Pages: 1 [2]   Go Down

Author Topic: Spectro Software Feature Request  (Read 13847 times)

MHMG

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 1285
Spectro Software Feature Request
« Reply #20 on: May 15, 2009, 09:39:07 am »

Quote from: papa v2.0
Hi

I* sounds quite interesting.
So how would it be used in the real world. Is it a proposed metric for image difference between original and reproduction?

The I* metric was originally conceived as an algorithm to objectively evaluate and score color and tonal accuracy for imaging systems that exhibit or have to contend with large shifts in color between the source image and a second mage of the same scene which is to be visually compared to the source image.  This happens almost all the time when trying to reproduce full dynamic range digital images containing rich highlight and shadow details onto a reflection print material that by its physical nature cannot support as large a color and tonal range.  Strong tonal compression and color remapping must be invoked to make a visually coherent "translation" or rendering of the image onto the paper.  Color difference models work really well when source and output have similar gamuts and a near perfect match can be achieved, but they aren't very useful in the situation I just described where significant color remapping must take place. That's where a metric like I* is needed.

A couple of real world examples:  1) we might want to judge initial appearance and aged appearance in image permanence studies where the source image is the print in its original condition and the comparison image is the print in greatly faded condition.  2) A printing company may want to evaluate two different papers for color and tonal output quality on a particular printer and be able to share the results objectively with its clients.  By scoring on a percentile basis, the I* metric gives even non color geeks a fighting chance of understanding what the scores mean.  If a printmaker tells the average customer "the print I just made deviated from your digital camera file on average by 20 delta E", that means nothing to most people and is in fact outside the useful range of perceptual scaling significance even to color scientists. If I said, the print I just made retains 90% color accuracy (hue and chroma) and 85% tonal accuracy (lightness and contrast) compared to your digital image file", most people will intuitively have a feel for what the percentile rank scores mean. They could probably guess  that 100% would have been a perfect score, whereas most customers would have no idea that zero delta E would have been a perfect score.

Note that the case of the "perfect match" is the one point in color reproduction terms where both delta E models and I* metric score the reproduction with equal success.  As the errors get larger the scoring advantage becomes increasingly in favor of the I* metric.

Is I* a proposed metric?  I guess so. The I* metric was published as a technical paper in the NIP 20 Conference proceeding of the Society for Imaging Science and Technology in November 2004.  Other researchers can try it out and use it if they think it has value.  The color science community hasn't given it a "blessing" yet, but researchers, especially in the museum and archives community that deal with works of art on display, are beginning to take serious interest in it.  Also, both WIR and AaI&A use it, but WIR so far has not switched its image permanence testing services over to I* yet.  There are undoubtedly legacy issues to face whenever companies move from one test standard to a newer one.  AaI&A is a new company with no such legacy issues, so all of its testing is I* based.  Because I invented the I* metric, I'm obviously biased about it's usefulness, but after five years of throwing lots of visual appearance problems in image reproduction at it to see if I can trick it into generating nonsensical scores, I haven't been able to do it. I can demonstrate nonsensical delta E results with incredible ease. And as for densitometric test models, let's not even go there!

If you'd like to see I* in action with a simple well known image target (macbeth colorchecker colors) in light fading studies of systems that exhibit significant color changes and catalytic fading problems as they are exposed to light, visit my website and download some of the examples for dye-based systems like the Epson 1270 (blue colored links have public access).

http://www.aardenburg-imaging.com/acceleratedagingtests.html

Follow the link for Light Fade Test Results

cheers,

Mark
http://www.aardenburg-imaging.com
« Last Edit: May 15, 2009, 09:42:34 am by MHMG »
Logged

papa v2.0

  • Full Member
  • ***
  • Offline Offline
  • Posts: 206
Spectro Software Feature Request
« Reply #21 on: May 15, 2009, 10:09:23 am »

Hi
thanks for the reply

Yes I agree we need a way of telling how accurate the reproduction is as a whole compared to the original.

Do you include some sort of viewing standard when calculating the I* value.

I see that the system is image dependent and is there a mechanism for reporting which elements of the image are in error. If for example I* reported a 85% colour accuracy - which colours are out? (Is it the sky or the models jacket or the product colour).

If at the end of the day the reproduction goal is to produce a 'pleasing' reproduction (for arguments sake) of the scene and not a colormetric reproduction, how would I* fit in?
Logged

MHMG

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 1285
Spectro Software Feature Request
« Reply #22 on: May 15, 2009, 01:17:03 pm »

Quote from: papa v2.0
Hi
 
Do you include some sort of viewing standard when calculating the I* value.

I see that the system is image dependent and is there a mechanism for reporting which elements of the image are in error. If for example I* reported a 85% colour accuracy - which colours are out? (Is it the sky or the models jacket or the product colour).

If at the end of the day the reproduction goal is to produce a 'pleasing' reproduction (for arguments sake) of the scene and not a colormetric reproduction, how would I* fit in?

The I* metric uses the CIELAB color model as the underlying architecture, so viewing standards get handled by the illuminant assumption you make when measuring the LAB values before processing the I* math.

