Texture characterization, representation, description, and classification based on full range Gaussian Markov random field model with Bayesian approach

Abstract

A statistical approach, based on full range Gaussian Markov random field model, is proposed for texture analysis such as texture characterization, unique representation, description, and classification. The parameters of the model are estimated based on the Bayesian approach. The estimated parameters are utilized to compute autocorrelation coefficients. The computed autocorrelation coefficients fall in between –1 and +1. The coefficients are converted into decimal numbers using a simple transformation. Based on the decimal numbers, two texture descriptors are proposed: (i) texnum, the local descriptor; (ii) texspectrum, the global descriptor. The decimal numbers are proposed to represent the textures present in a small image region. These numbers uniquely represent the texture primitives. The textured image under analysis is represented globally by observing the frequency of occurrences of the texnums called texspectrum. The textures are identified and are distinguished from untextured regions with edges. The classification analyses such as supervised and unsupervised are performed on the local descriptors.