Scanner accuracy improved to average dE .33

[Doug Gray]

My V850 scanner accuracy, profiled with a large patch set (4k patches) using Argyll, is now producing .33 dE average errors compared against an i1iSis scan of independent patches. The largest part of the improvement by far is the substantial reduction of reflected light contamination using the program I posted in the Digital Image Processing sub forum.

It's to the point I can make excellent printer profiles with the, gasp!, scanner. Further, I can fit 4k patches on a single letter size page using 4mm square patches.

Of course all of this, in addition to the required reflection correction, requires that the scanner is profiled to a specific printer initially measured with a spectro otherwise metameric error would be material. But that's a one time effort.

There are several possible scanner issues that could have, but didn't, rear their ugly heads.

1. Illumination evenness. The scanner's light could vary across the page.
2. Hue consistency. The scanner uses a row of LEDs to illuminate the paper. Slight differences in each one's "white" would be a problem.
3. Drift with temperature. LEDs change their white point as they warm up. Not much, but some.

However, these 3 error sources are still low enough to not matter significantly. At least on my sample of one scanner.

Since it may not be obvious given the complexity of color science,  this level of accuracy is only because the scanner was profiled against a specific printer. Getting a good colorimetric scan of a print made with the same printer that was used to generate a target does not do much in regard to scanning other materials like silver halide prints, artwork, or even other printers than the one I used.

That said, it is very slightly better than the small IT8 chart using I1Profiler but still has average dEs just over 2. for things like a real colorchecker or colorchecker CG. This is to be expected as the LED lighting and, I presume, color filters can't be expected to have metameric results with different reflection spectra than that the printed charts I used.

But my main goal here is to have colorimetically accurate  scans of printed material. Largely to test smoothness of gradients which are not possible with a spectro within any reasonable timeframe and set of resources. And that now looks very doable.

[Alan Goldhammer]

It's to the point I can make excellent printer profiles with the, gasp!, scanner. Further, I can fit 4k patches on a single letter size page using 4mm square patches.

But my main goal here is to have colorimetically accurate  scans of printed material. Largely to test smoothness of gradients which are not possible with a spectro within any reasonable timeframe and set of resources. And that now looks very doable.

Surely this beats scanning pages of patches with an i1 Pro!!!

[digitaldog]

Surely this beats scanning pages of patches with an i1 Pro!!!

Only if you don't believe in the capture of spectral data.

[Doug Gray]

Only if you don't believe in the capture of spectral data.

Yeah. While it works great for D50 profiles but only with that printer which requires initial spectral data for the scanner profile. And you can't correct for other illuminants. Also, it only works for M2 profiles since the scanner has no uV.

Not real practical for any sort of normal use. Mostly just interesting that it's possible.

[Doug Gray]

There's a few interesting notes about the really good scanner profile made wuing 4k patches printed with a 9800. Metameric error is small, if even present, in scans made of prints from that 9800. This is because the inks are the same ones used for creating the patches used to make the scanner profile.

However, if I look at scanning prints from the Canon 9500 I get substantially higher dEs. This is due to the 9500 having different spectral characteristics as well as additional red and green inks.

The 4 histograms shows the dE00 distribution of scanned v i1iSis measured colors which consist of 957 colors that are in gamut of both printer and, widely distributed. The top histograms are the those of the scanned 9500 patches. The left side is without reflection correction. The right side is with reflection correction (see scannerreflfix.exe).  The bottom histograms are those of the 9800 with the same conditions as the top row.

These are the average dE00 results:

9500 print scanned with 9800 based profile:
-Reflection Corrected:1.04
-Uncorrected: 1.33

9800 print scanned with 9800 based profile:
-Reflection Corrected:0.35
-Uncorrected: 0.66

One clear factor is that, in this set, the errors due to metameric failure are larger than those from the uncorrected reflection when scanning prints from the 9500.  This shows the value, when scanning inkjet prints, of making a special patch set with the same printer.



One interesting thing about scans, they have a fairly extended dE00 error tail in the histogram. The tails extend to about 4 dE00 and are on all histograms. Examining the cause, I've found that the scanner RGB space has some sensitive areas where a single bit change in one channel corresponds to more than 3 dE00!  When I then looked at the space of all possible RGB values from the scanner, only 20% of them corresponded to printable colors. Of those, the average dE sensitivity was .4 but with significant percentages exceeding 2. Generally, the worst areas of dE00 error in the above histograms are in the population of highly sensitive RGB areas.

