Dempster-Shafer Parzen-Rosenblatt Hidden Markov Fields for Multichannel Image Segmentation

Theory of evidence has been successfully used in many areas covering pattern recognition and image processing due to its effectiveness in both information fusion and reasoning under uncertainty. Such notoriety led to extension of many existing Bayesian tools such as hidden Markov models, extensively used for image segmentation. This paper falls under this category of frameworks and aims to propose a new hidden Markov field that better handles nonGaussian forms of noise, designed for multichannel image segmentation. To this end, we use a recent kernel smoothing- based noise density estimation combined with a genuine approach of mass determination from data. The proposed model is validated on sampled and real remote sensing images and the results obtained outperform those produced by conventional hidden Markov fields.