Study On Key Technologies Of Online Vision Measurement For Large Scale Parts

A computer visual measurement technique owns advantages of non-contact, no damage and with the adaption of dangerous occasion and etc., but few researches are conducted on the theory of online vision measurement for large scale parts. Supported by the foundation for the research on the measurement method with high-precision of large scale mechanical parts in the production line of the NSFC(50805023), this paper is dedicated to study the key technology of online vision measurement for large scale parts. And that provides theory and technology foundation for improving the degree of automation in the industrial production.The main contributions of this dissertation are as follows:(1) A set of practical online vision measurement system is designed.According to the function demand and technical process of online measurement for large scale parts, a set of practical online measurement system based on monocular vision is designed. The platform is set up by selecting light reasonably, designing illuminating system correctly and considering the ratio of quality and price of camera, camera lens and etc. In addition, a set of software is developed at the same time.(2) A fast clarity evaluation algorithm based on minimum span of edges is presented.According to the online measurement demand for large scale parts, an edge span function evaluating image clarity is defined. The value of edge span is calculated by a dynamic focusing and a fast searching strategy. And the minimum value corresponds to the clearest image. The comparison results show that the new algorithm is faster, which is over 30% faster than that of the fastest absolute gradient function and the new algorithm also is higher sensitivity, better unimodality and better anti-noise than those classical optimal focusing functions.(3) A fast edge detection method based on automatic threshold is proposed.A fast edge detection method based on automatic threshold by fusing gray values of background and foreground is proposed on the basis of a large number of experiments, analyses and verifications for images of large scale parts. Target points are those gray values of background being greater than threshold for the first time. The comparison experiment results show that not only the accuracy of the novel algorithm is better than that of Roberts algorithm, and is close to those of subpixel methods, but also the speed of it is much faster than those of existing methods.(4) A fast algorithm based on chain code tracing is presented to eliminate rotation deviation of images.According to the limitation that eliminating rotation deviation is the prerequisite of image matching, the length of each continuous code of parts’ sequence images with straight contour line is obtained by chain code tracing. Then different algorithms are selected by the length. The angle between the line and the axis is obtained. Thus the rotation deviation can be eliminated. Experiment results show that not only the precision of the new method is enhanced but also the speed of it is over 20 times faster than that of Hough transform.(5) Three kinds of improved algorithms for template matching for different kinds of parts are proposed.Algorithm 1:Objects with the same linear texture on their surfaces and the texture is in the same direction with axis, a novel algorithm comparing neighboring pixels is proposed. A matching template is obtained by accumulating the number of same values of comparing neighboring gray values. Then the fast matching is completed by sun of absolute differences (SAD). Experiment results show that the precision of the new method is the same as that of normalized cross-correlation (NCC), but the speed of it is about five times faster than that of NCC.Algorithm 2:For large scale parts without distinctive characteristics on their surfaces, a novel template matching method by calculating difference values of those neighboring pixels respectively is presented. Matching templates are obtained by that. Then the fast matching is completed by SAD with automatic threshold. The noise immunity of this algorithm is consequently enhanced.Algorithm 3:A fast matching algorithm by adding a manual label in vision while outside parts is proposed. The periodic label easy to be extracted is designed according to transmission speed of parts, acquisition speed of image and overlapped region of images. And the label is added in vision while outside large scale parts. Then the panoramic image is obtained by SAD with automatic threshold after the label is segmented and is enhanced and the number of periods is calculated. Thus no matter there is character on the surface or not, the matching for large scale parts is converted into the matching for the label. Then the matching speed and the matching probability are greatly enhanced, and objects are broadened without additional processing.