| Face recognition technology is widely used in Intelligent security,Internet finance and other scenes.In the process of face recognition,various kinds of spoof faces,such as faces in photos,faces displayed on the electronic screen,may be mistakenly recognized as live faces,which has a huge security risk.Therefore,as an important step to verify the authenticity of human identity in the process of face recognition,Face anti-spoofing technology has great research significance and application value.Although Face anti-spoofing technology has made some progress in the academic and industrial applications,with the development of attack technology and the diversification of Face anti-spoofing scenarios,Face anti-spoofing algorithms need to have high generalization and robustness.In multi-modal Face anti-spoofing,how to effectively mine facial details and fuse multi-modal features to improve the performance of multi-modal Face Anti-spoofing is a difficult problem.In order to solve these problems,this thesis conduct research on multi-modal Face Anti-spoofing.The main work of this thesis is as follows:(1)In view of the problem of the existing multi-modal Face anti-spoofing dataset MFA,such as low attack image quality and single data acquisition scene and narrow age distribution of the subjects,this thesis constructs a large-scale Face anti-spoofing dataset MFA2 with higher attack image quality,richer acquisition scene and wider age distribution of the subjects.MFA2 contains 2 168 videos and more than 1.6 million pictures.(2)In order to fully mine face details and fuse multimodal features,a multi-scale multi-modal context selection Face anti-spoofing algorithm is proposed.In this algorithm,attention mechanism is used to mine multi-modal context to promote multi-modal fusion;At the same time,the modal features are fused on multiple scales to mine more facial details.The performance of the proposed algorithm is verified on MFA and MFA2.(3)In order to improve the generalization of multi-modal Face anti-spoofing methods,a multi-modal spoof cue generation Face anti-spoofing algorithm based on anomaly detection is proposed.This algorithm maps the spoof samples to outliers in the feature space to improve the generalization to unknown spoof samples.In order to make effective fusion of multi-modal information,two weight strategies are proposed to adjust the importance of different modes in decision-making.Experiments on MFA and mfa2 verify that the proposed algorithm has high generalization.Experiments on CASIA-Ce FA,the largest cross ethnicity multi-modal Face anti-spoofing dataset at present,verify that the proposed algorithm still has high generalization in Cross-ethnicity and Cross-PAI. |
