Research On Face Anti-spoofing Algorithm Based On Deep Learning

Face recognition as one of the most extensive biometric technologies,is used in security authentication system such as online payment,access control,and smartphone authentication.A practical face recognition system requires not only high recognition performance,but also the ability to detect anti-spoofing attacks.Face anti-spoofing is to identify and judge the authenticity of the face images obtained by the face recognition system,and is a new research field that has only emerged in recent years.The thesis conducts research on anti-spoofing of human faces,and the main works are as follows:(1)For the face anti-spoofing technology of visible RGB images,an LGNet algorithm framework is proposed which fusing local information and global information.The feature at the tail of LGNet is divided into multiple local feature blocks for binary supervision,and the pixel level supervision is used for global features.Finally,the local prediction and global prediction are fused.The method is tested on multiple datasets in this thesis,and the final result shows that the algorithm has strong robustness and cross-dataset capabilities.(2)The face anti-spoofing method for structured light pictures is a hardware-based method.This thesis uses the FM810 structured light camera as the research basis.First,this thesis adopts zero-value filling,hole mean filtering,median filtering,and normalization to grayscale images as the preprocessing methods.Then,in order to improve the robustness of the algorithm and reduce the labor cost of data collection,this thesis proposes an algorithm for generating negative samples.Finally,by designing a face anti-spoofing data collection system,a large amount of data was collected,the Caffe Net model was trained with the generated negative samples together,and integrated into the face anti-spoofing data collection system for testing,the accuracy of the system is 99.75% and the operation is stable.(3)Face anti-spoofing method based on multi-modal images.In order to make the model run quickly in computing power-limited devices,an extreme light network architecture(Feather Net A/B)is proposed with a streaming module which fixes the weakness of Global Average Pooling and uses less parameters.Our single Feather Net B trained by depth image only,provides a higher baseline with 0.00168%ACER,0.35 M parameters and 83 M FLOPS.Meanwhile,the MMFD dataset is collected to provide more attacks and diversity to gain better generalization.Furthermore,a novel fusion procedure with “ensemble + cascade” structure is presented to satisfy the performance preferred use cases,which reaches to 0.0013%(ACER)in the CASIA-SURF test set,0.9814%(TPR @ FPR = 10e-4).Finally,different normalization methods are proposed to eliminate the style differences of different races,so as to improve the performance of cross race test in CASIA-Ce FA multi-modal dataset.