An edge-based description of color images

Abstract

This paper presents a method of describing multispectral images, for computer vision applications, in terms of contour elements. Contours are detected, at different scales of resolution, as the zero crossings of a second-order differential operator that represents an extension of the second directional derivative to the multispectral case. A fine-to-coarse analysis of contour behavior in scale space is then used to compute the attributes needed for the description of the image. Subsequent contour segmentation, based on both geometric and photometric features, allows for a further increase in description compactness without significant losses of information. In order to assess the faithfulness of the description, it is shown that an approximate reconstruction of the original image can be obtained from the coded contours. The method has been tested on several real-world color images. Two examples, in which images are described and reconstructed at different degrees of compactness, allowing for an objective and subjective evaluation of the performance of the method, are presented.