Moments of size distributions applied to texture classification

Abstract

This paper is mainly concerned with texture classification. A texture has to be classified in a class from a well defined and complete set of classes. The texture features proposed in the paper would be moments of different probability distributions associated with the texture. These probability distributions correspond with different formalisations of the concept of size distribution. Firstly, the well-known granulometric size distribution proposed by Matheron (1975). Secondly, a non-granulometric size distribution related with that proposed by E. de Ves et al. (Sept. 1999) and, finally, the spatial size distribution given by G. Ayala and J. Domingo (Dec. 2001). From the image, the density or the cumulative distribution function can be estimated and, from them, the moments of the distribution. Moments of first and second order constitute the texture features used for comparing the performance of these functions. A typical experimental set up has been performed on a small texture database. The results show the superior behaviour obtained for combining the size and the spatial components, and also give clues to determine the minimum number of features that can provide good percentages of correct classification.