW, with a compact model size of only 117 MB. In addition, the proposed CNN requires 33.63M parameters and 0.73 GFLOPs, which is substantially lower than ResNet101 (42M, 3.27 GFLOPs) and VGG-16 (134M, 3.09 GFLOPs), highlighting its efficiency in terms of both model size and computational complexity. Compared with state-of-the-art (SOTA) architectures such as ResNet101, GoogLeNet and VGG-16, the customized CNN delivers a favorable trade-off between accuracy, efficiency, and computational complexity. These results demonstrate that carefully designed lightweight CNNs, when combined with local and public datasets, can achieve robust face recognition in unconstrained environments, making them suitable for deployment in resource-limited real-world applications.