Represent image contents using curves and chain code

This paper presents a new approach to extracting and representing structural features of images. The approach is based on both a region-based analysis and a contour-based analysis. An image is first segmented into regions, after which an area of each region is computed using a number of bound pixels. A contour of the region is approximated by a B-spline curve to get its control polygon. A convex hull of each region is then computed from the obtained control polygon. Ellipses are used to intuitively represent structural components in the image. Each ellipse is centered at its corresponding convex hull’s centroid and is approximated from the convex hull. Color and spatial information of each region is also used. In a multi-scale setting, the Chaikin’s algorithm is iteratively applied to the obtained B-spline control polygon in order to get finer control polygons, which in the limit results in the B-spline curve. To compare similarity between the control polygons and their higher-level refined control polygons, we use the chain code representation in concert with the Chaikin’s algorithm. Our method could be applied in many applications including image retrieval, image classification, image clustering, image understanding, image manipulation, pattern recognition, and definitely in machine vision.