The residual neural network is prone to two problems when it is used in the process of face recognition: the first is "overfitting", and the other is the slow or non-convergence problem of the loss function of the network in the later stage of training. In this paper, in order to solve the problem of "overfitting", this paper increases the number of training samples by adding Gaussian noise and salt and pepper noise to the original image to achieve the purpose of enhancing the data, and then we added "dropout" to the network, which can improve the generalization ability of the network. In addition, we have improved the loss function and optimization algorithm of the network. After analyzing the three loss functions of Softmax, center, and triplet, we consider their advantages and disadvantages, and propose a joint loss function. Then, for the optimization algorithm that is widely used through the network at present, that is the Adam algorithm, although its convergence speed is relatively fast, but the convergence results are not necessarily satisfactory. According to the characteristics of the sample iteration of the convolutional neural network during the training process, in this paper, the memory factor and momentum ideas are introduced into the Adam optimization algorithm. This can increase the speed of network convergence and improve the effect of convergence. Finally, this paper conducted simulation experiments on the data-enhanced ORL face database and Yale face database, which proved the feasibility of the method proposed in this paper. Finally, this paper compares the time-consuming and power consumption of network training before and after the improvement on the CMU_PIE database, and comprehensively analyzes their performance.
展开▼
