Research On Surface Defects Quantitative Detection Based On Machine Vision

In recent years, researchers both at home and abroad carry out extensive research on surface defects detection work with the development of machine vision and image processing technology, and the level of detection and recognition of surface defects is continuously improved. However, the main function of the surface defect detection system is to identify or classify the defects at present, how to accurately and effectively realize the quantitative detection of surface defects is still a problem worthy of further study. In this paper, the quantitative detection technology of surface defect detection is studied based on the existing system. Including the following two parts:In the process of surface defect size calibration of optical components, due to the diffraction broadening phenomenon of smaller sized standard defects when they imaged in the optical surface defects detection system, so the relationship between the pixel size to actual size of surface defects will form a nonlinear range, and the existing least square method cannot evaluate the nonlinear range accurately. So we put out the least square support vector machine modeling instead of the traditional least square method. The optimal parameters of the LSSVM model are obtained by grid search and cross validation. The LSSVM model could reduce the absolute error of fitting to 0.4?m and the relative error to 17% in comparison with the LS model. Using the actual defect data to evaluate the predictive ability of the two models, the experimental result shows that the predictive error of LSSVM model of small sized defects is half of LS, the predictive ability of LSSVM is better than that of LS. Therefore, LSSVM modeling method can improve the detection accuracy of small sized defects, and further improve the quantitative detection accuracy of the system.Combining the line and circle detection principle based on Hough transform and image registration and difference method, we could extract the common glass defects feature quickly. Using the improved pixel distance algorithm to detect the standard line and circle edge in hole crack images, then reconstruct the corresponding reference image, the method can detect the small defects in images accurately. After all the images without defects registration and composition, we obtain an unique reference image for the hole chamfer images, then we can achieve rapid positioning and quantitative evaluation of the defects through difference between the reference image and the image to be detected. In the experiment, we compare the detected results with the optical microscope results of the defects, it is proved that the error of hole crack defects is within 10?m, and the error of hole chamfer defects is with 30?m.While the detection error requirement of the system is 20?m ?30?m, so the testing results meet the requirements of the system. Finally the two kinds of defect grade criterion is given based on the testing results.