Improving Performance of a Pre-trained ResNet-50 Based VGGFace Recognition System by Utilizing Retraining as a Heuristic Step

Deep learning has remodeled the research aspect of facial recognition throughout the last decade by utilizing multiple processing layers to extract significant facial features. Although this emerging technology has achieved high performance for the face recognition problems, the dilemma of achieving low performance while training with a few samples per class has not been resolved yet. In this study, it has been shown that by utilizing retraining as a heuristic step, ResNet-50 based VGGFace architecture can enhance the performance of the face recognition scheme significantly. Multi-task Cascaded Convolutional Neural Networks have been utilized to crop faces first. The first training phase was completed by considering train samples from a combined dataset of 5-celebrity dataset, Georgia tech database, and three variants of KomNet datasets. The retraining of individual datasets further produced 94.41% test accuracy for the KomNet social media dataset and 100% test accuracy for the other four datasets.