Research On 3D Point Cloud Supervised Face Anti-spoofing Algorithm

As an important security link of the face recognition system,face anti-spoofing has developed rapidly in the field of biometric technology in recent years,and is widely used in scenarios such as mobile payment,access control systems,and financial authentication.However,face recognition system is vulnerable to attacks from different methods such as printed images,digital images,and playback videos,making the security of the face recognition system seriously threatened.Therefore,face anti-spoofing plays an important role in the face recognition system.It is an important role and has important research value.Face anti-spoofing is a research direction that has attracted much attention in the field of biometrics technology,but there are still the following problems.First,whether it is traditional or deep learning face living detection algorithms,these algorithms either directly use RGB images as supervisory features,or use RGB images to assist 2D depth maps or temporal information as supervisory information.These methods are suitable for different lighting conditions,attack equipment and attack methods are sensitive to environmental changes,and cannot learn the invariance characteristics between real samples and attack samples,and are not robust;Second,these methods require another additional network to estimate and generate 2D depth images or r PPG information.The network model is complex,and the supervision features are complex and redundant,which is not conducive to deployment in the actual application environment;Third,the existing methods the feature difference between real samples and attack samples is not considered from the perspective of fine learning features.Aiming at the above three issues,this paper conducts research on the face anti-spoofing problem based on feature supervision.The main work of this paper is as follows:(1)Lightweight face anti-spoofing algorithm based on 3D point cloud supervision.Because 3D point cloud has rich three-dimensional spatial information,it is not easy to be affected by the changes of lighting and other environmental conditions.In view of the complexity of the previous methods and the redundancy of the supervised features,this paper proposes an efficient and effective 3D point cloud supervised lightweight face antispoofing network model(3DPC-Net)by using 3D point cloud as the supervised information for the first time.The point cloud corresponding to the real face sample has three-dimensional structural information,while the point cloud of the attacking face sample is on the same plane.The 3D point cloud data is generated by a dense face reconstruction algorithm,and then subjected to a series of preprocessing to obtain the final usable label.3DPC-Net is composed of an encoding and decoding network,which uses the chamfering distance loss function to learn 3D point cloud supervised feature mapping.Compared with the previous methods,this method achieves better detection results in inter class experiments,and improves the efficiency of model operation.(2)A 3D point cloud based dynamic graph convolutional cascade face anti-spoofing algorithm is proposed.Aiming at the problem of failing to finely learn the 3D feature difference between real samples and attack samples,this paper proposes a face antispoofing algorithm based on 3D point cloud supervision of dynamic graph convolution cascade(DGC-FAS).The algorithm is divided into two stages.The first stage uses the encoder proposed in Chapter 3 and introduces a spatial attention mechanism to generate coarse 3D point cloud features.The second stage uses a dynamic graph convolution cascade model for fine-grained learn 3D point cloud features.Throughout the training process,a dynamic loss function is designed to dynamically optimize the learned 3D point cloud features.Experiments show that the detection performance of this method on the four general data sets is better than most current mainstream methods.