Abstract: Softmax loss is usually used in convolution neural networks for feature learning, but features obtained are not discriminative enough in some cases. Traditional methods for trying to solve this problem come with additional computational complexity. A face recognition method for integrating the nested residual convolution neural network and angle metric is proposed. First, a novel feature extraction network based on the nested residual block is designed to extract various features by integrating feature maps. Then, a method of angular metric based on weight normalization is utilized. The discrimination of features is enhanced by normalizing the weights of the last fully connected layer. The learned features can satisfy the condition that the maximum intra-class distance is less than the minimum inter-class distance by combining two methods mentioned above for training. Experimental results indicate that this method leads to an accuracy of 99.03% on the LFW(Labeled Faces in the Wild). The proposed algorithm only contains a single network and provides a higher accuracy and a lower computational cost than those methods using softmax loss and other metric learning.

Key words: nested residual convolutional neural networks, weight normalization, angular metric, face recognition