Research And Implementation Of 3D Model Segmentation Algorithm Based On Convolutional Neural Network

3D model segmentation is the basis for further understanding and analysis of 3D models.It is not only widely used in research fields such as deformation,simplification,and retrieval of 3D models,but also plays an important role in model processing and practical applications such as ancient cultural relics protection,medical detection,and geometric compression.At present,the 3D model is mainly represented by triangular grid and 3D point cloud.Point cloud segmentation is the difficulty in segmenting 3D models.With the continuous development of deep learning and the neural network,domestic and foreign research scholars have also used it to conduct a lot of researches on point cloud segmentation.According to the analysis of the three-dimensional shapes,the neural network is used to efficiently process the 3D data,and the Convolutional Neural Network is outstanding in the segmentation task.Aiming at the problem of how to improve the accuracy of segmentation and reduce the memory consumption and running time during training,a point cloud segmentation algorithm based on Convolutional Neural Network is researched in this paper.The main work completed is as follows:(1)Introducing the characteristics and differences of point cloud segmentation in traditional algorithms and deep learning algorithms,discussing four types of segmentation methods for 3D data based on voxel,popular structure,multi-view and original point cloud combined deep learning,and expounds related instructions such as deep learning and Convolutional Neural Network.(2)Proposing a point cloud segmentation algorithm based on octree with Convolutional Neural Network,constructing a Convolutional Neural Network based on the Caffe framework for point cloud segmentation algorithm,having a voxelization of the point cloud and constructing an efficient octree structure,so that the multi-dimensional input features after the feature fusion and octree information of the point cloud are entered into the Convolutional Neural Network for training.By constantly adjusting the parameters of the multi-layer in 3D convolution structure,the higher-dimensional information of the point cloud models are extracted,and the prediction labels of segmentation algorithm are obtained,in addition,optimizing the results of segmentation algorithm by CRF algorithm,and realizing the segmented visualization results by PCL.(3)The sub dataset of ShapeNetCore is used for experimental verification,the accuracy is compared with the existing deep learning algorithm by the IOU indicator.This algorithm is superior to full voxel algorithm in terms of the GPU memory consumption and running time.(4)This paper implements the demand analysis and overall design of the point cloud segmentation system based on the convolutional neural network,and uses VS2015,Qt and PCL to achieve it,and finally meets the system requirements and completes the design of each functional module.After functional testing,a three-dimensional point cloud segmentation system was realized.