Research Of Vectorization Methods On Industrial Ct Images Based On Chain Code

The CAD model reconstruction of reverse engineering has becomed a mainstream design method which has been wildly applied in industrial areas such as aviation, aerospace, die, automotive and ship. It can greatly shorten the design cycle of modern product, improve market competitiveness. It’s also a good way for technology introduction and technological improvement. Industrial CT is an advanced non-destructive testing technology, the vectorization technology of industrial CT images is the basis of the automatic detection and3D reconstruction of parts. This dissertation mainly researches the vectorization methods of industrial CT images, which mainly include:1. Research image thresholding methods. A fast single threshold segmentation method is proposed. We extend the between-cluster variance in continuous domain and prove that the between-cluster variance curve has only one peak, the threshold can be quickly got by applying one-dimensional search. Since the trough points of histogram curve corresponding to the threshold for the segmentation, we analyze the multi-scale gray difference histogram and get the candidate sets. Combination of the two methods then a fast image segmentation algorithm can be achieved. For multi-material images, an adaptive multilevel thresholding algorithm based on immune genetic algorithm and Gaussian mixture model for image segmentation is proposed. The method allows the determination of the appropriate number of thresholds as well as the adequate threshold values. The threshold candidate set with limited valley points can be attained by means of transforming the histogram with the right scale continuous wavelet. Then, the number of thresholds and the quasi-threshold values are determined by using the immune genetic algorithm. The parameters of Gaussian mixture model can be received by the way of fitting the histogram with the quasi-threshold values. Last, the segmentation thresholds can be attained by using of the minimum error criterion. In order to enhance the anti-noise performance of segmentation algorithm, multi-dimensional histogram segmentation methods are researched. For reducing the computational complexity of multidimensional extension, the dimensionality of the two-dimensional and three-dimensional histograms is reduced. The thresholding method of the histogram after the projection can effectively suppress random noise and Gaussian noise.2. A binary images identification method is proposed. The run-length structures and run-length chains are dynamically obtained in the binary images. The run-length chains are the elements of seed growth, and a rapid identification of the binary image method is achieved. The contour extraction and tracking method of identified regions with the chain code technology is introduced. Using the FC4for contour extraction and the FC8for contour tracing, then an orderly and single pixel contour of region is achieved, and the chain code sequences of the region are obtained. The attributes of chain code sequences are calculated, and the hole filling method based on contours bound is proposed. Using the contour obtained by the chain codes as the initial contour, the application of Snake Model can get the precise contours of the region iteratively. 3. The algorithm of contour feature points extraction and contour feature recognition of chain codes is proposed. The contour feature points can be got by threshold segmentation and non-maxima suppression with difference of N chain codes histogram, then the contours are automatic divided by the feature points. The contours are identified as line or circle (arc) by the characteristic of the sum and difference of N chain codes. The identified contour is fited by using the minimum quadratic method of the geometric distance, and the fitting parameters can be calculated by Lagrange multiplier method.