Data Augmentation for Face Recognition with CNN Transfer Learning

Face recognition is a challenging Computer Vision task. In this paper, we propose a method for face recognition by applying data augmentation and transfer learning in pre-trained Convolutional Neural Networks (CNNs). Our main focus is to analyze the power of data augmentation for face recognition systems with CNN transfer learning. We extracted features from the images and trained a KNN classifier using the VGG-Face CNN. For the input dataset, we applied several transformations to generate 12 different versions of datasets used to evaluate which combination produces better results. We ran experiments using data augmentation on the LFW dataset. We also created a proprietary dataset composed of 12 subjects. The tests have shown that the classifier trained with the dataset Saturation presented the best results with an accuracy of 98.43%. For the proprietary dataset, the best accuracy was 95.41% obtained with the Brightness, Contrast, and Saturation Combined version. In this way, we conclude that the data augmentation is essential for the face recognition task. However, a study needs to be performed for each application, since the augmentation operations that improve the classification results are dependent on the input set. Furthermore, we show that with a few samples available, it is possible to have a face recognition system with high accuracy.