| With the progress of technology,face recognition systems have been applied to daily life.At the same time,there are more and more attacks for face recognition systems.Many criminals deceive the face recognition system to obtain users’ privacy and rights for illegal acts through photos or videos.Therefore,face anti-spoofing,as the key to the security of the system,has become the focus of many researchers.At present,most face anti-spoofing algorithms distinguish live and spoof faces through deep learning methods,and the effect of detection is better than traditional feature representation methods.However,in practical applications,the collection scene is complex,the attack methods are diverse,and the hardware resources are limited.Thus,face anti-spoofing still has problems in terms of detection accuracy and model complexity,which needs to be further studied.In practical applications,image acquisition devices mostly use monocular RGB cameras,so this thesis mainly studies face anti-spoofing on monocular RGB images.In order to solve the problem of low accuracy and high complexity,this thesis studies and analyzes face anti-spoofing from three perspectives: designing high-precision networks,designing lightweight networks,and improving the performance of lightweight networks.The main researches are as follows:(1)In order to cope with complex scenes and changing attacks in face anti-spoofing,and make full use of the feature extraction capacity of traditional feature description operators,a two-stream face anti-spoofing method combined with local gradient information is proposed.This method consists of a fully convolutional network and a local gradient information network.Among them,the fully convolutional network mainly extracts the spoof information and depth information of the face;the local gradient information network mainly uses the texture extraction module to extract the gradient information of the face and supplement the texture features of the image.In addition,an attention fusion module is designed for the fusion of the two-stream network.This module can highlight important features and remove redundant information.The high-precision network achieves 0.9% ACER on protocol 1 of OULU-NPU dataset.(2)In order to meet the needs of practical applications and obtain a lightweight network with fewer parameters and less computation,a face anti-spoofing method based on mutual learning mechanism and fuzzy matching strategy is proposed.This method introduces the mutual learning mechanism into face anti-spoofing and designs a central difference auxiliary network and a residual lightweight network.Firstly,the output features of the student networks are aggregated to obtain soft labels that contain rich information.During training,soft labels are used together with dataset labels to supervise each network,allowing the student networks to learn from each other.In order to make the lightweight network learn more information,an intermediate feature learning process is added.In addition,due to inaccurate depth map labels,a fuzzy matching strategy is designed to further improve the detection performance of the network.This method achieves 1.5% ACER on protocol 1 of OULU-NPU dataset.(3)In order to enhance the feature extraction capacity of the lightweight network and balance the performance and complexity of the network,a multi-supervised face anti-spoofing method that integrates multi-scale features is proposed.Based on the lightweight network,it designs a multi-scale feature fusion module to explore image features.The multi-scale feature fusion module is divided into a detail texture branch and a group receptive field branch.Among them,the detail texture branch mainly extracts the detailed information of the image,and the group receptive field branch mainly extracts the multi-scale spatial and semantic information of the image.In addition,a multiple supervision strategy is proposed to make the network pay more attention to the face.The lightweight network achieves 0.9% ACER on protocol 1 of OULU-NPU dataset.The above methods are carried out and analyzed on OULU-NPU,Si W,CASIA-FASD and Replay-Attack datasets,and the models are evaluated by a variety of evaluation indicators.The experimental results verify the effectiveness of the methods in this thesis. |
