Research On Technology Of 3D Face Recognition Based On Deep Learning

As the most widely used identification technology in the field of biometrics,face recognition has wide application prospects in public security,interactive entertainment,mobile payment and other fields.Although there are many solutions for the research of face recognition technology based on two-dimensional images,in actual application scenarios,there are still great challenges in the recognition accuracy under the environment of posture,lighting,expression changes,etc.In particular,face anti-spoofing has never achieved a breakthrough.Compared with two-dimensional images,three-dimensional face data contains the spatial geometry and information of the face model,which can greatly improve the recognition accuracy in complex environments.Hence this paper proposes a hierarchical feature network and a rigid area-assisted3 D face recognition method for 3D face feature extraction and classification based on the 3D point cloud classification network.A series of research and experiments have been carried out on the three-dimensional face recognition technology under the influence of posture changes and expression changes.The main research work is as follows:1.Aiming at the problem that the incomplete 3D face area information caused by pose changes affects the recognition accuracy,a 3D face recognition algorithm based on hierarchical feature network is proposed.Firstly,use Set Abstraction module of PointNet++ network and Directional Spatial Aggregation module of spatial aggregation net to extract 3D face features.Then,based on SA module and DSA module,design a feature cascade network to achieve the fusion architecture of 3D face features.Finally,input the extracted features to the cascade network to perform the fusion of features,thereby achieving 3D face recognition.Relevant experiments were carried out on CASIA and its' sub-pose datasets.Simultaneously,conduct experiments on different input points.The results show that the hierarchical feature network can effectively improve the recognition rate on the pose datasets.2.Focusing on the problem of deformation of local face area due to the change of expression,a three-dimensional face recognition method assisted by the rigid area is proposed.Firstly,the T-shaped rigid area composed of the forehead,eyebrows-eyes,and nose in the 3D face model is segmented using the coordinate relationship between nose tip and rigid area.Secondly,based on the structure and characteristics of the three-dimensional face data,fine-tune the sampling and grouping mode of PointNet++ to realize feature transfer learning ofPointNet++.Finally,the fusion features of face and T-shaped region are used to perform 3D face recognition on the fine-tuned PointNet++ classification network.Compared with the results of single feature recognition,the accuracy of 3D face classification based on fusion features is improved on both expression and neutral expression datasets.In summary,this paper proposes a hierarchical feature network and a rigid area-assisted 3D face recognition method to address the problem that pose changes and expression changes affect the performance of 3D face recognition by using the cascading network,fused features respectively.The experimental results on the pose dataset and the expression dataset show that the proposed algorithms effectively improve the accuracy of 3D face recognition.