xref #360 (comment)

xref #360 (comment)

Thank you Patrick but as you can see I already posted in that thread. That is a separate and somewhat minor issue. The issue I discuss in THIS thread is specifically about the minimum contrast 1.4.3, and is not minor as it has far-reaching consequences such as a ton of apps that now incorrectly present colors as "accessible" when in fact they are not. And this issue has led to a great deal of misunderstanding regarding color choices and contrast.

The thread you linked to deals with a minor error in the relative luminance equation, and while the W3C picked the wrong equation that is not what is causing the much more serious problem that I outline in this separate issue.

The issue HERE is using a simple contrast (L1/L2) on linear luminance to define color & luminance contrast. But this does not provide any meaningful value for perceived contrast.

I posted this as an issue for discussion while I continue my research (search for an accurate programatic contrast assessment) and pull request separately.

Some additional thoughts regarding 1.4.3

FWIW MY BACKGROUND: I work in the film and television industry in Hollywood as an editor/colorist/VFX & Title Supervisor. I work with color and visual perception issues every day.

I am going to continue to post in this thread while I delve further into this before generating a pull request. It is a concern for me because this W3C document is considered authoritative, and has made its way into government regulations. It is important that it be correct, and it is not at present. PLEASE COMMENT if you have thoughts or insights as to why some of these choices were made. Thank you.

On Terms:

Simple contrast "is not useful for real-world luminances, because of their much higher dynamic range and the logarithmic response characteristics of the human eye."[1]

and using simple contrast seems to have led to the higher (4.5:1) contrast specifications:

What is the cite and specific justification for the claimed need for a 4.5:1 contrast ratio? Studies by Legge, Rubin, Bangor etc. found that "Contrast by itself had no significance for either vision group." [unimpaired or impaired] [2]

However font size and polarity are very important, and contrast does interact with very small font sizes to a degree, especially in negative polarity. The Bangor study indicated that font sizes below 18 px resulted in a need for increased contrast, but these study participants were ether legally blind (20/200) or very impaired (20/100).

It is NOT about contrast as much as size and possibly polarity. While it is true that increasing contrast can help legibility for small fonts for visually impaired, increasing the font size offers a better improvement.

As I mentioned in my first post Weber or Michaelson are possibly better here, and one of those are what is used in nearly all the research & standards. However, I am working with Bartleson-Breneman perceptual contrast at the moment. EDIT: they are used in research for consistency and for the at-threshold JND, but are not perceptual uniform at supra-threshold which is what is needed for design guidance.

To make this point more clear: The "simple" ratio of #FFF to #808080 is 4.6:1 (3.95:1 if you add in the W3C's 5% bonus luminance). But #808080 to #040404 is a ratio of 178.88:1 (5:19:1 using the 5% extra).

#FFF is luminance 100, mid-grey #808080 is 21.59, and #040404 is 0.12

So ignoring the oddly-applied/misapplied "flare" value, white to mid grey is a ratio of 100:21.6 and mid grey to black is a ratio of 21.7:0.12

BUT because the first much smaller ratio is also associated with high luminance it is much easier to read and has much better PERCEPTUAL contrast:

4.6:1 (3.95:1)

And the black one with 179:1 contrast (LOL, 5.19:1 with the cheat)

I find no justification for the 4.5:1 contrast ratio for 20/40 vision as indicated in the W3's standard. Is it set that way (along with the excessive flare luminance add) to attempt make up for the other deficiencies?

See also reference [3] below, a EU paper on this subject.

Contrast Sensitivity is a separate measurement from visual acuity. From the referenced Arditi paper: "visual acuity measurements alone are insufficient to characterize basic spatial visual function..." But I don't see where you get multiplying the ISO well established 3:1 standard by 1.5 ?? Looking at acuity vs contrast graphs is see a difference in CS as low as 5% for a 20/40 person. And as I recall the common 3:1 luminance contrast ratio included near normal vision (20/40 is near normal).

Here's a graph, (for reference logMAR 0.3 is approximately 20/40.)

In short, it appears to me the 4.5:1 contrast standard is somewhat arbitrary, and there are other more important means to improve accessibility, namely font size, appropriate polarity, and total luminance.

LUMINANCE:

NITS TO THE RESCUE! (by nits I mean cd/m^2, 1 cd/m^2 is 1 nit ... but maybe I also mean ME, nit-picking on this issue, LOL).

The sRGB spec states an 80 nit monitor, however people commonly adjust them to 120 nit to 160 nit, even more (300+ is common. Some phones do 1200). If the monitor is brighter, and the material is black text on white, the light from the monitor results in pupil contraction which improves perceived sharpness.

I'll opine that it is more important to have a monitor that is adjusted bright enough for its environment. In fact it would be a good idea to lobby the ISO for an amendment to the sRGB spec to adjust away from 80 cd/m2 to a specific luminance based on the environment. 1996 was a long time ago, and display technology has changed substantially — we shouldn't have to adjust the ROOM lighting to match the monitor, it's easier to adjust the monitor. A standard stating the max display luminance for a given ambient light would go a long way toward real accessibility/accommodation.

SUMMARY:

The main thing I am lobbying for here is a revised programatic contrast assessment that is perceptually correct. But as I research this, I see there are other concerns that should be considered.

• New contrast assessment method, with a revision of the standard to accurately model perceptual needs. I.e. no simple ratios, etc.
• Among the contrast revisions, provide guidance on excessive contrast — too high of a contrast causes reading problems and eye strain as well.
• Add guidance on display polarization (light on dark vs dark on light) to the standard.
• Along with polarization, guidance for local adaptation which is part of the issue with current contrast assessments. (Local adaptation is one reason that white on a darker background looks more contrasty than grey on a dark background even if the luminance ratio is the same).
• Emphasize font size for accessibility, with guidance to avoid ever overriding the user agent's root font size.
• Lobby ISO to amend sRGB to provide a standard for adjusting a display to the room. 80 nits is really a bit dim in a lot of environments.

Thank you for reading. I hope to have a solid contrast assessment model this week.

Andy

Refs:
[1] https://www.schorsch.com/en/kbase/glossary/contrast.html
[2] https://pdfs.semanticscholar.org/4f9f/f4bcbc0eb8d1040228e5d84cd0c0e75962c7.pdf
[3] https://www.anec.eu/images/Publications/technical-studies/ANEC-final-report-1503-1700-Lenoir-et-al.pdf

Edited May 22 for typos and some clarity.

Hi @Myndex

• I agree that there are some combinations that pass that should fail and that there are some situations that fail that should pass. However, in general the algorithm works well and has made a big difference to ensure that text has better contrast.
• While I agree that lower acuity doesn't necessarily mean less contrast sensitivity -- once you get into the 2070 20200 levels folks tend to have eye conditions with multiple factors and there is a larger correlation between the two
• 20/40 may not be visually impaired at least by most US standards. 20/70 is generally considered visually impaired by many and I believe the criteria were written to help many of us who use vision in the range of 20/70 - 20/400. For someone with 20/200 contrast very much tends to be an important issue and these guidelines were written for folks in this range. 4.5 is very much a tipping for point for many of us and I do not agree that 4.5 is too high for someone who is visually impaired or legally blind. Many legally blind people use their vision and require this type of minimum contrast. So if the understanding text is not clear around this area then we need to update it.

Hi @mraccess77

Thank you for commenting, it helps me to see when I am not explaining or describing completely. But perhaps more important it leads to some new discoveries. In answering your post I did some experiments that add insight. More below.

