Research On 3D Face Recognition Based On Convolutional Neural Network

Face recognition has become an important biometric technology for identity authentication and has been widely used in military,financial,public security,daily life and other fields,such as human-computer interaction,access control,video monitoring and so on.At the same time,face recognition has been a long-term research topic in the field of computer graphics.As early as the early 1990s,the introduction of feature face method in the history indicates that face recogni-tion has become popular.In the past few decades,however,considerable research has focused on the development of reliable automated face recognition systems us-ing official 2D images.Although 2D face recognition research has made significant progress in recent years,its accuracy is still highly dependent on lighting condi-tions and body posture.When the light is dim or the face posture is not properly aligned in the camera view,the recognition accuracy will be reduced.Therefore,more and more researchers have focused their research on 3d face recognition,which makes this field a new trend.However,the efficiency of 3d face recognition algorithm is still a difficult problem.This paper analyzes and summarizes the existing 3d face recognition algorithms and proposes an end-to-end 3d convolutional neural network suitable for face recognition.The main work of this paper is as follows:(1)The paper proposes a complete set of 3D face preprocessing process,which can effectively deal with the absence of depth images and can complete basic 3D shapes without the aid of 3D reconstruction technology.This method is based on the common ima.ge processing program and has no prior assumption for human face.Although it cannot deal with large-angle side faces and large-area occlusion,it can obtain high-quality three-dimensional information on common images at a faster speed.This process can be used not only to process face information,but also to any picture taken by a three-dimensional camera,so it can be applied in the fields of object recognition,three-dimensional reconstruction,etc.(2)The paper proposes a face recognition framework based on octree Con-volutional Neural Network.The network can effectively extract the three-dimensional features of the face,realize the signal level fusion of color and depth images,and complete the face recognition end to end.The network effectively combines the advantages of traditional two-dimensional Convolutional Neural Network and a three-dimensional object detection neural network using octree Memory Spot cloud,and can process large-scale point cloud data without manu-ally designing features.Experiments show that the network can effectively adapt to different data sets with only fine tuning,and requires less input pictures.(3)In order to adapt to the problem that most of the current three-dimensional face data sets have few samples,on the one hand,the paper uses the central loss function as a neural network learning index,which can effectively reduce the differences between classes and improve the face recognition degree.On the other hand,through the unique structure of octree,the original point cloud is resampled from coa.rse to fine.By combining the discrimination under different resolutions,the face recognition accuracy is effectively improved.Experiments are carried out on several face recognition databases,and the recognition efficiency equivalent to the current best results is obtained on almost every data set,which proves the effectiveness and robustness of the proposed method.