Research On Face Anti-spoofing Algorithm In Face Recognition

With the rapid development of society,more convenient and efficient biometric recognition technology has been widely used.The human face is a key and challenging item in biological characteristics,and it is the research focus of researchers.In recent years,with the emergence of emerging technologies such as deep learning,face recognition technology has made great progress and has been more and more widely used in life.Face recognition has the advantages of low cost,no contact,long distance,etc.At the same time,the privacy of face recognition is worse,and the face anti-spoofing technology is the key part to ensure that face recognition can be used on the ground.Aiming at the problems such as the small number of training samples,limited performance,and poor generalization ability in the existing face anti-spoofing algorithms,this paper introduces auxiliary supervision information and attention mechanism to design a multi-task convolutional neural network,and uses transfer learning to effectively improve the algorithm generalization capabilities.The main work of this article is as follows:(1)This paper analyzes the problem of insufficient supervision information in the 2-class classification of face anti-spoofing.Considering the obvious spatial difference between real and fake faces,the face depth information is introduced as auxiliary supervision information.Based on the residual network ResNet-50,a multi-tasking convolutional neural network is designed to perform face depth label regression while performing real-fake face classification.In addition,this paper also introduces channel attention mechanism in specific task branches to further enhance network performance.In this paper,sufficient experiments are performed on three public face anti-spoofing datasets of CASIA-FASD,NUAA,and REPLAY-ATTACK to verify the performance of the algorithm.The results show that the performance of this algorithm exceeds or reaches the current research in the field.(2)This paper focuses on the problem of poor generalization ability in face anti-spoofing algorithms.By extracting the characteristics of different data sets for visual analysis,it is found that there are obvious domain differences between them.Later,this paper proposes to use the image style conversion algorithm based on PhotoWCT to perform domain adaptation on the face anti-spoofing dataset to reduce the domain difference between the two datasets.After the image style conversion is performed on the dataset,the multi-task convolutional neural network designed earlier is used for experimental verification.The results of inter-library experiments on the data sets NUAA,CASIA,and REPLAY show that the solution proposed in this paper can reduce the EER by 13%-18%,and effectively improve the domain differences between different datasets.