Research On Face Anti-spoofing Algorithm Based On Multi-modal Multi-scale Fusion

Face anti-spoofing means to judge whether the captured face image is a real face or a false face,which is an important guarantee for the security of face recognition system.Traditional face anti-spoofing methods mainly use handcrafted features,such as LBP,HOG,SIFT,SURF and DOG,to depict the different feature distribution between real and fake faces,but it is difficult to adapt to the anti-spoofing problem in unconstrained environment(such as the change of lighting and background).In view of this,the convolutional neural network model is designed mainly from two aspects of multi-scale and multi-modal fusion,and more robust face anti-spoofing is realized by integrating features of different modes and different scales.The main work of this paper is as follows:1)Based on feature pyramid network,a multi-scale fusion face anti-spoofing algorithm is proposed.To solve the problem of discriminant difference at different scales,first implements the fusion of different scales features through the feature pyramid network.Then,the channel attention fusion network and the spatial attention network are combined to fuse the features of different modes.Experimental results on CASIA-SURF datasets show that compared with the mainstream multi-modal face anti-spoofing method Multi-Scale Fusion,the proposed method can reduce APCER and ACER indexes by 1.2% and 0.5%respectively.The experimental results show that the above method can effectively integrate the features of different scales,but there are still problems such as lack of interaction between different modes and poor multi-mode fusion effect.2)Based on multi-modal shared branch network,a multi-modal fusion face anti-spoofing algorithm is proposed.In order to solve the problems of lack of interaction between different modes and poor multi-mode fusion effect,based on the channel attention network,a multi-mode shared branch network is designed to realize the information interaction among different modes in the process of feature extraction.Based on the channel attention fusion network,a multi-mode channel attention fusion network is proposed to fuse features of different modes for classification.Experimental results on CASIA-SURF datasets show that compared with the mainstream multi-modal face anti-spoofing method MultiScale Fusion,the proposed method can reduce APCER and ACER by 1.1% and0.4% respectively.Experimental results show that the proposed method can effectively integrate the features of different modes and improve the robustness of the model.