WAI/WCAG Contrast Research

Polarity & Uniformity Testing — Updated Dec 15, 2019

This is a demonstration of the "SuperSymmetry" of the SAPC-7 algorithms regarding Contrast Polarity. SAPC generates the same contrast percentage (as a signed negative) for reverse polarity for the same perceptual contrast, even when inverted. In other words, a -40% is light text on dark, but with a similar contrast look/feel as 40% for dark text on light. (At least in a general sense — contrast is a function of so many variables such as light adaptation absolute prediction is challenging).

SAPC In A Nutshell

SAPC is a set of contrast assessment algorithms, and the product of substantial research into vision and color science, derived from the latest in image and color assessment models (CAMs). This page and the other CE17 experiments demonstrate some of the key features of the new contrast method, which are:

• Perceptual Uniformity at a given contrast percentage for luminance in the range of Photopic vision (i.e. daylight),
• Perceptually Symmetrical reporting of contrast for positive and negative contrast polarities,
• Contrast Standards are an easy to remember set of percentages, and will seem familiar to the previous WCAG standard, and
• This (Basic) version easily tests two colors, in a manner similar to the old version (but SAPC uses correct math to avoid creating false fails and false passes.)

In addition, SAPC is extensible and can have features added (several already have the inputs set as constants, pending a future revision). For instance, future versions will be capable of scalar compliance, so designers will not be restricted to a single level. And important psychophysics such as spatial frequency and relative light adaptation are being added to the predictive model, as well as exciting new technologies that are in current development.

This is a chart of the contrast results for the test patches on this page. The solid lines are SAPC, blue is normal polarity, and magenta is reverse polarity (light text on dark BG, or WoB). As you can see the positive and negative contrast values are tightly tracking each other (reported error is less than ±0.5%) until we run into dark/black and mesopic vision.

This is the simplest version of SAPC, without any compensation or adjustments in the high end, i.e. no soft clip or boost for disability glare. On the low end this simple version also has no added/separate black level compensation for mesopic vision. But in a way this can be seen as beneficial for content authors because it results in a very subtle contrast increase for low contrasts of very dark colors.

Notice also the Mod Weber (green dotted) and WCAG (red dotted) plots for reference. WGAC is clearly over-reporting the contrast, an unfortunate cause of false-passes. Both normal and reverse SAPC hug tightly to the target 40% contrast line.

This shows the power curves for SAPC. The yellow line shows luminance (light) for the brightest color of a pair. This chart is linear as to light, and thus non linear as to perception. The SAPC perception curves relate to Steven's Law which posits a psychophysical relationship where the psychological magnitude of a sensation is proportional to a power of the stimulus producing it. (S.Stevens circa 1957)

Steven's law is similar to Weber–Fechner law, (Fechner circa 1860) and Weber Contrast is commonly used in research. But while Weber contrast may work well for reflected stimuli (i.e. signs), it fails to accommodate the myriad of perceptual & context related issues in viewing a computer monitor.

Steven's law indicates different exponents for different sensory types and contextual conditions (i.e. point source light vs large field). We use that premise in a manner similar to Fairchild's R-LAB, such that the exponent for a particular curve is assigned based on certain contextual conditions.

This shows the perceptual uniformity of SAPC. Normally, light (Y or L luminance) would be a straight line, but this chart is linear to perception, so the luminance (yellow) line is curved. For a perceptually uniform environment, the linear XY plot of the perception curves should show them as a straight line as they are here, and for perceptually uniform contrast the (now straight) lines for the two colors should be parallel as shown.

In short, the simplest description is to linearize sRGB using a 2.218 exponent^[1], then apply the sRGB spectral weighting to get Y luminance. The appropriate SAPC exponents depend on context, such as if your are using dark text on white, or light text on dark. Then convert to perceptual lightness using the appropriate SAPC exponent:

SAPC Ultra Simplified:
   LSbackgnd = (sRGBbackgnd2.218 * SpectralWeighting)SAPC_backgndExponent
   LStext  = (sRGBtext2.218  * SpectralWeighting)SAPC_textExponent

For the pair of luminance values plotted as straight & parallel lines, it is easy to determine contrast by the simple Euclidian distance between the two colors. In this case (two greyscale values) it is just the simple mathematical difference, as in

   CPredictedContrast = LSbackgnd - LStext

Order of operations is set per the use case, subtracting text from background luminance. For dark text on a light background, this generates a positive percentage, and for reverse contrast of light text on a dark background, the percentage generated is negative.

The Test Panel (Panel Four) Below:

In this test, the center point of contrast perception revolves around the brightest color in each pair, so that the BoW and WoB both share the same lightest color, the dark colors are of course different to match contrast. The target contrast is a low 40% to make contrast variations more evident.

The first two digits next to each test patch is the color of the text, then the background. For instance, 5A:C3 means the text color is #5A5A5A, and the background color is #C3C3C3. The Advanced Perceptual Contrast percentage is listed immediately under the CSS color HEX values.

It is important to keep in mind that the simple version of the SAPC used in this test does not directly consider some psychophysical contexts, such as spatial frequency, which is in development as a near-future feature.

There is more discussion after the test panel(s)

FOOTNOTE: 1) Because we are not doing "image processing" but instead seeking to predict what a monitor will do, it is more correct to use a simple exponent and not the piecewise sRGB linearization. This is partly because monitors use a simple exponent as provided in the IEC specification and the piecewise method was developed for image processing applications to prevent math problems with an infinite slope at zero. The very low level linear section of the sRGB TRC is not useful for what we are doing, which is to predict contrast on an sRGB monitor. Using a simple exponent is more accurate and also simpler for this use case.