Research On Face Anti-spoofing Algorithm For Dual Camera Image Data

In recent years,face recognition technology has been closely related to our life and plays an important role in many fields such as mobile payment,distance education,unmanned security,and intelligent medicine.However,the development of face recognition technology has also brought new risks.Criminals can steal face information which leaked by users in daily life to attack the face recognition system,causing heavy losses to the individual and society.Face anti-spoofing is a technology to recognize whether a face is real or fake.It is usually used as a front task of face recognition to ensure the security and reliability of the face recognition system.Through several years of development,researchers have proposed a series of face anti-spoofing algorithms and achieved corresponding results by using the technology of feature engineering,classification algorithm,and deep learning.However,there are still some problems:(1)the face anti-spoofing algorithm based on depth map relies on deep camera or stereo vision algorithm to obtain the depth information of face,which increases the cost and computational complexity of the algorithm.(2)The supervised face anti-spoofing algorithm needs a large-scale labeled dataset to train,which increases the cost of labeling data.(3)The multi-modal information of the face(such as texture information and depth information)has different structures,which makes it difficult to fuse them effectively.In view of the above three problems,this thesis studies the face anti-spoofing algorithm for dual-camera image data,including(1)Unsupervised face anti-spoofing algorithm using dual camera-based feature matching: the algorithm extracts and matches face image pairs,obtains the number of successfully matched points in each face pair.It then uses an unsupervised clustering algorithm to obtain the data centers of successfully matched points in the real face and spoofing attacks and finally classifies the data according to the distance from the data center.(2)A dual-camera face-antispoofing algorithm based on multi-task learning and adaptive feature fusion:the algorithm obtains texture style and similarity features of the face by multi-task learning,then weights and fuses feature according to the importance of them,finally uses the fusion feature to classify.In this paper,we create a face anti-spoofing dataset Dual Cam which is based on a dual-camera system to verify the effectiveness of the algorithms.The dataset includes three kinds of face image data,which are the real face,print attack,and replay attack.Different camera positions,lighting intensity,and facial expression are set to increase the data diversity.Compared with the existing face anti-spoofing algorithms,the two algorithms proposed in this paper express the depth information of face by the similarity between face pairs rather than depth map,avoid the use of depth camera and complex stereo vision algorithm,and effectively reduce the cost and computational complexity of the algorithm.Besides,the two algorithms solve the problems respectively that existing face anti-spoofing algorithms rely on large-scale labeled data and multi-modal features are difficult to effectively fusion.The final experimental results show that the two algorithms not only reduce the cost but also have good classification performance,the HTER of each algorithm in Dual Cam is 5.1% and 2.4% respectively.In practical application,the corresponding algorithm can be selected according to different needs for face anti-spoofing.