Face Anti-spoofing Based On Robust Loss Face Alignment And Binocular Measuring

In recent years,face recognition technology has been widely used in areas such as entry barriers and public security because of its safety and stability.Face recognition system consists of Face detection,Facial keypoints detection,Face Anti-spoofing and face recognition.Face detection and face recognition technology are already very mature.While the training process of the facial keypoints detection task is vulnerable to unusual facial gestures or expressions,illumination changes and occlusion,so it still needs improvement.The development of the Internet makes it easier for offenders to obtain victims' photo and video information.It is possible to use this information to impersonate the victim to attack the login system of the bank or residence,then the life and property of the victim will be threatened.Therefore,with the increasing vigilance on the security of personal information,the heat of Face Anti-spoofing is also increasing already.However,current Face Anti-spoofing systems mostly require users' cooperation and can only resist one of photo attacks or video attacks.This paper aims at the above problems,and studies facial keypoints detection,binocular distance measurement,and Face Anti-spoofing.The depth information of each key point of the face is used as a feature to distinguish human face from face photo or video attack.The main work is as follows:(1)This paper trains a convolutional neural network(CNN)model based on robust loss function to achieve the task of detecting facial keypoints.Face feature detection tasks can be viewed as a regression problem.In most cases,when convolutional neural networks are used to solve regression problems,the L2 loss function is usually used during CNN training.But its robustness is low.This paper applies the Tukey estimator in Robust Statistics as loss function to the training of the model,which makes the training of the model make the best use of the contribution of each sample,and avoids the excessive guidance to the training by the outliers.The process greatly enhances the robustness and generalization ability of the model and makes the training model converge faster and the convergence value is lower.In the experiment,the algorithm was compared with three advanced methods and two commercial software on two public datasets,and achieved good results.(2)In order to complete Face Anti-spoofing,it is necessary to implement distance measurement first.At present,the method of distance measurement is mainly divided into monocular and binocular distance measurement.However,the monocular camera's perception of the position and distance of the object may cause deviations,so this paper chooses binocular to achieve distance measurement.Calibrate the two cameras using the classical Zhang calibration method to obtain the internal and external parameters of the binocular camera.These parameters are used to perform stereo correction and stereo matching on the images obtained by the left and right cameras.In this paper,we use the BM block stereo matching method to obtain the disparity map,and then obtain the three-dimensional point cloud,and the three-dimensional coordinate information of the human face feature points can also be obtained.(3)Real-time Face Anti-spoofing based on facial keypoints detection and binocular distance measurement.After obtaining the depth information of each f keypoints of the face,the classic two-class classification classifier support vector machine(SVM)is used to classify the real living face and the fake face.Since there is no face depth information database for Face Anti-spoofing at present,this paper obtained a self-built small sample database through experiments,including depth information of photos and videos and real faces.The video sampling of attack attempts are taken from videos which are played on mobile phones,tablets,and computer screens.The model is capable of distinguishing photo and video of attack attempts in a variety of situations in the test set.The test pictures are accompanied by artificially distorted photos that simulate the changing depth of real face,and the models can still accurately identify photos or faces.