Extracting global structures using perceptual grouping

One of the problems in the study of image recognition is to extract the global structure, e.g., the contour of the object, with a high reliability. There already exists a technique in which the model figure is pre-determined. This gives a fairly satisfactory result. It is essentially desirable, however, that the system can handle the unknown figure. This study considers a approach to such a problem, where the perceptual grouping as pointed out by Gestalt psychologists is applied. More precisely, the method is as follows. The connected component elements are extracted from the input image. Then, based on the factors for perceptual grouping (such as proximity, similarity, good continuity and closure), which are quantitatively given expressions by the psychological experiment conducted by the authors, elements which have high possibilities of forming a global structure are determined and connected, based on such information as relative position of elements. Using the evaluation function, the optimal candidate for the figure is extracted from the result. When observing overlapped multiple figures, a human can separate the figures (figure segmentation). This aspect also is considered in this study. The foregoing idea is implemented on a computer. Simulation experiments are made for several patterns, and the operation of the proposed system is verified.