Design And Realization Of Multi-modal Face Anti-Spoofing System Based On Feature Branch Optimization

In recent years,face recognition technology has become increasingly popular.However,there are many concerns about its safety and effectiveness.Most of face recognition technology can verify the identity of a person only through an image,while there is a maj or threat to gain illegitimate access if it can not determine whether the subject in image is live.So,the face anti-spoofing is needed to prevent security breaches.However,the existing face anti-spoofing algorithms using single modal suffer low detection accuracy when the face anti-spoofing are complex,and the algorithms using multiple modals tend to have large model size and slow running speeds.In order to solve the above problems,a face anti-spoofing detection system that can run quickly on Android terminals is designed in three aspects:algorithm design,system design,and terminal deployment.The main work of this paper is as follows:First of all,striking a balance between the size and accuracy of model,the lightweight multi-modal face anti-spoofing algorithm based on feature extraction branch optimization is proposed.For Infrared and Depth modals,the feature extraction branch based on full convolutional network is introduced.For RGB modal,the feature extraction branch using a lightweight structure of the Ghost module is presented.And the deep convolution and flow module are used to get the result through the decition network.The proposed algorithm was tested on the CASIA-SURF dataset and achieve 99.85%accuracy,which is better than FaceBagNet.Moreover,the model size is only 21.7 Mb,while FaceBagNet is 178 Mb,which is reduced by 87.8%.Second,in view of the insufficiency of the existing face spoofing dataset and the challenge of lightweight embedded,the hierarchical face anti-spoofing strategy and builds an efficient detection system are designed.And a dataset containing variety types of spoofing and real samples is established in various environments.In order to overcome face detection in complex illumination.Moreover,a hierarchical face anti-spoofing strategy is designed.The face detection based on IR image is used firstly to prevent electronic screen attacks,a single-modal network based on depth images for rapid detection is used secondly to prevent two-dimensional spoofing attacks,and then the above-mentioned multi-modal face anti-spoofing detection is used to prevent threedimensional spoofing attacks.Combined with multi-frame detection strategy to improve the accuracy of overall detection system.In this paper,the system is deployed on an Android terminal,tested in the built BCTC face anti-spoofing attack detection detection environment and passed the standard-level index test.The average time of detection is 0.55 second,which proves that the system can quickly and accurately achieve the task of face anti-spoofing.Out-of-distribution generalization,zero-shot face anti-spoofing and illumination balancing by image processing will be further explored in future research.