The Research And Implementation Of Face Liveness Detection Technology

Face liveness detection technology can judge whether the human face is a living body or not before identity authenticating by facial feature,so as to prevent malicious people to forge and steal the biometric characteristics of others for authentication.As one of the key technologies to ensure the security of face recognition,face liveness detection has been widely used in various authentication and security scenarios.This thesis designs and accomplishes an interactive liveness detection system based on key points of human face.The system requires the users to make corresponding actions(smile,shaking head,winking)according to random instructions,then through the corresponding detection algorithm to recognize face movement,so that it can effectively judge whether it is a living body or not.The system adopts SDM(Supervised Descent Method)algorithm to detect the key points of human faces.The algorithm employs the regression analysis to effectively reduce the timeconsuming iteration operation and splendidly improves the speed of operation.On this basis,it implements detection to three kinds of behavior patterns,i.e.shaking head,smiling and winking.For the head shaking mode,the rotation matrix and the shift matrix are calculated,which maps the key points of the nose tip from 2D to 3D space,so as to determine whether the head is shaking or not.For the smiling mode,the distance between the mouth key points is calculated,once the distance variation exceeds the specified threshold,the smiling action is detected.For the winking mode,the distance between the key points of the upper and lower eyelids is calculated,once the distance variation exceeds the specified threshold,the winking action is detected.The liveness detection system is implemented in the Windows system based on C/C++ programming language.The experimental results show that the passing rate of living face is up to 100%,while for the non-living face(face in the picture),the passing rate is only 1/300(still)and 11/900(simulated action attack).The average time of the system to detect the three behavior patterns is: 0.28±0.02 s for shaking,0.75±0.02 s for winking,0.81±0.01 s for smiling,which can meet the requirements of real-time and accuracy in practical applications.In addition,the system is transplanted to the Android operating system to satisfy mobility and portability.