Multiresolution Non-local Means Filter for 3D Image Denoising

In this work, we present a new adaptive multiresolution Non-Local Means filter for 3D MR Image Denoising. Our work is based on the recently proposed Non-local (NL-) means filter (Buades2005). As for all the denoising filters, the choice of the filtering parameters is crucial in NL-means-based restoration. The balance between structure preservation and noise removal is a difficult task and different sets of parameters can be optimal for different space-frequency resolutions of the image. In this paper, we propose to implicitly adapt the among of denoising according to the underlying structures thanks to a new adaptive soft wavelet coefficient mixing. Moreover, a Rician formulation of the NL-means filter has been incorporated within our multiresolution framework in order to deal with the specific nature of noise in MR images. Quantitative validation was carried out on Brainweb datasets with Gaussian and Rician noise. The results show that the proposed multiresolution filter outperforms optimized versions of the NL-means filter. Finally, qualitative experiments on anatomical and Diffusion-Weighted MR images show that the proposed filter efficiently removes the noise while preserving fine structures.

Details can be found in:

Coupé P., Manjón J.V., Robles M., Collins D.L. Adaptive Multiresolution Non-Local Means Filter for 3D MR Image Denoising. IET Image processing 6(5): 558-568, 2012.

Demo data and the source code of the proposed method can be found here

Version 1.0 (September 2008)