Research On Face Anti-spoofing Algorithm Based On Video

In recent years,face recognition technology has been widely used in daily life,but face information can easily be accessed by others and made into fake faces.Illegal users use fake faces to attack the face recognition system,which poses a great hidden danger to user information security.The existing face anti-spoofing algorithms are gradually decreasing in detection capability as the attack methods keep changing,and the security of face recognition systems is facing serious challenges.Aiming at printing attacks,video replay attacks and mask attacks,this thesis combines face motion features with deep learning to propose a face anti-spoofing algorithm with high accurac.The specific work is as follows.(1)Aiming at the difference of facial motion between real faces and fake faces,a face anti-spoofing algorithm based on transient motion is proposed.Firstly,we perform motion amplification algorithm to the video to make tiny motion more obvious.Secondly,the transient motion feature maps are generated using transient motion features in RGB colour space to describe the transient motion.Thirdly,a VGG16 network is adopted to distinguish the real faces from fake faces.We use multiple tests and calculate the average to obtain the final results.Experiments on two datasets verify the effectiveness of the proposed algorithm.The half total error rate on the Replay-attack dataset is 1.5% and the equal error rate on the CASIA FASD dataset is 3.0%,which are respectively 0.75% and 1.2% lower than the result of existing algorithms.The accuracy on self-built mask attack dataset is 95.1%.It proves that the algorithm is effective and can provide effective protection to the face recognition system.(2)Aiming at the problem that the transient motion feature cannot effectively use the time domain features in the video,a face anti-spoofing algorithm based on tiny motion is proposed.Firstly,we perform motion amplification algorithm to the video to make tiny motion more obvious.Secondly,a motion feature map is created by fusing motion intensity and motion direction to describe the tiny motion.Thirdly,a VGG16 network based on attention mechanism model is adopted to distinguish the real faces from fake faces.Experiments on two datasets verify the effectiveness of the proposed algorithm.The half total error rate of our algorithm on the Replay-attack dataset is1.35%,the equal error rate on the CASIA FASD dataset is 1.2% and the accuracy on self-built mask attack dataset is 97.3%.Compared to the test results of transient motion feature,the half total error rate on the Replay-attack dataset decreased by0.15%,the equal error rate on the CASIA FASD dataset decreased by 1.8%,the accuracy rate on the self-built mask attack dataset increased by 2.2%.It proves the effectiveness of the tiny motion feature that combines the time domain feature and the spatial domain feature on face anti-spoofing detection.Aiming at the difference of facial motion between real faces and fake faces,this thesis proposes two face anti-spoofing algorithms,which can effectively against printing attacks,video replay attacks and mask attacks,and improve the security of face recognition systems.