Unsupervised multiscale image segmentation

Abstract

We propose a general unsupervised multiscale feature-based approach towards image segmentation. Clusters in the feature space are assumed to be properties of underlying classes, the recovery of which is achieved by the use of the mean shift procedure, a robust nonparametric decomposition method. The subsequent classification procedure consists of Bayesian multiscale processing which models the inherent uncertainty in the joint class and position domains via a multiscale random field model. At every scale, the segmentation map and model parameters are estimated by sampling using Markov chain Monte Carlo simulations. The method is applied to perform colour and texture segmentation with good results.