Research On Point Cloud Enhancement Method Based On Deep Learning

3D point cloud is a massive collection of points acquired by 3D sensors(e.g.3D cameras,LIDAR)that can describe the surface information of an object,and each point contains 3D coordinates,colors and other information.Due to its powerful 3D spatial characterization capability,3D point clouds are widely used in fields such as unmanned vehicles,satellite maps,robotics assistance and medicine,where dense and uniform point clouds are often required for a variety of 3D vision tasks such as identification and classification.However,the acquired point clouds are usually sparse,noisy and non-uniform due to the sensor acquisition characteristics and the properties of the objects themselves,while the compressed transmission of point clouds often loses feature information and degrades the quality of point clouds,which brings great difficulties to the subsequent processing of vision tasks,therefore,it is important to study the enhancement of low-quality point clouds.Traditional point cloud enhancement methods often rely on human a priori knowledge and have limited enhancement performance for complex point clouds,while point cloud enhancement based on deep learning uses a data-driven approach to learn the distribution characteristics of point clouds themselves without relying on human a priori knowledge,and has achieved better results.Therefore,this paper is selected to study 3D point cloud enhancement methods based on deep learning,aiming to study methods to enhance the effect of point cloud enhancement,mainly accomplishing the following tasks:(1)Design Point Upsampling with Neighborhood Graph Convolution(PU-NGC),a point cloud upsampling network consisting of four parts: Patch extraction,feature extraction,feature expansion,and coordinate regression.First,Patch extraction is performed on the point cloud to make it suitable for network training,and then in the feature extraction network,a Graph Convolutional Network(GCN)-based neighborhood feature extraction module is proposed to design the neighborhood GCN module to extract the neighborhood features of the point cloud,and a dense connection is used to connect multiple neighborhood feature extraction modules to extract the feature information under different Secondly,we use the neighborhood map convolution module in series with Shuffle operation in the feature expansion module,which can avoid the local point stacking phenomenon caused by direct replication and expand the number of points without increasing the computation.coordinates of the point cloud.The experimental comparison with existing methods on PU1 K dataset shows that the point clouds generated by the proposed PU-NGC network are closer to the original point clouds than other methods in both evaluation metrics and visualization analysis.(2)To address the problem that the point clouds generated by the upsampling network will generate holes after surface reconstruction,a self-attention-based correction network is proposed to further correct the dense point clouds.The network consists of a local feature correction module and a global feature correction module,respectively.The local feature correction module focuses on the local geometric structure and corrects it by calculating the spatial weights of the regression of feature differences between the center point and the surrounding neighboring points;the global feature correction module generates a global attention map through the self-attention module to correct the point cloud by constraining the overall shape of the object.Finally,this paper further connects the upsampling network with the correction network in series to optimize the data training strategy and design the joint loss function for end-to-end training.The experimental results show that compared with the single upsampling model,the point cloud enhancement model reconstructs the point cloud surface with significantly fewer holes after cascading the correction network.