First, just to provide a little more background, I want to mention that I have personal experience with 20/200 vision. Several years ago (in my late 40s) I developed early onset cataracts which brought my vision to worse than 20/200 in one eye, the other diminished a bit less. I now have IOL implants, but those surgeries caused vitreal detachments and retinal detachments, requiring a vitrectomy in one eye, and in the other, continuing issues due to large vitreal floaters that still can interfere with reading. Also, I still need glasses (i.e. it is trivial for me to remove them to introduce poor vision).

As such WCAG is a topic I have a close personal interest in.

As I mentioned in an above post, I am an imaging professional in Hollywood with a career that spans decades and a background that includes broadcast engineering, colorist, VFX Supervisor, and perhaps most relevant, title designer.

I came across this WCAG issue while developing a CSS framework. For the color module I am trying to create a simple color subset that is both easy to read and aesthetically pleasing. This led me down the color and vision theory rabbit hole, where I stumbled on a contrast calculator and saw an odd comment by the coder who mentioned that his "calculate" button did not pass the WCAG standard. The button is completely readable with more than adequate contrast, thus started my present research journey.

I have since been doing nothing but research this issue in depth. My posts here are based on that research and my extensive experience in digital imaging.

• mraccess77 said: I agree that there are some combinations that pass that should fail and that there are some situations that fail that should pass. However, in general the algorithm works well and has made a big difference to ensure that text has better contrast.

I cannot agree here. Estimating roughly, 40% of the color pairs the WCAG math calls "PASS" are poor in quality and should fail. And somewhere in the area of 51% of colors they fail could conceivably pass.

Wrong nearly half the time is one huge fail. I believe it is the result of incorrect assumptions and cherry picking various bits of standards and cobbling them together into something that is truly inaccurate and unsuitable for the purpose.

I'm going to quote Whittle from his paper [1] (emphasis added)

With regard to the mathematical description, an important distinction is between ratios and contrast expressions that also involve differences. This concerns what is meant by ‘contrast’. Weber contrast and Michelson contrast are the commonest expressions for it, not the simple ratio L/Lb. Weber contrast = (L- Lb)/Lb, usually written ÆL/Lb. When the backgrounds are weak a ‘dark light’ constant must be included: ÆL/(Lb+L0). Michelson contrast = |ÆL|/(L+Lb). When people talk of contrast coding in the early stages of the visual pathway, they usually mean that the firing rate of sensory neurons is a function of, perhaps proportional to, Weber contrast or Michelson contrast. Such a code combines differencing (calculating L-Lb) with normalisation (attenuating by some function of the absolute stimulus level).

And then from Pelli's paper [2] (emphasis added)

The contrast of the target quantifies its relative difference in luminance from the background, and may be specified as Weber contrast Lmax−LminLbackground, Michelson contrast Lmax−LminLmax+Lmin, or RMS contrast LσLμ, where Lmax, Lmin, Lbackground, Lμ, and Lσ are luminance maximum, minimum, background, mean, and standard deviation, respectively. Weber contrast is preferred for letter stimuli, Michelson contrast is preferred for gratings, and RMS contrast is preferred for natural stimuli and efficiency calculations (Bex & Makous, 2002; Pelli & Farell, 1999). Threshold contrast is the contrast required to see the target reliably. The reciprocal of threshold is called sensitivity.

The WCAG ignores this wholesale, despite it being prominent in most research and even noted in the ISO and ANSI standards. WCAG is using simple contrast (Lw/Lk) and that is one of the errors.

Among my findings, some of the sites I have the most difficultly reading are "compliant" with the WCAG. There are countless contrast checkers and other automated or semi automated accessibility checkers that use the algorithm as written, and they all fail to detect poor perceptual contrast.

And there are a LOT of combinations that pass but are hard to read. An accessibility site has this on one of their pages indicating the problem:

The contrast checkers are all using this flawed model which ignores many important aspects of perception. The results are all over the place and inconsistent. I can see why designers are ignoring the contrast recommendations: they are ambiguous and inconsistent at best. At the moment, visual judgement does a better job than the contrast calculators.

I have a page where I am conducting live experiments in this regard, aloing with some commentary on my findings as I go along. This link is:
https://www.myndex.com/WEB/W3Contrastissue

Here are some examples from today's experiments:

Today, I've been looking into luminance — it is well known and researched that increasing luminance improves readability (within a range). One thing the WCAG lacks is a specification on minimum lightness for the lightest element in a color pair.

I have more on this on the page, but as you can see, setting a minimum lightness of the lightest element to #AAA results in a consistent, readable, block of text. On the page you'll see examples where pages with 3:1 contrast and minimum #AAA on the brightest of the pair is more readable than 4.5:1 contrast on darker pairs.

• mraccess77 said: While I agree that lower acuity doesn't necessarily mean less contrast sensitivity -- once you get into the 2070 20200 levels folks tend to have eye conditions with multiple factors and there is a larger correlation between the two.

Yes, I am aware, there is definitely correlation to contrast sensitivity due to a number of vision impairments. Indeed, one can have good visual acuity and bad contrast sensitivity. The WCAG does not discuss CS though, and only lists some Snellen numbers like 20/40.

I was talking mainly of 20/40, which is what the WCAG talks about regarding AA. For the portion of the standard that relates to profoundly impaired, still the math they provide does not create useful numbers for guidance.

And that is the point I am getting at. The math is essentially wrong. Lw/Lk is not well suited for determining contrast in this context. The vast majority of research on contrast sensitivity uses WEBER or MICHELSON or both. Rarely simple contrast. But there are other math mistakes in WCAG related to sRGB and computer displays that also need to be addressed.

(EDIT by Andy: May 2019: My recent experiments and research indicates that a "classical, unmodified Weber contrast" is really not "substantially" better than the WCAG math, though there is a modified Weber from Hwang/Peli that is much better than the WCAG math, and other more modern contrast equations such as PCL).

• mraccess77 said: 20/40 may not be visually impaired at least by most US standards. 20/70 is generally considered visually impaired by many and I believe the criteria were written to help many of us who use vision in the range of 20/70 - 20/400. For someone with 20/200 contrast very much tends to be an important issue and these guidelines were written for folks in this range. 4.5 is very much a tipping for point for many of us and I do not agree that 4.5 is too high for someone who is visually impaired or legally blind. Many legally blind people use their vision and require this type of minimum contrast. So if the understanding text is not clear around this area then we need to update it.

I never said that 4.5:1 is too high, particularly in regards to profound impairment, 4.5:1 is TOO LOW when using the WCAG math. WCAG indicates 7:1 for the more visually impaired (AAA). And yes, the "understanding" text is all over the place and not clear.

To be clear, I am not "condemning" the 4.5:1 ratio per se, but I am questioning where it was derived from and the basis of the equations when those equations are not supported by standards nor research. I am also pointing out that luminance is a much bigger factor than contrast yet it is not mentioned, nor is local adaptation unless I missed seeing that. My statements here are from published research as well as my own research.

The California Council of the Blind (Lozano, 2009) states, and Federal ADA guidelines also state, that contrast for signs be 70% (note it is a percent not a ratio). The math used for the Federal standard is (B1-B2)/B1 (B1 is the lighter LRV, and B2 is the darker).[3] Now, if I use the WCAG math, 70% equals a ratio somewhere around 2.3:1 to 3.2:1. WCAG math is all over the place and does not relate to Weber's law nor anything else useful.

Nevertheless, California Council of the Blind (Lozano, 2009)[4] is on record stating that the Federal equation is flawed when B1 is less than 45. I just came across this a minute ago — I'm slightly amused as it is closely mirroring what I have been saying about the WCAG.

