Research And Implementation Of Face Anti-Spoofing Algorithm Based On Deep Learning

The face live detection task aims to ensure the security of the face recognition process,that is,before face recognition,it is determined whether the input user is a real live face rather than a fake face.According to whether user cooperation is needed,face live detection can be divided into interactive face live detection and non-interactive face live detection.The demand for detection algorithms is increasing.There are various forms of face forgery attacks,and there are many types of face background changes.Algorithms trained on a specific or a few specific data sets generally have poor generalization performance.Real application scenarios are often very complicated.This thesis summarizes the research results of human face liveness detection.In view of the fraud problem in face liveness detection,this thesis studies the features of border texture,the features of moiré texture and domain adaptation,and designs a live detection algorithm based on color single-frame images.The main work of the thesis is as follows:First of all,dataset is make annotated for the characteristics of non-living data with black frame,and a multi-scale multi-fusion-based black frame detection algorithm is proposed.which mainly includes edge detection module and the Hough line detection Module.Different from the previous face liveness detection algorithm,this thesis believes that the background area of the image also contributes to face liveness detection.The edge detection module is an end-to-end network based on multi-scale multi-fusion,which fuses the low-layer features and high-layer features of the network to obtain the edge detection image of the image to be detected.Besides,to solve the sample imbalance problem during the training process,the method uses a class-balanced cross-entropy loss.The method is compared with the traditional algorithm on the dataset.The experimental results show that the method in this thesis has better detection ability than the traditional algorithm.Then,according to the characteristics of non-living data with moiré pattern,dataset is make annotated And a moiré pattern detection algorithm based on lightweight deep neural network is proposed.Which mainly includes threshold segmentation processing module and classifier module.Moiré pattern is a kind of high frequency irregular stripes,this thesis believes that high-frequency information of the image contributes to face liveness detection.In order to better utilize the high-frequency information in the image,in this thesis,adaptive thresholding is applied to the images to be detected,and a strong classifier dedicated to moiré detection is obtained by using a lightweight deep neural network trained on the dataset.The experimental results show that the method can effectively detect non-living face images on the moiré dataset.Then,a face liveness detection algorithm based on domain adaptation is studied,which mainly solves the problem of poor generalization ability of face live detection algorithm.This thesis believes that chromaticity component contributes to face liveness detection,introduce YCb Cr color space to increase the texture differentiation between living and non-living bodies.This thesis treat different live detection datasets as different domains,and learn more representational features under the same category by cross-domain training.Besides,the genuine faces have a three-dimensional spatial structure,while fake faces are often flat structure.To further improve the generalization of the model,this thesis uses three-dimensional face reconstruction technology.Face depth information is used as auxiliary information for the network.The experimental results show that our method further improves the generalization of the face liveness detection algorithm.Finally,aiming at the requirements of 2D camera equipment for living algorithm and the characteristics of engineering dataset,a fusion algorithm is proposed combining black frame detection,moiré detection,and living detection based on domain adaptation.This algorithm solves the above engineering problems and is practical.The experimental results show that black frame detection module and moiré detection module can greatly improve the performance of the face liveness detection algorithm.