Likelihood Ratio based Loss to finetune CNNs for Very Low Resolution Face Verification

Abstract

In this paper, we propose a likelihood ratio based loss for very low-resolution face verification. Existing loss functions either improve the softmax loss to learn large-margin facial features or impose Euclidean margin constraints between image pairs. These methods are proved to be better than traditional softmax, but fail to guarantee the best discrimination features. Therefore, we propose a loss function based on likelihood ratio classifier, an optimal classifier in Neyman-Pearson sense, to give the highest verification rate at a given false accept rate, which is suitable for biometrics verification. To verify the efficacy of the proposed loss function, we apply it to address the very low-resolution face recognition problem. We conduct extensive experiments on the challenging SCface dataset with the resolution of the faces to be recognized below 16 × 16. The results show that the proposed approach outperforms state-of-the-art methods.