I describe this in more detail on the experiments page, and there are more examples.

In closing I just want to say that simply switching the equation to Weber is not the complete answer. I think we can do better, and that is the focus of my research.

Thank you again for the comments.

Andy

REFS:
[1]http://aardvark.ucsd.edu/color/whittle.pdf
[2]https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3744596/
[3]https://segd.org/sites/default/files/SEGD_2012_ADA_White_Paper_Update.pdf
[4]https://www.documents.dgs.ca.gov/dsa/access/Pt2_Final-SOR.pdfHi @mraccess77

Edited May 2019 for some minor clarity fixes.

Hi @Myndex I appreciate your efforts in addressing the shortcomings of the current algorithm. In your examples, I personally found some of the first 4.5 items easier to read than the ones with values greater than #AAA -- thus I know there will always be differences in interpretation by different people as we all see differently. Along those lines with the adaption you were discussing, halos may technically be used to meet the requirement but when you take into account the width of the stroke and surrounding colors haloed text can actually be harder for me to read. I agree that we want more people to use contrasting colors that meet users needs and if we can change the algorithm to meet those needs without lessening it get more adoption that would be a good thing. Personally, I see these changes as something that can't be changed with the current standard as the method is too normative to change with an errata but would be a great opportunity to address for the next version of the accessibility guidelines (silver). It would be good to socialize this with some other folks such as Jared Smith from WebAIM who also would like to change the future direction of the contrast calculations and the Low Vision Accessibility Task Force which is part of the Accessibility Guidelines Working group. Adding @WayneEDick and @allanj-uaag.

Thanks @Myndex for writing this up so thoroughly!

2023 edit for clarification: Weber is not the "gold standard" today, and the link to this post is an inappropriate reference made by Bengfort, as this post was part of the very early due diligence looking into the problems of WCAG 2 contrast failures. (In fact this post was made literally within the first two days of looking into and discussing this problem, four years later the body of knowledge in this area is substantially improved).

Weber is not perceptually uniform for the purpose of supra-threshold text on self-illuminated displays. Further, deeper investigation into the origins of the WCAG 2.0 contrast math shows that it is a derivation of Weber, but it is extremely important to point out that the derivation was originally intended only for 7:1, and not intended for lower contrasts such as 3:1 or 4.5:1. At 7:1, the dark color failures of WCAG 2 contrast were not as obvious.

Nevertheless the poor results of WCAG2 were used as part of the misguided 1.4.3 SC, despite the objections of major stakeholders including IBM.

Original 2019 message continues below:

Hi @mraccess77

Hi @Myndex I appreciate your efforts in addressing the shortcomings of the current algorithm. In your examples, I personally found some of the first 4.5 items easier to read than the ones with values greater than #AAA.

If that is based on the images in my post above, I should mention that the SIZE of the first set is 30% larger, and therefore easier to read (I just realized the scaling error due to how this site handles images as I looked at the post, post edited to correct).BUT ALSO, three of the first 5 have the brightest color well above #AAA. For an apples to apples comparison, please see the live experiment on the website: https://www.myndex.com/WEB/W3Contrastissue
Text scaling on the site for the samples is equal in each group (the first group is general contrast assessments, and the second is body text).

thus I know there will always be differences in interpretation by different people as we all see differently. Along those lines with the adaption you were discussing, halos may technically be used to meet the requirement but when you take into account the width of the stroke and surrounding colors haloed text can actually be harder for me to read.

Indeed, for instance, research shows that most people do better with dark text on a light background (Positive Display) but with my vision, I much prefer light/colored text on a black background (Negative Display). Right now I am having difficulty with THIS site due to the bright background (L* 98 in the text area) yet for most people this is the ideal presentation.

I agree that we want more people to use contrasting colors that meet users needs and if we can change the algorithm to meet those needs without lessening it get more adoption that would be a good thing. Personally, I see these changes as something that can't be changed with the current standard as the method is too normative to change with an errata but would be a great opportunity to address for the next version of the accessibility guidelines (silver).

Yes I see it was just tagged as WCAG 2.2 which I was somewhat expecting. Correcting the algorithm also means changing the standard, as far as I can tell the current standard(s) seem to be compensating for the math issues — I don't know the complete history, but that is how it appears based on the reverse engineering/analysis.

For an errata, it might be useful to place a note to the effect of: "Current contrast algorithms may overvalue contrasts with pairs of darker colors. Designers should be cautioned not to rely on contrast numbers in these cases/"

I should note that problems and controversy on this very subject are visibly present in the research and some standards. It is partly why I am being so proactive here. I hope to cut through the clutter to bring some clarity (puns more or less intended).

It would be good to socialize this with some other folks such as Jared Smith from WebAIM who also would like to change the future direction of the contrast calculations and the Low Vision Accessibility Task Force which is part of the Accessibility Guidelines Working group. Adding @WayneEDick and @allanj-uaag.

Excellent. What are the deadlines for 2.2? As I mentioned in one of my posts while there is much research on simple monitor displays (i.e. black on white, white on black), there is not much in the way of research on complex, graphically rich content (that I've found anyway). I'm thinking some empirical studies would be illustrative.

ON ALGORITHMS:
Weber has been the "gold standard" for text contrast. But I am presently leaning toward a modified version of Perceptual Contrast Length (PCL) which "essentially" uses a modified L* type curve (perceptual lightness) in calculating contrast. I believe I have a way to adapt it so the contrast results are similar to what people are used to using (i.e. 3:1 etc).

• 2022 EDIT for clarity: Weber or Michelson are used in research, but neither are "perceptually uniform over the visual range" as neither takes adaptation into account. To grossly oversimplify, this means only the stimulus can change, but not the background as that directly impacts adaptation. Weber may have been the "gold standard" once upon a time, back in the day when everything was printed on white paper—it is not even close to accurate when adaptation and spatial frequency are added variables. Stevens in the 1960s was critical of Weber and found various power-laws better defined perception, and I believe Stevens was the first (or among them) to recognize the importance of spatial frequency. More recently Poynton observed that the lightness perception curve is better described as part log and part power-law. In 2019, the author found further variables that are the subject of a paper (WIP).

Original post continues:

Another idea is using the difference between a brighter/darker L* values (as in CIE L* a* b*).

Those are all fairly simple models for contrast determination. A more advanced approach is a true color or image model like CIECAM02 or ICAM. ICAM is the work of Mark Fairchild at Rochester Inst. of Technology. A model like that could (I believe) analyze an overall page, as opposed to a pair of colors in isolation.

On the experiments page there is an example of local adaptation issues do to surrounding colors. But here's a quick example:

The blue text on grey is WCAG 4.5:1 contrast, and both bits of text are identical. But the one centered on black is more readable because the black allows local adaptation to the darker colors. So among other things, minimum padding for elements against a high contrasting color are important.

(I'll mention in passing that pair of blue on grey is a fail in my modified PCL algorithm).

And this is where things are more complicated than a simple contrast — web content is graphically rich. Text on a background may be a pass, but if it is far different than the overall page, adaptation will affect perceived legibility.

Thank you again,

Andy

Thanks @Myndex for writing this up so thoroughly!

Thank you @bruce-usab — I consider this a particularly important issue, partly because W3C standards are used not just for web but for app design and other applications as well. Because it is a freely distributed standard, it has a very wide reach. Myself I spend 12 hours a day in front of monitors, while that may be more than average displays are certainly integral to the lives of so many — we are inseparable from our technology — standards like this have a very real affect on people's lives. This standard in particular has become part of government regulations, for instance.

