"convert -colorspace lab" output out of range

[displayguy]

Hi all,

I'm trying to convert sRGB pictures to CIELAB by running:

Code: Select all

convert original.tif -colorspace lab output.tif

When I extract the values with:

Code: Select all

convert output.tif output.txt

The L value is a LOT higher than the values that I get from resources like Photoshop and colormine.org. More worryingly, the a and b values routinely exceed 250. As far as I know, a and b should be between -110 and 110.

Could anyone please point out what I missed (if anything)? It seems important since both Inkscape and Gimp users are told to use Imagemagick to do RGB to CIELab conversions.

Thanks in advance!

P.S: I run Imagemagick from the Ubuntu repos.

[snibgo]

If you are running integer (non-HDRI) ImageMagic, the values can never be negative. The "middle" value of the a* and b* channels is 50% of quantum.

Does that answer the question?

[fmw42]

IM biases the A and B values to 50% as snibgo says for normal Q8 and Q16 IM compiles because all image must range from 0 to 256 or 0 to 65535. To get negative values you need to use HDRI compiled IM and tell tiff to save negatives. see quantum:format at http://www.imagemagick.org/script/comma ... php#define

[displayguy]

Thanks a bunch!

To sum up, I installed libtiff* on my system and compiled IM from source with the following (just being thorough for the benefit of future inquirers):

Code: Select all

./configure --enable-hdri --with-libtiff
make clean
make
sudo make install

Then I convert the images with:

Code: Select all

convert input.tif -colorspace lab -define quantum:format=floating-point output.tif

The L* looks good, but the a* and b* deviate somewhat from the results I get with Photoshop and colormine.org. Then again, the latter don't match either. I'd trust IM and Colormine to do a better job than Photoshop since the conversion algorithm is open. Many eyeballs and all...

P.S: Increasing quantum-depth to 32 helps.

[snibgo]

If you can publish the input, and the versions converted by IM, PS and colormine, someone can take a look. (Upload to somewhere like dopbox.com and paste the result here.)

I would expect a trivial difference due to rounding errors. There may be larger differences due to different "zero" points in a* and b*, and/or full-scale differences (eg one program might use only 100 out of 256 possible values).

[displayguy]

Here comes the files (plus screenshots from Colormine and Photoshop):

https://drive.google.com/folderview?id= ... sp=sharing

It's the original sRGB file and the output of:

Code: Select all

convert original.tif -colorspace lab -define quantum:format=floating-point quatum-depth=32 output.tif

When I convert the original to text I get: 0,0: (45,65,87) #2D4157 srgb(45,65,87)
The CIELab one gives: 0,0: (26.6667%,-0.39292%,-6.27527%) #440000 cielab(26.6667%,-0.39292%,-6.27527%)

When I feed the original's sRGB values (i.e. 45,65,87) into Colormine it gives: CIELab (26.8, -1.09, -15.63)
Opening the original image in Photoshop gives: RGB (45, 64, 87) and CIELab (25, -6, -19)

As you can see, the a* and b* difference is fairly substantial.

Thanks again!

[snibgo]

But what are the units? http://colormine.org/color-converter has units of -128 to +128. To see this, try converting from "CIE-L*ab". Why do they call it that? "CIE-L*a*b*" or "CIE-Lab", but having * beside just L is weird.

Colormine says it is open source. In theory, someone could look at the code.

[fmw42]

IM uses code from http://www.brucelindbloom.com/. Photoshop is notorious for not following standards. See http://en.wikipedia.org/wiki/Lab_color_space for the difference between Lab and La*b*