Perceptual Rendering Argyll CMS

[Doug Gray]

It would be nice if it were that way in the real world, but there are a couple of confounding factors. One is the fuzziness of the ICC spec. There is enough ambiguity in it and the various attached commentaries that one could be forgiven for thinking that conversion to PCS values (even for the colorimetric table) should involve an appearance transform, since PCS has defined appearance parameters. I don't agree that such an interpretation is useful - I think RC should be firmly tied to instrument measurable values, but different factions within the ICC seem to be pulling in other directions at times - witness the ICCV4 spec. change that switched from display profiles being based on contact me asurements, to be being based on less well defined tele. measurements.

I don't like the changes re monitors but can understand the rationale. In general people don't understand why prints, in general, have lower Dmax than monitors and this causes a lot of folks to complain about washed out looking prints compared to what they see on the screen. Perceptual intent is designed to somewhat offset this by increasing contrast.

OTOH, V4 specs do clarify RI and require it to be measurement based. Of course instrument and printer variability are significant factors limiting how close RI comes to fulfilling the goals. I've see some profiles that rendered RI to scale the black point. In some cases I've seen a dE of 10 around L30 or so between what is printed or reflected in AtoB1. Apparently V2 specs have enough wiggle room that this is allowed though most of the V2 profiles I've looked at do not do this. This makes them useless for Absolute I the clear goal of which is to render a specific color reasonably accurately so long as it is within the printer's gamut range. It's rather pathetic to see a matte patch print in Absolute that renders a requested L25 as L35. Both PM5 and I1Profiler will render it within 1 dE and usually around .5dE.

The other confounding factor is a repeatability/statistical issue. Neither the printing nor the measuring process are perfectly repeatable. Every reading is a point sample in the colorspace with added print, measurement and instrument uncertainty added to it. So the mathematical fact is that the profile that best represents the underlying device behavior, is not the one that minimizes the delta E to the measurement point set that it is created from. The latter in fact risks being a case of over-fitting.

Ditto that. One of the functions of smoothing in generating interpolation tables is to look over many nearby patch colors and create LUTs that best re-create the desired color. This can interact with the CMM which may use something more than just a linear fit - or not. And the number of LUTs can be increased which will provide a better match using linear interpolation if the printer curves have larger second derivatives.

In fact there are some printers where the best results I've had were using a relatively small number of patches, down to 480 or so, due to very high gradient changes. Higher numbers of patches produce worse profiles on these beasts. I had to do this to get reasonable results from a laserjet. Generally much less a problem on the printers today than 10 years ago. They seem to have made significant progress in print gradient smoothness.

[DaveRichardson]

Graeme

Yes, the device value curve shape should make little or no difference - the white & black points and the shape of the gamut surface should be the same.

Thanks for the explanation - very helpful.

You can also supply an image gamut, although I'd recommend a Device Link workflow rather than creating image optimized Device Profiles - you probably want to build one for each image or group of images, so you may as well build just one specific table rather than 4, and a Device Link will give a noticeably smoother conversion

I agree - and I like to keep things simple. That said - I also like the idea of having the option to build something specific if faced with a really awkward image to print. As an example - I kept marine tropical fish and corals -  the combination of marine LED lighting and coral colours resulted in some image colours being captured that were well outside the gamut of my printer.

All - thanks for the input and discussion - I raised the original question to enhance my understanding and the thread is certainly helping that.

Dave

[digitaldog]

Well, I don't have two profiles made with the same spectral data handy so I'll do something better. Actually print the chart, measure the colors, and report the results with dE2k compared to Andrew's CC image.

It's not better, it's far worse. After pointing this fact out to you, I'm shocked you still don't understand how flawed your methodology and thus conclusions are about all this!

