Research On Deep Learning Based Face Recognition

As one of the most practical technologies in the field of biometrics,face recognition technology has been the focus of many researchers.In recent years,relying on deep learning,face recognition technology has been widely used in many fields such as finance and public safety.The thesis has studied some state-of-the-art deep face recognitions by exploring model tuning,code acceleration and face reconstruction,The main work and contribution are reminded as follows:First,Deep learning model is difficult to obtain global optimal solutions,because it has complex solution space and the limitations of current optimization methods.To solve this problem,the thesis uses the DSD training method and the loss function Triplet Loss to fine-tune the parameters of the existing deep face model.While increasing the robustness of the model,better face recognition accuracy is achieved.The experimental results show that the accuracy of the two methods on the LFW and Megaface is increased by 1.66% and 0.4% respectively.Second,The deep learning model generally relies on the CUDA environment to accelerate,but the application of the original code will encounter problems such as excessive memory loss and the acceleration effect can not reach the theory,so we optimized the code of Caffe framework.we optimized the common network layers such as BN layer and Prelu layer,and accelerated the depth separable convolution.The experimental results show that the memory loss of the BN layer is reduced by one-half,the memory loss of the Prelu layer is reduced by three-quarters,and the speed of the deep separable convolution structure is increased by more than five times.Third,Due to solve the problem that the existing face recognition model is sensitive to light and image quality,The thesis design a new network X-GAN to reconstruct the image which will be sent to the network,and mitigate the problem of light and image quality from the perspective of vision and model.