Research And Application Of Face Liveness Detection Based On Multi-feature Fusion

With the development of the comprehensive national strength and the progress of science,face recognition technology is developing rapidly under the trend of artificial intelligence.Face recognition technology is widely used in online identity authentication due to its directness,friendliness and convenience.It is important to quantitatively analyze individual differences and self-stability through image processing,thus realizing the identification of biological individuals.However,there are also great challenges.The identity authentication system based on face recognition is vulnerable to be attacked under the complex and changeable cases,including photo attack,video attack and 3D mask attack.Nowadays,unmanned supermarkets and secret-free payment have become common in society.Therefore,liveness detection in face recognition has turned into a crucial research subject.This paper focuses on the security of face recognition.Based on different feature extraction methods,this paper studies the differences of true face and fake face images in terms of texture,structure and depth.This research contents of this paper mianly include the following aspects.(1)We develop a face liveness detection algorithm based on texture feature fusion.This algorithm contains multi-modal information of face images,such as structure,texture and spatial location,thus increasing the differences of texture features between true and fake faces.We fuse the different types of features to train a discriminate classifier,improving the accuracy in the discrimination stage.(2)We propose a face liveness detection algorithm combining dynamic texture features and depth textures.A shallow convolution neural network model is designed,which combines the differences of texture information and depth features.This makes the feature representation forms of images in different regions become different.And the representation errors of probability distribution are used to distinguish between true and fake face images,which not only increase the efficiency of model detection,but also improve the accuracy of face liveness detection.Compared with traditional and deep learning methods,the accuracy of face presentation attack detection is improved.(3)We adopt a face liveness detection algorithm based on multilevel fusion network with Laplacian embedding.Based on the geometric meaning of Laplacian matrix(graph theory),the distance constraints between inter-classes and intra-classes are embedded in the shallow convolutional neural network.In addition,the multilevel fusion structure is introduced to supervise the intermediate decision by using the original features.That makes the network model not only becoming discriminative,but also maintaining the local neighborhood structure of the original images,significantly inceasing the accuracy of face presentation attack detection algorithm.(4)We design a face presentation attack detection system.Its main function is to detect face presentation attacks based on the proposed liveness detection algorithm.And the basic function modules mainly include image capture,image saving,feature extraction,model training,face anti-spoofing detection,comparison of algorithm accuracy and so on.