Texture classification based on moments of the spatial size distribution

Abstract

This paper reviews the concept of spatial size distribution, proposed by Ayala and Domingo (2001) and states its applicability for the analysis and classification of textures. The result of this method is a distribution function associated to each texture; since the most usual approach in statistical pattern recognition consists of assigning a vector of characteristics to each sample, instead of a function, a method for reducing the amount of data, keeping the relevant information, is needed. The main goal of this communication is to assess the validity of using several moments of the (1, 1)-spatial size distribution as texture features in the problem of texture classification. The performance of these texture features are compared with the results obtained by using the whole cumulative distribution function sampled at regular intervals. Preliminary results on a small texture database show that a similar or even better performance is achieved with a few texture features, due to the noise reduction that statistical expectations always provide.