Face Patches Designed Through Neuroevolution for Face Recognition With Large Pose Variation

Face Recognition (FR) has been a widely used biometric technique for identity authentication in various domains. Despite the remarkable progress in the field of FR during the past few years, there are still challenges that need to be addressed, including recognition of faces with large pose variation. Following the idea of organic face patches found in and used by human and macaque monkey brains, in this paper we report a set of convolutional neural networks (CNNs) that we have designed to represent face patches specialized in a particular face pose orientation range. We use neuroevolution, employing genetic algorithms (GAs), to define the structure of three CNNs. Each CNN is evolved through a genetic algorithm, to a particular face pose orientation range for small, medium, and large face rotations. Each CNN was evolved to improve FR accuracy using the training partition of the VGGFace2 dataset, split into three sets for small, medium, and large face rotations. The GAs generated new specialized CNNs tuned to the three face orientation ranges: small, medium, and large face rotations. Following standard procedures, using extended training for a longer number of epochs, the best CNN individuals were trained with VGGFace2 (training set), and MS1M datasets. The performance of our proposed method was assessed on several datasets that contain a significant number of faces with large pose variation: VGGFace2 (test set), CPFLW, and CFP_FP. The results reached accuracies greater than those of the state-of-the-art of 95.91%, 95.73%, 94.60% and 99.18% on VGGFace2(test), VGGFace2_FP, CPFLW and CFP_FP, respectively.