Research On Face Anti-Spoofing System Based On Multi-Spectral Decision Level Fusion

With the development of Internet and the arrival of digital age,convenient and intelligent new lifestyle is increasingly pursued.Under this huge background,face recognition,as a technology that can automatically detect and classify biometrics,is very suitable for the needs of social development.However,the face recognition system can only authenticate the identity of the subject,but cannot distinguish and judge the true and false image,which will lead to a series of security loopholes that can't be ignored.Because face recognition system has major security risks,therefore,the face anti-spoofing detection technology came into being.As the front field of face system,this technology ensures the efficient and reliable operation of the whole living identity recognition module,which is of great significance to the safety of people's life and property.In this paper,we make a lot of analysis and research on the hot topic of human face anti-spoofing detection.The main work is as follows:Firstly,aiming at the problem of high false detection rate and great limitation of single modal RGB in detection,the binocular camera is selected by us to analyze and study the problem of living detection from the perspective of multi spectrum.Then we propose a living detection algorithm based on NIR+RGB multi feature fusion.The experimental results show that the accuracy of the proposed algorithm is 94.37%in the self built dataset,which is 5%higher than that of other mainstream feature extraction algorithms.In addition,following the principle of "combination of theory and practice",a complete set of real-time face detection platform based binocular camera is built by us.The test shows that the system has a very good classification effect.And all of these system is successfully delivered to the project company.Secondly,in the field of deep learning,this paper takes facebagnet as a reference,and optimizes the network model for its shortcomings.A lightweight end-to-end deep learning face anti-spoofing detection algorithm is proposed in this paper.Compared with the original network model,two improvements are proposed in this algorithm as follows:first,considering the greater limitations of facebagnet model,this paper redesigns the network using deep separable convolution,and optimizes the size of the classify model;second,this paper creatively introduces the central loss function in the problem of living detection,which interacts with the cross entropy loss function,tightens the class internal distance,and expands the class Spacing.The experimental results show that the two improvements proposed in this paper can effectively reduce the size of the network model without losing too much accuracy.Finally,the accuracy of the improved network in casia-surf multimodal data set is 99.78%,and the final model size is only 26.6MB(only one seventh of the original network model size).In this paper,we mainly study human face detection from the perspective of multispectral and multimodal,which has a certain cutting edge and timeliness.In the future,we will further highlight the weight of different spectra in classification,consider different feature fusion schemes,and further optimize and output the whole face application system according to the actual needs.We believe that the application of face related technology is bound to be more mature and reliable.