I illustrated the fact that measuring isn't necessary and introduces noise into the data. That you don't have a way to build two profiles with same spectral data illustrates you don't understand all the unnecessary variables into evaluating this 'issue' that I have eliminated. So I guess I'll have to discuss this as if you're a novice: You have two profiles. Both built using exactly the same data and the ONLY difference is the software that built them. You take a set of RGB values in any RGB working space you desire. Ideally a real image something like the Macbeth but doesn't matter to clarify this simple process in your mind. You have a set of RGB values, you convert them through each profile. You get NEW RGB numbers. Still with me Doug? You produce a dE report in ColorThink Pro. This is exactly what I did to illustrate the differences (average dE over 1, max just over 3) with two profiles passing data through their RelCol tables to produce new numbers.


Now you say you want to move those pristine numbers through a driver, make ink hit paper, then measure it. And you formed the opinion this is better? Are you certain Doug? Because all I see is some desperate attempt to spend more time, introduce more noise into the data that's totally unnecessary. Maybe it will produce values you feel are closer to your unproven idea that two profiles using a RelCol conversion will produce the same results. That's wrong to believe just as your methodology to test this is wrong.


I've gone out of my way to reduce anything in the testing that will pollute the results. Sorry if you don't like the science and if so, knock the methodology, not the numbers that clearly show that the two profiles made from the same measurement data don't produce the same colorimetric values through their RelCol tables.

Oh, and I'll do it from Lab. No point in clipping the cyan patch. Also, I don't have any recently made profiles from different S/W as I'm using I1Profiler for printer profiles and have been for some time.

All textual digressions that are unnecessary and will not produce any better results.

After all, it's the print that counts.

NO, it's what the profile produces that counts. Try to get that into your head. We're talking about how two profiles convert RGB values through a RelCol table, NOT what you measured after introducing false 'data' into a process!

I'm curious what color is producing the dE2k max of 3.  That's outside what I normally see.

Because what you normally see is viewed though flawed testing! You can't by your own admission even build two profiles with same measurement data!

Be happy to look at how your two profiles render the CC image if you want to share them.

Until you can warp your head around sound testing, it's rather pointless. You're just wasting your time, my time and anyone else here who's considering the unproven statement (just the opposite): AtoB1 tables are supposed to accurately colorimetrically represent the printed color. Always.

[digitaldog]

More proof of evidence, this time using the fabric image from Gamut Test File and running that through the same two profiles built identically other than the software used.
Here's your visual dE difference thanks to ColorThink. And the dE report from a mere 220,000 colors (device values) using Extract ALL colors in CTP.
As I had said, using a Macbeth in sRGB was being kind....
Doug, think you can see a dE of 9?
--------------------------------------------------


dE Report

Number of Samples: 220000

Delta-E Formula dE2000


Overall - (220000 colors)
--------------------------------------------------
Average dE:   1.11
    Max dE:   9.37
    Min dE:   0.01
 StdDev dE:   0.90


Best 90% - (197999 colors)
--------------------------------------------------
Average dE:   0.87
    Max dE:   2.30
    Min dE:   0.01
 StdDev dE:   0.47


Worst 10% - (22001 colors)
--------------------------------------------------
Average dE:   3.29
    Max dE:   9.37
    Min dE:   2.30
 StdDev dE:   0.93


--------------------------------------------------
--------------------------------------------------

[Doug Gray]

More proof of evidence, this time using the fabric image from Gamut Test File and running that through the same two profiles built identically other than the software used.
Here's your visual dE difference thanks to ColorThink. And the dE report from a mere 220,000 colors (device values) using Extract ALL colors in CTP.
As I had said, using a Macbeth in sRGB was being kind....
Doug, think you can see a dE of 9?
--------------------------------------------------

    Max dE:   9.37

Holy moley.

A dE2k of 9 is a very visible difference. Even allowing for the colors to be at the gamut boundary that's a really big difference. Something's broken. I've never gotten a print with colors that far off. It's probably one of the profiles simply has some very bad math creating the 3D LUTs. Almost  certainly at a gamut edge. That's the region most interpolation algorithms have the most error. It's also possible it's something like special handling of OBs. Some profiling software "adjusts" for this differently than others but use the same spectral data. I have no idea what your profile vendors are doing.

