Idea: Implement Retinex Algorithm

[mikeytown2]

Was browsing around and found this
http://www.fmwconcepts.com/imagemagick/retinex/index.php
Looks like it requires Gaussian blur and log operations on the image data.

Use this for blur
http://forum.doom9.org/showthread.php?t=88645
What about log operations?

[Sagekilla]

mask tools perhaps? I know it has a log operator. So you could do mt_lut("x log")

Edit:


src = last

G1 = src.Blur(0.3)
M1 = mt_lutxy(src, G1, "x y / 1 + log")
G2 = src.Blur(0.7)
M2 = mt_lutxy(src, G2, "x y / 1 + log")
G3 = src.Blur(1.5)
M3 = mt_lutxy(src, G3, "x y / 1 + log")

Avg = Merge(Merge(M1,M2,weight=.50),M3,weight=.33)

G0 = src.Grayscale()
CB = mt_lutxy(src, G0, "x y / 1 + log")

MI = Overlay(Avg, CB, mode="multiply")
Merge(Avg, MI)


That would work. If it worked with RGB top to bottom. Unfortunately we can't :( Will post update in a few minutes.

Update: You have to do this through a plugin because through scripts, it would be -extremely- unwieldly. I can try fudging together something as my first attempt at a plugin but no guarantees there.

[tin3tin]

Was browsing around and found this
http://www.fmwconcepts.com/imagemagi...inex/index.php
Nice find. It would really be great if usage of IM scripts was possible in the (old) ImageMagick Avisynth (http://forum.doom9.org/showthread.php?p=612335#post612335) plugin. :)

[Mug Funky]

it seems to do the same as HDRAGC? or am i missing something?

btw, in most of those examples the original looked most natural, and would have probably only needed a vanilla lift/gain/gamma adjustment to achieve the same goal in a more aesthetically pleasing way...

i'm not a huge fan of performing contrast and colour balance adjustments in YCbCr space though - it always looks wrong

[Sagekilla]

I don't think you can do this in YV12 anyway.

Unless you jump through a lot of hoops (ShowRed() etc), you can't do this filter in the current state of things.

The only function I know of that provides a log function is mask tools. Unless you find a way to make it support RGB, it's just going to be painfully annoying trying to get it done.

Merge, Overlay, GaussianBlur, they all exist. Only problem is the "log" portion :(

[Didée]

Sagekilla - there is still MaskTools v1.5.8, which hands RGBLUT() at your disposal. :)

However, "insufficiency of YV12" is no excuse for your script being faulty. It produces a black output. Hang on - don't you see the mistake(s)? ;)


BTW, I didn't know of that "Retinex" algorithm, but no surprise, 've been there before. It was around late 2005 / early 2006 when I was first playing with that idea. A small hint was given here (http://forum.doom9.org/showthread.php?p=760935#post760935): "Todo: replace [additive] sharpening with [multiplication] sharpening".
That was when I was at a somewhat serious attempt to do LimitedSharpenEX(), which should implement combined multiscale nonlinear sharpening. It used sqrt(), however, not log(). What I found was that, while results on selected scenes with selected parameters can be great, the method is rather hard to control so that it gives a consistent result on content with varying characteristics. Showing some great result on selected still images is one thing. Achieving a convincing result on moving pictures, aka "video", is another story. Characteristics change each scene, and even the slightest compression artifacts get blown up into nasty prominence.

[mikeytown2]

NASA appears to be using this on video
http://dragon.larc.nasa.gov/retinex/servo/
And seeing how its designed for a live feed, they didn't take into consideration compression artifacts.

This is strangely similar to HDRAGC now that you mention it.
http://forum.doom9.org/showthread.php?p=917911#post917911

Did some more searching... looks like a dud. It's in gimp, play around with it. It's SLOW
Colors -> Retinex (at the bottom of the menu)
http://www.sfr-fresh.com/unix/misc/gimp-2.6.5.tar.gz:a/gimp-2.6.5/plug-ins/common/contrast-retinex.c

[Sagekilla]

@Didee: Oh I know, I went back and I realized I screwed up quite a bunch of things in that original script, namely the mt_lut calls. Didn't know there was a RGBLUT() though ;) Thanks for pointing that out.

[Fizick]

GaussianBlur from tsp plugin should be faster (it uses fftw)