Deep Face Recognition with Cosine Boundary Softmax Loss

To improve the accuracy of face recognition when there are wrong-labeled samples, a new deep face recognition model with cosine boundary loss is proposed in this paper. First, the proposed model uses the cosine similarity to determine the boundary that divides training samples into easy samples, semi-hard samples and harder samples, which play different roles during the training process. Then, an adaptive weighted piecewise loss function is developed to emphasize semi-hard samples and suppress wrong-labeled samples in harder samples by assigning different weights to related types of samples during different training stages. Compared with the state-of-the-art face recognition methods, i.e., CosFace, CurricularFace, and EnhanceFace, experimental results on CFP_FF, CFP_FP, AgeDB, LFW, CALFW, CPLFW, VGG2_FP datasets demonstrate that the proposed method can effectively reduce the impact of the wrong-labeled samples and provide a better accuracy.