What is the timeline/deadlines for 2.2? I'm hoping to have some candidate contrast models soon, but also thinking there is one giant rabbit hole to crawl down considering how page complexity affects perception (adaptation) etc.

Thank you!

Andy

There is a mathematical problem with this discussion line. L1 and L2 are computed using weightings that take color receptivity into account. R, G, and B have distinct weights in the relative luminance formula.

Hi @WayneEDick , thank you for commenting. I’m away from the studio, on location, so I can’t comment in depth, but:

Yes, luminance is spectrally weighted. However luminance is a linear measure of light.

Light is linear (additive) but human vision is NOT linear (essentially a power curve). So while the sRGB coefficients adjust for spectral sensitivity, luminance is NOT relative to PERCEPTION of lightness. L* is perceptual (CIELAB), and gamma encoded transfer curves are somewhat perceptual (such as luma, the Y‘ of Y‘IQ) but not luminance (Y).

But that’s not even the most relevant part. L1/L2 is called “simple contrast” and it is wrong in this context. The “standard” for contrast for TEXT is Weber, which is based on ∆L/L typically as
(LMax-Lmin)/Lbackground or (LMax-Lmin)/LMax
And is in fact still using linear luminance. But there are some better more modern implementations now such as Perceptual Contrast Length (PCL) which essentially converts to a modified L* (perceptual lightness). EDIT FOR CLARITY: WCAG2 simple contrast is essentially Weber+1, regardless, values become increasingly incorrect and lacking in perceptual uniformity as the color pair becomes darker relative to the adaptation level. Therefore they both fail for self-illuminated displays for similar reasons. This was clear a short time after making this post, but as this thread ranks highly on search engines it's useful to amend errata inline, here. End edit.

This issue came to my attention when I saw the contrast equation was wrong (as is the sRGB threshold WCAG lists), as I have outlined in my posts above. But now that this is being discussed, we can do better. I am currently investigating PCL and other methods.

For further details, I suggest Charles Poynton’s GAMMA FAQ and COLOR FAQ. Here’s a link: https://poynton.ca/GammaFAQ.html
but please feel free to ask me any other questions.

-Andy

@Myndex, you asked:

What is the timeline/deadlines for 2.2?

The formal/approved Project Plan has a goal of this time next year for the first public working draft.

I'm hoping to have some candidate contrast models soon, but also thinking there is one giant rabbit hole to crawl down considering how page complexity affects perception (adaptation) etc.

I am not optimistic about the chances for wholesale replacement formulas for 2.2. That is possible for 3.0.

This standard in particular has become part of government regulations, for instance.

Yes. I am one of the actors in helping that happen.

There was some user testing associated with the validation of the 2.0 formula. I could not quickly find a cite for that. My recollection is that the hard data pointed to a ratio of 4.65:1 as a defensible break point. The working group was close to rounding that up to 5:1, just to have round numbers. I successfully lobbied for 4.5:1 mostly because (1) the empirical data was not overwhelmingly compelling, and (2) 4.5:1 allowed the option for white and black (simultaneously) on a middle gray.

I am sorry to say that I will offline for the next ten days or so, but I will be circling back to this!

@Myndex, this one assertion leapt at me:

The California Council of the Blind (Lozano, 2009) states, and Federal ADA guidelines also state, that contrast for signs be 70% (note it is a percent not a ratio). The math used for the Federal standard is (B1-B2)/B1 (B1 is the lighter LRV, and B2 is the darker).[3]

That formula was only ever intended for reflective light, not luminescence. It was promulgated in the 1991 ADAAG and was sufficiently problematic that is was dropped in the 2004/2010 ADAAG/ADAAS.

Your citation [3] clearly states (more than once) that 70% “is no longer a requirement”.

Hi @Myndex,

Just upfront - I strongly suggest we come to a resolution on this issue before you spend time creating a PR.

estimating roughly, 40% of the color pairs the WCAG math calls "PASS" are poor in quality and should fail. And somewhere in the area of 51% of colors they fail could conceivably pass.

This doesn't match my testing with people over the years. Not a large scientific study, but 100s of tests (since the early 2000s) with people with low-vision. Whenever there was a color combination that people struggled with it virtually always failed on the contrast level checks.

I've also found there are huge differences between people and the particular colors that were an issue for them. E.g. Some participants couldn't see a strong pink on white, which others couldn't ignore as it was so intense.

Broadly I think the context that you need to account for is what the guidelines are for, and how they are used. A method to measure contrast for the web content accessibility guidelines needs to:

• Work across many types of visual impairment. Low vision is a general category, various forms impact color perception in different ways.
• Work across many devices. As a web designer your work will be viewed on cheap & expensive phones, expensive Macs (or Chomebooks), cheap Chromebooks (or Windows laptops), TVs, projectors etc. Therefore it cannot account for the brightness of the display, let alone the surroundings. It has to be a measure from the way the web page is setup.
• Be simple to test. I.e. a web designer needs to be able to test their own work easily.

A lot of the factors you added in the summary above cannot be accounted for in a web standard (e.g. display polarization, nits).

Also, at least some of the examples you created have the same 'background bias' effect I mentioned here, perhaps you know the name for that effect? I.e. having a different general background behind the area of interest affects the perceived readability. Reading on, I guess this is the 'local adaptation' issues?

In short, I don't think there is such as a thing as a "revised programatic contrast assessment that is perceptually correct", but I'd love to be wrong. A change would need a lot of real-world testing to ensure it provides better results.

What is the timeline/deadlines for 2.2?

Given the scale of change this would require (including the research), I suspect it would be a 3.0/Silver type of thing to do.

@Myndex, this one assertion leapt at me:

The California Council of the Blind (Lozano, 2009) states, and Federal ADA guidelines also state, that contrast for signs be 70% (note it is a percent not a ratio). The math used for the Federal standard is (B1-B2)/B1 (B1 is the lighter LRV, and B2 is the darker).[3]

That formula was only ever intended for reflective light, not luminescence. It was promulgated in the 1991 ADAAG and was sufficiently problematic that is was dropped in the 2004/2010 ADAAG/ADAAS.

Your citation [3] clearly states (more than once) that 70% “is no longer a requirement”.

This is the section in reference [3] I was referring to (I did not read the entire document, I was mainly pointing out the continued controversy and unsettled nature of the issue):

Historical and current studies of contrast sensitivity it is typically about 1% to 1.6% over a wide range from 7 or 8 cd/m2 to over 500 cd/m2. NASA also found that under 8 cd/m2, contrast sensitivity fails increasingly.

But on the subject of reflected light vs emitted light: both can be measured in luminance. Luminance is proportional to both illuminance and reflectance.

And this is one of the HUGE ENORMOUS PROBLEMS facing us in the present conversation, I have seen two COMPLETELY DIFFERENT definitions of LRV. The correct definition of LRV is based on luminance (Y or L) which is linear light, yet some sources state it is based on lightness (L* as in CIELAB, L* a* b*) which is perceptual lightness NOT linear light.

YIKES. It appears this stems for the error in the 1991 ADAAG, which from what I have been reading was using Weber on L* and not Luminance?

Hi @Myndex,

Just upfront - I strongly suggest we come to a resolution on this issue before you spend time creating a PR.

Hi @alastc I agree and said as much in one of my posts, it is why I am posting an issue instead of a pull request first.

estimating roughly, 40% of the color pairs the WCAG math calls "PASS" are poor in quality and should fail. And somewhere in the area of 51% of colors they fail could conceivably pass.