We still don't have really great artificial intelligence algorithms that will take an original scene and capture and process it for color and tone in a way that pleases everyone (though many digital camera companies do have proprietary ways to produce "pleasing" color that they think the majority of their customers will like "out of the box"). In fact it's obviously an impossible task to please everyone. Long live custom layer edits in photoshop!  You may like skintones warm in a particular scene, for example, while I may prefer them cooler.  Anyway, I* enters the workflow at the point where you have decided what a good source image should be. That image becomes your reference image for the I* calculations.  In image permanence testing, for example, the assumption is that the original print (whether it's the most pleasing print or not) contains the color and tonal qualities you are trying to preserve. One could find a situation, for example, where a print that is too dark will fade and lighten in a way that becomes more pleasing to most observers over time before it fades too far and becomes less pleasing than the original. The I* metric would not compute it getting "better" then getting worse.  The reference print was dark to begin with, so accurate color and tone scores means it should stay that way.

Thus, the basic assumption with I* is that you now want to bolt down your preferred color and tonal relationships and reproduce them with as much colorimetric accuracy as possible.  Once you've got your aimpoint reproduction in mind and have edited your preferences into your digital image file, then I* can take it from there and tell you downstream how the subsequent reproduction choices are stacking up in terms of retaining said chosen color and tonal quality. In other words, once you've created your "pleasing" image, then the whole process at that point becomes one of accurate colorimetric matching to the extent that it is possible. There's the rub. It often not possible, so I* can tell you how far you have strayed, and as you suggest, even what localized parts of the image are suffering more in the reproduction than others. The I* analysis is all about color and tone distributions in an image. Specific images have specific colors. They get sampled as an ordered array of locations (ie. spatial frequency analysis) within the image and then summed and averaged by the I* method to produce the overall score.

The whole "pleasing versus accurate" endeavor is why, for example, many printmakers build a "master" file with carefully chosen edits that gives them their "ideal" color and tone, then turn on softproofing in Photoshop and add final edit layers to try to "pop" the profile-translated color and tone back into better visual alignment with the source image before committing to printer output.  An I* plug in in photoshop, for example, could help to objectively quide your visual edits to tell you you when are getting closer or farther away. It's possible that the I* metric could even give profiling applications some feedback that could help produce a "smart CMM".  There's a lot of research potential for further development of the I* metric. We've just scratched the surface.

Finally, just to make sure I've fully answered your question about tracking specific regions of interest in the image with the I* metric, a robust I* software application can do what your are suggesting and track specific colors or even localized areas within the image and give you selective I* scores about these specialized regions. Some "colorgeek" apps have this image-specific tracking capability now only using delta E.  The WIR iStar comparative image analysis software was designed to perform this type of evaluation using I* math as well as having the conventional delta E methods available to the user.

best regards,

Mark


Logged

papa v2.0

  • Full Member
  • ***
  • Offline Offline
  • Posts: 206
Spectro Software Feature Request
« Reply #23 on: May 15, 2009, 03:54:06 pm »

Hi Mark

How does it  take into account an image that is in one colour space and printed to another colour space using perceptual as the rendering intent.
Would that not give poor I* values as all the colormetric values could be changed?
« Last Edit: May 15, 2009, 05:01:08 pm by papa v2.0 »
Logged

MHMG

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 1285
Spectro Software Feature Request
« Reply #24 on: May 15, 2009, 05:43:46 pm »

Quote from: papa v2.0
Hi Mark

How does it  take into account an image that is in one colour space and printed to another colour space using rel col as the rendering intent.
Would that not give poor I* values as all the colormetric values could be changed?


Ideally, in a perfect reproduction the measured LAB values for the reference image in one colorspace match the LAB values of the comparison image in another colorspace because under both colorspace viewing conditions the assumption is that you have fully adapted to the illuminant. LAB theory says, for example, that neutral gray under one illuminant still looks like the same neutral gray under another illuminant unless you have not adapted to each illluminant when you viewed the samples. The I* metric is indeed comparing reference image to output image in absolute colorimetric terms. That is by definition what colorimetric accuracy is all about.  The metric is not trying to compensate for relative colorimetric translations. On the contrary, it is intended to show you how much accuracy loss occurred as you made the translation.  For example, all working colorpace models (sRGB, aRGB, prophoto, etc) and all monitor display profiles map monitor white to LAB 100, 0, 0 which of course no reflection print can attain and monitor black to LAB 0,0,0 which again no reflection print media can attain. That's why the need for an I* metric.  So, when you render monitor white relative to the paper white you get no colorant being formed and your actual paper white LAB value may be very different not only in the fact that L* won't =100 but the color may not have a*=0 or b*=0 either. Yet the I* metric is looking at your reference image aimpoint (LAB 100,0,0). You didn't get there so there is an error and perfect I* scores aren't obtained. That's reality. So yes, the more your media deviates from neutral white and is darker than L*=100 and the more the max black in the system deviates from LAB 0,0,0 and the more compressed your color gamut is, then the more constrained the system is and the more translation must take place in terms of printing your digital file to paper, especially if your digital file contains perfect whites and perfect blacks. I*scores do indeed go down. That said, a good ICC profile will do a better job than a bad one even under relative colorimetric mapping translations, so the better your relative colorimetric rendering, the higher the I* score will be.

