I figured it would be interesting to see how people at pixls.us handle extremely noisy images with open source RAW developers.
Here is a rather challenging and problematic shot I took a few days ago: a black dog taken at night under sodium light (hello narrow spectrum!) at ISO 12800. This is certainly pushing the limits of current APS-C sensors, and the result is, of course, a heavy dose of chroma noise everywhere on the image:
After adding my standard gamma and contrast stretch, knowing the image would be regarded here (not printed, etc.) I resized it to display, added a nlmeans denoise (sigma=68,local=3,patch=9), and minimal sharpen. The point being that reduction resizing does a certain amount of noise elimination, making the residual nlmeans operation less onerous.
I’m not color-managed yet, so that part is not addressed.
I had a go at it, but I wanted to use my custom GMIC filters, so I imported the file into GIMP with Photoflow. I did a few things to minimise noise in the raw conversion. I knew I was going to go to B&W so I set the white balance multipliers a 1. Then I set the colour profile to RAW so there was no profile applied.
When in GIMP I used some custom GMIC filters to reduce noise in several steps and added local contrast to just the dog.
Then I used GIMP on this JPG file. First used levels tool and then used Nik Dfine tool (via shellout) only on the background and then GMIC sharpen texture tool with final curves touch-up.
Thanx @PkmX for the image and @ll for sharing your views =)
I went a couple different directions. Both started in Gimp and Dfine sorry wrong, those were of an eartlier version; these 2 had a bath in ACR to follow rightaway to photoflow, no GIMP.
Both finished in PS. A Tmax with custom curves version. I though it was quite done already in PF where I had manage an ok separation of pet’s head and BG, but then I remembered Jack London and in my head I visualized its face merging into the darkness, I know silly, anyway that was the reason I ended up in PS, the truth is that it needed more work, much more grittyness… so half way to darkness
And an orto rollei, aslo with custom curves and framed in a darkish polaroid repli. All heavy stuff was PF, I already had the frame action in PS where I only needed to tweak it.
Quite a noisy picture! Using Lightzone, Dfine and GIMP (noise reduction, colour balance (auto) and wavelet sharpen (luminance)).
Struggled with the colour cast and the noise. DFine is great at noise-reduction, and I added some more in GIMP. Colour cast caused by the sodium lamps!! I’m going to make the excuse that the limited spectrum of the sodium light (2 bands in the yellow part of the spectrum around 589nm) some colour information will be missing. (If all else fails - blind 'em with science!! )
@james This is my favorite so far! You did a very nice job on removing almost all luna noise while keeping most of the detail on the fur. You also somehow managed to correct the extremely problematic narrow spectrum from sodium lamps and recovered the highlight detail in the name tag. I’d like to see the pp3 if you are willing to share.
Have you tried to setting the quality to “high” in the noise reduction tab? In my experience, it deals with the colorful blotches left with normal chroma denoise rather well.
@chroma_ghost See the first post, it is a Canon FD 85mm f/1.8 adapted to E-mount.
And yeah, I’d like to give her a bone, but she is getting a little heavy these days.
@PkmX Thanks. Of course, I have added pp3 to my original post. I discovered the high quality setting whilst playing with this image, it definitely made a big difference.
I used a lot of tools I don’t normally touch in RT. Some because others didn’t have a strong enough effect, others to counter unwanted effects from other tools/settings.
It’s interesting that you both mention highlight recovery, because I haven’t recovered any. Although HR is enabled (I always leave it enabled), exposure remains at 0, I don’t use the compression slider, but prefer to reduce exposure to bring in recovered details, and then use curves to compress them to give me control of the compression shape.
I have used a slight amount of tonemapping, maybe this has “recovered” the details in the tag.
Regarding colour. I used auto white balance, and then a slight reduction curve on the green channel (RGB curves). I boosted chromaticity with Lab CC curve, and then selectively reduced some colours with the CH curve (which ended up being all colours except red). I corrected the colour of the heart with HH curve.
Regarding detail. I normally leave luminance noise (I’ve learnt to accept and eventually like it), but here I used high settings and also the median filter (3x3 with 3 iterations) and also high impulse reduction. To combat the softness, I have used high micro contrast and some CBDL, and some mild tonemapping. Sharpened with RL deconvolution and post resize USM sharpening.
@james Your comment is exactly what I’m looking for in these PlayRaw threads that goes into detail about why/how you use each tool and what they achieve!
Here is my second attempt in RT that goes for a more neutral B&W look:
There are still some noise in the background but to my eyes the noise looks fairly non distracting already, so I will leave them there in favor of not erasing too many details from the fur. This is probably one of the image that I hope RT supports masked editing so I can apply stronger denoise on the background, as well as bring down the highlight from the tag without affecting the white part of the harness.
Here is another photo of “Volno”, aka the most majestic dog in our park:
RawTherapee with LMMSE demosaicing & some basic white balance then gimp. Nuked the color noise from orbit using a gausian blur. Processed the areas with more detail with an anisotropic filter which seems to preserve the structure of the fury bits quite well. I wanted a smoother look for the background so I used a bilateral filter there. I messed up the transition between the two a bit but I think the final result is quite good given how noisy the image was.
[quote=“McCap, post:17, topic:2778”]
Demosaicing choice based on what? Knowledge or trial and error?
[/quote]Based on knowledge that lmmse is quite robust in the presence of noise. It’s also quite good when you have problems with moire. Definitely something I miss in darktable. Might be worth porting one of these days. I’m not sure how much of a difference it makes in an image that’s that noisy.
Gaussian blur, anisotropic filter and bilateral filter used how and in what programm?
Sorry forgot to mention that. I used gimp & g’mic for the second step.
I’m afraid there is no style in my ‘style’, I was actually just throwing tools at the image, trying different things. (Something like Hegel’s drunken arrows of spirit …) It is a mess.
You should be able to open the text-only sidecar file.
)
