Research Of Face Anti-Spoofing Method Based On Convolution Neural Network

In the past several years,with the development of science and technology,biometric is becoming a hot research topic.Among then,biometric based on face is widely used in various scenes due to its particularity,but some videos and photos related to the faces are easily forged or maliciously stolen by others,the face recognition system faces a series of security issues,and the face anti-spoofing technology emerges at the moment.Face anti-spoofing is mainly used to determine whether the acquired face is liveness or not.This technology is used to enhance the security of the system.For the face anti-spoofing,there are problems such as susceptibility to the external environment,the diversity of deception types,the single detection mode,the author made in-depth research on it,and the main work content are as follows:(1)Aiming at the problems of unstable face alignment,complex lighting,complex network architecture and feature design in traditional face anti-spoofing,a method combining brightness equalization and convolutional neural networks is proposed.This method fully reflects the advantages of convolutional neural networks that don't require manual design features and the feature that brightness equalization algorithm can perform brightness compensation in different brightness areas of the face.At the same time,the MTCNN using multiple network cascades is used to achieve face alignment and accurate positioning.The accuracy of the final experiment reaches99.32%,which is improved by 7% on the original basis.It proves that the proposed method can effectively distinguish face liveness and not face liveness such as photos and videos,and the accuracy of the detection has certain advantages compared with traditional methods.(2)Aiming at the problem of single input mode and many training parameters,and the detection performance is limited in convolutional neural network,a method of combining multiple modal input and mobilenet V2 is proposed.This method replaces single mode input with three modes of HSV,depth map,and infrared map to improve the diversity of features.At the same time,the advantages of the inverted residual block and the squeeze and excitation module in the mobilenet V2 network are fully utilized,reducing the number of parameters and improving the effectiveness of the features,and finally add global average pooling in the M-CNN network to prevent overfitting.In the final experiment,the average classification error rate ACER reaches 0.4%,having a significant reduction.It proves that the proposed face anti-spoofing based on M-CNN network has better performance.(3)Using tensorflow and Pytorch two deep learning frameworks,and a face antispoofing system based on the two algorithms is developed on pycharm.The system can choose one of the two algorithms for face anti-spoofing.The system has the functions of image import,model selection and face live detection.