This doesn't match my testing with people over the years. Not a large scientific study, but 100s of tests (since the early 2000s) with people with low-vision. Whenever there was a color combination that people struggled with it virtually always failed on the contrast level checks.

That's good to know, though I am concerned about the large number of sites I encounter that pass the test yet I find very hard to read. I am less concerned with false fails and more concerned about the false passes in other words.

I've also found there are huge differences between people and the particular colors that were an issue for them. E.g. Some participants couldn't see a strong pink on white, which others couldn't ignore as it was so intense.

A "strong" pink on white should have a fairly low luminance contrast, and should fail with proper math though the WCAG math might pass it when it should fail in some cases. The problem with hue is how people with color deficient vision rely on luminance contrast. But also, a light background changes perception of text & contrast vs a dark background. I'm wondering how those who had a hard time with pink on white would have seen the white on pink.

Broadly I think the context that you need to account for is what the guidelines are for, and how they are used. A method to measure contrast for the web content accessibility guidelines needs to:

• Work across many types of visual impairment. Low vision is a general category, various forms impact color perception in different ways.

Yes the main issue is adequate luminance contrast. A useful tool for designers might be a tool that captured a website and converted it to greyscale based on luminance so the designer could see the luminance contrast without being influenced by hue.

• Work across many devices. As a web designer your work will be viewed on cheap & expensive phones, expensive Macs (or Chomebooks), cheap Chromebooks (or Windows laptops), TVs, projectors etc. Therefore it cannot account for the brightness of the display, let alone the surroundings. It has to be a measure from the way the web page is setup.

Cheap or expensive, displays are built to sRGB standards, and often with better brightness. But not the point, as the eye adapts to various conditions of light. What is important is to consider how adaptation affects readability (more on that below).

• Be simple to test. I.e. a web designer needs to be able to test their own work easily.

Most that I am discussing is simple to implement. (It's not harder to use correct math, for instance).

A lot of the factors you added in the summary above cannot be accounted for in a web standard (e.g. display polarization, nits).

It can — when I talk of polarization, I talk specifically of web DESIGN. light text on a dark background is "negative polarization" (or confusingly, positive contrast) and vice versa.

As for nits (cd/m^2) I'm not saying the web standard should specify any particular "absolute" luminance output, but the standard IS already trying to take environment into consideration

Also, at least some of the examples you created have the same 'background bias' effect I mentioned here, perhaps you know the name for that effect? I.e. having a different general background behind the area of interest affects the perceived readability. Reading on, I guess this is the 'local adaptation' issues?

Local adaptation, and adaptation in general, need to be part of the design considerations. Dark text on on a grey background may pass via the math, but if the grey background is a div with no padding on a white background, they eye adapts to the white making the dark text on grey hard to read. There is a demonstration of this on the experiments site.

In short, I don't think there is such as a thing as a "revised programatic contrast assessment that is perceptually correct", but I'd love to be wrong. A change would need a lot of real-world testing to ensure it provides better results.

There are definitely better choices than using incorrect math, which is the current state. And while vision and perception are complicated, it is also mostly academic in terms of things like contrast. There is a wealth of research on vision perception and contrast over the last several hundred years that can and should be used to guide this standard. In other words, yes there is such a thing as "programatic contrast assessment that is perceptually correct." It's just a matter of implementing it.

There may be added challenges in modern webpages due to the graphically rich content AND the variety of environments due to mobile devices. But the W3C provides the standards and guidelines not just for web design but for browser software.

(edited for spelling and some clarity issues)

False fails/passes & ‘incorrect math’: Yes there is lots of research, but any model using Mathematics is a mapping of how light is measured to how it is perceived.

There are individual differences in perception, so there cannot be a perfect model (that’s what I meant by ‘no such thing’). Otherwise the pink example wouldn’t vary by person.

A different model may improve the fit across a range of visual impairments, but it is not an absolute right/wrong. One model will not fit everyone perfectly, and we should be optimising for people with visual impairments rather than the general population.

If there is a better industry standard model to use for measuring contrast, great, let’s test it across a range of people.

Secondary notes

There are already tools for greyscaling a screenshot, but we need to be able to assess text (and certain graphics) and show a pass/fail individually.

Web content can be defined to use color spaces other than sRGB, but we are planning to standardise testing of contrast to sRGB as a lowest common denominator.

I am not optimistic about the chances for wholesale replacement formulas for 2.2. That is possible for 3.0.

That makes sense, there are probably some incremental changes that might be helpful as well as "leading a path" to a larger change.

This standard in particular has become part of government regulations, for instance.

Yes. I am one of the actors in helping that happen.

Ah excellent! However, that also means that the standard needs to be solid and unimpeachable. I'd like to help to get to that point.

There was some user testing associated with the validation of the 2.0 formula. I could not quickly find a cite for that. My recollection is that the hard data pointed to a ratio of 4.65:1 as a defensible break point.

Hmmm. I'd love to see this data. I believe you that the ratio from the data is higher that other standards as the equation being used overstates the contrast ratio in addition to being perceptually incorrect.

The working group was close to rounding that up to 5:1, just to have round numbers. I successfully lobbied for 4.5:1 mostly because (1) the empirical data was not overwhelmingly compelling, and (2) 4.5:1 allowed the option for white and black (simultaneously) on a middle gray.

I'm not sure it does, as written the equation is overstating contrast for darker colors. It appears the equation does not take system gamma gain into account, nor the floor of 8 cd/m2 in terms of minimum luminance for contrast (NASA). More discussion to come on these issues.

QUESTION: it would be helpful to get online access to certain ISO standards, as well as papers used in the current specification — is that possible?

Thank you!

Ah excellent! However, that also means that the standard needs to be solid and unimpeachable. I'd like to help to get to that point.

Just for context though, the formula was already in WCAG 2.0, and that's been out since 2008 https://www.w3.org/TR/WCAG20/ ... so just a word of warning that it's something deeply enshrined and not something that can be changed quickly or easily. It would take a few years at least...

Just for context though, the formula was already in WCAG 2.0, and that's been out since 2008 https://www.w3.org/TR/WCAG20/ ... so just a word of warning that it's something deeply enshrined and not something that can be changed quickly or easily. It would take a few years at least...

HI @patrickhlauke, Yes, I do realize this and recognize the issue. I'm certainly not expecting any overnight major change! As I mentioned in one of my posts, I am looking at potential incremental changes that can lead to a more solid solution.

At the same time there are a lot of other related standards that use different models and compliance parameters. Nearly all of them are using Weber, but there are newer more useful models emerging.

I mentioned some of the reasons I'm motivated for some positive changes here — and to be clear, my intent is to assist in finding easy and workable solutions to the issues I've outlined, and perhaps others.

As CSS, Java, HTML develop into greater feature sets, I've noticed a disturbing trend toward sites that are "fancy but less useable." So much so that many browsers now have the reader view to turn off all the crap!!!

There are individual differences in perception, so there cannot be a perfect model (that’s what I meant by ‘no such thing’). Otherwise the pink example wouldn’t vary by person.

Pink/white relates to hue contrasts. The more perfect accepted model is luminance contrast as it's connected to contrast sensitivity. CS threshold is 1%-1.6% over the wide range of 8 cd/m2 to over 500 cd/m2, as shown in study after study, including many visual impairments. There are of course some impairments that directly affect CS/CSF, but contrast sensitivity is separate from visual acuity.

Visual acuity is helped more by size than contrast. perceived contrast is more complex than a ratio between two colors as it is substantially affected by adaptation, local adaptation, chromatic aberation, and other issues.