Still a dE2k of 9 is a strong indicator something is amiss. Unless the difference can be explained by reasons stated above I wouldn't use either of those profiles because there is no way of knowing which is better and their is strong evidence that one, at the very least, is quite flawed.

Try printing the color that shows a dE of 9 using both profiles and see which one is broken. It's the only way to tell where that dE is coming from.

[digitaldog]

Holy moley.
A dE2k of 9 is a very visible difference. Even allowing for the colors to be at the gamut boundary that's a really big difference. Something's broken.

No, only your reaction to the data 
Do you realize that that is ONE patch (device value) of 200,000? Examine the report again. Examine the average, max and min dE values and consider that we're talking about one color, versus one other among 200,000! It's why colorimetry tells us one thing and viewing an image within context tells us something different. Colorimetry and the dE testing is about color perception. It is not about color appearance. It can be subjective or colorimetric (the only way we should define color accuracy in dE) or both.

I've never gotten a print with colors that far off.

You may have, you just didn't do the testing correctly as I've attempted to point out repeatedly! It's very possible that one of 200,000 or one of a million solid colors were indeed a dE of 9 off when you compare 1 sample to 1 other, which is what the report illustrated (max dE 9). You cannot measure a print and come up with that kind of data. None the less, the data is still colorimetrically sound.

Try printing the color that shows a dE of 9 using both profiles and see which one is broken

A print where 1 pixel is off or 10,000? There's a difference which is what I'm attempting again to teach you. It illustrates how your testing (printing and measuring) isn't valid expect for a difference you can see on a print. We're not talking about that, we are talking about colorimetry and differences using deltaE which is a comparison of simply two samples!
This boils down to what you're trying to test! In this discussion, it's the false concept from false testing that suggests: AtoB1 tables are supposed to accurately colorimetrically represent the printed color. Always. YOU brought up colorimetry in that sentence Doug.


Colorimetry and the dE testing is about color perception. Of one set of solid colors. It is not about color appearance. The reason why viewing a print is more valid than measuring it is because measurement is about comparing solid colors. Color appearance is about evaluating images and color in context which measurement devices can't provide. Decide which you're talking about, maybe we can come to an agreement. But in terms of colorimetry, something you brought up, suggesting the report tells us a profile is 'broken' because you've never measured a print and seen this value appear, illustrates you're talking apples and oranges. It's why I keep telling you that measuring a print when what we're debating, the Colorimetry of what a profile produces, pixel by pixel is so flawed.

It's the only way to tell where that dE is coming from

No, just the opposite! Do you now understand why? The dE of 9 is coming directly from the differences in the two profiles, something you continue to refuse to accept.

[Doug Gray]

Do you now understand why? The dE of 9 is coming directly from the differences in the two profiles, something you continue to refuse to accept.

Of course the dE of 9 is coming from the two profiles. It's the dE between the two Lab colors that each profile, as interpolated from the 3D LUT AtoB1 tables, indicates.  It's a rather large error. If a large portion of the image consisted of that color with that large a dE then images printed using each of the two profiles could look quite different. If the vast majority of image pixels had a low dE then it probably wouldn't be noticed.  That's why a histogram rather than just a limited set of stats, is more useful.

[digitaldog]

Of course the dE of 9 is coming from the two profiles. It's the dE between the two Lab colors that each profile, as interpolated from the 3D LUT AtoB1 tables, indicates.  It's a rather large error. If a large portion of the image consisted of that color with that large a dE then images printed using each of the two profiles could look quite different. If the vast majority of image pixels had a low dE then it probably wouldn't be noticed.  That's why a histogram rather than just a limited set of stats, is more useful.

I give up. You still can't seem to connect the dots about colorimetry of two profiles and output to a print from the profile. Or that the idea: AtoB1 tables are supposed to accurately colorimetrically represent the printed color. Always, is incorrect. I've proven it, you can't.

