Study On Some Key Problems In Visual Inspection

Vision inspection is a new technique that applies the computer vision to the inspection area. American National Standards Institute (ANSI) has predicted on the basis of the survey that 80% even 90% of the inspection tasks will be completed with visual inspection, so it has very important research value. According to the need of visual inspection for modern advanced manufacturing technology, the thesis firstly builds the imaging model of single-eye visual inspection and multiple-eye visual inspection on the basis of view geometry theory, then extracts the inspecting object through image segmentation, finally puts the emphasis on two key questions in visual inspection: distortion correction and the inspection of depth, puts forward some new detection methods and verifies the effectiveness of these methods.On the basis of solving the imaging question of camera for single-eye vision inspection and multiple-eye vision inspection using view geometry theory, the thesis analyzes two significances for visual inspection combining the parameters inpsection of Petroleum Tube Thread Combs, the counting inspection of paper and the uncertainty question of stereo matching for depth inspection: in the aspects of extracting of inspection objects, the thesis uses the OTSU global threshold value algorithm to segment the Combs image, puts forward the segmentation algorthm for paper images based on peak-valley morph morphological feature, which can adaptively reflect better the local property of paper images; in the aspects of taking segmentaion areaes as mid-level symbolic representation, the thesis uses Meanshift algorithm to segment color images and analyzes the action of the homogeneity of segmentation area in overcoming the uncertanity of stereo matching.The actual imaging system of camera is a nonlinear system, therefore, images are usually distorted and image distortion rectification becomes necessary to ensure the accuracy of visual inspection. Multi-eye visual inspection is only need to restore the linear imaging properties of the imaging system; while single-eye visual inspection is not only need to restore the linear imaging properties of the camera, but also need to eliminate the linear distortion to restore the measure nature of the image. On the basis of the non-linear distortion coefficients obtained by camera calibration, the paper puts forward two methods to solve the distortion rectification of images in single-eye inspection. One method is to separate the linear distortion from non-linear distortion, which are eliminated respectively by position control device and the camera calibration. The other method is to build the differential equation of image distortion field, the complete and accurate description of distortion field can be obtained after the differential equation is resolved, this method makes a uniform treatment of linear distortion and non-linear distortion. For the latter, the paper puts forward the determination method of ideal translation parameter and local translation parameter of the differential equation of image distortion field and gives their theoretical results. The first method needs four freedoms, so the mechanical device is complicated but may be non-accurate; the second method needs only two freedoms, the mechanical device is simplified greatly but want to be accurate.The extraction of depth information is the most research content for three-dimension visual inspection. Stereo matching is the basic method that extracts the depth information, which is usually classified into the WTA-based local optimal algorithm and global optimal algorithm based on the minimizing of energy function. For local optimal algorithm, the thesis improves the Yoon adpative support-weight stereo matching algorithm: Firstly, analyzes, according to disparity assumptions based on color segmentation, the different influences of color similarity and spatial proximity on the weight (disparity) determination, on the basis of which the thesis puts forward a new weight determination model of spatial proximity; Then, extends the Birchfield dissimilarity measurement function from 2-neighbord to 8-neighbord, which solves the question that the AD dissimilarity measurement function adopted by Yoon algorithm is sentive to row sampling and collum sampling , and the matching cost is more accurate. The experiments show that the proposed algorithm has improved effectively the matching performance of Yoon algorithm; For the global optimal algorithm, the paper studies the particularity of general plane matching, the differential geometric description of disparity was introduced on basis of image segmentation to hereby rebuild the energy function and the message transmission process of belief propagation stereo matching, which can reflect better the depth structure of visual scene and can meet better the needs of stereo matching for vision scenes of general planes even curve surfaces.