On The Irregular Planar Fragments Matching

Irregular fragments matching and restoring is an important issue of pattern recognition and computer vision etc. It widens the application area of pattern recognition. Furthermore, it is of widespread practical use in the archaeology, paleontology, medicine etc. Based on Image Processing, Differential Geometry, Pattern Recognition and shape matching, this thesis studies some key problems of the irregular planar fragments matching. The fragments' shape is described by the unchangeable characteristic of contours. Accordingly, fragments matching can be converted to curve matching. The research work can be summarized as follows:1. The methods about fragments digital, extracting and pre-processing fragments' outlines are introduced. For the scrapes of the paper studied in the thesis, the image boundaries of fragments are first extracted, and then the contours are filtered and resample.2. On the basis of analyzing the previous algorithms of corner detection, a staged method of detecting corner points of contours is presented. The method overcomes the shortcoming of single-scale technique which can only detect feature points on some special scale and of the multi-scale algorithm which lacks geometric visibility and needs selection of the scale parameter.3. Due to that the interpolating methods have large calculating quantity, and the precision of the least-square methods is comparatively low, an algorithm fitting thecontour with constrained B-spline curve is presented. The fitting curve interpolates the corners and approximates other points in the sense of least-square. The examples indicate the algorithm is effective.4. Two methods of computing discrete curvature are presented. One is based on the B-spline fitting to the contours;the other computes the curvature directly taking advantage of the relation of curvature and osculating circle. The algorithm is easy and very precise.5. A fragment matching algorithm is given. First, all matching corner points between two fragments are detected by computing the edit distance of the two curvature strings with the method from Robert A. and Michael J[90l Then detect the longest common sub-string between any two adjoining corner points with Chen Hong-ming's'39' method.