CHROMATIC ABERRATION: So this relates to an optical issue with any lens system, including human eyes. Light at different wavelengths are "bent" differently through a prism, which is why a prism creates a "rainbow". Lenses are a form of prism, and blue light through a lens lands in a different spot than red light as a result. (It is theorized that this is why our eye evolved to have red cones in the center and blue cones on the periphery). But this is a reason that red (#F00) and blue (#00F) look wacky together - the shared red/blue edge focus to different places on the retina. The hot pink you mention is #FF00FF - so red and blue with no green, the red portion of the text edge focusing differently than the blue edge. So for instance, working with colors that have a high blue content needs care, as NASA discusses:

https://colorusage.arc.nasa.gov/blue_2.php

That NASA site covers a lot of related material, and it is all about user interface design considering adverse viewing circumstances.

If there is a better industry standard model to use for measuring contrast, great, let’s test it across a range of people.

The "standard' has been Weber the 1800s. There are better models now, and particularly as web pages are a "unique environment" in that they are displayed using certain standards, there are definitely better ways to assess perceptual contrast.

Secondary notes

There are already tools for greyscaling a screenshot, but we need to be able to assess text (and certain graphics) and show a pass/fail individually.

Okay, but as I have demonstrated and discussed, the ratio of two colors by themselves in isolation will not give you a complete answer.

Web content can be defined to use color spaces other than sRGB, but we are planning to standardise testing of contrast to sRGB as a lowest common denominator.

Hmmm, no you can't. The standard is still sRGB. the CSS 4 working draft does list additional working color spaces as something desired for future implementation, but that's a pretty horrible idea at today's level of technology. sRGB is ideal for 8 bit. Any larger colorspace and you start needing at least 10 bit pretty quick. I see talk of linear_sRGB or linear_Rec2020 - then you need at least 16bit_HALF(FLOAT). Double bit depth and you double data size, and pages are ALREADY overbloated and slow. And you'll NEVER see the benefits under typical ambient conditions and cheap devices.

To wit: bigger color spaces do absolutely ZERO to assist impaired vision, There is ZERO luminance contrast difference (and it's worse if you stay in 8 bit: A super-big space like ProPhoto is GARBAGE on 8 bit, and will provide WORSE contrast gradation (i.e. causes banding) due to the ginormous delta E errors. NOT TO MENTION the fact that ProPhoto uses IMAGINARY primaries, meaning that values like #00FFFF DO NOT EXIST in ProPhoto as something you can see.

Most mobile browsers and many desktop browsers still do not support any form of color management. sRGB is the standard, and is expected remain that way for the foreseeable future. The CSS tag for alternate colorspaces is not.

FOR ACCESSIBLE: 8 bit and sRGB (and Rec709) is the ideal standard at present technology.

Yes, there are some emerging color spaces like Rec2020 that are bound to make a difference someday but all these alternate color spaces have different transfer curves and different primary coordinates. Converting between spaces is computationally expensive, which is why most mobile browsers are NOT color managed and instead are sRGB "compliant".

I have high end $$$ wide gamut monitors (which are probably what caused my early cataracts), but those are rare — sRGB/Rec709 define nearly all distributed content be it Web or Broadcast worldwide, and in a non-color managed way. If you use monitors OTHER THAN sRGB/Rec709, then you MUST have color management to transform colorspaces, and that is computationally expensive. I discuss some of this in some articles I've written over the years reprinted here: https://www.generaltitles.com/helpfiles/13-q-a-blog/colorspaces-and-file-types
but still, Poynton's Gamma and Color FAQ are a good and easy read first: https://poynton.ca/GammaFAQ.html

Thank you for the comments!

Andy

Here are my questions. When you use the term spectral in the sense of functional analysis?

Hi @WayneEDick !

This is part of the CIE 1931 standard on luminance, the Y in CIEXYZ. The standard is spectrally weighted relative to the LMS cones (red green blue cones) that make up human trichromatic vision.

The coefficients 0.2126, 0.7152, and 0.0722 are part of the Rec709 standard for HDTV, and sRGB is derived from that standard. Both Rec709 and sRGB use the same color primaries and white point — the only practical difference is the transfer curve (effective gamma) is a little different between sRGB and Rec709, the reason being that Rec709 gamma is relative to a dark living room and sRGB is relative to a brighter office type setting.

Charles Poynton's Gamma FAQ is really the best crash course on this.[1]

While gamma is not a linear function it is differentiable and can be represented piecewise by line segments without?

Gamma is one form of a "transfer curve" to transform a particular color value from one colorspace to another. It is often represented "piecewise by line segments" as what we call a LUT (look up table). LUTs are very common in the film/television industry because the color represented in negative film is sufficiently complicated that it can't be accurately represented with a simple equation or matrix. 3D LUTs are used to create accurate transforms through various color spaces in the post production process.

Some color spaces like Adobe98, use a pure exponential transfer curve.

BUT ALSO: the sRGB and Rec709 transfer curves in their "correct" implementation use a combination of an exponential curve attaches the a linear region near black. The linear region has a number of purposes and motivations, including reducing camera noise near black and math issues with pure exponential curves near black.

Are you saying the W3C representation does not use enough line segments in it's approximations?

It uses "none" because luminance is linear, as in a straight line. Luminance has no gamma (or technically, the gamma is 1.0). Luminance is proportional to light, and light in the real world is linear.

The human eye is NOT linear, photopic vision has a gamma of around 2.4 to 2.5 (though vision is more complicated due to adaptation, scoptic (rod/dark night) vision, etc.)

This is the CIE L* curve

L* is based on human perception. Luminance (not shown) would just be a straight diagonal line from 0,0 to 100,100.

Or are you saying that gamma does not come into the equation?

Right now human perceptual contrast is not represented in the WCAG "Understanding 1.4.3."

Luminance is derived by first applying the reverse transfer curve to each of the R´G´B´components, then multiplying them by the coefficients, and then summing them for the total luminance (Y but sometimes shown as L but NOT to be confused with L*).

THEN they use a simple contrast ratio Yhi/Ylo or as they print it L1/L2. They also add 0.05 flare to each term. ((L1 + 0.05)/(L2 + 0.05)).

So here is the point I was getting at: the use of L1/L2 is only useful for very absolute black & white values, because it ignores a lot of what happens with perception of in-between values.

The "standard" math for contrast of TEXT is WEBER CONTRAST which uses the Weber fraction, which is ΔL/L — Weber has been around for a very long time, and most contrast standards and research are based on Weber or Michelson, not simple contrast. Simple contrast is used for example for the contrast of a monitor from maximum black to maximum white, but not for the in between values.

EDIT: Weber contrast is often stated as (Ybg-Ystim)/Ybg, but this can result in odd results. For monitors/displays, try (Ylightest- Ydarkest) / Ylightest

I am NOT saying that Weber is the ultimate solution, but it is what jumped out at me when I was investigating why web contrast calculators were presenting "weird" numbers relative to legibility. This led me on this path of "how did we end up here" which has now morphed into "what is the most useful modern perceptual contrast calculation."

Other notes on WCAG math: The sRGB conversion to luminance is using some incorrect values. The problem is minor and likely has little effect on the contrast issue, but I will show to the correct sRGB formula below.

