Deep Face Recognition Based on Penalty Cosface.

Shuoyan Lin,Jianxiong Tang,Zhanxiang Feng,Jianhuang Lai

DOI: https://doi.org/10.1007/978-3-030-60639-8_38

2020-01-01

Abstract:Face recognition has achieved great progress because of the advancement of deep convolutional neural networks (CNNs) techniques. The Softmax loss is one of the most popular loss function for deep learning models. In many situations, face images are captured in the unconstrained environments with changing poses and illuminations, making face recognition very challenging because of the dramatic appearance variations. The models trained with the Softmax loss may fail to extract discriminative information for the face images with extreme illumination or pose conditions. Recently, Cosface has been proven effective for improving the generalization ability of the Softmax loss. Derived from Cosface, we propose a novel method named Penalty Cosface to address the unconstrained face recognition challenges and learn discriminative features. Specifically, we design a variant of Cosface that remove radial variations by penalizing \\(\\ell _2\\)-normalized constraints of the features and weights. Therefore, the discriminative ability of the Penalty Cosface is guaranteed by the large margin of the Cosface, and the penalty term is beneficial to simplifying the gradient calculations. Experimental results show that the Penalty Cosface improves the discriminative power of deep networks and outperforms the other variants of Softmax loss.