Research On 3D Detection Method Of Steel Plate Surface Defect Based On Point Cloud Data

With the rapid development of computer,digital image processing,sensor and other technologies,surface defect detection based on computer vision has become one of the research hotspots in the field of machine vision.At present,most of the related researches use image processing or machine learning to detect and analyze 2D images,which are restricted by the information dimension of the planar images.These methods are usually difficult to accurately identify defects with depth information,and this type of defects often plays a decisive role in the quality of the steel plate.Therefore,the defect detection method based on 3D point cloud data has important research value.This paper studies the problems of point cloud data reconstruction and defect segmentation based on low-parallax multi-view image sequences.The main work includes:1.A steel plate test block with surface defects was prepared and a multi-view image data set was obtained on this basis.A Structure from Motion(SFM)algorithm to optimize the graph matching process was proposed for the low-parallax characteristic of the image data set.This method uses Fisher vectors to generate clusters of similar graphs,thereby reducing part of the repetitive and unnecessary matching process in a parallel computing method of intra-class matching and inter-class fusion.On this basis,a dense cloud reconstruction method was used to obtain a point cloud model on the surface of the steel plate.Experiments show that this method significantly improves the running efficiency of the algorithm and obtains higher reconstruction accuracy than traditional methods.2.A RANSAC plane segmentation algorithm based on hierarchical adaptive Octree optimization is given.This method uses bilateral filtering to eliminate noise points in the initial point cloud,and then establishes the Octree data structure of the point cloud by dividing the hierarchy by curvature judgment.Without affecting the characteristics of the point cloud structure,the data redundancy is reduced,and the parameter range estimated by the RANSAC model is determined through parameter experiments.Compared with the conventional RANSAC point cloud segmentation algorithm and other major point cloud segmentation algorithms,this method greatly reduces the running time of the algorithm and obtains good segmentation results.