3D Face Occlusion Determination Algorithm Research

Face occlusion determination and processing technology is an important technology in the field of computer vision,which has a wide range of application prospects.Three dimensional face recognition technology breaks through the bottleneck of two-dimensional face recognition technology and it has high robustness to external interference such as illumination and posture.And it has become a research hotspot at home and abroad.Different from the common threedimensional face occlusion determination algorithm based on curve,model and multimodal methods,algorithms based on the original features of point cloud such as abrupt edge,normal vector and shape structure and so on can achieve occlusion determination more efficiently.Based on the deep analysis of 3D face occlusion,3D face occlusion determination based on the features is studied in this paper.The main research work and innovations are as follows:1)A 3D face occlusion determination method based on Multiscale Uniform LBP features of Multi Bit-Plane Slicing is proposed in this paper.In order to solve the problem of large amount of point cloud data and the long time-consuming process of feature extraction and training in 3D face occlusion determination,the method of cutting the bit plane of face depth map to get the cutting plane of each figure is used as the important part in this method.Then synthesizing the multi bit planes to obtain the feature extraction plane.At the same time,the faces are divided into 12 sub regions according to the horizontal and vertical coordinates.The mu LBP features are extracted,and then the feature vectors of each sub region are concatenated and the mu LBP-MBPs are obtained.It is shown that this method has strong detection and classification performance,simple principle and easy implementation,and a competitive 3D face occlusion determination rate is achieved.2)In order to solve the problem that the general methods of 3D face occlusion determination need not only point cloud data but also additional features such as texture and colors,a method based on Local Entropy of Normal Vector Azimuths is provided.Only point cloud normal vectors are used as the initial feature to describe the shape change of 3D face occlusion in this algorithm.In order to separate the occluded regions accurately,the face division method based on the abscissa and ordinate coordinates is used the.Then the entropy value of the normal vector azimuths of each point cloud sub region are calculated.Finally,the entropy values of the three azimuth angles,pitching angle,deflection angle and roll angle,are concatenated to obtain the training required eigenvector.It is shown that the occlusion situation can be described correctly with this algorithm,and a better result can be gained as well.3)A 3D face occlusion determination algorithm based on RS-CL-CNN model is proposed in this chapter.Point cloud learning and analysis is very challenging,because the difficulty to capture the hidden shape features in disordered points.In order to solve the problem of geometric topology information loss between point clouds in the classical regular mesh convolution method,RS-CL-CNN model extends regular mesh convolution to irregular configuration for point cloud learning and analysis.The core of RS-CL-CNN is to learn from the geometric relationship of point cloud,that is,the geometric topological constraints between points.With the help of this learning method,the feature description of local point set shape can be obtained,and the explicit reasoning of point spatial layout can be realized,which has good robustness.In the same time,the Circle Loss function is used in the model.The Circle Loss has the advantages that increasing the distance between features and reducing the distance within features,which enhance the ability to distinguish occlusion of this model.It can be seen that the superiority of the model for point cloud convolution learning is obvious.The experiment results proved that the good performance of 3D face occlusion determination can be acquired by the RS-CL-CNN model.