Shape Analysis Based On Curvature Scale-Space

Curvature scale-space (CSS) technique is selected in MPEG-7as a standard shapedescription technique owing to its outstanding properties. This paper mainly studies theimprovement of shape analysis algorithms based on the CSS technique, including shaperetrieval and corner detection.Shape retrieval is also called shape similarity retrieval, which generally contains twosteps: shape description and shape matching. In the first part of this paper, a new shapedescription method based on the CSS technique is proposed, called the parabolical fittingdescription method. Firstly, the shape evolves with the Gaussian function and the CSS mapof its curvature zero-crossing is constructed. Then a parabolical fitting of the CSSarc-shaped contours is conducted. Finally, the shape is described with the fitted parametersets. This part also provides a shape matching and retrieval technique based on theproposed CSS description method. Because of the improvement on shape description andmatching steps, the CSS retrieval efficiency increases greatly when it is applied on twostandard shape databases.CSS corner detection usually detect points whose curvature keeps local extremum andcan exist in a high scale during the curve evolution. In the second part of this paper, weproposed a succinct CSS corner detection algorithm. Firstly, the curvature extremum pointsare extracted during the process of curve smoothing. Then a backword-correction strategyis applied to correct the location of the detected corners. Finally, the detection result isobtained. Numerical results indicate that the new CSS corner detection algorithm is muchbetter than the existing algorithms in the aspects of computation efficiency and theaccuracy of corner location.