Live Face Verification Based On Video Sequence

This thesis discusses the live face verification algorithm,which is an important part of modern face recognition systems(such as online banking,online shopping,etc.).The accuracy and robustness performance of live face verification technology still need to be enhanced,although many institutes have made so much effort on it.The main reason includes model's over-fitting various variables(such as illuminations,different condition of scenes and equipments of collecting images).We improve the robustness performance of live face verification,based on homography analysis between video frames.The main contributions show in the following aspects.First of all,in view of the limitation of performance in live face verification task from the artificial features,such as Optical Flow,LBP,HOG,we propose a video classification method based on the the homography analysis between video frames.We observe that local facial patches from the recorded video present significant motion pattern,e.g.,local homographic property between adjacent frames,while real ones do not possess this nature.For a "fake" face,there exist multiple patches which are easily recognized as "fake"while all local patches of a real one must be recognized as "real".The features extracted from the homography parameters show strong accuracy and robustness performance on multiple datasets,especially in the cross-validation experiments on different datasets.The model shows very good generalization ability.Secondly,this thesis discusses the feature extraction method based on homography analysis.In homography regions,the video frames can be transformed from one frame to another by a homography transformation matrix.In this thesis,a Spatial Transformation Network(STN)is designed to extract the homography transformation matrix between two frames.We combined the parameters extracted by STN with the images to extract the motion feature.Finally,since in "fake" face videos,the relative motion between video frames and can be decomposed into the relative motion component(homography component)and the user's motion component(non-homography component).Therefore,in "fake" face video,some regions do not show significant homography due to the user's motion.Since all regions in a "real" face video are not homography,some regions of a fake face video show strong homography.Therefore,we propose a novel framework for applying multiple instance learning(MIL)algorithms to multiple Patchs of an image.To sum up,in this thesis,the live face verification algorithm has been widely discussed,and the key issues have been deeply studied.In view of the limitations of the generalization ability of traditional feature,we propose a novel algorithm based on homography analysis.In addition,this thesis proposed a method to extract the homography transformation parameters.Finally,it proposed to introduce the multi-instance framework into the live face verification,which improves the recall rate of the model.Based on a large number of theoretical analysis and experimental results,the proposed method is stable and has strong generalization ability,which significantly improves the performance of the live face verification algorithm.