Generating stable structure using Scale-space analysis with non-uniform Gaussian kernels

Abstract

Coarseness and directionality are important sources of information for texture image recognition. They are especially important for distinguishing between the textures and understanding the characteristics of similar textures. So we propose a new Scale-space analysis with non-uniform Gaussian kernels in order to find a stable image in which coarseness and directionality are not affected comparatively. We analyze zero-crossing surfaces to generate non-uniform Gaussian Scale-space from observations of a limited number. Singular points, where the topology of zero-crossing surfaces change are plotted on a new Scale-space. A filter parameter, which correspondeds to the maximum-size chunks enclosed by topology change surfaces was selected as the optimal parameter of a pixel. Optimal surface parameters for all pixels are calculated through observation of a limited scale, using parameter surface analysis of an optimal filter in Scale-space. The optimal filter and the image description are calculated by this approach to achive a natural image.