Research On Point Cloud Feature Extraction And Enhancement Methods Based On Deep-Learning

With the rapid development of 3D scanning equipment,3D point cloud data has been widely used in various fields.However,during the 3D scanning process,environmental factors or limitations in the scanning accuracy of the equipment itself may cause the sharp features of the scanned point cloud data to become smoothed out.Therefore,in order to improve the quality of point cloud data,further processing of the initial scattered point cloud data is necessary to achieve feature enhancement.In recent years,deep learning techniques have been widely applied in various fields.This article explores methods for estimating point cloud normals,extracting feature lines,and enhancing features using deep learning.The specific contents are as follows:(1)A new method for estimating point cloud normals was studied,which uses self-attention mechanism and neural networks to process point clouds,aiming to solve problems such as poor robustness and inaccurate estimation of sharp feature parts in existing methods.Specifically,this method introduces a self-attention mechanism based on the PCPNet neural network to encode the input point cloud and accurately estimate the normal vector of each point by enhancing the feature extraction ability of the network.Experimental results show that this method can effectively estimate the normal vectors of point clouds and has significantly improved accuracy and robustness compared to PCPNet.(2)A point cloud feature line extraction method based on feature fusion was studied.This method first uses a local feature fusion algorithm to extract feature points,and then uses an improved feature line extraction algorithm based on these feature points to obtain the feature lines of the point cloud model.Specifically,this method concatenates neighboring features and expands the feature input to better extract feature points.Based on the existing feature line extraction algorithm,an angle constraint for growing points was added to limit some outlier feature points,making the feature lines smoother.Experimental results show that this algorithm can extract feature points of the point cloud model completely and accurately extract the feature lines of the point cloud,and the extracted feature lines are smoother compared to existing methods.(3)An algorithm for point cloud feature enhancement based on spatial and normal constraints was studied.To address the problem that the feature enhancement effect using only spatial constraints is poor,a neural network-based algorithm was studied that improves the performance and accuracy of the network by adding normal constraints and spatial uniform distribution to achieve feature enhancement.Experimental results show that this algorithm can effectively enhance the edge features of CAD point cloud models and has strong robustness.