Research On 3D Point Cloud Face Recognition Technology Based On Deep Learning

Face recognition as a popular biosafety verification method,has been used in many fields with the increasing emphasis on safety information.Because two-dimensional face recognition is greatly affected by light,expressions,gestures,etc.,the overall face recognition is low in robustness,and the security of two-dimensional faces is low.Therefore,3D face recognition has become a popular upgrade solution for current research.3D face recognition technology can not only provide traditional 2D face information,but also add the texture and geometric features of the face,and also provide the depth of the real scene.Information,recognition effect has improved to a certain extent compared to two-dimensional recognition.However,there are three problems in 3D face recognition that need to be solved urgently: first,the lack of a large amount of training data;the second is the large amount of single 3D face data sample data,which is not convenient for deep learning;the third is that currently there is no suitable depth model suitable for 3D people Face recognition.For the problem of insufficient training data,after comparing the 3D face data set,this thesis adopts BFM2017 face data set.For the problem of insufficient training data,this thesis uses the BFM2017 face dataset after comparing the 3D face dataset.This dataset provides parameters such as average face,shape,and expression.The data set is used for 3D people by using Gaussian probability The shape and expression of the face are randomly fitted to the face so as to expand the training data.For the problem of large sample data of a single 3D face data,this article extracts the point cloud information in the data set and calculates the curvature of the point cloud.On the one hand,it can reduce the amount of data,on the other hand,it can unify the data format,which is convenient for other data sets.Testing to improve the versatility of the method.In view of the lack of a unified deep learning model for 3D face recognition,the Pointnet ++ network for object classification and segmentation is used for 3D point cloud face recognition.This network aims to extract the entire object point cloud globally based on the Pointnet network model.Features are improved to extract local features hierarchically.Through the pointent ++ network,and improved the sampling layer and grouping layer in its SA module: For the sampling layer,this thesis introduces curvature features instead of the original iterative farthest point sampling method to quickly mine feature points;for the grouping layer,this thesis uses KNN neighborhood queries it replaces the original network radius query grouping method.Finally,this article modified and compared the loss function in the pointent ++ network.In this thesis,Softmax loss function,Centor loss function and ArcFace loss function were selected for experimental comparison and analysis.The experiment proved that the classification of the face in the ArcFace loss function works best.At the same time,in order to verify the effectiveness of the method used in this thesis,preprocessing was performed on the CASIA 3D dataset and the recognition rates under different expressions and poses were analyzed.