Perceptual Colour Tolerancing: From Delta E to Press-Ready Files

When a designer, a brand owner and a press operator disagree about colour, they are rarely seeing different things. The trouble is usually a specification that never said, in numbers, how close counts as close enough, how that should be measured, or under what light. The sections below explain why distances in device-dependent RGB and CMYK do not match perceived difference, why CIELAB and CIEDE2000 were developed, and how perceptual colour tolerancing turns a stated tolerance into an acceptance test that a pre-press pipeline can run automatically before a file ever reaches a plate.

Device colour values versus device-independent colorimetry

An RGB triplet or a set of CMYK tint values tells a device what to do. It does not describe the colour stimulus that comes out. Send the same four tint values to two offset presses and you get two different colours, because ink pigmentation, ink film thickness, screening, tone value increase and substrate all sit between the number and the light that leaves the sheet. The number is correct as an instruction, but it carries none of the information you would need to verify the result. Device-independent colorimetry exists to close that gap. It describes the stimulus itself, instead of the machine setting that produced it (Commission Internationale de l'Éclairage, 2004).

Two more effects make informal colour communication unreliable, even between people acting in good faith. The first is metamerism. Two samples with different spectral reflectance curves can match under one illuminant and visibly differ under another. A proof made with a dye set and a production sheet printed with process inks are a classic metameric pair, which is why a match agreed at a window in the afternoon can fail under the shop's fluorescent lighting. The second is that colour appearance depends on viewing conditions: adaptation state, surround luminance, the size of the sample and the colour of whatever borders it. Colour appearance modelling is a separate discipline for this reason (Fairchild, 2013), and production work has to take it into account.

Stages where colour ambiguity accumulates in a print job

A job goes through a design file, a soft proof on a monitor, a contract proof, a plate, a press sheet at make-ready, sheets sampled across the run, and finally the copy that lands on the client's desk under their office lighting. Each step changes at least one variable. Without a numeric criterion tied to a defined measurement condition, every one of those steps gives two competent people a chance to reach opposite conclusions, with neither able to prove their case.

CIELAB and the motivation for a perceptually uniform space

CIELAB is a nonlinear transformation of CIE XYZ tristimulus values relative to a specified white point. It was designed so that Euclidean distance in the space approximates perceived colour difference. Lightness L* runs from 0 to 100, and a* and b* are opponent axes running roughly red–green and yellow–blue. The transformation is:

L* = 116·f(Y/Yn) − 16
a* = 500·[f(X/Xn) − f(Y/Yn)]
b* = 200·[f(Y/Yn) − f(Z/Zn)]

where f(t) = t^(1/3) above a small threshold and a linear segment below it, and Xn, Yn, Zn are the tristimulus values of the reference white. Chroma is C* = sqrt(a*² + b*²) and hue angle is h = atan2(b*, a*). These matter because tolerancing decisions are usually easier to express in lightness, chroma and hue than in the Cartesian axes.

The important word in the design goal is approximates. CIELAB was adopted as a working compromise. It is far more uniform than any device space, it is cheap to compute, and it has been the common currency of industrial colour specification for decades (Judd & Wyszecki, 1975; Hunt & Pointer, 2011). But it is not uniform enough for a single Euclidean distance to mean the same thing everywhere in the space. That leftover non-uniformity is why the Delta E family has more than one member.

Delta E formulas from ΔE*ab to CIEDE2000

The 1976 formula is the plain Euclidean distance:

ΔE*ab = sqrt(ΔL*² + Δa*² + Δb*²)

Its failures are systematic rather than random. Perceptual tolerance grows with chroma, so a given ΔE*ab is far more visible on a near-neutral grey than on a saturated colour. Sensitivity varies with hue, and the blue region behaves particularly badly. Perception also weights lightness differences and chromatic differences unequally, while the formula weights them the same.

Later formulas dealt with this by dividing each component by a weighting function that depends on where the pair sits in the space. CIEDE2000 (Luo, Cui & Rigg, 2001) is the current recommendation for small colour differences and takes the form:

ΔE00 = sqrt[ (ΔL'/(kL·SL))² + (ΔC'/(kC·SC))² + (ΔH'/(kH·SH))²
             + RT·(ΔC'/(kC·SC))·(ΔH'/(kH·SH)) ]

SL, SC and SH are the lightness, chroma and hue weighting functions. kL, kC and kH are parametric factors set by the application, conventionally 1:1:1 in graphic arts. RT is an interaction term that rotates the tolerance ellipse in the blue region, which corrects a failure the earlier formulas had no way to express. And a* is rescaled for near-neutral colours so that behaviour along the grey axis is predicted better.

In practice this means that ΔE00 and ΔE*ab give different numbers for the same physical pair, and you cannot swap one for the other. A tolerance written as "ΔE 2" cannot be verified. A tolerance is only complete when it names the formula, the parametric factors, and the measurement conditions discussed below.

Setting colour tolerances per element class

The reflex is to find the threshold of visibility and adopt it. That approach is wrong for two reasons. First, difference thresholds are statistical properties of observer populations and have no sharp edge. A difference of around one unit is commonly described as near the limit of detection for a trained observer under ideal side-by-side conditions, and those are rarely the conditions that matter commercially. Second, and more important, Taguchi's loss function (Taguchi, 1986) is the better model. Quality loss grows continuously as a characteristic drifts from target, instead of appearing as a step at the specification limit. So a tolerance is a commercial decision about where accumulated loss stops being acceptable. The same accounting governs the cost of quality in short-run production, where the cost of tightening a limit has to be weighed against the cost of the failures it prevents.

