Performance Analysis of Deep Learning based Face Recognition Model for Video Surveillance

Abstract

Face detection, face alignment, and face recognition processes are all included in a full-fledged face recognition system. Due to many unconstrained elements, such as pose fluctuation, lighting, aging, partial occlusion, and low resolution, facial identification has proven to be difficult. A resilient and robust solution is currently required because existing face recognition approaches are only partially effective at identifying faces in unrestricted contexts. Therefore, the authors suggested a modified VGG19 model for face recognition using deep transfer learning. The model is trained to utilize the two-phase training approach employing a one-cycle policy's differential learning rate to enhance the recognition accuracy of the model. State-of-the-art (SOTA) competent outcomes are produced in the suggested rectification by concatenating global max pooling and global average pooling, batch normalization, and dropout layers in the classification layers. The experiments are conducted on three benchmark datasets such as Georgia Tech (GT) Face database, Labelled Faces in the Wild (LFW), and YouTube Faces (YTF) to illustrate the supremacy of the proposed work. The proposed work achieved 99.6 % recognition accuracy on the GT face database, approximately 1.35% improvement of recognition accuracy in comparison to other deep learning models on LFW, and 100% recognition accuracy on YTF.