Shape and structure for image matching and retrieval

Abstract

We report on methods for using object shape and structure for matching a pair of images or an image and a drawing. These methods can be used in combination with the several methods using color and texture already in use in content based image retrieval (CBIR). Matching methods are based on edges, corners, and ribbons and the relations among them and can be controlled for rotation, translation, and scale invariance. Good results are shown on a small but diverse sample of images with man-made structure. Although we have been oriented toward CBIR because image parts are put into correspondence, the methods are also of general use in object tracking, recognition, registration, structural stereo, or perhaps even detection of shot boundaries in video.