A fast structural matching and its application to pattern analysis of 2-D electrophoresis images

Image pattern matching is essential and important in such applications as registration, stereo-matching and model-based recognition. Although a variety of pattern matching algorithms have been proposed so far, almost all of them are time-consuming and sensitive to geometrical distortion of an image. Thus, from the practical point of view, we have investigated a fast and reliable structural matching algorithm. Our approach to pattern matching deals with the correspondence of two kinds of structured graphs, that is, the Delaunay net and relative neighbourhood graph with attributes, rather than two sets of points. The algorithm estimates a matching merit and uses thresholds to cut off the redundant path and therefore perform a fast matching. Then, graph traverse is carried out by using not only real arcs of the relative neighbourhood graph but also virtual arcs which are virtually extended to the look-ahead node in order to deal with geometrical distortion and thus improve the reliability of matching. We apply the algorithm to the pattern matching problem of 2-D gel electrophoresis images often appearing in the area of biochemistry. As a result, it takes about ten minutes to detect thousands of spots and compare them with those on a reference pattern even with a cheaper DOS/V computer running with Linux, while it takes about two or three hours by visual inspection