[Doug Gray]

I give up. You still can't seem to connect the dots about colorimetry of two profiles and output to a print from the profile. Or that the idea: AtoB1 tables are supposed to accurately colorimetrically represent the printed color. Always, is incorrect. I've proven it, you can't.

No, you have proven no such thing. The only thing that can be said with certitude about them is that they will soft proof that color differently with a dE2k of 9 using the two profiles made with exactly the same spectral data. If you think that's an acceptable level of soft proofing, fine, but the entire point of soft proofing is to show you on your monitor what a print, properly illuminated, will look like. You just have to hope that there aren't large numbers of colors that are over dE2k of 3 or so. I've seen far worse profiles. I try not to use them.

But let's get on to the real world, the color errors in printing your colorchecker image. The 24 patches all have identical Lab colors except for a small amount of variation, typically well under .1. It's possibly a small amount of dithering done during conversion from some other space like aRGB.

So I printed the CC image on Epson Ultra Prem. Glossy but using a 9500 II, then spot read each of the patches with PatchTool. I used I1 Pro 2, XRGA which was also what was used making the profile. Compared against the image, average dE2k was .7, max dE2k was 2.8, which was the white square. Next worse was the dark blue at dE2k of 1.2. The others were under 1.0.

[digitaldog]

But let's get on to the real world, the color errors in printing your colorchecker image. The 24 patches all have identical Lab colors except for a small amount of variation, typically well under .1. It's possibly a small amount of dithering done during conversion from some other space like aRGB.

That's in your imagined world of printing and measuring which isn't the same as inspecting the RGB values the profiles produced. I told you why. Graeme told you why. Again, I'm sorry the facts are ruining your reality. Got nothing to do with soft proofing, the colorimetric values I produced where from RGB values from a RelCol conversion. Those ARE the values the profiles produce, not some measured print with all nature of noise. It's kind of shocking you still don't seem to understand that fact.

So I printed the CC image on Epson Ultra Prem. Glossy but using a 9500 II, then spot read each of the patches with PatchTool.

Pointless and filled with data errors. But then as I suspected, your methodology was built to produce a result you wanted, not that was colorimetrically accurate.

[digitaldog]

It's possibly a small amount of dithering done during conversion from some other space like aRGB.

Nope, sorry. There's no dither in high bit images when converting. Photoshop's Color preferences clearly indicate this if you'll take the time to examine them...

[Doug Gray]

That's in your imagined world of printing and measuring which isn't the same as inspecting the RGB values the profiles produced. I told you why. Graeme told you why. Again, I'm sorry the facts are ruining your reality. Got nothing to do with soft proofing, the colorimetric values I produced where from RGB values from a RelCol conversion. Those ARE the values the profiles produce, not some measured print with all nature of noise. It's kind of shocking you still don't seem to understand that fact.Pointless and filled with data errors. But then as I suspected, your methodology was built to produce a result you wanted, not that was colorimetrically accurate.

Utter nonsense. The process was designed to see how colorimetrically accurate my printing workflow would print your colorchecker image using my standard profiles.  The results are very good, excluding the white patch. Out of gamut colors should be excluded, of course, as they will have higher dE's by definition.

How well do you print defined colors? Or do you ever?

