JPEG Artifact removal

[Kort]

Hello everyone.
I just joined but have been using NI for a while.

I heard somewhere that it is also handy for removing JPEG articfacts, but have not had great success with that, is there a trick to it?
Thanks.

~Kort

[NITeam]

There is no special trick, just let NI sample those artifacts in a flat featureless area, for example in a sky area. Build a noise profile, just like it is done in the quick start guide, step 4, and apply filtration. The should clear the skies.

Hope this helps.
Vlad

[Kort]

Thanks Vlad.

I know that process, I should have been more clear. I am putting a catalog together and unfortunately some of the products ONLY are of low rez JPEG images. I upsize with Genuine Fractals but the artifacts are still pretty bad. And they cannot be re-shot.

Am I asking too much of NI? Often there aren't areas for it to build a profile from, and it either smears it all to heck or makes the artifacts worse.

Edit: Side note, I tried the trick of adding a little uniform noise to the image and then running NI, and that helps, but by no means solves the issue.

[arkhangelsk]

Thanks Vlad.

I know that process, I should have been more clear. I am putting a catalog together and unfortunately some of the products ONLY are of low rez JPEG images. I upsize with Genuine Fractals but the artifacts are still pretty bad. And they cannot be re-shot.

Am I asking too much of NI? Often there aren't areas for it to build a profile from, and it either smears it all to heck or makes the artifacts worse.

Edit: Side note, I tried the trick of adding a little uniform noise to the image and then running NI, and that helps, but by no means solves the issue.

Small pictures of any sort are a problem for NI in my experience - its minimum requirements for a profiling is relatively big. I think for it to work well you need at least a 3-megapixel image - it sometimes has problems finding a suitable profiling area in 1600x1200. If your pictures are like 640x480 then things get a bit ... tight ... if they are like 320x240, forget it.

JPEG artifacts are mostly HF and some MF noise, so that helps you - you have a chance of getting the best out of a horrible situation. Don't bother expecting miracles - there are none, at least not with a GIMP Quality 25 picture. The below is tried out with such a picture.

You presumably want to reduce two things - those annoying 8x8 JPEG checkerboarding squares and those darn slivers near the edges (high contrast). The auto profiler is no good to you - it's programmed to avoid those squares because they are non-uniform.

So, first, the squares. Get the largest uniform spot, making sure you cover some of the obvious squares (you want them recognized). Don't worry too much if you engulfed a LF detail region - that may be impossible to avoid - just avoid HF and MF. Make your Rough Profile out of that. Do the Auto Fine Tune, then bracket some smaller uniform areas with the squares and select Manual fine tuning - just watch those Equalizer bars rise.

Head over to the Noise Filtering page. Turn on Smooth Edges, Disable VLF and LF filtering (-100%) if you don't want your picture to be completely destroyed. Now, use the component viewer and increase HF noise reduction until the squares are muted to a tolerable (to you) level. You may need some MF as well, but do that more cautiously. If a square is entirely the wrong color there is nothing you can do about it (try careful cloning with Photoshop / GIMP or whatever), but otherwise you should be able to mute the squares...

Now, see if the slivers have been reduced to something more tolerable - try and mute them with the HF (read, raising HF NL even further). The slivers can be the worst...

If that doesn't work out: Return to the Noise Filter, Save the current profile and try to select squares that include the annoying slivers and then punch MFT. Start with larger boxes and work your way down to smaller boxes that include only the slivers - see if you can force the Equalizers to higher and higher settings as they measure higher noise averages - at least you'll be limiting the destruction to a certain combination of R, G and B levels, and you effectively go beyond the +150% allowed in the NF controls. Otherwise, figure out what brightness bands they are in and try using the sliders.

Keep playing around with the NF and profile. Even then, don't hope to annihilate the JPEG slivers - though attenuation would be pretty good (you can try to use Photoshop or GIMP or something with a clone tool to squelch those slivers manually, which is probably the better path). And settle for keeping maybe half the image details.

By the way, if you are thinking about trying another program, for the reference I've tried that with Noise Ninja. The auto-seeker locked onto a flat piece and gave me an estimate of 2 for luminence noise, which is of course no good for muting the squares, so I manual profiled. It was going pretty well until I tried to annihilate the slivers by forcing NN to profile them. The luminance estimate rose to >100 (NN's way of saying maybe I should try Gaussian Blur instead), and because NN lacks any form of frequency control, the end result was that the image was annihilated. Just to provide a sense of perspective.

[Kort]

Thank you for the comprehensive reply arkhangelsk.
I will give that methodology a try.
Happy New Year!
~Kort