This is a "weighted" median filter. This means that pixels in different parts of the median kernel have different sizes on the X axis of the sorted pixel values, when intensity is mapped as Y and X moves from pixel to pixel. You read that right. More-central pixels take up more space in the list.
You will find that it results in much sharper images than using a 3x3 flat median kernel, yet still having the property of being centered. Also, you can run it recursively without blowing up all edges. You can also "blur" the result (to hide kernel jaggies) in a much more powerful way than you can with a flat kernel.
Read the source code. I did my best to keep things clean.
Sure thing.
- Original:
- Gimp, 1px radius "Despeckle", -1 black, 256 white, not adaptive, not recursive (normal 3x3 median filter):
- Above, at 50% wetness (blend with original):
- Wimg median, normal:
- Wimg median, blurry:
- Wimg median, blurrier:
- Wimg median, special (dynamic blur):
Yes, "normal", "blurry", and "blurrier" look extremely similar in real world examples. However, you will find that "blurrier" shows slightly less aliasing than the other two. The difference between "normal" and "blurry", on the other hand, is invisible except for in certain pathological pixel patterns.
"special" is a mode that sorts a median set, but then blurs it with a triangle filter. This results in an image with almost as little noise as a median, but with almost the same smoothness as a blur. The implementation here is almost as sharp as a 2x2 box blur, but with much less noise, and the property of being centered.
There is a setting called '--split' for sorting individual RGB channels, which can work wonders on pixelated images with lots of sharp color changes. Before: http://i.imgur.com/fllo2mP.png -- After: http://i.imgur.com/dxZ0Ict.png -- The improvement is much more noticable with --normal mode. These pictures were done with --special mode, which you should use for thin pixelated things. If you don't use --special mode, you get bad smudging: http://i.imgur.com/aJ2MZpr.png
It's better than most simple ones, but you really want to get into the advanced stuff for that. Check out https://www5.cs.fau.de/research/software/idaa/
This is because the method I use gives 16 samples of which to calculate a median. The "correct" way to select a median between an even number of elements averages the central two samples. If you dislike this, you can patch the method and add an extra center sample. Don't remove one; you'll crash on one-dimension images. Being biased towards the higher or lower center sample will result in even more bright or dark fringing on certain edge patterns.
There's a little aliasing. ====================== That's a drawback of denoising with a median filter. Certain images, including this one, are anti-aliased very sharply, but also have ringing. Medians share properties with surface blur type effects, and when you filter over these sharp ringed edges, you get the same aliasing behavior. The "blurrier" variant should cause less of this behavior, though.
That's a drawback any 3x3 median filter will have with high-but-not-high-enough frequency patterns in the image. For example, http://i.imgur.com/rdCQajw.jpg is filtered into http://i.imgur.com/Bq2YGUs.png -- you can use --special to reduce the moire pattern at the expense of a lot of sharpness. You could also try running the filter multiple times.
You can try to reduce other forms of moire by using nonlinear srgb with --srgb, but this will cause dark fringing when hue and saturation change.
Upscale your image to 300% with whatever algorithm you want. Now run your normal flat median with a large kernel (2px radius? 3px radius?) -- you will get less of the dot pattern outline effect than if you ran the median on the native resolution of the image. Now downscale to the original resolution.
Use waifu2x's jpeg denoising. It's not general purpose, but it might work.