There are also other ways to analyze a system with the I* metric. You could for example convert the source image to the printer profile color space using relative colorimetric rendering, then convert to LAB using absolute colorimetric rendering. This would produce a digital file containing the predicted LAB values of the actual print (including predicted paper white and max black values). Use that as the reference image data for I*, and now you have a theoretical shot at obtaining the 100% scores when you measure the comparison image (ie. the actual print) whereas in the previous example above, there is no possibility of a perfect match between reference and comparison images. In this latter example, the more the color and tonal accuracy scores as computed by I* drop from 100%, the more your printer is not printing the way the ICC profile is predicting. This is a really great way to validate ICC profile performance.

I know we are in a color management discussion thread, but we are also starting to push hard into color geek territory. You've raised very good questions. I hope my answers make sense. It took me about three years to develop the I* metric, so it's understandable that a first read through the published I* papers are going to leave you with questions.

cheers,
Mark
« Last Edit: May 15, 2009, 05:52:09 pm by MHMG »
Logged

digitaldog

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 20630
  • Andrew Rodney
    • http://www.digitaldog.net/
Spectro Software Feature Request
« Reply #25 on: May 16, 2009, 09:00:50 am »

Quote from: MHMG
Thus, the basic assumption with I* is that you now want to bolt down your preferred color and tonal relationships and reproduce them with as much colorimetric accuracy as possible.  Once you've got your aimpoint reproduction in mind and have edited your preferences into your digital image file, then I* can take it from there and tell you downstream how the subsequent reproduction choices are stacking up in terms of retaining said chosen color and tonal quality. In other words, once you've created your "pleasing" image, then the whole process at that point becomes one of accurate colorimetric matching to the extent that it is possible. There's the rub. It often not possible, so I* can tell you how far you have strayed, and as you suggest, even what localized parts of the image are suffering more in the reproduction than others.


This is what we've been trying to do with iStar Mark. We have a group of hero images that we output on an Epson 3800, including the wonderful Roman 16s (http://www.roman16.com/en/), our preferred images and even synthetics. The final output is to a digital press. We know that there's a huge difference in gamut and so forth between the two but the Epson output represents our idealized, preferred output. Its what the client looks at and says "that is our preferred output from these images). With iStar, I would expect that as we tweak our output profiles and press behavior, we now have a metric that tells us, and client how much closer we are getting to the idealized output (Epson reference prints). Using deltaE would show huge differences due to the vast differences in the two devices and of course, not be weighted to imagery but lots of solid colored patches. Does this seem like a reasonable use of the iStar technology?

Once we get as close as we can to our goal using iStar, we can also use it as a trending target.
Logged
http://www.digitaldog.net/
Author "Color Management for Photographers".

Jonathan Wienke

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 5829
    • http://visual-vacations.com/
Spectro Software Feature Request
« Reply #26 on: May 16, 2009, 09:30:32 am »

So what are the chances of I* being incorporated into EyeOne Match anytime soon?
Logged

MHMG

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 1285
Spectro Software Feature Request
« Reply #27 on: May 16, 2009, 12:27:41 pm »

Quote from: digitaldog
This is what we've been trying to do with iStar Mark. We have a group of hero images that we output on an Epson 3800, including the wonderful Roman 16s (http://www.roman16.com/en/), our preferred images and even synthetics. The final output is to a digital press. We know that there's a huge difference in gamut and so forth between the two but the Epson output represents our idealized, preferred output. Its what the client looks at and says "that is our preferred output from these images). With iStar, I would expect that as we tweak our output profiles and press behavior, we now have a metric that tells us, and client how much closer we are getting to the idealized output (Epson reference prints). Using deltaE would show huge differences due to the vast differences in the two devices and of course, not be weighted to imagery but lots of solid colored patches. Does this seem like a reasonable use of the iStar technology?

Once we get as close as we can to our goal using iStar, we can also use it as a trending target.

Yes, what you describe is a perfect situation for the I* metric. It should work extremely well. Thanks for sharing.

Mark
Logged

MHMG

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 1285
Spectro Software Feature Request
« Reply #28 on: May 16, 2009, 12:39:43 pm »

Quote from: Jonathan Wienke
So what are the chances of I* being incorporated into EyeOne Match anytime soon?
Probably not good at the moment. I've been trying to get an answer to a simple question for some time now about Xrite's Net Profiler software. Does it or doesn't it support the Spectrolino?  No follow-through from anyone I've spoken to at Xrite so far.  I really don't know the right people to talk to about the I* metric, so that seems even more challenging to open a dialogue about it.  I think it would make a nice fit with Measure Tool. That said, Digital Dog just gave me a good contact name, so you never know, it could happen.    
Logged
Pages: 1 [2]   Go Up