Also, just FYI the coefficients must be applied only after the gamma is removed, but there is an interesting wrinkle here: even the "correct" luminance math does not account for system gamma gain. There is an additional 1.1 or 1.2 exponent applied to the signal by the monitor/display. This is common even in older systems like NTSC, which used a 1/2.2 exponent at the camera, but the CRT display was actually ~ 2.5 resulting in a system gamma gain at final display. Final display gamma can in fact be adjusted by the user with the monitor controls (that adds the uncertainty aspect to all of this). But I did notice that when I added a 1.2 exponent to the resultant luminance, it improved the perceptual uniformity of the resultant reported contrast (at least it seemed to, I have not run a real controlled study yet).

Finally what is your formula precisely including visual factors.

I suggest looking at Weber contrast, Michelson contrast (aka modulation), but also two modern methods that I am investigating and experimenting with, Bartleson-Breneman Perceptual Contrast Length[2], and one I just recently found that apparently is the basis for the Australian accessibility standards, the Bowman-Sapolinski Equation [3], though I;m not certain it can be used on CIE Y (Luminance). And then there are methods using L* (perceptual lightness from CIE L* a* b*) instead of luminance, in that case, it's not usually a ratio, but a difference (L*1background - L*2 foreground) as L* is perceptually uniform.

So for normalized values of 0 is min and 100 is max:

Y (luminance) 0 = L* 0 = sRGB 0 and Y 100 = L* 100 = sRGB 100
BUT
Y 18.4 = L* 50 = sRGB 46.7

This is because the perceptual halfway point between black and white is not a luminance of 50, but a luminance of 18.4 but on the perceptually uniform L* curve, the halfway point is 50. On sRGB its about 46.7 because as I mentioned earlier, sRGB has additional system gamma gain. Adding an expoinent of 1.102 to Y will put Y 18.4 at sRGB 50 for example (and I'm not saying that necessarily "should" be done, just that's how the math is for comparison).

I would like to analyze this. I am a mathematician. Sincerely, Wayne Dick PhD.

That is REALLY awesome to hear, I was hoping a mathematician would get involved. I have some planned experiments this weekend, I'll post more as I progress.

Note: the correct luminance calculation for sRGB -> D65 Y is:

From 8 bit R´G´B´, divide each channel R´G´B´/ 255 to get them 0-1, then

This reverses the gamma resulting in linear RGB, then

R * 0.2126 + G * 0.7152 + B * 0.0722 = Y (D65 luminance)

For your cut and paste convenience, here is the gamma-to-linear portion from my OO spreadsheet:

=IF( R1 <= 0.04045 ; R1/12.92 ; POWER(((R1 + 0.055)/1.055) ; 2.4) )

ALSO if you are looking at other color transforms, we are only concerned with D65. Some CIEXYZ and L* a* b* transforms use a D50 whitepoint, which should not be a part of anything we are doing with monitor contrast, it's D65 only.

[1]https://poynton.ca/Poynton-color.html
[2]https://icdm-sid.org/downloads/idms1.html
[3] https://bcdg.hoop.la/fileSendAction/fcType/0/fcOid/330620847500650652/filePointer/330902322495002398/fodoid/330902322495002396/Color_Contrast_Scientific_Report.pdf
[4]http://www.brucelindbloom.com/index.html?Math.html

MORE USEFUL CONTRAST MATH

So today I came across this recent research at NIH (just a couple years ago) that directly states what I have been attempting to explain in the above posts. While they don't mention the WCAG, they do use the WCAG simple contrast ratio (CR) equation as a comparison to their modified Weber equation, including the WCAG's 5% ambient component.

The paper states specifically (emphasis added):

The contrast ratio (CR) based visibility predictions fail because it does not consider the function of the observer’s visual system. Human vision achieves high dynamic range through luminance (retinal) adaptation. As the overall scene luminance increases or decreases, the viewer’s adaptation level normalizes the target to background luminance difference, and perceives the same contrast (contrast constancy). For example, a large absolute luminance difference displayed under high overall luminance condition is perceived to have the same contrast as a lower absolute luminance difference displayed under low overall luminance condition. This characteristic is embedded in the Weber contrast definition.

NIH PAPER: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5489230/

They do modify Weber a little differently than I have, and their results are interesting and provide a further demonstration of the problems with "simple contrast" (CR). There are a couple small caveats I'll discuss after the summary. The paper is a short read, but here's a synopsis:

The WCAG contrast ratio (CR) is (Llight + 0.05)/(Ldark + 0.05)

The modified Weber is: (Llight - Ldark ) / (Llight + 0.05)

Hwang-Peli Modified Weber for Realistic Contrast for Monitors

CS Log plots of contrast versus ambient light:

• CR Plots suggest that when the display viewed indoor is moved outdoors under bright sunlight condition, the visibility reduction of the high contrast letter is much larger than the low contrast letter (the opposite of what happens in the real world).

• CR Plots show patterns of contrast ratio change under the same changes in ambient luminance condition are different for the two contrast polarity cases, again opposite the real wolrd case.

Modified Weber log plots of contrast versus ambient light. Note that Weber is not a ratio, but a value ov 0 to 1 (which can be *100 and described as a percentage, like Michelson):

• The Weber contrast of target decreases rapidly to zero (0), as the target luminance approaches near to the background luminance level, while it converges slowly to contrast of 1.0 as the target luminance depart from the background luminance level at all ambient light conditions.

• the impact of ambient light variation is much stronger for positive contrast polarity than negative polarity contrast. (Note: this report uses contrast polarity, so positive contrast polarity is light text dark BG, this is opposite from "screen polarity" where positive means dark on light BG, which I used in another post above).

• The contrast threshold can be drawn on the plots as a black horizontal line. Any target with contrast (including the impact of ambient light) below the threshold line will not be visible to the viewer.

• An image composed of both contrast polarities letters (Fig. C), it is expected that even for letters of the same contrast, those darkTX-on-lightBG letters disappear first (becomes lower than contrast threshold), as ambient light increases. These results on the mixed polarity are particularly important as it suggests an effective way to compensate for the ambient light.

Some thoughts:

1. TERMS: I have seen in the research two different polarity types. Contrast polarity, where black text on a white BG is called NEGATIVE, and screen polarity where black text on a white BG is called POSITIVE. Gee can we be more confusing?

For the purposes OF THIS DISCUSSION THREAD, I want to offer these (hopefully more clear) terms based on acronyms:

WOB: for white on black, any light text on any darker background.
BOW: for black on white, any dark text on any lighter background.

And maybe:

WOG and BOG, where the background is near a middle grey value.

2. The cited report describes what I have found regarding the perception difference of WOB vs BOW page designs. The implication as I have stated earlier is that there need to be a separate accounting or specification depending on design polarity.

3. The report also shows that BOW is the most stable under ALL ambient conditions, while WOB is more susceptible to brighter ambient conditions, and the middle-grey-background version (C) emphasizes this even more. This directly supports my earlier suggestions that the minimum value for the BRIGHTEST element be around #AAA (or somewhere between Various substantive and editorial rewrites including to resolve #993 and to resolve #986" #999 and #AAA, TBD).

Next Post: Path Forward.

