Image denoising using the TV-means algorithm

General description and related papers

On this page you will find examples illustrating the TV-means image denoising algorithm. It is a patch-based image denoising using Total Variation regularization according to the following theoretical scheme:
• apply Total-Variation regularization to all image patches at all scales
• for each pixel, select the minimum scale ensuring that a sufficient number of patches similar to the current patch are found
• average all these patches to obtain a "denoised patch"
• aggregate all image estimates coming from these denoised patches to compute the denoised image
This algorithm combines two famous (and very different) image denoising methods: Total Variation denoising [1] and NL-means denoising [2]. It exploits the strenghts of both methods and manages to produce better results than each of them, as illustrated below. For mode precise details about the method and the algorithm, see

 C. Louchet, L. Moisan, "Total Variation as a local filter", SIAM Journal on Imaging Sciences, vol 4:2, pp. 651-694, 2011. download: published version

References:

[1] L. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algorithms, Physica D, vol. 60, n. 1-4, pp. 259-268, 1992.

[2] A. Buades, B. Coll, J.-M. Morel, A review of image denoising algorithms, with a new one, SIAM Multiscale Modeling and Simulation, vol. 4, n. 2, pp. 490-530 (electronic), 2005.

Examples

Examples below compare three methods: Total Variation, NL-means, and (aggregated) TV-means. The noisy images are classical images corrupted with a white Gaussian noise with standard deviation 20. Click on the links in the PSNR table below to see the images.

 Barbara Lena Boats House Peppers noisy 22.1 22.1 22.1 22.1 22.1 TV [1] 26.69 30.89 29.21 31.22 29.62 NL-means [2] 29.59 31.50 29.32 32.05 30.12 TV-means 30.93 32.48 30.00 33.10 30.63

History

february 2010: first version
april 2011: link to revised preprint (to appear in SIIMS)