What is Wrong with the CIE L*a*b* Color Space Specification

The CIE Lab specification (like many other color specifications and color research) is quite old, it dates back to year 1976. It was pushed rather strongly by the industry, the color cons(ortium), who wanted something simple for the "high end" "color management" applications of those days. So the CIE Lab color-space was born, it was said to be perceptually uniform color-space.

Today we have true and accurate ICC color-management, we know more about color, more about the human vision and we have more accurate tools and methods to characterize and measure both color and the behavior of the vision. Even then the misinformation is still the same, being repeated over and over by some color "specialists" as well as companies.

The problem with the CIE Lab is that it is nowhere near perceptually uniform color space. Well no, that is not yet the problem. One of the the main problem for digital imaging is that some people, e.g. those who write ICC profiling applications, truly believe that the CIE Lab would be a perceptually uniform space, so they use the CIE Lab deltaE* (dE*) weighting in their profilers and those profilers then perform just miserably.

The dE* is being used in gamut mapping algorithms that various ICC color-conversion "engines" do when profile converting e.g. from the color-space of the image-file to the viewing screen or to the printer or just to another color-space.

Some people are using the CIE Lab dE* for the purpose of reducing the colors when creating GIF images from continuos tone photographs, this btw is what e.g. Photoshop does.

There are many other people and businesses who rely on the same ultimately incorrect dE*, it is is being used e.g. in paint, textile etc industries quite largely.

What is the CIE Lab deltaE*

The CIE deltaE*, shortly dE*, is defined as an measure of color difference, namely the Euclidean distance or the straight line distance between two colors that are expressed by the Lab coordinates.

Because it is said that the CIE Lab color-space would be a perceptually uniform color-space then the Euclidean distance between two colors in this space would be a perceptually uniform measure for color-difference. The dE* is calculated as:

dE* = [(L1-L2)^2 + (a1-a2)^2 + (b1-b2)^2] ^(1/2) where:

L1, a1 and b1 are the Lab values of color 1 and
L2, a2 and b2 are the Lab values of color 2.

Introduction to CIE Lab dE* errors

A quite simple evaluation shows that the dE* is no where near a perceptual measure of color difference; in the below image there are 48 comparison patches that all have exactly 10dE* difference between the upper and lower square. So in each of the 48 comparison patches the perceptual difference between the upper and lower square should be "similar" in "amount" or should give "sensation" that appear to be of "equal strength".

10dE* difference in every comparison patch

The above image is in the sadRGB color-space so it is not colorimetricly accurate in your Web browser and it has the JPEG damage over it, if you like to evaluate the Lab values yourself please download the original Photoshop Lab mode PSD file. Please note, some versions of Corel Photopaint (v8 and v9 at least) show the a* and b* values differently even if the image appears the same as in Photoshop.

The above 48 comparison patches are divided into 12 groups, in each of these 12 group the lower square has the same Lab value and the upper squares are selected colors around that base color, exactly 10dE* away from it.

It is also said that 1dE* would the about the threshold for the human vision so that difference below 1dE* would not be discernible. Now note again that the difference in each of the above 48 comparison patches is as much as 10dE*. And to reiterate, the perceptual difference in each of the above 48 comparison patches should appear to be similar in "amount", should give equally "strong sensation". Does your vision comply with that?

A closer look at the huge errors of the CIE Lab dE*

Yes, in case it happens that you extremely easily discern a difference of 10dE* somewhere in the color-space but have trouble to discern that same 10dE* difference elsewhere in that color-space then yes, the errors are just huge, rendering the dE* totally useless; it is said that 1dE* is the threshold of human vision so this demonstrates an error of 10x in magnitude.

Below are 7 evaluation images, each evaluation image has 30 comparisons patches, in each comparison patch the color difference is 10dE* between the upper and lower square. In each evaluation image the lower square of each the 30 comparison patches has the same Lab value so the upper squares are 10dE* apart from the lower squares.

What does it mean: The perceptual difference in all the 7*30 comparison patches should be "similar" in "amount" or should give "sensation" that appear to be of "equal strength" IF the CIE Lab is a perceptually uniform color-space and IF the dE* is an perceptually uniform measure of color-difference.

The evaluation images are in the sadRGB color-space so they are not colorimetricly accurate in your Web browser and they have the JPEG damage over them, for evaluating in Photoshop please download the original Lab mode Photoshop file.

The evaluations are overlaid on a quite typical photograph, this is because the aim of this evaluation is to assess the usefulness of the Lab color-space and the dE* as a color-difference measure for photographic imaging purposes. Please download the original Lab mode Photoshop file, the image is on a separate layer in it so you can easily change the background.

As can be seen (in case you have not already done so please download the original Lab mode Photoshop file) when the 10dE* difference is coming from the L* value alone it appears to cause extremely large perceptual differences.

So the transfer-function (or the tone reproduction curve) of the Lab space, the L* channel, is way incorrect. The L* is quite close to the transfer-function of the native CRT monitor so in the early days of color-management it was convenient to claim that the L* would be perceptually uniform, that way no special attention was needed in order to take the transfer-function of the native CRT (gamma 1/2.5) into account.

Good reading about the issue: Gunnar TONNQUIST; choose the link Colour order systems there.

See what Adobe Inc has to say about the Lab color-space in their TIFF specification (section 23) and have a good laugh.

Continue to Part II after discussions on the Usenet newsgroup sci.engr.color the same test charts with only 3.6dE* color difference.