Technique for precision anaylsys of Profiles

[Doug Gray]

I've been working on a precision technique for analyzing the accuracy of profiles. This technique allows one to quantify relative performance of two or more profiles to an overall accuracy of close to .01 dE00 while avoiding such issues as warm up and occasional printer jiggling which the Pro1000 tends to do at irregular intervals.

The approach is:

1. Collect the various patch sets that are to be compared relative to each other.
2. Randomize the complete set then print the patches w/o color management.
3. Scan the target pages.
4. De-randomize the patches and create profiles for each of the sets under test.
5. Create a large set of evenly distributed, in-gamut colors along with a set of neutral colors in Lab space.
6. Convert these to device RGB space using Abs. Col. for each profile created from step 4.
7. Randomize these so all RGB patches from the profiles are distributed throughout.
8. Print these patches then scan.
9. De-randomize these patches then accumulate dE00 accuracy statistics on the neutral and color patches.

The color/neutral patches tested this way have a dE00 distribution average that has a mean accuracy proportional to the square root of the size of the sets. With 700 color patches the repeatability of the mean between sets differently randomized is improved to approx. .01dE for any given profile.

One of the more interesting results is that profile accuracy improves markedly up to around 1k patches but after that the improvements become quite small up to about 3k patches. Beyond that there are no measureable improvements as the patch counts increase. There is some evidence that really large patch sets ( over 5k ) may even produce worse profiles though the differences are small (no more than .03dE00)  enough that statistical noise makes this uncertain.

There is some variation between printers. The Pro1000 reaches near optimum dE00 at about 2K patches and is quite good even with 1K while the Epson 9800 needs about 3k to reach the same, near optimum on the color patch sets.

The neutrals are a different story. The improve continuously as the patch count is increased but the major improvements of neutrals occurs from the patch set near neutral counts. This verifies the observation that I1Profiler patch sets should be chosen at the break points where the near neutral patch sets are the largest just before the count shifts the RGB grid size to the next larger value. However, the largest improvement in the neutrals occurs when "tracking" neutrals are added to patch set. This, unlike generic near neutrals, is paper/printer specific and the large improvement only applies to that printer/paper combo.

I've recently looked at the i1iSis default patch set of 957. This set contains 900 evenly spaced grid points in a 9x10x10 arrangement with an additional 57 near neutrals. I've long considered to a pretty good patch set for single, US letter page profiling.

I've just completed testing this default patch set against the same, but with 957 additional patches created using the I1Profiler "optimize" process. This yields 1914 patches in total with about 40% of the additional patches along near neutrals and 60% distributed color patches. I also included the "packed" RGB set which is an 8x8x8 grid with a 7x7x7 grid overlaid at the center points of the 8x8x8 grid producing a total of 855 patches. This was filled out to 957 patches by adding tracking neutrals.

On each of my 3 printers (9500II, 9800, Pro1000) the packed RGB w tracking neutrals totaling 957 patches outperformed the default i1iSis 957 patch set in both the color patches and neutrals with the latter by a wide margin. The 1914, I1Profiler optimized patch set slightly outperformed for the color patches but the 957 packed RGB and tracked neutrals still had a significant advantage for all 3 printers in neutrals.

[Doug Gray]

Here's the measured results for my Canon 9500 II, Pro1000, and Epson 9800

Each printer has 3 lines for the 3 profiles being compared. This first is the default i1iSis 957 patch set. The second is the default patch set with an additional 957 patches created by I1Profiler's optimize process. About 40% of these are near neutrals. This set is specific to the printer/paper.  The third line is from the 957 patches which use packed RGB values for 855 of them and additional RGB values along the LAB neutrals to fill out the 957 patch set.

Values, left to right:

Average dE00 Color patches
Standard Deviation of the mean of Color Patches

Average dE00 of worst 10% Color Patches
Standard Deviation of the mean of Worst 10% of Color Patches

Average dE00 Neutral Patches
Standard Deviation of the mean of Neutral Patches

Average dE00 of worst 10% Neutral Patches
Standard Deviation of the mean of Worst 10% of Neutral Patches


Colors              Worst Colors                     Neutrals         Worst Neutrals

Canon 9500 II, Color Patch Count:752, Neutral Patch Count:205
0.608  0.011      1.235  0.035                 0.597  0.015      0.976  0.049
0.589  0.011      1.214  0.034                 0.526  0.013      0.875  0.042
0.589  0.010      1.181  0.033                 0.387  0.012      0.718  0.037

Canon Pro1000, Color Patch Count:732, Neutral Patch Count:225
0.449  0.010      1.067  0.033                 0.552  0.026      1.491  0.083
0.421  0.010      1.012  0.031                 0.406  0.017      1.029  0.056
0.427  0.010      0.990  0.031                 0.329  0.011      0.678  0.036

Epson 9800, Color Patch Count:741, Neutral Patch Count:216
0.476  0.009      0.989  0.029                 0.494  0.015      0.918  0.049
0.400  0.008      0.869  0.025                 0.472  0.014      0.828  0.044
0.438  0.008      0.917  0.027                 0.355  0.011      0.681  0.036

What stands out is how effective adding the tracked neutrals is. The neutral accuracy is significantly better than either the default i1iSis set or the extended 1914 set that has hundreds of extra, near neutrals.

Also interesting is how close the packed RGB 957 patch set is to 2x larger, optimized set on the color patches. It's always better than the default set which uses the slightly larger color grid of 900 patches and for the Canons, quite close to the larger, optimized set.

[Doug Gray]

Here's bar graphs of in gamut L* neutral response each profile on the three printers