Pixon-based image denoising with Markov random "elds

Image restoration is an essential preprocessing step for many image analysis applications. So far, the majority of works have been devoted to image denoising. For this issue, the most common problem is that some interesting structures in the image will be removed from the concerned image during noise suppression. Such interesting structures in an image often correspond to the discontinuities in the image. In this paper, we propose a novel pixon-based multiresolution method for image denoising. The key idea to our approach is that a pixon map is embedded into a MRF model under a Bayesian framework. The remarkable advantage of our approach over the existing works in this "eld is that restoring corrupted images and preserving the shape transitions in the restored results have been orchestrated very well. A simulated annealing algorithm is implemented to "nd the MAP solution. A measure of the performance of various algorithms has been de"ned and some quantitative comparisons between our approach and two typical "ltering techniques have also been done. A lot of experiments illustrate that our method is much more e!ective and powerful in the noise reduction than the Wiener and median "ltering techniques, two typical and widely used techniques. 2001 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.