A Modified NLM Algorithm Motivated with a Non-uniform Bayesian Prior Distribution

Nicholas Dwork, Nikolaus West, Christian Elder, Thomas Walther and Michael Digman

This is the companion website for the paper A Modified NLM Algorithm Motivated with a Non-uniform Bayesian Prior Distribution. From the abstract of the paper:

Abstract - This paper modifies the NLM algorithm to retain more of the structure of the true image. We use a Bayesian rationale for the algortihm, and then modify the prior distribution to create a new weighting term. We show that several variants of the NLM algorithm can be modified in a similar way. Our technique is shown to improve the PSNR and observed image qualities of these variants on several images. Finally, we show beneficial results with MRI data, and audio data.

Downloads

The paper is currently in review at IEEE Transactions on Image Processing. If it gets accepted, we will post a link to the IEEE Transactions here. The source code for our algorithms is freely available. It requires Matlab, and optionally supports the Matlab Parallel Toolbox.

Download Source

Example Images

The following set of images demonstrates the performance of the algorithm. While the black and white image series allows comparison of the details recovered by the various algorithms, the color images demonstrate the improvement of the modified prior on the image as a whole.

From left to right: NLM, NLM Euc, NLM+. Top row: original. Bottom row: with modification

Performance

We analyzed the effect of our modification on the Peak Signal To Noise Ratio (PSNR) for various NLM algorithms and noise levels. We plotted the resulting PSNRs against the PSNRs of the original algorithms. The diagonal line visualizes an equal PSNR level of the original and the modified algorithms. Every point in the lower triangle demonstrates an improvement of PSNR due to the modified prior.