Pages: 1 [2]   Go Down

Author Topic: I1Isis 2 XL, the good, bad, and ugly  (Read 10559 times)

Doug Gray

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2197
Re: I1Isis 2 XL, the good, bad, and ugly
« Reply #20 on: August 30, 2017, 01:49:39 am »

You will find color variations from different sources. Photo inkjets will, as you've seen, exhibit variation depending on how the printer/paper combination works. Your hypothesis that the paper hold-down is playing a part sounds reasonable. We see this frequently on wide format Epsons, where there are distinct areas of the page that print slightly differently than others. Laser-exposed silver halide printers (Fuji Frontier, Noritsu QSS, etc.) can show a difference usually from one side of the page to another.
Yes, I see large, repeatable variations, mostly horizontal, on my 9800. These are not present on my 9500 II desktop. OTOH, the 9500 exhibits a sort of "warm up" phenomina where the dE changes almost 3 dE when printing a fairly high saturation red. This occurred after the 9500 had been unused for 24 hours. Checking it against after a 2 hour pause showed about a 1 dE "warm up" period. Nothing odd is showing in the image and nozzle checks have always been clean unless it sits more than a few weeks. The 9800's horizontal patterning is quite consistent, and changes only slightly when altering the vacuum or single/dual direction modes. It did improve about 20% or so using the narrow platen selection. The 9500's ink flow warmup issue is visible in the Patchtool image I attached. As you can see, the first rows have pretty significant dE variation.
Quote

Our targets have a number of repeated patches to detect this variability. We also track the behavior of similar colors across the page (e.g. target values with consistent, small offsets) to see if they are consistent. We preprocess each file before feeding to the profiler. The key is to determine whether variations are (a) significant and if so (b) is the printer setup within normal parameters.
My initial test was with a 512 patch set that included 64 distributed RGB patches randomly placed and repeated 8 times. The fact there was significant variation in patches of the same color compared to reading the same sheet and comparing readings of the same patches kicked me off down this rabbit hole.

Quote
Looking at variation, it is worth first checking the level of variability from the iSis itself. A straightforward test is to measure a blank page. Simply print the positioning bar and side markers and measure. Using an OEM paper (Canon or Epson) with a lustre surface will give uniformity well within what the iSis can resolve. We also printed the iSis fiducial marks on pvc sheets of various colors to see how the iSis behaves on colors other than white. Gather short-term repeatability data as well as measurements a day or two apart. You'll derive a baseline for instrument performance.
Funny you should mention that. I did just that earlier today out of curiosity as to whether there was some sort of trend from the start of reading patches as the sheet scan progresses. The I1 Pro shows a significant drift (.2 to .3 L*) and drift from the illuminant has the largest impact at high luminance. Turned out to be less than .1 from start to end. Frankly, I was rather pleasantly surprised.
Quote

Start customizing the canned targets and preprocessing. We toyed with the idea of creating targets with identical patches per page but different layouts. You'll need to average the data first (stay in the spectral domain!) before feeding the summarized data file to i1Profiler. We ended up not going this route for a couple of reasons. First, we typically did not read variations sufficiently far from the iSis noise floor to be worth the doubled measurement times. Also, we have the luxury for most of the profiles we build of having hundreds of the same model of printer to build a reference form. At that point, a significant portion of the profile generation hinges on eigenvectors rather than only the raw data.
Averaging the data is pretty straightforward. I use Matlab as I find it's easier but I understand I1Profiler will average files if you select both and drag and drop. Haven't tried it though.

Curious as to how you use eigenvectors in profile generation. I've kicked around doing some principal factor analysis to see if the spectral functions could be reduced (mostly - SVD here we go) to a linear group in 3 space. and then one should be able to find a matrix to convert linearized scanner RGB to XYZ. I doubt it would work with the 9500 since that has a green and red ink as well as the CYMs. OTOH, the 9800 just uses CYM and lightened version of C and M so it might have a better shot
[/quote]
Logged

Doug Gray

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2197
Re: I1Isis 2 XL, the good, bad, and ugly
« Reply #21 on: August 30, 2017, 11:30:47 am »

Looking at variation, it is worth first checking the level of variability from the iSis itself. A straightforward test is to measure a blank page. Simply print the positioning bar and side markers and measure. Using an OEM paper (Canon or Epson) with a lustre surface will give uniformity well within what the iSis can resolve.

I wanted to test both the consistency of the Isis and get some idea of the production consistency of the (cheap) Costco glossy media I'm using to refine processes before expanding into OEM or specialty media. So I pulled two sheets from the new Costco package and rotated 1 relative to the other 180 degrees which flips top/bottom and right/left. These were then printed with the registration marks and form a 841 patch set 29 by 29 with the default 6 mm patch sizes.