Image (Lab): 37.65 15.58 16.57   actual print (Lab): 37.28 15.42 17.11    dE2k:0.54
Image (Lab): 61.96 35.64 61.75   actual print (Lab): 62.03 36.26 63.98    dE2k:0.63
Image (Lab): 28.61 19.61 -53.73   actual print (Lab): 29.71 18.43 -54.05    dE2k:1.24
Image (Lab): 96.08  0.52  0.50   actual print (Lab): 95.94 -0.03  3.42    dE2k:2.84
Image (Lab): 66.67 16.57 18.58   actual print (Lab): 66.88 16.56 19.22    dE2k:0.45
Image (Lab): 39.99  9.55 -42.69   actual print (Lab): 40.94  9.77 -42.10    dE2k:0.94
Image (Lab): 55.28 -38.65 33.63   actual print (Lab): 55.44 -39.36 34.43    dE2k:0.36
Image (Lab): 81.16  0.51  0.50   actual print (Lab): 81.60  0.84  1.21    dE2k:0.88
Image (Lab): 50.19 -4.50 -21.60   actual print (Lab): 50.45 -5.04 -20.40    dE2k:0.78
Image (Lab): 51.76 48.70 17.57   actual print (Lab): 51.70 49.40 18.69    dE2k:0.56
Image (Lab): 41.96 56.70 29.63   actual print (Lab): 41.99 57.66 30.18    dE2k:0.28
Image (Lab): 66.65  0.52  0.50   actual print (Lab): 66.60  0.70  0.44    dE2k:0.28
Image (Lab): 43.13 -13.55 22.60   actual print (Lab): 42.52 -13.37 21.96    dE2k:0.65
Image (Lab): 30.98 21.57 -20.59   actual print (Lab): 30.90 20.88 -19.56    dE2k:0.57
Image (Lab): 81.56  5.53 78.79   actual print (Lab): 82.08  5.26 79.53    dE2k:0.43
Image (Lab): 51.75  0.52  0.49   actual print (Lab): 51.73  0.89  0.33    dE2k:0.56
Image (Lab): 55.29  9.57 -24.62   actual print (Lab): 55.64  9.39 -23.44    dE2k:0.75
Image (Lab): 72.15 -22.59 57.73   actual print (Lab): 72.10 -22.50 57.95    dE2k:0.10
Image (Lab): 51.36 50.70 -12.55   actual print (Lab): 51.93 50.51 -13.02    dE2k:0.61
Image (Lab): 36.07  0.48  0.51   actual print (Lab): 36.21  0.64  0.34    dE2k:0.31
Image (Lab): 70.59 -31.63  0.50   actual print (Lab): 70.76 -31.69  1.27    dE2k:0.53
Image (Lab): 72.94 19.61 68.77   actual print (Lab): 72.94 20.25 68.58    dE2k:0.41
Image (Lab): 49.79 -27.59 -28.63   actual print (Lab): 50.57 -27.64 -28.22    dE2k:0.80
Image (Lab): 20.38  0.52  0.48   actual print (Lab): 20.38  0.85 -0.35    dE2k:0.95

[Doug Gray]

Nope, sorry. There's no dither in high bit images when converting. Photoshop's Color preferences clearly indicate this if you'll take the time to examine them...

Ok, so the color patches just have small variations in them. These are tiny ones. Just surprised they are there.

[digitaldog]

Just surprised they are there.

Based on so many of your ideas, and the actual facts presented, I'm not surprised you are surprised! 

“There are two ways to be fooled. One is to believe what isn't true; the other is to refuse to believe what is true.”
-Søren Kierkegaard

[digitaldog]

Out of gamut colors should be excluded, of course, as they will have higher dE's by definition.

What out of gamut colors (other than white)?
More assumptions on your part, the sRGB Macbeth image fits fully inside the printer profile:

[Doug Gray]

Based on so many of your ideas, and the actual facts presented, I'm not surprised you are surprised! 

“There are two ways to be fooled. One is to believe what isn't true; the other is to refuse to believe what is true.”
-Søren Kierkegaard

Did you miss breakfast or something?

I didn't add dithering so it has nothing to do with Photoshop.

The small variations in 16 bit files didn't get their by accident. Noise is intrinsic in scanning and photography but these noise levels are far below those produce by pixel shot noise. And I rather doubt they were put their pixel by pixel. It's common when a small amount of dithering is added. It's just that I didn't add it. The low noise level is in your image.

[digitaldog]

I didn't add dithering so it has nothing to do with Photoshop.

And I never said you did! I pointed out the facts of which you seem to have difficulty accepting or understanding that the data I provided had not a lick of dither! In fact the data provided you still can't wrap your head around are values produced by each profile without any noise or measurement errors I've tried to point out your methodology produces.

