A modified contrastive loss method for face recognition

Contrastive Loss is frequently used as loss function in CNN for face recognition, but it can result in the overfitting and low sampling efficiency for the positive samples. In this paper, a Modified Contrastive Loss (MCL) is proposed to overcome the shortcomings of contrastive loss. MCL and ResNet are combined with a Joint Bayesian technique to develop a ResNet-Modified Contrastive Loss-Joint Bayesian (ResNet-MCL-JB) model. First, ResNet is used as the basic network structure, and several ResNets are trained to use the MCL. Then, the ResNet with the Joint Bayesian for metric learning is integrated. The state-of-the-art performance of ResNet-MCL-JB attests to its effect. For further improvement, a Progressive Soft Filter Pruning method (PSFP) is applied in the neural network. PSFP can effectively diminish the size of the network while maintaining high accuracy. This method gradually prunes the filters on each layer by the weight of each filter. We combine the MCL and PSFP together with ResNet, and thus, we achieve considerable much improvement in both accuracy and computational cost.