Generative Adversarial Nets for Cost-Sensitive Face Recognition

Most face recognition studies are based on standard frontal face databases, but in real life, the images we obtain are profile face images of any angle in most instances. In this case, the traditional face recognition methods cannot achieve the lowest recognition cost. Therefore, how to use the obtained profile face images to synthesize the corresponding frontal face images is important in the face recognition system. Besides, most traditional face recognition systems are try to find an accurate classifier to achieve the lowest error rate, implicitly assuming that all misclassification costs are equal. It is an unreasonable assumption because almost in all face recognition systems, different types of misclassification errors often lead to different misclassification costs. To address the two issues, we propose a cost-sensitive face recognition method based on generative adversarial nets. First, generate frontal face images using the two-channel generative adversarial nets, and then introduce cost-sensitive learning in the recognition process to consider the cost imbalance problem. The experimental results demonstrate the effectiveness of the proposed method.