Structural analysis of natural textures by Fourier transformation

Abstract

A new method for the structural analysis of texture is presented. Most of the structural analyzers of texture so far have used a “bottom-up” analysis procedure: first, they extract texture elements, and then describe their attributes and mutual spatial relations. In these methods, however, segmentation algorithms for extracting texture elements tend to be sensitive to various degradations of images such as random noise, shading, blur, geometric distortion, and so on. Here a “top-down” analysis procedure is proposed where the placement rule of the texture elements is recognized by using the Fourier transformation. At the first step of the analysis, two spatial vectors representing the placement rule of the texture elements are extracted by examining the energy distribution in the Fourier power spectrum. Then with the help of the extracted vectors and the phase information of the Fourier transformation, texture elements are extracted by a region growing method. Since the global structure (placement rule) of the texture is given, the region growing process can correctly extract texture elements without being affected by degradations of the image. The method has been evaluated in several experiments, in which natural textures taken from P. Brodatz ( Textures, Dover, New York, 1966) were analyzed. Although the class of textures to which this method is applicable is limited to those with a two-dimensional lattice arrangement (not necessarily orthogonal), the hybrid analysis in both the frequency domain and the picture space has proved to give fairly good results even for texture patterns of poor quality.