Research On Surface Defect Recognition Based On Machine Vision

With the development of modern technology, the demands for precise surface quality are also higher and higher. Light will be scattered if there're scratches and other defects on the surfaces. Light Scattering will result in light energy dissipation, even damage optical elements seriously, which will affect the performance of optical systems.And for different uses optical system, different defects have different impact on the system .The automatic inspection and classification of surface imperfections of precise elements are problems expected to be solved.In light of detection requirement of large-caliber optical element, I proposed a defects detection system based on defects image which is bright in black background theory after analyzing the existing defects detection technology and advanced detection technologies.The image processing system software is designed by the combination of VC++ and MATLAB, which consists of information acquiring and classification . Defects are in random distribution and random shapes on the surface of optical element ,so in order to implement the inspection of large-aperture surface, it is necessary to stitch images by sub-aperture scanning techniques. The system stitches images fast and accurately by using image stitching ratio matching principles.During image preprocessing, in order to preserve the edges of defects, I studied Noise Smoothing Algorithm based on Partial Differential Equations(PDE)and improved its insufficiency, and the result is very satisfactory by using the algorithm. In Edge Detection, because Defects identification is required to achieve um magnitude, zernike moment subpixel edge detection algorithm is proposed. Zenike moment has shortcomings, so an improved zernike moment is proposed and it is proved that the subpixel which coordinates the edge calculated by the improved algorithm is more precise. Then implement binary image feature extraction to acquire digital information and eigenvectors. The different defects size feature (e.g. length, width, area) could be analyzed and evaluated by criterion. Finally, eigenvectors are fed into a fuzzy classifier for the identification of defects classification. Experimental results has proved that the system has good performance to identify defects.