Multiple resolution image segmentation using four QP supports of 2D autoregressive model

Olivier Alata, Pierre Baylou, Mohamed Najim

Abstract: In the framework of a model based approach and using Bayesian estimation techniques, one can improve the results of image segmentation algorithms. In such cases, the texture field is modeled by a 2D autoregressive model. In previous works, segmentation algorithm derivation was based on the prediction error calculated from the first quadrant quarter plane support [Bouman and Liu, 1992]. In this paper, we introduce information extracted from the estimation of the four linear prediction errors calculated from the four quarter plane supports in order to solve boundary problems and to propose isotropic local criteria. Simulation results using the multiple resolution segmentation algorithm [Bouman and Liu] with single quarter plane and four quarter plane criteria are provided.