Research On Face Recognition Based On Deep Learning

As one of the most popular research topics in pattern recognition and computer vision,face recognition has broad application prospects in the fields of finance,security,and business.With the rapid development of deep learning and the improvement of computer hardware performance,the face recognition ability of machine has begun to exceed the human eye level.However,how to better use deep learning technology to further improve the performance of face recognition model is still a problem.Because network structure and loss function used in natural image recognition in deep learning is not suitable for training face recognition model,this thesis has deeply explored the influence of face recognition network and loss function on recognition performance.The main research contents are divided into the following parts:(1)Discussion on two face detection algorithms.The Multi-task Cascaded Convolutional Networks(MTCNN)algorithm calculates face classification,bounding box,and five key points(two eyes,nose,and mouth corners)regression by cascading three sub-networks.Because Single Stage Headless Face Detector(SSH)algorithm has characteristic of fast speed,low memory consumption and no distortion of scale,this thesis uses three detection modules with stride of 8,16,32 for large,medium and tiny face detection.The experimental results show that these algorithms can get good detection results in both simple and complex backgrounds.(2)Research on deep face network and facial feature extraction layer.Based on the theory of deep learning,the classic and the latest neural network structure are analyzed.The network based on natural image classification is not suitable for face recognition tasks.This paper improves the network structure that can simultaneously recognize the recognition speed and recognition accuracy.Then the influence of face feature extraction layer on recognition performance is discussed,by comparing advantages and disadvantages of inner product layer and global average pooling layer and finally the face extraction feature layer with the inner product layer is used.The experimental results show that it can effectively enhance the facial feature expression ability and achieve the effect of improving face recognition rate.(3)Research on loss function.The research analyzes various loss functions suchas Softmax Loss,Center Loss and Angular Softmax Loss which are difficult to train,converge and operate for face recognition.The improved loss function is proposed,which can shrink the inter-class distance and increase extra-class distance.Feature normalization and weight normalization are introduced.The experimental results show that this method makes the face recognition network training simpler and can better improve the face recognition rate.(4)The construction of face recognition system.Based on the deep learning framework TensorFlow,PyQt5 is used as the interface development tool to integrate the related algorithms of face detection,facial feature extraction,classification and recognition.The face recognition system built in this paper can get excellent recognition effect under the posture change or occlusion.