Computer Vision Based Precision Measuring Technologies For Geometric Parameters Of Mechanical Parts

The process of computer vision inspection is as follows, two-dimensional image(s) of the target are captured through a variety of image acquisition devices, then the interested information in the image(s) are extracted and analyzed by image processing software, which will be used to coduct image recognition, image measurement, etc, and finally the inspection results will be obtained. With its merits, such as non-contact, run fast, reproducible, high flexible and reliability, computer vision inspection technology has been widely used in production practice as electronics, machinery, automobile, agriculture, timber and textiles fields, and which is becoming a important inspection method in production of various modern industries.Measurement of geometric parameters such as dimension and tolerance of mechanical parts is a important part of manufacturing process. A number of special measuring instruments are not only expensive, but can only test unitary item. Three coordinate measuring machines and other precision measuring equipment for general-purpose need a long time to test a part and with complexity process. It was taken into account in the dissertation to improve the accuracy and the speed of geometric parameters measurement of mechanical parts, and theories and key techniques of using computer vision inspection technology to measure mini-structure and complex outline of mechanical parts were studied.Hardwares of inspection system are the guarantees in obtaining high-quality digital images, so a reasonable selection must be conduct in accordance with inspection mission requirements. A typical visual inspection system structure was discussed and hardware characteristic parameters and selection principles of vision inspection system were described, meanwhile, the importance of the software for improve inspection accuracy and speed was explained clearly.The sources and characteristics of noises in digital image were explained, and noise elimination methods were discussed, they are mean filtering, median filtering and edge keeping filtering method, by comparative study, their advantages and disadvantages are clear. Gray level threshold method and maximum variance automatic thresholding method for image segmentation were explained, and edge-detection technologies were illiminated in detail, mainly concern first-order differential edge detection operators, second-order differential edge detection operators and Canny edge detection operators, and then these operators were compared with digital image of real mechanical part, the experiment made sure the advanteges and disadvanteges of them. As for extracting image geometric features, the dissertation discussed Hurris and SUSAN corner detection methods, and the applications of Hough transformation method for extracting segmentations and curves in a image.Sub-pixel positioning technology is a software means to enhance the accuracy of visual inspection in an effective way, the dissertation discussed the commonly used sub-pixel positioning methods, and then on the basis of illuminating the least squares support vector regression theory and wavelet theory, put forward a sub-pixel positioning method for image curve named Least Squrare-Wavelet Support Vector Regression(LS-WSVR) method. The influence regulations of parameters of LS-WSVR method, penalty-factor and wavelet scale parameter, for regression precision were found through experiment, when panalty-factor gradually increases, the regression error change on a contrary way, and when the value of panalty-factor is 1000 or more, the regression error is almost constant; wavelet scale parameter has the optimum value, a value more than or less than it can make regression error increase.Calibration is also an important part in computer vision based inspection. Different calibration method should be used aimed at different inspection task. In the dissertation, the detection task is to measure geometric parameters, to make sure how many melemeters each pixel represent is enough, it need not to acquire all of the parameters of the camera, as we all know, inner-parameters and world coordinate parameters, through complex calibration proccess. In order to ensure calibration procission under a ralatively easy way, at first, a small optical distortion lens CCD was made use of, and then self-made model having standard dimension was used to conduct relative calibration.Arc is typical structure of mechanical parts, and center point is a basic parameter of it, based on the coordinates of the center point, radius of the arc can be obtained through simple compute. Least square method usually be used to solving out center point coordinates of a complete circle, but it can not solve out that of a non-complete circle. So aim at arc outline of a mechanical part image, the dissertation put forword a three-point method to solve out the center point coordinates of a arc and then radius of it. Firstly, edge detection method was used to extract pixel level points coordinates of a arc image outline, then these points were used as input parameters of LS-WSVR to get their fit points coordinates. Secondly, choose 1 point from every 15 adjacent points of those fit points, and each 3 points made a group, the coordinates of which were input to the circle function, here 3 points and 3 functions, through solving the 3 functions, a set of arc center point coordinates and radius be acquired. Finally, 3σprinciple was used to eliminate gross error, redid the first and the second step, until there was no gross error, the final arc center point coordinates and radius were the results.A mini-structure of a part was measured precisely based on computer vision inspection technology. A vision based inspection system for measuring Charpy V-notch impact test piece was established in the dissertation, and measuring software was developed, and then the parameters of the V part of the test piece were measured. The inspection steps are composed of image collection, noise elimination, image segmentation, outline extraction, sub-pixel positioning, calibration and necessary computing, while the parameters measured include the angle between the two tilt segments, center coordinates and radius of the little arc, and bottom thick of the V-norch. The dissertation solved the problem that Charpy impact test piece cannot be detected quantitatively.A inspection method based on computer vision technology for mechnical parts having complex curve outline was also proposed. A vision based cam measuring inspection system was built, and the importance of man-machine coorperation in the inspection process was emphasized, meanwhile, the corresponding measuring software was developped. LS-WSVR was made use of to get the sub-pixel level fit points coordinates of the outlines of the center hole and outer contour. For the center hole, the three-point method mentioned above was used to compute the center point coordinates of the center hole, and then roundness error can be obtained. It is a little more complicated to measure the boundary curve of the cam, because the 0o angle position must be found at first, while the dissertation proposed a two-step method to handle this problem. The first step was roughly locating which was realized by hardware, a positioning sign was set on the loading platform, and the operator put the target on the right location, so the place of the cam in the whole image was relatively stable, it meant that the 0o angle position was roughly located, and there would be a start point in rise curve and a end point in return trip curve. The second step was fine positioning, which was carried out by software. According to the derivative characteristic of the outline fitting function and the feature of the curvature the outline, the zero location was finally find on the outerline. Thus the polar radiuses, lift error and profile error were acquired. The results contrast with that from CMM proved that the method of the dissertation can satisfy inspection accuracy need, spends much less time and has a high efficiency. The dissertation provided a non-contact precision measuring method for mechanical parts with complex outline.