Research On Point Cloud Super-resolution And Completion Algorithm Based On Graph Convolution

Point cloud is a collection of sampling points with spatial coordinates obtained by lidar,which is called "point cloud" due to its large number,denseness and disorder.As a new type of dataset,point cloud has been widely used by researchers in various fields such as autonomous driving and cultural relic restoration in recent years.However,it is not easy to obtain high-quality point cloud data.Lidar may be limited by terrain,weather or acquisition equipment when acquiring point cloud data,resulting in sparse and incomplete point cloud data.As a result,the super-resolution and completion technology of point clouds came into being.Traditional point cloud super-resolution algorithms have strong assumptions on reconstructing surfaces,and cannot perform high-quality point cloud super-resolution reconstruction tasks,and traditional point cloud completion algorithms are even rare.With the introduction of deep learning technology into the research of point cloud super-resolution and completion,the field has made great progress and achieved excellent super-resolution and completion effects.The main contents of this thesis are as follows:(1)Based on deep learning,combined with theories such as channel attention mechanism,multi-head attention mechanism and graph convolution,a graph convolution point cloud super-resolution network model based on attention mechanism is proposed.The network takes low-resolution point cloud data as input data and encodes it through a graph convolutional layer.Further,this paper designs an improved dense connection module that combines attention mechanism and graph convolution,and uses it as the feature extraction layer of neural network,so that the network can not only pay attention to the useful information under different channels,but also enhance the learning ability of the model,and the position information between points can be learned through the graph convolution module.Experiments show that this method can effectively improve the point cloud super-resolution performance of the network.(2)This paper proposes a point cloud completion network model that combines a multi-scale point cloud encoding network and a pyramid decoding network.The network takes the residual defect cloud as input,and obtains three residual defect clouds of different scales through two downsampling,which can retain the shape information of the residual defect cloud to the greatest extent.The residual defect clouds of three different scales are respectively input into the feature extraction network composed of the vector attention-based graph convolution module and the channel attention mechanism,so that the network can make full use of the low-level,middle-level and High-level shape information.Finally,the fused feature vectors are used to generate incomplete point clouds of different scales through the pyramid decoding network.The residual defect cloud is used as the final point cloud completion result.Experiments show that the algorithm proposed in this chapter can effectively improve the point cloud completion performance of the network.