Unsupervised textured image segmentation using 2-D quarter plane autoregressive model with four prediction supports

In the context of the model-based methods for image processing, we propose some improvements for an unsupervised textured image segmentation algorithm using a 2-D Quarter Plane Autoregressive model. The segmentation algorithm works in two stages: The first stage consists in an estimation of both the number of textures and the model parameters associated with each existing texture. The estimation is achieved by minimizing a probabilistic criterion which comprises a penalty term such as those used in Information Criteria (IC). The second stage deals with a Maximum a Posteriori estimation of the label field by a Simulated Annealing method. In a former work, Akaike IC (AIC) and a 2-D First Quarter Plane Autoregressive Model with fixed (1,1) order were used. In order to estimate the number of textures and the model orders, we propose to use Bayesian IC (BIC) and $phi_beta$IC. Moreover, during the two stages of the algorithm, the four Quarter Planes prediction supports have been used in order to solve problems at image and region boundaries. The results are given on images containing synthetic and natural textures.