I think it should also be compared to non-local means (with various patch size and strength settings) with some image pairs, with a reasonably faithful image comparison metric (https://github.com/kornelski/dssim for example) And AFAIK the kind of noise which is removed is not really gaussian, so perhaps it'd make sense to test with the kind of noise which the real-world cameras produce (for example take a pair of images of exactly the same scene with ISO 100 and ISO 1600 with different exposure time to compensate for brightness difference, and use the first image as the noise-less reference; for better results, first image could be an average of multiple shots, because most cameras still produce some visible noise at ISO 100) When simply looking at the demo, it's not immediately clear if it's really better than NLM.
(For an image metric to be meaningful, it should have good correlation with human vision tests, for example http://www.ponomarenko.info/tid2013.htm PSNR is obviously an outsider, and while RMSE is not listed on that page, I guess it's not much better). > On 8 Jan 2018, at 07:38, Aurélien PIERRE <rese...@aurelienpierre.com> wrote: > > Hi, > > I would like to propose a new denoising method for dt : the Total Variation. > It plays in the same category as the bilateral denoising : > > • it smoothens surfaces while retaining edges > • it's very efficient to recover chroma noise and noise in bokeh areas > without affecting too much the in-focus areas > • but it could be more computing efficient. > A researcher has published an article comparing several sub-methods to do so > (with various regularizations) and released his C++ code under the GPL. > Regular people can test the algorithm online with their own pictures. The > rest of the details (and more candies) are on his page : > https://joandurangrimalt.wordpress.com/software/ > The background of this method is to minimize the energy of the picture, hence > the noise, defined as the integral of the gradients over the picture. I have > succesfully achieved a faster gradient estimation using a 2D convolution with > separable filters (same way as the Sobel operator) with this research. > > I don't have time now to work on integrating this myself, as I'm already > buisy working on the blind deconvolution and stuggling with C, but I believe > it would be a great add-on to offer a more efficient alternative to the > bilateral filter, and all the maths and C libs are already there, so most of > the work would be UI. > > Anybody willing to work on that ? > > Thanks ! > > PS : there is also a non-local total-variation regularization available, able > to both denoise and reconstruct details by inpainting, but it's just maths > and no GPL code for now : > • https://hal.archives-ouvertes.fr/hal-01342111/ > • > https://joandurangrimalt.wordpress.com/research/novel-tv-based-regularization/ > -- > Aurélien PIERRE > aurelienpierre.com > > > ___________________________________________________________________________ > darktable developer mailing list to unsubscribe send a mail to > darktable-dev+unsubscr...@lists.darktable.org ___________________________________________________________________________ darktable developer mailing list to unsubscribe send a mail to darktable-dev+unsubscr...@lists.darktable.org
