1
\$\begingroup\$

I'm having images with a visible color noise taking form of color blocks:

original

One can reveal the patches with modular multiplication (x -> x*d%255, x a 8 bits RGB value of a pixel, d some value, eg., 1.2):

modular

I tried to remove the noise with the state of the art denoiser block-matching 3D BM3D (exactly BM3D with collaborative hard-thresholding and the seconds step is a regularized Wiener inversion using BM3D with collaborative Wiener filtering, as described in Image restoration by sparse 3D transform-domain collaborative filtering, 2009, Dabov et al):

bm3d

The BM3D filtered with the modular reveal:

bm3d_revealed

I lost a significant amount of details (shoe, pant folds, grass).

Can someone recommend a denoiser algorithm that would do a better job correcting this noise while preserving the details ?

The image is from a raw file, developped then into a 10bits tiff file, and never went through any jpeg compression, even though, it looks like it did; a jpeg algorithm seems to be applied to the raw image, at least the blocks make me think about it.

Funnily enough, the denoising with BM3D is revealing those blocks, which are less visible in the original.

\$\endgroup\$
4
  • \$\begingroup\$ Are you applying noise reduction before or after transforming the raw image file to a 10-bit TIFF? \$\endgroup\$
    – Michael C
    Mar 30, 2021 at 0:04
  • \$\begingroup\$ @MichaelC I apply the noise reduction on the 10bits (scaled to 8 bits) tiff. \$\endgroup\$
    – Soleil
    Mar 30, 2021 at 0:37
  • \$\begingroup\$ Have you tried using NR with your raw convertor before exporting the image as a TIFF? Many of the most popular NR tools can be used as plugins with the most popular raw convertors. The best NR tools have context awareness built in, so that they can tell the difference between an area that is supposed to be uniformly dark (and thus apply more NR in that area) and an area that has a lot of local contrast (i.e. image detail) to which much less aggressive NR is applied. \$\endgroup\$
    – Michael C
    Mar 30, 2021 at 0:58
  • \$\begingroup\$ @MichaelC This is an experimental camera and I do not have NR for the raw format, only a conversion to pixels. \$\endgroup\$
    – Soleil
    Mar 30, 2021 at 1:01

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.