Low Resolution Face Recognition Based on Contrast Representa Distillation

Face recognition is a challenging problem in computer vision, but most mature face recognition algorithms focus on the high resolution (HR) case and suffer from severe performance degradation at Low-resolution (LR), owing to the difficulty for getting information. In this paper, we propose a training method based on contrast representation distillation (CRD). The proposed method firstly feeds the features extracted from the HR and LR into two linear layers, allowing both of them to be mapped into a uniform space. Based on the idea of contrast learning, the cosine similarity is used as a loss to increase the similarity of the same person at different resolutions and decrease the similarity of different people. Experiments on real and simulated datasets with different resolutions have been used to verify the effectiveness of the method, all experimental results achieved SOTA.