Scale-space approach to shape-based image retrieval with application to image thinning

Recent developments in acquisition and storage technology facilitate large collections of digital images where existing database technology captures image content with textual descriptions. Textual descriptions do not adequately capture the subtleties of image content. Therefore, there has been interest in methods based directly on image content. Together with color and texture, shape is an important feature commonly used for image retrieval. Better performance retrieving similar shapes typically requires more features at the expense of processing speed. No universally-accepted shape definition exists, therefore, shape-based retrieval systems have been proposed employing a variety of shape representations, many with refinements to improve performance anticipating application to larger image databases.; Scale-space allows coarse to fine shape characteristics to be used in defining feature sets. Coarse-scale shape features reduce the number of features required to capture shape characteristics, reducing computation and storage. We make two contributions applying scale-space to shape-based image retrieval. The most prominent ridge-line (MPRL) locates topographical ridge-line points on an intensity-surface image representation in scale-space based on a shape region, such that ridge-line points have greatest contrast to adjacent points. MPRL points' scale corresponds to the coarseness of the structure it represents. An image pyramid implementation captures shape characteristics with significantly fewer features, and a ridge-following locates ridge-line points avoiding topographical computation. A unique feature of the MPRL method is the ability to retrieve images using shape features over a scale range. The MPRL has been used to develop algorithms for binary and grayscale image thinning.; The difference between the shape radius of a boundary point with respect to the shape centroid and the radius of the shape's primeval circle defines the Radius Difference Function (RDF). Extrema of the RDF in scale-space comprise a feature set with fewer features at coarser scale. Results show that shapes may be aligned at coarser scales with little or no performance loss, and that using coarser scales for retrieval gives equal or better performance. Perceived boundary shrinkage associated with increased scale is eliminated with the RDF. The RDF is able to represent convex shapes that cannot be represented by methods that use curvature points of inflection.