Study On The Theory And Methodology Of Workpiece Surface Defect Detection Based On Image Processing

Workpiece surface quality is one of the important indicators to measure the quality of the products,which not only determines its value in the market,and directly affect the subsequent processing or the security and stability of mechanical parts in the work.As a result,more and more attention has been paid to the quality detection of workpiece surface.Currently,the methods used to detect the surface quality of workpiece contain traditional manual testing and non-destructive testing technology,and these methods exist some limitations,such as low efficiency,high cost,and the low degree of intelligence.In recent years,with the rapid development of computer technology,the advantages in surface defect detection of machine vision technology on the basis of image processing are more and more obvious,so its application in the industry has been rapid development.The defect detection technology of workpiece surface based on image processing mainly includes the following procedures:image acquisition,image preprocessing,defect segmentation and target feature extraction and recognition.However,some problems still exist in the process of detection for the traditional calculation.Firstly,there is a contradiction between the image details and noise filtering for the traditional filtering algorithms.Secondly,the traditional segmentation method,such as threshold segmentation method,is powerless for weak target segmentation under complicated background.Thirdly,it is difficult for the traditional feature extraction methods to take into account the global and local structural information of data set.Therefore,how to further improve the comprehensive performance of the traditional algorithm,and applied to industrial production practice is the main purpose of this dissertation.The researches mainly include the following several aspects:(1)Aiming at giving consideration to the noise suppression and image details protection in the filtering process,an improved filtering algorithm is proposed based on the weighted median filtering and mean filtering method.The weight of proposed filtering algorithm can be adaptively determined according to thedensity changes of the salt and pepper noise,which make the algorithm be adapted to any density of salt and pepper noise.In addition,small window filteringand integration of the advantages of median filter and mean filter algorithm are used in the presented filtering algorithm,which makes the algorithm can get better image detailpreserved while suppressingthe noise.The results from the application of the proposed algorithm in Lena and engineering images filtering show the feasibility of the proposed method in noise suppression.(2)Aiming at overcoming the shortcomings of the single threshold algorithm in the segmentation of weak targets under complex background,a pixel searching segmentation algorithm based on block is proposed.Firstly,the image is divided into small blocks,and then excludes those image blocksno containing defect targets.Secondly,calculate the gray difference of adjacent pixels in the rest image blocks to divide the defect targets and background.In the segmentation experiment of the weak defect,the defects separated successfully have verified the effectiveness of the presented algorithm.(3)In order to overcoming the lower segmentation accuracy of hard segmentation method for weak target segmentation under uneven distribution of background gray,and improve the robustness of segmentation algorithm to noise,a segmentation method based on the improved data field and FCM clustering algorithm is proposed.The established image data field using the concept of improved data fields,which not only achieve the purpose of balancing the background gray level of image,but also provide a basis for FCM clustering algorithm to determine the initial cluster centers.In addition,in order to avoid the problem of discontinuous region appeared after segmented,combined the potential value of image data field and grayscale variance to improved the membership function of FCM,to a certain extent,which can ease the internal discontinuity problem.Furthermore,in the process of clustering,a new clustering model,which is image data field as the main model jointing assistant model of gray image,is proposed.At the same time,the restrictive relationship betweenneighborhood pixels isalso introduced to the image data filed,which can strengthen the algorithm robustness to noise.Moreover,Compared to other related algorithms,the experimental results show that the proposed segmentation algorithm has certain advantages under complex background of weak target segmentation,and has some immunity to noise.(4)Aiming at the difficult problem of selection neighbor number kin LPP algorithm,a LPP feature extraction algorithm based on Fuzzy k nearest neighbor is proposed.By calculating the membership value of the two samples,the nearest neighborkcan be adaptively determined,which not only provides the basis for the selection of the nearest neighbor number in the LPP algorithm,but also optimized the LPP algorithm in feature extraction.The feasibility of the proposed algorithm is verified in the experiments of grayscale images and two valued images.(5)In order to further improve the feature extraction algorithm robustness to noise in the image space domain,and give attention to both global and local structural information of high-dimensional data set,a new feature extraction method named global-kernel local preserving embedded based on multi-scale transform(MGKLPE)is proposed.Taking the curvelet method to transform the gray image into frequency domain space,which can reduce the algorithm sensitivity to the noise to some extent.Aiming at overcome the shortcomings of LLE algorithm without clear projection matrix and LPP algorithm no finding the nonlinear structure of high-dimensional data sets,a method of the fusion of LPP and LLE algorithm is proposed,which can keep neighborhood structure information of samples.In order to descript the targets informationcomprehensively,the fusion of PCA,LPP and LLE method is proposed,and the method stillemployed the kernel function K and coefficient matrixS,whichcan make the algorithm adapt to nonlinear data sets,moreover avoids the problem of matrix singularity.Through the Swiss-roll experimental data set,the effective property of GKLPE algorithm in keeping the global and local structural information is verified.The results from engineering image database and binary image database show the validity of the algorithm in the feature extraction and recognition.