Identification of similar shape contours based on the curvature extremes description

The use of the curvature extremes of the smoothed contour for a reliable shape comparison suitable for shape retrieving is analysed. The shape contours are smoothed using linear diffusion of the contour. A diffusion scale proportional to the contour dimension is chosen, improving the uniform scaling invariance. The relative positions of the smoothed contours curvature extremes are compared, resulting in a shape similarity measure. This method is compared with the resulting from comparing the maxima of the scale-space maps of the contours' curvature zero-crossing (CSS), used in MPEG-7. It is shown that those maxima always coincide with a curvature extreme at the maximum scale. Retrieving examples using the proposed method, are compared with the results of a CSS implementation. A visual improvement of the retrieving results was found. The results were also compared with a polygonal approximation based method.