Research on Face Recognition Algorithm Based on Improved Residual Neural Network

• DOI: 10.11648/j.acis.20210901.16
• Publication Date: 2021-07-20T07:06:07Z
• Authors: Tang Xiaolin; Wang Xiaogang; Hou Jin; Han Yiting; Huang Ye
• Keywords: Softmax function; Overfitting; Dropout (neural networks); Activation function

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