Curious.

The image was corrected: the top and bottom rows were mislabeled. The top rows shows metameric failure causing higher dE when scanning prints made with a Canon 9500 II but a scanner profile made with an Epson 9800.

[GWGill]

Of course all of this, in addition to the required reflection correction, requires that the scanner is profiled to a specific printer initially measured with a spectro otherwise metameric error would be material. But that's a one time effort.

Yep. It's this that makes using a scanner as a "poor man's" colorimeter pretty useless in practice, since you need an instrument to create the scanner reference file, while standard scanner targets are for media that is too different to the printer to be accurate.

[ It seems a pity that scanners don't have their own stray reflection reduction as a matter of course. ]

[Doug Gray]

[ It seems a pity that scanners don't have their own stray reflection reduction as a matter of course. ]

Indeed. It's highly consistent on the V850 scanner. In extremis, where a white patch is surrounded by black v all white, the differences are on the order of 20% from light being reflected off the paper white back onto and then from the white, translucent plastic strips that are backlit by rows of LEDs.

Funny thing is you don't really easily see the effect because the black surround makes the white patch look pretty white even with an L* change of over 7. Under normal scenes the reflectance average is much lower so instead of a white getting a 20% (linear) boost from a white surround on average it's only about 20% of 20% or about 4%.  That's just over 1 L*. That lower average reflectance makes profiling scanners possible though one probably gets somewhat better results averaging several pages of the same patches with different randomization.

It certainly would be fairly easy for the scanner OEMs to compensate. Just takes some ram and a little processing power. It's a linear process. From what I see, it's highly repeatable for any given physical scanner design.

[Ethan Hansen]

Doug - How are you measuring profile accuracy? Theoretically via roundtripping or validation with measurements of a print on a different instrument?

[Doug Gray]

Doug - How are you measuring profile accuracy? Theoretically via roundtripping or validation with measurements of a print on a different instrument?

Interesting question and it's similar in some ways, but different in other areas when looking at printer profile accuracy.

First, let's look at "accuracy." It's a somewhat misleading term because, while widely used when referring to things like dE reports from monitor and printer profiling what is really meant is consistency. When a monitor is measured, whether with a spectro or colorimeter, few instruments have specified accuracy levels with the exception (sometimes) of white. Even then it's usually spec'ed in a repeatability sense, not an absolute accuracy in the sense of tolerance limits to 1931, 2 degree, CIELAB. Even within a single company, there are variations between instruments and especially instrument models usually exceeding 1 dE.

So pretty much invariably, when the term accuracy is used, people are really referring to measurements from the same instrument. And then there's the issue of people who want "accuracy" from scene to camera to print and display and have no idea what sort of rabbit hole they invoke.

However, both printers and profiles should use an independent set of colors to measure "accuracy" from those used to create the patches from which profiles were made.

For scanner "accuracy" tests for prints, I typically print a patch set of 957 colors evenly distributed in LAB space that are in the printer's gamut but are uncorrelated (both in color and physical position)  with the patches used to create profiles.

These are read with an i1iSis and scanned. The RGB values from the patches are extracted from the scan by first removing 15% from the 6mm square boundaries where diffraction overlap occurs, then discarding the largest 10% and smallest 10%  in magnitude (total RGB) and taking the average of all remaining pixels.

My gold test for combined printer and scanner profile accuracy, is to:

1: Rescale to 720 DPI (printer and scanner resolution).
2: Convert the image to the printer/paper profile using any desired Intent (Perceptual typically).
3. Print the image with color management disabled. I use null transform in Photoshop.
4. Convert the image to ProPhotoRGB using Abs. Col. Intent. (essentially creates the equivalent of a soft proof with show paper color)
5. Scan the image at 720DPI on the V850 with all color management disabled.
6. Fix the resulting Tif image by running scannerreflfix.exe
7. Attach the scanner profile then convert to ProPhoto RGB.
8. Load the scanned image into a layer behind the image retained from step 4.
9. Auto align them and trim the result as needed.
10. Sometimes moving the second layer up or down a few pixels can further  improve alignment.
11. Optional: Compensate for printer (mostly) and scanner (a bit) rolloff by sharpening.

Assuming one has good color management and profiles for both the printer and scanner are made by the same instrument, the two layers, when toggling visibility between the two, should appear virtually identical.

If all has been carefully done!