Research Of Face Recognition Based On Deep Learning

With the development of technology, biometric technology has become an import way of the personal identification or authentication technology, which has been a very active research topic. Face recognition, as an important branch of biometrics, which is non-invasive and most natural for users, more easily accepted for the most intuitive means of identification, and has showed a wide range of applications in information security, criminal detection, access control and other areas.Architectures are applied in many existing machine learning algorithms including neural networks with only one hidden layer, support vector machine and many others are using shallow architecture, those shallow architectures are incapable of representing the complex function. At the same time, in the past though the multilayer perception can represent complex function relation, but since these is no good learning algorithm, it often easily to fall into local optimum. Therefore, this dissertation proposed a face recognition based on the deep learning, the first adopts the method of layer-by-layer pre-training learn network initial weights, later fine adjustment network, could make networks fall into global optimum, can effectively avoid network model into the local optimum.The main works in the dissertation list as follow:(1) Application of extraction the key areas of face to the deep neural networks, the experimental results show that the key parts of the face separately extracted input to the neural network can effectively improve face recognition rate.(2) Proposed a deep neural network based on fine-tuning principal component analysis (PCA) parameter. This dissertation reduce dimension through PCA and extract feature. Last, fine-tuning PCA feature vector by back propagation algorithm, this method can make whole deep neural network more robust.(3) We have built up face recognition system based on deep learning, and designed four different experiments, comparison of recognition rate with different basic module and amount of layers, experimental results show that the effectiveness of the algorithm based on fine-tuning PCA parameter deep neural networks.Improved the basis of the original module, this dissertation proposed a deep neural networks based on fine-tune PCA parameter, can improve recognition rate.