Re: [Gimp-developer] Retinex color operation

[gg <gg catking net>]

On 04/25/12 10:43, Michael Grosberg wrote:
> I'd like to Discuss Retinex. I decided to do it here before submitting a
> bug report.
>
> In theory the retinex operation is very useful in fixing badly lit
> images. In practice, I've had little success with it. So I decided to
> fiddle around with it. My test picture was a Chinese temple on a
> background of blue sky (I can upload the XCF on request). usually you
> want to just brighten up some darker areas or bring detail in an area
> that's too brightly lit, without changing the rest of the image. In my
> test case, The sky's color was ruined when retinex was applied and most
> of the well-lit areas lost most of their color.
>
> I'll break down what happened:
>
> 1. The properties that can be changed by the sliders in the retinex
> dialog are not clear at all. The documentation doesn't explain the first
> two in terms understandable to a non-mathematician. Whatever it is they
> are doing, they are not helping in correcting images. If it is opssible
> to give them more descriptive names this would be great.
>
> 2. The result is too strong. I find I have to do a "fade" every time I
> use it. My suggestion: add a multiplier slider that would do an internal
> fade. The default should be about 0.5.
>
> 3. The algorithm is indiscriminate: it is used to fix bad lighting but
> it also works on well-lit areas. I believe it should have less effect on
> pixels (or areas) whose average luminosity value is closer to 0.5, and
> more effect on pixels with values closer to 0 or 1.
>
> I managed to achieve a selective retinex using the following method:
> 1. copy the layer
> 2. paste into quick mask
> 3. blur the mask using gaussian blur (8.0 radius)
> 4. invert the colors (only if you want to affect just the darker areas).
> now your light areas are masked out.
> 5. use levels to also mask out the mid values: in the input scale, push
> the black triangle to about midpoint. This part may require a different
> setting for each image.
> 6. go back from quick mask to image and perform Retinex.
> 7. Fade the result
>
> The above procedure can probably be turned into an integral part of
> retinex so that retinex only works on those parts. Perhaps adding
> another slider that emulates what I did with the levels operation, to
> choose how much you want to affect midtones in the image.
>
> Lastly a comparison to photoshop:
> Photoshop's "shadow/highlight" operation (basically retinex) has two
> sliders in its basic settings. One for choosing the strength of the
> effect on lighter parts of the image, and one for choosing the strength
> of the effect on darker parts of the image. So basically it's like doing
> what I did with masks twice, one for dark areas and one for light, and
> combining the result. There are also dual extended options, one set for
> dark areas and one for light areas. In it there is a "tonal width"
> slider which according to the help file does pretty much what I did with
> the levels operation.
>
> Comments?

I agree with your initial reaction, I found it fairly useless as it was 
and consigned it be yet another interesting but practically useless toy.

You experiments with masking etc  are very helpful. If the filter could 
be made to act in a more useful manner that does not involve having to 
discover all this , it would be a valuable addition.

thanks for the input.