Two Isis read scans were done with each paper. The variation, small as it is between the two papers, is far smaller comparing two scans with the same paper.  Yes!

This says volumes about the excellent Isis repeatability and good positioning to the registration marks. Very low noise baseline.

About .11 ave dE between the two papers and .02 dE on the same paper. See attachments.

The 4 scans were done back to back with the fast, 1 pass (M2) mode over a 10 minute interval.
Logged

Doug Gray

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2197
Re: I1Isis 2 XL, the good, bad, and ugly
« Reply #22 on: August 30, 2017, 01:16:43 pm »

>> I see large, repeatable variations, mostly horizontal, on my 9800. These are not present on my 9500 II desktop. OTOH, the 9500 exhibits a sort of "warm up" phenomina where the dE changes almost 3 dE when printing a fairly high saturation red. This occurred after the 9500 had been unused for 24 hours.

Test protocol: For each printer, the Canon 9500 II and the Epson 9800, a page of 512 patches containing 8, randomly located, groups of 64 max spaced RGB triplets was printed. The pages were scanned and each set of 8 patch samples of the same colors were read for both printers. The RGB triplet with the maximum average deltaE on the 9500 and another on the 9800 was used to create a 841 patch set in a square, 29x29, of 6 mm identical patches. These were printed and scanned with the Isis. In both cases the major contributor to Delta E was b* variation so that was the variable studied.

The scans, a matrix of 29 rows by 29 columns was averaged horizontally to determine vertical variation and averaged vertically to determine horizontal variation. Each of these produced a length 29 vector. Attached charts labeled "Horizontal Variation" is a plot of the 29 points, (left to right paper in profile) from averaging each column. "Vertical Variation" charts were row averaged producing a vertical column vector. The 29 points is the average  b* values from top to bottom.

A note about the error bars. These are the standard deviation of the set of 29 samples from which that point on the chart was determined.  Thus, they represent the error bars of samples orthogonal to the plot line. When the horizontal variation is greater than the vertical variation the error bars shown on the vertical plot will also be higher and vice versa. Please note that these error bars are per sample that makes up the point. The error bars of the averaged point itself would be a bit over 80% narrower.


Note the strong shift in *b in the top portion of the Canon 9500 chart representing the average  change vertically. It's useful to compare that to the Patchtool 9500 chart dE chart posted earlier.

Also, the Epson 9800, in a reverse of the Canon 9500, shows more variation horizontally than vertically. This is consistent with what Ethan noted about wide format Epson printers.
Logged

Doug Gray

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2197
Re: I1Isis 2 XL, the good, bad, and ugly
« Reply #23 on: August 30, 2017, 05:35:00 pm »

More good stuff from the I1 Isis 2.

In order to rule out whether the Isis was reading some portion of adjacent patches I made a special target file with 841 (29x29) 6 mm black patches. Then I added white stripes .25 mm inset from all patch edges. This changes the actual black patch sizes to 5.5 mm square with white filling the gaps between patches. This maximizes light leakage and makes it easy to detect since the reflectance of the black patches is very low. At DMax of L*=4 (my 9800's BP) leakage of .01% would produce an increase of .9 L*. Running this chart, even with a slight angle tilt, showed no signs of adjacent patch light leakage.

I've included a tif image of the target which anyone can use with an Isis. Just scan it into a target of 841 patches using the defaults and a standard letter size profile page. Save the results as a CGATS files and read it with a text editor or graph the L* values in Excel.

I am way more comfortable now that any variations I see are actually on the media and not leakage from adjacent patches.
« Last Edit: August 30, 2017, 05:58:33 pm by Doug Gray »
Logged

Doug Gray

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2197
Re: I1Isis 2 XL, the good, bad, and ugly
« Reply #24 on: September 02, 2017, 05:14:01 pm »

Building on some of the observations and comments, I created two, 9800 glossy profiles using the Isis default target of 957 patches with the following differences:

Target 1: The patches were duplicated then randomized producing a two page target but with the patches randomly scattered.

Target 2: The patches from one default target were duplicated then the columns were relocated. Each column was placed at the max and min horizontal deviation (sorted col extremes). As these columns were filled the two extremes were removed and the process repeated. This also yields 2 target pages where each of the default patches are duplicated though at positions where the horizontal inking variation tends to cancel.

