Research On 3D Face Recognition Based On Local Curve And Point Cloud Convolutional Network

Three-dimensional face recognition technology has achieved satisfactory results after more than 20 years of research,but there are still some challenges in practical applications.Because 3D face can be represented by point cloud data,the point cloud data not only contains the coordinate information of points,but also contains texture and other features,which results in more calculation in recognition.Moreover,when the facial expression changes,the face surface will produce certain deformation,which will also affect the accuracy of recognition.With the development of deep learning,many scholars have proposed models for point cloud processing,but there is no mature 3D face recognition network model yet.There are two main reasons: one is that there is not enough 3D face data at present,and the other is that the current point cloud convolution model is mainly used for object classification and cannot be applied to the classification task of face recognition with small differences between classes.The main research content of this dissertation includes two aspects.One is to propose a face representation method based on the radial curve of the nose point and the isogeometric contour line to simplify the face data and improve the speed of face recognition.At the same time,shape analysis is applied to curve analysis,and other descriptors are introduced to analyze the geometric invariance of the curve.Then divide the face into a rigid area and a non-rigid area,set different weights according to the influence of different areas on face recognition,and design a weighted classifier.Experiments are performed on a commonly used data set and compared with other experimental methods on this data set.The experimental results show that the algorithm still has a high recognition rate when the facial expression changes.The second is to apply the Point Net++ network to three-dimensional face recognition.For the problem of insufficient training samples,this dissertation expands the BFM2017 data set according to the 3DMM model,and generates different types of samples by setting different shape parameters and expression parameters,and generates enough training samples according to this method.On the other hand,some improvements are made to Point Net++.In addition to three-dimensional coordinates,the input of the network also introduces curvature and normal information.Improve the sampling layer,combine the farthest point sampling algorithm with curvature sampling,so that more points are collected in areas with greater discrimination,and sampling points in areas with less discrimination are more uniform.In the grouping layer,KNN is used to replace the original sphere radius query method,so that the model can obtain overlapping sub-regions based on the face data information.Finally,several loss functions are compared,and their effectiveness under 3D face recognition is verified.At the same time,in order to verify the effectiveness of the method proposed in this dissertation,experiments are carried out on the Bosphorus data set.The results show that the method has a higher recognition rate and robustness of expressions,and speeds up the recognition speed without decreasing the recognition rate.