The small variations in 16 bit files didn't get their by accident. Noise is intrinsic in scanning and photography but these noise levels are far below those produce by pixel shot noise. And I rather doubt they were put their pixel by pixel. It's common when a small amount of dithering is added. It's just that I didn't add it. The low noise level is in your image.

You're going down another rabbit hole of your own creation. If you want to look into noise that affects data, it's been proven and agreed upon by another poster** that you are the one introducing the noise by measuring a print when that isn't necessary at all. The two profiles produced colorimetric values from actual pixels that were directly compared with as little variations possible! Even from the same measured data you by your own admission can't produce. Again, you're just wasting my time. Your tests are flawed, it's been pointed out too many times now. My methods are sound and you've completely failed to point out any issues with that data; you just ignore them! The two profiles produced 220,000 values that were directly compared. The results are clear. IF you don't want to accept them, fine.


**http://forum.luminous-landscape.com/index.php?topic=106407.msg875857#msg875857



I often say that when you can measure what you are speaking about, and express it in numbers, you know something about it;but when you cannot measure it, when you cannot express it in numbers, your knowledge is of a meagre and unsatisfactory kind.
-Lord Kelvin

[Doug Gray]

What out of gamut colors (other than white)?
More assumptions on your part, the sRGB Macbeth image fits fully inside the printer profile:

No, you are the one making assumptions. The colorchecker image's white square is outside the printer/paper gamut since the paper's white point is Lab (97.3, 1.86, -7.0).

As you fully well know but failed to consider, Epson Ultra Premium Glossy Photo Paper, which I earlier stated I used, has a great deal of OBs resulting in an large white point shift. As a result, the profile/printer/paper combination is unable to print the requested color, Lab (96.08,.52,.50) though it succeeded somewhat in more than halving the "b*."

[digitaldog]

The colorchecker image's white square is outside the printer/paper gamut since the paper's white point is Lab (97.3, 1.86, -7.0).

Perhaps English is your 2nd language? Read what I wrote: What out of gamut colors (other than white)?

As you fully well know but failed to consider, Epson Ultra Premium Glossy Photo Paper, which I earlier stated I used, has a great deal of OBs resulting in an large white point shift. As a result, the profile/printer/paper combination is unable to print the requested color, Lab (96.08,.52,.50) though it succeeded somewhat in more than halving the "b*.

The profile plotted and the colorimetry reports I proved are not Ultra Premium Glossy Photo Paper.

[Doug Gray]

Perhaps English is your 2nd language? Read what I wrote: What out of gamut colors (other than white)?

When you said this:

What out of gamut colors (other than white)?
More assumptions on your part, the sRGB Macbeth image fits fully inside the printer profile:

Apparently you were referring to the image when you ran it through your two profiles. I mistook "the" to refer to a profile in the singular and presumed you were referring to my printer profile and the results I was discussing and that somehow it was inside the gamut after clipping to sRGB. I didn't convert it to sRGB.

The profile plotted and the colorimetry reports I proved are not Ultra Premium Glossy Photo Paper.

Ah, back to your two profiles that display poor profile programming.  What exactly does that prove other than one or both profiles are defective? And they obviously are. If they can't run the extremely color limited sRGB clipped colorchecker image through them without getting a max dE2k of 3 from exactly the same spectral data thus eliminating instrumentation variables there is something seriously flawed between the two.

But thank you for agreeing that the white patch is out my printer/paper gamut. So that leaves the actual results, added noise and all  , as:
Max dE2k: 1.24,
Ave dE2k: .6

What you refuse to see is that your profiles are flawed in some way in their AtoB1 conversions. Having seen individual profiles that were also far more flawed, it doesn't surprise me and, frankly, I don't see what conclusions you can draw from them other than one or the other of them had some sloppy programming.

The real question is how well do these profiles that show a difference dE2k of 3 on a mere colorchecker image actually print it? Mine gives dE2k max of 1.24 and an average over the patches of .6.  Does either of yours come close?