Face Liveness Detection Based On Nonlinear Diffusion And Convolution Neural Network

With the development of face recognition technology,face recognition technology has been gradually applied to identity authentication system.But people's face photos and face video are getting more and more easy to be obtained,and the quality is getting higher and higher,so the problem of using face photos or videos to attack face authentication system has appeared.This kind of face forgery attack has brought great challenge to the security of face authentication system.Therefore,face liveness detection in face authentication system is of great significance.The real face is different from the face photo.The real face is three-dimensional,but the face photo is flat.The face photo lost a lot of details in the process of shooting and printing,for example the facial features and the contour edge in the face photo are more smooth than the real face.The material of real face is different from the face photo,so the reflection of light from the real face is different from the face photo.Because of these differences,face images taken from real face are different from face images taken from face photo.According to the difference between real face and face photo,this paper presents a real-time,efficient and non-intrusive face liveness detection method to judge whether the face image is taken from real face or face photo.First,this paper uses nonlinear diffusion technology based on additive operator splitting scheme to process images.Then this paper calculates the diffusion speed maps base on diffused images and builds a training and testing data set.But beyond that,this paper proposes a convolution neural network,which can learn the discriminative and high-level features representation from large-scale training data.After training,the neural network can effectively determine that diffusion speed maps correspond to real face or face photo and realizes the purpose of face liveness detection.Experiments are executed on the NUAA photograph imposter database and the accuracy reached 97.8%.The experimental results show the effectiveness of the proposed method in this article.