Segmentation of Textured Satellite and Aerial Images by Bayesian Inference and Markov Random Fields

摘要

We investigate Bayesian solutions to image segmentation based on the double Markov random field model, originally proposed by Melas and Wilson. Inference on the number of classes in the image is done via reversible jump Metropolis moves. These moves, usually implemented by splitting and merging classes, can be very slow, making them impractical for large images. We investigate simpler reversible jump moves that are quick to implement but show that they may mix very slowly. We propose a more complex split and merge scheme and compare its performance. Tests are conducted on satellite and aerial images.