Texture segmentation by least squares filters

Abstract In this paper a new approach for texture segmentation, the least squares filter method, is proposed. By calculating the least squares value of the correlation over the original image with each background textured image which contains only one kind of texture, the least squares filter (LSF) for each kind of texture is obtained. After filtering the original image with LSF, the scalar label is adopted, finally, to segment the image into many regions. All filters can be interpreted as adapted filter masks for which the designation of the least squares filters (LSF) is proposed. With the LSF method, we have r energy output images for any original image composed of r kinds of textures, which can be used to segment efficiently the original image. For some application the LSFs could replace the empirical filters in Laws' texture energy measures and theoretical filters in Ade's eigenfilters.