LinearSharpen, Conversion Engines and Data Loss

[Roger_Cavanagh]

Pigasus and I have been looking at LS D60 and possible implications for changes to LS with D30 images. Sally sent to me a revised action for the D30 to test yesterday, but I immediaely noticed a problem with the reds and magentas in my Q13 test image.

We have both investigated and reached the same conclusion that converting from Lab mode into RGB mode using the Microsoft ICM conversion engine causes loss of colour data.

I set up a test with a document containing a fill colour with RGB values 200, 40, 25. The document was in 8-bit Adobe RGB. The following transformations were applied to the image:

1. Image>Mode>Lab
2. Image>Mode>Convert to Profile
2a. using Microsoft ICM with relative colorimetric
2b. using ACE with relative colorimetric

In both cases, the colour numbers were changed, the histogram showed a spread a numbers:

2a. MS: 198-201, 26-36, 20-25
2b. ACE: 199-202, 37-43, 24-28 (oops, typo 24-38 was wrong)

As you can see, errors have been introduced in both cases, but at least with ACE, the median values were unchanged from 200, 20, 25, but with the MS version showed significant colour loss: 199, 32, 21.

I repeated the test using a gray path 128, 128, 128 and again found data loss:

MS: 125-127, 126-128, 126-128
ACE: 127-129, 127-129, 127-129

And finally for anyone who doesn't think 16-bits is a good idea... I converted the red patch to 16-bit and repated the profile conversions. The MS version still showed data loss, but there was no variability. RGB values were consistent across the image at 199, 31, 21. The ACE version showed no data loss or error RGB values were still 200, 40, 25.

The Microsoft ICM conversion engine looks like something to avoid.

Regards,

[Persio]

Roger,

I agree with your final statement although my conclusion is based on a less scientific process.

I had developed a print-image workflow for my Epson Stylus 980 printer using Adobe's ACE engine to convert my images to sRGB, with very acceptable results.

I recently tried to apply the same workflow on an image converted using Pekka's D60 Beta2, which, by the way uses MS conversion engine, and the end results I got out of the printer were disastrous. Not because of LS_D60_Beta2 but because I used MS engine to convert my image to sRGB for printing.

I will perform some further tests but my inclination is to use ACE, even though Pekka has concluded that the MS engine is better for blacks (if I understood correctly).

Regards,
Persio.

[Timo Autiokari]

Roger,

the MicrosoftICM engine is *far* more accurate in doing the conversion, especially in the dark end of the range.

That you found some more quantization (very small so nothing to worry about) so that the AdobeACE seems to be in this test better is because the AdobeACE damages the dark end quite a lot (does not apply the conversion accurately) and it does some noise dithering (adds noise) even if you uncheck all the related dither checkboxes in Photoshop.

[Roger_Cavanagh]

Timo Autiokari wrote:
Roger,

the MicrosoftICM engine is *far* more accurate in doing the conversion, especially in the dark end of the range.

That you found some more quantization (very small so nothing to worry about) so that the AdobeACE seems to be in this test better is because the AdobeACE damages the dark end quite a lot (does not apply the conversion accurately) and it does some noise dithering (adds noise) even if you uncheck all the related dither checkboxes in Photoshop.


Timo,

I much more worried about the data loss than the quantization. The Micrsoft ICM engine is reducing the RGB values. I noticed the effect in a test image when a red patch (linear TIFF processed using LinearSharpen method using ACE) with average values of 167, 41, 44 was in, the MS ICM version (also linear TIFF with similar - but not identical - processing), changed to 171, 0, 27.

That seems to me much more destructive than the dark area damage you have highlighted with ACE. Can you be more precise about the point (i.e., RGB values) at which these errors start to occur?

Regards,