Targets were printed, scanned, and profiles made. Then my 389 patch set of in-gamut LAB colors was printed using the two profiles. This differs from things such as the dE report from ColorThink which reports on the reverse lookup of RGB values and compared them to the target's scanned LAB values. This uses the AtoB1 tables in ICC profiles. Printing photos (colorimetric) does not use AtoB1 tables and only uses the BtoA1 tables.

Here are the statistics on the two, 9800 profiles which use exactly the same patches and page size but with the patches re-ordered to cancel the effect of the predictable portion of horizontal positioning inking sensitivity. Results in Delta E 2000.

CG BASE: dE's-> ave:0.46  median:0.42  <90%:0.75    chart max: 1.36
CG OPT:   dE's-> ave:0.39  median:0.35  <90%:0.67    chart max: 1.16

I consider both of these results to be well within what I would expect for a high quality profile given the relatively small number of unique (957) patches.

The BASE profile benefits from averaging targets with randomly scattered patches while the OPT benefits from strategically placing the patches such that most of the observed horizontal inking variation in the wide format 9800 cancels.


Side Notes:
The inking variation occurs most frequently on channels with a lot of added yellow. Magenta and cyan dominant colors show less pattering. Speculation is that there is only one "Y" ink whereas there are light versions of cyan and magenta.

Also, this technique of compensating for horizontal variation only improves the profile accuracy to the extent the profile is generated from affected patches . Variation in the printed image itself will exhibit this variation but no longer has the risk of compounding variation in creating the profile and variation in printing using the profile.
« Last Edit: September 02, 2017, 05:27:04 pm by Doug Gray »
Logged

Doug Gray

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2197
Re: I1Isis 2 XL, the good, bad, and ugly
« Reply #25 on: October 02, 2017, 05:26:32 pm »

I am way more comfortable now that any variations I see are actually on the media and not leakage from adjacent patches.

No longer!

I scanned a single sheet of 2920, 6mm x 6mm patches on a glossy 13x19"  3 patches were way off. 8, 10 and 11 dE compared to 3 scans I had made the day before respectively. All others were reasonably close. Ave dE was .12.

Those three patches were all on the right most column and each spaced with a good row in between. Interestingly, those patches were printed with RGB 78,0,0  0,78,0,  0,0,78 and are quite dark on the Espon 9800 with L* of 7 to 10. The patches to the left were all similar in L*. Doing the math, it appears that light was leaking from the white margins. It would take a tiny misregistration of about to produce that sort of shift. As large as it is, it's only the effect of about .5% light contamination from "white." Misregistration errors are far more damaging to darker patches than lighter ones. The same error at L*=50 would only produce about 1 dE difference.

The difference between this and the other scans, which match nicely, is that I enabled dual scan mode. This measures 2 rows with a white light, then, backs up a row and measures the same two rows with a uV LED. Also, each row is measured in alternating directions, right to left, step a row, left to right, step ... repeat.

There are other factors that should be considered. The I1Isis does not use the color changes from one patch to another to determine boundaries. They are purely a function of dividing up the total row length into precise, segments and building a window that should exclude leakage from adjacent patches. This makes the patch spacing linearity extremely critical. This is magnified when the patches range over a long distance as is the case of the 13x19, single page, 2920 patch print.

Recommendations:

When profiling larger than US letter size and you have the choice, select Profile, not Landscape, when using letter size. The longer the rows, the more error potential when interpolating where those patch boundaries are.

Use 6.5 or 7.0 mm spacing, in steps of .5 mm. This is because other fractions produce patch sizes that vary. The patch pixel width on the print is .25 mm and so .2 or .3 mm slider settings will result in different patch sizes across a row. Dumb.

Let the print dry. There could be subtle dimension changes due to variations in inking and drying. While the colors stabilize in a few minutes, the dimensions may not. I've seen some limited evidence of this.

I need to look into this more closely but I believe these recommendations will avoid most of the registration issues.

Edit for minor dim corrections.

Edit: Explanation found!
I tried to repeat scans on the target that showed the dE > 10 deviations and was not successful until I noticed that the position of the Isis XL together with the use of very a large target was the problem.

I had placed the Isis XL on a table about 28" in width, closer to exit edge. I had also select dual scan mode - I normally scan with just M2. Because the paper is fairly long at 19 inches, a portion of it would bend when extending past the table edge. This increases the drag and makes the drag asymmetrical. When in the dual scan mode the scanning motion is to read with a white lamp left to right then step forward and read right to left.  Then the device reverses paper direction and steps back to the previous scanned row where the process is repeated with the uV LED. This exacerbates the effect of asymmetrical drag gradually producing an offset that resulted in enough leaked light from adjacent patches to produce the larger error. The asymmetrical drag is made worse because the 13x19" 250 gsm paper's weight and inertia is quite an increase over my normal targets that are US Letter size and lie flat on both sides of the Isis.

