Liveness Detection On Face Anti-spoofing

With the increasing number of electronic devices,there are more demands of biological identity verification system,of which face recognition systems enjoy the most popularity.Face anti-spoofing technique,an extension of face recognition,is mainly utilized to prevent photo attack or video attack.Most existing face anti-spoofing systems distinguish spoofing samples by extracting texture pattern information.However,due to the complexity of face anti-spoofing,different environment,illumination and materials will lead to various spoofing texture patterns,which is almost impossible to generalize,and these algorithms normally tend to fail when facing samples from unknown domains or in a cross-database test.To solve this problem,the thesis proposed a generalizable model with a combination of temporal motion information and traditional spatial texture information,and inspired by the human eye visual pathway system,this paper put forward an innovative three dimensional convolution mimicry model based on human visual pathway,and verified the performance and robustness of the method through challenging comprehensive test.More specifically,the paper mainly carried out the research as the following:Firstly,a study on three-dimensional neural network was carried out to explore its application effect in the field of face anti-spoofing detection,and explored the extraction of temporal sequence motion feature information,then an approximate optimal three-dimensional convolution structure was proposed.Secondly,a motion amplification algorithm based on phase shift was introduced to improve the efficiency and capability of extracting motion features from face.Comparing with the linear motion amplification algorithm and Lagrange motion amplification algorithm,the phase translation based motion amplification algorithm was established as the front module of the dynamic motion cue branch,and the boundary of motion amplification and the noise processing of motion amplification algorithm were also explored.Finally,the paper proposed a three dimensional convolution mimicry model inspired by human visual system.It utilized different convolution kernels and constraint skills to lead two pathways learning the characteristic information of the different dimensions.And with the application of Mahalanobis distance loss function in the dynamic motion branch,a more fine-grained constraints was further added.Finally,softmax was used for classification,and generalization experiments across databases proves the validity of the model.