This also explains why one number for a whole sheet cannot be defended. Different elements carry different loss functions, so they need different tolerance zones.

Two more decisions are often left out of specifications, and they are the ones that end up in disputes. A tolerance should say whether it applies to the mean of the sampled sheets, to the worst single sheet, or to both with different limits. A run has variance, and a specification that says nothing about it cannot be tested. A tolerance should also state a sampling plan, because a number checked on one sheet at make-ready says very little about sheet nine thousand. That is the sampling question statistical process control in the age of machine vision exists to answer, and colour is one more characteristic with a mean and a variance to chart.

Measurement conditions in a colour tolerance specification

A ΔE reported without its conditions cannot be checked by anyone else. A complete specification names the illuminant and observer (D50 and the 2° standard observer for graphic arts), the instrument geometry, whether a polarising filter was used, the backing behind the sheet, the aperture size, and how dry the ink was when the reading was taken. Any one of these can move a reading by more than the tolerance being argued about. Substrate has the largest effect. The paper white is part of the stimulus, and a colour printed on two different papers cannot measure the same, which is why it makes more sense to evaluate against a substrate-corrected aim than against an absolute one.

ISO 12647 codifies the instrument side of this for offset process control. ISO 3664 covers the human side: it specifies viewing conditions for graphic technology. You need both. Even a fully instrumented workflow ends with a person looking at a sheet in a booth, and a specification that covers the instrument but leaves out the booth has only moved the dispute somewhere else.

Automating the colour tolerance check in pre-press

Once a tolerance names a formula, a target, a condition and a sampling rule, a machine can check it, and the natural place to do that is before the file becomes a plate. Our own automated QC pipeline for a collectible card producer runs 90+ validation checkpoints covering typography, spelling, image placement, colour accuracy and print specification. Roughly 95% of the QC process is automated. An inspection pass that took eight hours by hand now takes about 15 minutes, and production is around 300% faster. A follow-on pre-press tool converts approved designs into press-ready montage files, so the artwork that passes the gate is the artwork that gets imposed.

The colour part of such a gate is simple in structure:

for element in artwork.tagged_elements:
    aim      = spec.aim_lab[element.class]
    measured = to_lab(element, profile=spec.icc, illuminant="D50")
    dE       = ciede2000(aim, measured, kL=1, kC=1, kH=1)
    if dE > spec.tolerance[element.class]:
        record_failure(element, dE, aim, spec.tolerance[element.class])

Four design rules decide whether people trust the gate or work around it. The tolerance table is data, versioned alongside the job, and not a constant compiled into the checker. Every failure reports the measured value, the aim and the limit, so whoever reads the report knows what to fix. The gate has three outcomes: pass, fail, and refer to a human. In the ambiguous band the automation should ask instead of deciding, and human-in-the-loop design under automation bias is what keeps that referral a real judgement and not a rubber stamp. Finally, elements have to be tagged with their class in the file itself, since a checker cannot work out on its own that a particular patch of ink is a brand mark.

What a colour difference metric cannot decide

A colour difference formula reduces a pair of stimuli to one scalar, and that scalar throws away information production decisions depend on. The scalar has no direction, and a warm and a cool deviation of the same size are not equally acceptable on a face. It carries no spatial structure either, so banding, mottle, ghosting and colour fringing from misregistration can all coexist with excellent per-patch numbers. Gamut limits are outside its scope. When a brand colour lies outside the achievable gamut, the smallest attainable difference is above zero and the real question is which rendering compromise to accept. Metameric pairs, gloss and finish differences, and appearance effects such as simultaneous contrast are invisible to it as well. Use the metric as a gate, and expect the questions above to still need someone's judgement.

Limitations

We do not claim that any specific tolerance value is right for any specific product. The tolerance postures in the table are reasoning about element classes, not measured thresholds, and they assume trained observers, controlled viewing and offset process printing, which will not hold everywhere. We also do not claim that CIEDE2000 predicts perceived difference well for large colour differences, textured samples or self-luminous displays. The formula was not developed for those conditions. The first-party pipeline figures above describe our own work on one product family in one production context, and should not be read as what automated QC will do in general. And none of this settles the organisational question that usually comes before the technical one: whether the parties to a print job are willing to agree on a number before production, instead of arguing about a sheet afterwards.

References

• Commission Internationale de l'Éclairage. (2004). Colorimetry.
• Fairchild, M. D. (2013). Color Appearance Models.
• Hunt, R. W. G., & Pointer, M. R. (2011). Measuring Colour.
• International Organization for Standardization. ISO 3664: Graphic technology and photography — Viewing conditions.
• International Organization for Standardization. ISO 12647: Graphic technology — Process control for the production of half-tone colour separations, proof and production prints.
• Judd, D. B., & Wyszecki, G. (1975). Color in Business, Science and Industry.
• Luo, M. R., Cui, G., & Rigg, B. (2001). The development of the CIE 2000 colour-difference formula: CIEDE2000.
• Taguchi, G. (1986). Introduction to Quality Engineering.