Development of Robust Face Recognition System using Transfer Learning and Fine Tuning

Abstract

Face recognition using deep learning method has achieved exceptional results in the past few years. Face recognition by convolution neural network (CNN) using Deep learning techniques is very complex in nature mainly due to size of the dataset and the prerequisite of high-performance computing power for dataset training and testing. Sometimes, deep learning-based methods fail to recognize a person if dataset size is small. As a consequence, transfer learning and fine-tuning can be used to solve this problem. After outcomes of learning are transferred, implementation of different applications can be done based on the deep learning model's pre-trained weights, which reduces training time with significant improvement in accuracy. Therefore, in this study, we have explored the development of a face recognition system in real time based on deep learning using transfer learning and fine-tuning. For this purpose, we train ResNet18 deep architecture by the Caltech face database. The architecture is fine-tuned by freezing all internal layers except the fully connected layer and modifying the number to class in the SoftMax layer. Precision, recall, f1-score, and accuracy are used to determine the model's performance. Results are analysed by plotting the ROC curve and precision-recall curve.