Face Recognition Based On Lightweight Neural Networks

Face recognition technology is one of the most widely used applications in biometrics.The traditional face recognition technology usually consists of two modules:acquisition and processing.The acquisition is done by the camera deployed in the corresponding site,and the processing is performed by powerful computing nodes on the server side.In this mode,usually one server needs to support multiple cameras inputs,and with the current increasing of cameras,relying on the central server is hard to scale.At the same time,the face recognition technology currently used is generally built on the framework of deep learning.Although the deep learning network has strong fitting ability,it also faces the problem of large amount of computing.How to reduce the computation of face recognition technology based on deep learning and deploys it to end devices becomes an urgent problem to be solved.Considering the amount of computation and running memory,the main research topic of this thesis is to improve the performance of the recognition system as much as possible.On the issue of the huge amount of computation for deep learning,the current academic community is actively exploring and designing lightweight networks.Since deep learning for computer vision problems mainly depends on convolutional networks,and the calculation of convolution operations accounts for the absolute dominantness of the network.How to reduce the amount of convolution computation of operations is the main research direction of lightweight networks.At present,there is a certain lightweight networks for object classification.In this thesis,based on these lightweight networks,we will explore the use of them in face recognition scenarios and improve it.The direct training lightweight convolution networks will make the performance of the model much worse compared with large models because of its small model capacity and high training difficulty.Focusing on this phenomenon,two scheme,knowledge distillation and deep mutual learning,are used to reduce the training difficulty of the model and increase the generalization ability of the model for the face recognition task.The knowledge-based distillation scheme requires us to train a large network in advance and transfer the knowledge from the large network to lightweight networks to reduce the difficulty of small networks learning directly from the training set's labels.Deep mutual learning requires us to build multiple small networks and use them to learn from each other to compensate themselves.In this thesis,based on the two schemes we improve the recognition accuracy of lightweight networks by 2% on multiple benchmarks.In face recognition tasks,the recognition of large poses has been an insurmountable problem.For lightweight networks,from frontal face recognition to profile face recognition,performance degradation is much greater than for larger networks.Aiming at the problem of profile face recognition of lightweight networks,a multi-task generative adversarial network framework based on WGAN-GP is proposed.By utilizing this network framework and the distillation,we have improve the 2% performance of the multiple large pose benchmarks.