Three Dimentional Face Recognition Based On Local Texture Features On The Mesh

Texture is an important visual cue that reflects human visual perception of the surface information of objects.The texture reflected by the 2D face image is not the true texture of the 3D face surface,and this 2D texture is considerably affected by the variations of illumination and make-up.The concept of 3D texture is completely different from 2D texture,which reflects the repeatable geometric patterns on the 3D face surface,so it is not sensitive to the changes of optical information and effectively overcomes the deficiency of2 D texture.In this context,the research on face recognition based on the 3D texture features is carried out.The main research contents of this paper are summarized as follows:1)The precise position of the nib point is usually required in a 3D face preprocessing step.In this paper,the common three nib point positioning methods,namely the method based on the maximum z coordinate value,the method based on the shape index and the method based on the horizontal slice are detailed.This paper introduces and summarizes their advantages and disadvantages,and analyzes the final positioning results in the case of pose and occlusion changes of a 3D face.2)This paper improves the hand-crafted Mesh-LBP feature and proposes Mesh-tLBP,Mesh-MBP and Mesh-LTP.The first two improvements are designed for the robustness of the Mesh-LBP to noise or expression changes,whereas the last one improves the power of the Mesh-LBP in capturing facial details.Also,this paper summarizes three different statistical methods to count the calculated Mesh-LBP pattern,including the naive holistic histogram,the spatially enhanced histogram and the holistic coded image.Finally,using the proposed improved operators and statistical methods,the identification experiment was carried out on CASIA 3D face database containing expression and pose changes.The results show that Mesh-tLBP combined with the spatially enhanced histogram can further improve face recognition performance.3)This paper uses VGG16 network to learn the MSD3 DP texture feature from 3D face data directly.Aiming at the problem of the input format of 3D data for a neural network,the DAE image of a 3D face is proposed by the GridFit algorithm.For the problem of scarcity of 3D face data,three augmentation algorithms of expression,pose and occlusion are used.For the relatively smooth characteristics of the 3D facial DAE image,a multi-scale module is introduced.Finally,a multi-scale VGG16 network was used to extract the MSD3 DPtexture features using the AM-Softmax loss function.The identification experiment was carried out on CASIA and Bosphorus 3D face databases.The results show that the MSD3 DP features have good recognition performance and are very robust to expression changes.