Research On Face Liveness Detection Method Based On Physiological Signal And Deep Learning

With the maturity of face recognition technology,face recognition has been used in many areas,such as access control,face payment,and device unlocking.Therefore,the security of the face recognition system has attracted more and more attention.We mainly use image features to authenticate the user's identity.This authentication method is vulnerable to spoofing attacks,resulting in false authentication.Common attack methods include presenting fake photos,videos or 3D masks of real faces in front of the authentication system.Face liveness detection is an important technical mean for face authentication system to resist spoofing attacks by analyzing face information captured by the camera and sequentially to judge whether the object which is detected is liveness.At present,face liveness detection is mainly based on the texture information of the face image and facial motion information.However,due to the multiformity of attack forms,the complicacy of the detection environment and the lighting conditions,the detection performance of the present methods is unsteady.To perfect this algorithm,this paper designs a face liveness detection algorithm model which is based on the fusion of physiological signals and deep learning.Experiments have proved that this algorithm model has better performance on public face liveness detection databases,and further improved Robustness of the face liveness detection algorithm model in complex scenes.The main works of this paper are as follows:1.To summarise the current research results in the field of face liveness detection,and analyze the reasons why the existing algorithm models are not robust enough and have poor anti-spoofing capability.2.To finish the constraction of a detection branch model based on physiological signals,and use the remote photoplethysmography(r PPG)to analyze the heartbeat information which is included in the information of face image,thus it can achieve liveness detection.The signal processing technology is used to extract the power spectrum density feature and cross-correlation spectrum feature of the face r PPG signal,and integrate this two features by the feature matrix cascade method,then use the fused features to train the SVM classifier,then test and evaluate the effectiveness of the model on the dataset Replay.3.To Construct the detection branch model based on neural network,transform the face image to HSV and Ycb Cr color space through image processing technology,then use deep convolutional neural network to extract the appearance features of the face image,and train the classifier model,then test intra-class and inter-class performance on the CASIA and Replay data sets to evaluate the effectiveness of the whole model.4.Integrate the r PPG detection branch and the neural network detection branch,use a single-layer perceptron to train the weights of the output probabilities of two classifiers,then make judgments about identity of users based on the optimal weighted sum of output probabilities of the two classifiers,then test the performance of the fused algorithm model on the Replay data set.