Subsequently, I have changed the orientation so that the paper always lies flat and doesn't extend past the  tables with these long papers. Since then there has been no recurrence.
« Last Edit: October 05, 2017, 01:30:23 pm by Doug Gray »
Logged

Ethan Hansen

  • Full Member
  • ***
  • Offline Offline
  • Posts: 148
    • Dry Creek Photo
Re: I1Isis 2 XL, the good, bad, and ugly
« Reply #26 on: October 04, 2017, 05:04:10 pm »

Use 6.5 or 7.0 mm spacing, in steps of .5 mm. This is because other fractions produce patch sizes that vary. The patch pixel width on the print is .25 mm and so .2 or .3 mm slider settings will result in different patch sizes across a row. Dumb.

Dumb indeed. We build our iSis targets without i1Profiler. The resolution of 40 pixels/cm (101.6 px/in) produces a patch resolution of +/- 0.25mm. And i1Profiler does not accept values less than 0.1mm in normal configuration.

What we do instead is make targets at (usually) 300 ppi. This gives +/- 0.085 mm patch resolution. A typical printer will not introduce much additional sizing error. Note that we picked 300 ppi as this is the resolution used by the drivers for our main customers. Other printers vary, and you may well choose a target dimension of 360 ppi for an Epson inkjet.

Entering the resulting patch sizing into i1Profiler is easiest if you switch the units on the test chart page to inches. Magically sizing resolution switches to 0.001 inches (or 0.0254 mm), making it an order of magnitude better than the baseline behavior.

The total length of each row is sized to the number of patches * the patch size specified in i1Profiler. This means that individual patches may vary by 1 pixel; at 300ppi this is not enough to cause measurement errors. For example if you want a patch size of 6.69mm, this translates to 79 pixels at 300 ppi. Let's assume we want rows of 30 patches (7.9 inches). The total row length should be 30 [patch] * 6.69 [mm/patch] * 1/25.4 [in/mm] * 300 [px/in] = 2370 pixels (rounded down). One patch in the row will have only 78 pixels while the rest have 79.

The end result of these shenanigans is fewer read errors for a given patch size.

Doug Gray

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 2197
Re: I1Isis 2 XL, the good, bad, and ugly
« Reply #27 on: October 04, 2017, 05:53:43 pm »

Dumb indeed. We build our iSis targets without i1Profiler. The resolution of 40 pixels/cm (101.6 px/in) produces a patch resolution of +/- 0.25mm. And i1Profiler does not accept values less than 0.1mm in normal configuration.

What we do instead is make targets at (usually) 300 ppi. This gives +/- 0.085 mm patch resolution. A typical printer will not introduce much additional sizing error. Note that we picked 300 ppi as this is the resolution used by the drivers for our main customers. Other printers vary, and you may well choose a target dimension of 360 ppi for an Epson inkjet.

Entering the resulting patch sizing into i1Profiler is easiest if you switch the units on the test chart page to inches. Magically sizing resolution switches to 0.001 inches (or 0.0254 mm), making it an order of magnitude better than the baseline behavior.

The total length of each row is sized to the number of patches * the patch size specified in i1Profiler. This means that individual patches may vary by 1 pixel; at 300ppi this is not enough to cause measurement errors. For example if you want a patch size of 6.69mm, this translates to 79 pixels at 300 ppi. Let's assume we want rows of 30 patches (7.9 inches). The total row length should be 30 [patch] * 6.69 [mm/patch] * 1/25.4 [in/mm] * 300 [px/in] = 2370 pixels (rounded down). One patch in the row will have only 78 pixels while the rest have 79.

The end result of these shenanigans is fewer read errors for a given patch size.

Interesting you mentioned this. I'm in the process of creating Isis chart image files using matlab with arbitrary DPI (720 for my Epson, 600 for the Canon). I've finished the code for generating arbitraryf patch sizes from 6 to 7 mm width and height. I'm now adding another feature, creating a sliding window across the patch, vertically or horizontally, which will allow me to characterize the specific variability and effective scanned patch sizes. RGB values will be set to one value on the left and another on the right. The dividing line will be a ratio fro 0 to 1. Same for the vertical.

I am most curious what it will show. I expect it to provide clear results for maximizing comfort level of profile quality together with getting the most patch density.
Logged
Pages: 1 [2]   Go Up