Deep compact discriminative representation for unconstrained face recognition

Convolutional Neural Network has been widely used in pattern recognition community, especially face recognition. Loss function, as a supervisory signal to learn a CNN model, plays an important role in obtaining the desired facial features. However, how to design a loss function to make the features more compact and discriminative for unconstrained face recognition, is still an open problem. In this paper, we propose two novel loss functions, Compact Discriminative loss and Advanced Compact Discriminative loss. They supervise CNN to map the raw data onto the face feature space, where the intra-class space is compact and inter-class spaces have sensible gaps, by constraining the intra-class variations and the inter-class variations simultaneously. Three CNNs (i.e. LeNet, CNN-M and ResNet-50) are used to analyze the effectiveness of the proposed approaches, the obtained models are evaluated on several famous benchmark databases, such as MNIST, LFW, FGLFW, YTF and IJB-A. Experimental results show that the proposed losses are effective for face recognition, and can easily generate comparable results than related state-of-the-art methods.