Key Techniques Research On Data Processing For The Defect Inspection Of Product Surfaces Based On Machine Vision

Surface defect inspection is an important process to ensure the product quality in manufacturing of product. With the extensive application of detection techniques based on machine vision, surface defect inspections of products based on 2D image and 3D point cloud have attracted more and more attention. The rate and geometric precise of defect detection depend on the techniques of data processing used in these methods. Although dimensions of 2D image and 3D point cloud are difference, there are common areas in both methods considering their principles in the defect detection and methods in the data processing. Therefore, key techniques of images and point cloud data processing for the defect inspection of product surfaces are researched in this thesis.The main research and contributions of this thesis are as follows:1. For the problem of missing small defects in defective images and false inspections in defect-free images, the defect segmentation based on image methods is studied. A weighted object variance (WOV) of the between-class variance method is proposed for the surface defect inspection. An adaptive parameter is weighted on the object variance of the between-class variance. The weight ensures that the threshold always be a value that locates at the valley of two peaks or at the left bottom rim of a single peak histogram. It is essential to have a high detection rate and low false alarm rate for the defect detection. The segmentation results of steel, wood, fabric and rail images demonstrate the effectiveness of the WOV method to form thresholds close to desired values. The rail defect inspection system based on the machine vision is developed to validate the performance of the WOV method in the application of machine vision inspections. Comparing with other methods, the WOV method provides better segmentation results with a higher detection rate and lower false alarm rate of the detect.2. Deviation information of the curvature and normal vector estimation is the foundation of point cloud processing. In order to improve the accuracy of curvature of a point, a parametric quadratic surface fitting method is employed to estimate the curvature of point cloud. To deal with the problem that the estimated normal vectors of feature points in sharp feature surface is inaccuracy, an iterative weighting method is proposed to correct the inaccuracy normal vector. The proposed method is accurate, and less time-consuming to estimate normal vectors for the point clouds with sharp features and noise.3. Point cloud denoising and refinement usually make the sharp feature smooth to lead excessive processing errors. To address this problem, a point cloud denosing and refinement methods are proposed for preserving the features. The neighborhoods of a feature point are mapped to a Gaussian sphere, and the hierarchical clustering algorithm is executed in the Gaussian map to obtain an optimal sub-neighborhood. Then, the bilateral filter is employed for denoising in the optimal sub-neighborhood area. The proposed denoising method can effectively remove noise while preserving the geometry of original data. The K means clustering algorithm is used to refine point data in a flat area. The adaptive mean shift and the Gaussian map are employed to refine points in the non-flat area. The experiment results show that the refinement data provide less processing error to better preserve the original geometry compared to other methods.4. To detect concave, convex defects and manufacture deviations, the point cloud defect inspection based on data matching methodology is studied. To improve the efficiency and precision of data registration, parts of feature points are selected for the coarse registration and a modified iterative closest point methods proposed for the refinement registration. In order to speed up the convergence rate of the refinement registration, the optimal projection points of measured data in triangulation surfaces of standard data are used as corresponding points of the iterative registration. The minimum orientation-distance of a point to the triangulation surface is used to generate a color map to visualize the defect. The defect inspection method is validated by inspecting the manufacture deviation errors in auto parts and concave and convex defects of the rail.