Color Constancy test | Thomas Holm | Pixl Aps

Color Constancy test

Colour constancy on various printers papers and inks

This test is designed to illustrate the percieved change in appearence on various printed samples when viewed under different light. A set of colour patches, 28 in all, has been printed on various paper types on different printers. The patches has been measured spectrally, which allows simulation of appearence under various types of light. The reference light (D50) is the standard used within graphic arts. It is also the illuminant for which printer profiles are built to yield optimal colour.

The reference D50 daylight, 5004 K (Kelvin), CRI 100 (colour rendering index) is the black graph in the charts below.

The colour patches are comparing appearence between the reference and the following three light sources:

· Tungsten (Illuminant A) 2856 Kelvin, CRI 100

· TL84 fluorescent (Illuminant F11) 3997 Kelvin CRI 82

· Cool White Fluorescent (Illuminant F2) 4221 Kelvin CRI 64

Explanation of the numbers in the test
Delta E (dE) is a measurement method designed to quantify colour variance between two samples. 'Regular' Delta E has poor correlation with Human vision. For example will most humans define a difference in greys of around dE 0.6 as the maximum tolerable, but for a pure yellow the number could be as high as dE 20. This makes 'regular' dE numbers hard to use, if the goal is to quantify percieved colour differences or make 'Pass/Fail' decisions.

Delta E 2000 is a weighted formula which provides better correlation numerically, with what we humans see. My rule of thumb, based on empirical research, is that any number around 1.0-1.5 dE 2000 is acceptable, as differences below this limit is difficult to percieve, anything over 4 is easily percievable and generally unacceptable to most critical viewers. Your milage may vary of course. 
The numbers in the test should be read like this:

Total: This is the average deviance between all the samples (when comparing D50 and the other illuminant).
Best 90%: Similar to Total, but samples only include the best 90% of the data.
Worst 10%: This is the deviance between the least accurate 10% of the patches.

Sigma indicate something about how the distribution is between all the patches, the best 90% and the worst 10%, respectively. Sigma shows how big the 'spread' of the results is. A very small number indicates that the print changes, but all the patches change almost equally much (but not whether it is in the same direction). A low sigma value, say <1.0, usually makes it much easier for the eye to adapt to the other light source and thus percieve colour as being similar if not the same. A larger number indicates that certain patches change more than the others, which makes it impossible for the eye to adapt.

Total: How much does the very worst patch differ under the two illuminants.
Best 90%: When evaluating the worst 10% of the patches how does the best 90% of these perform.

What creates colour inconstancy
The main issues that affect colour inconstancy (apparent change of colors when viewing something under different light sources) is the composition of the inks. Certain inks reflect pretty much the same amount of colour under various light sources, others differ a lot (this is of course grossly simplified). Optical brightener in the paper can also be a source, as optical brightener will take portions of invisible ultraviolet light and re-emit it in the visual spectrum. On certain papers a yellow colour can be measured to reflect 120% of the light (which is a lot better that a perfect reflector like polished silver or what have we). Optical brighteners will affect almost all colours but particular pastels will suffer. When measuring optical brightened paper it usually reads as being blue.

In certain ink-sets, it is the presence of one or two of the inks that will tip's the balance between acceptable and unacceptable. When using certain RIP's (Raster Image Processors) one has the ability to limit inks very accurately. One can also build profiles that will drastically improve colour constancy in certain colours and in particular in neutrals. The optimal combination of inklimiting, RIP and Colour profiles can work wonders.

Optimizing color constancy
At Pixl we've spent a lot of time testing and trying to figure out how to make perfect colour AND perfect Black & White prints on ink-jet printers. We've found a formula which, in our own opinion, is very succesful. Judge for yourself though, both by the numbers and visually. We've included a sample with both an RGB profile and our own "Mono" profile concept, done on the same printer/paper/ink (and no we didn't cheat with the visuals or measurements in the test in any way). Pay particular attention the neutral colours which, in our "mono" profiles are superior to everything else in this test, and almost on par with a Macbeth ColorChecker!
We can and do build such profiles for other "Feinsmechers" who needs the best possible results. It requires a RIP and in many cases the profiles must be done on-site which makes the procedure somewhat more expensive than regular remote profiles. Please enquire Thomas,, if you are interested in this.

The techincal stuff
The patches printed are the Macbeth ColourChecker 24 + 4 additional colours from a ColorChecker DC. The colours are derived from measuring the chart (and a few other colours from the DC chart) with a spectrophotometer.
The measured LAB colours were then converted to DON RGB (a very, very large working space). Prints were then converted from DON RGB to the custom printer profiles for each printer, using Relative Colorimetric intent. The patches were then measured spectrally, and compared.
Lightsources used for comparison are standard illuminants, but they represent some the most common light sources used around.
The comparison work (somewhat simplified) like this:
When a certain colour is measured spectrally one knows how much of each wave-lengths (nm) of light wil be reflected. For the measurement filtered tungsten is used, and a mathematic model is applied to adapt the measurements to correspond with D50 daylight. Then, one can compare the reflectance of one light source with another (which has also been measured spectrally). Certain cromatic adaption formulas is applied to simulate the eye's ability to "white balance" under different light. If the cromatic adaption wasn't taken into account a white under tungsten would appear yellow and under Fluorescent light it would appear green. As you can see from the test they don't.
The cromatic adaption, for various reasons, isn't perfect but it's pretty close, and the same formula used on all samples so equal terms for all.

Credits and conclusion
I would like to extend a large thanks to the generous people who have helped with this test by printing samples on their various printers and papers - thanks a lot to all of you!
I've spent the better part of a couple of months trying to figure out exactly what the implications for various imagetypes are, and worked on possible soloutions. I'm happy to say I've discovered that by utilizing a rip (the important part is it's a CMYK printer driver) which will allow the nessesary control over the individual inks, in combination with a certain way of building profiles, color constancy problems can largely be eliminated (to a degree similar to a GretagMacbeth Color Checker. In essence people who need images that will hardly change appearence under various lightsources need look no further.