Texture Classification and Verification Using Bispectral Estimates at All the Frequencies on a Lattice

Digitized texture images can often be considered as realisations of stationary random fields and their estimated normalised bispectra at diagonal frequencies have been used for classification. In this paper, we devise a classifier and a hypothesis test using bispectral estimates at all the discrete frequencies on a lattice. For flexibility textures in each class are allowed to have a certain amount of variation in theoretical normalised bispectra from that of a single training sample for the class, and we also consider the case where the texture to be classified does not belong to any of the classes. The novelty of this paper is in a) the flexible formulation of the problem, b) the known asymptotic distribution of the classifier, c) the inclusion of a verification stage and d) the complete coverage of frequencies on a lattice. The methodology is extendable to 4th and higher order spectra and is applicable to random field data other than texture images.