PATH FORWARD

Based on all the research and discussion THUS FAR, I see the following general path forward as far as changes and pull requests for the WCAG:

WCAG 2.2 (and possible errata for 2.1/2.0)

1. Clearly, correcting the algorithm is too great a change for an incremental version as an algorithm change will necessitate a change in the pass/fail specification.
2. There are still incremental changes that can be added to 2.2 that will be helpful, such as: a minimum luminance for the lightest element, and maximum for the darkest element (essentially, depending on the A/AA/AAA level performance level, the brightest element be no darker than luminance 35% which is sRGB 63% which is #A0A0A0 and the darkest element as no lighter than 25% luminance which is 53% sRGB which is #888888. All while still maintaining the contrast as currently specified).
3. Another incremental change to address shortcomings of the current algorithm is to make WOB require a higher contrast than BOW, and similarly:
4. Address local adaptation, which can be helped by a minimum-padding spec, which is related to 1.4.8 and 1.4.12 for visual presentation and line spacing. (Essentially saying that if a containing div background and the text therein is substantially darker than the overall BG of the larger container, than padding equal or greater than the leading needs to be incorporated in the darker div element).
5. On visual presentation, emphasize user-adjustible size, which I have identified as being more important than contrast. Make it a FAIL if a webpage does not allow zooming in, and a fail if a webpage overrides user agent default font size or zoom level.
   5b. RELATED: Users should be able to set the default screen height of 1 REM. If page designers design around that base unit. it would allow users to set a larger base font (say 20 instead of 16) and have the rest of the page design maintain relative proportions. However, as a mandate there are issues due to responsive design concerns. The more important issue is disallowing zoom lock on pages!! Personally I find nothing more irritating than a page on my phone that has text too small and won't allow me to zoom in. That's a HUGE problem.
6. One final incremental change would be to use the correct values in the relative luminance equation so it is in line with the sRGB standard (which it current is not).

WCAG 3.0

1. At the very least, change the contrast algorithm to a modified Weber contrast, if not another perceptually based contrast assessment.
2. Modify the standard for contrast minimum & enhanced to use the values created by the new algorithm.
3. To help prevent confusion, the new algorithm should create contrast values as a percentage so the values are clearly different than those of the WCAG 2.0 standard which uses a ratio.
4. As a new algorithm/method may reduce or eliminate the need for some of the incremental changes listed above for 2.2, make sure those are altered as needed. For instance, a correct algorithm/method may take positive/negative polarity in design into account.
5. EVALUATION: If funding can be obtained, ideally a laboratory controlled study would be ideal as part of an evaluation. But also, a wide and public study using a web page based perception test should be included. While such a study would have less under "controlled" conditions, it would provide more under real world conditions, including how users have their browsers and devices setup. This wide study would be cheap and also create a huge dataset. I'd hope to see 50,000 or more responses, which should cancel out the far outliers and provide a real world basis for standards.

THREADS: Should we create a new thread for 3.0, and then set the discussion here just to the incremental changes I've proposed for 2.2?

Thank you all again for all comments and thoughts.

Andy

APCA npm Package

One end result of this thread has been the creation of the Advanced Perceptual Contrast Algorithm for WCAG_3 (silver).

While research continues, the public beta APCA npm package is now available: apca-w3

Install:

 npm i apca-w3

Usage:

And the API is trivially simple: send it two colors, it returns a signed INT of the Lc value (lightness contrast).

 Lc = APCAcontrast( sRGBtoY( textColor ), sRGBtoY( backgroundColor ) );

Implementation notes:

Color sent to the required parsing function sRGBtoY() can be a 3 or 6 character hex string #fff or a3b3d4 (hash # is optional but not required), a named CSS color (such as aquamarine), as rgb(99,122,14), or as a number 0x123456 or 15213412. Text color first, background color second — do not swap colors, as use case affects results. light text on a dark background returns a negative value.

For questions and comments please visit the main APCA repo discussion area.

Thank you!

Andy

Three Year Anniversary

Today is the three year anniversary of this thread, and of the research project into developing the technologies to improve readability on self-illuminated monitors and mobile devices.

Thread TL;DR

Here are the highlights & overview:

• Thread begins with a discussion of how the WCAG 2 contrast math and related SCs are not fit for purpose, and are not relevant to human perception of readability contrast on self-illuminated monitors.
• I investigated this issue pursuant to a book I am writing on modern color theory, with a focus on additive models, that book being motivated by my own personal loss of eyesight and continuing vision issues.
• I cover by background as a 35 year veteran of the Hollywood film and television industry, substantially involved in the industries transition from chemical imaging to digital imaging, and an expert in title design, typography, color spaces, visual perception, and related technologies.
• The thread covers much of the first round, first year research. As the scope of the project grew, I became an invited expert of the W3 and AGWG, and the repositories of the work were moved elsewhere, and continues.

Status of the work

This work became a very deep dive into the user needs for readability of text and discernibility of non-text on self illuminated displays. It has been a combination of scientific research and investigation here in our lab in Hollywood, and with the guidance and oversight of members of the AGWG in the Visual Contrast Subgroup where the practical application in terms of a reasonable guideline was formed for the future Silver/WCAG 3 guidelines.

APCA: Accessible Perceptual Contrast Algorithm

The key technology that resulted from the work is the APCA, an algorithm for predicting human perception of contrast on self-illuminated monitors and devices. APCA is rooted in decades of existing color appearance model technologies, including but not limited to CIELUV, CIELAB, CIECAM02, The Hunt model (Kodak), R-Lab (M. Fairchild, RIT), P.Barten work on contrast modeling, the work of M.Stone (PARC,NIST), and of L. Ahrens (NASA).

The models mentioned above are general in purpose, and often extremely complicated, being utilized on tasks such as image data compression. For APCA we worked to create a model as simple as possible and tuned specifically to the user needs of readability on displays and mobile devices.

Candidate WCAG 3 guidelines

By itself, a mathematical model of human contrast perception would lack utility without a set of guidelines on how to use the results of the predictions in terms of design choices. The candidate guidelines for WCAG 3 are based on the well established readability research of Bailey/Lovie-Kitchin, G. Legge, A. Arditi. and were developed in the Visual Contrast Subgroup of the AGWG/Silver project and that continues in the Low Vision Task Force, we have ongoing studies validating these guidelines.

APCA and the APCA guidelines promote a substantial improvement in readability and accessibility for sighted users of content on displays and devices, especially compared to those of the old WCAG 2 contrast SCs.

APCA and the APCA guidelines provide:

• Accurate assessment of readability over the visual range
  • Based on modern vision science
  • Considers spatial frequency (weight and size) as the primary driver of contrast, per human perception.
  • Considers contrast as viewed on self-illuminated displays and devices
• Recommended minimum font size and weight for a given contrast level
• Use-case related conformance levels to permit enhanced contrast where needed most, while improving design flexibility.
  • Body text is provided the highest contrast
  • Followed by fluent text, being content text that is critical to understanding
  • A level of "spot" readable text for less important features such as placeholder text, disabled items, copyright, etc.
  • Non-text elements with contrast requirements based on the wider range of spatial frequencies of these elements, and the different manner they are processed neurologically/cognitively.

Bridge PCA

A concern (voiced by a small minority) relates to conformance for WCAG 2 contrast SCs where they may be dictated by law ro regulation. While in the majority of cases, when APCA guidelines are followed correctly, the result is superior accessibility compared to 14.3 and 1.4.11, among others. In particular as WCAG 2 contrast math provides an unacceptable number of false passes as it is not correct relative to human perception nor impairments. One of the side effects of the WCAG 2 contrast math inaccuracy is that it also fails color pairs that should reasonably pass.

APCA may correctly pass a color pair in this situation, and the concern was raised regarding strict compliance. To be clear, this is a very narrow interpretation of the laws, and is applicable to a narrow set of use cases, if at all. Nevertheless, to provide a bridge method for those holding this concern, to develop something backwards compatible to WCAG 2, and also forwards comptible for WCAG 3, Bridge PCA was born.

Bridge PCA is using APCA technology, but has the following features:

• Cures the serious false-pass problem of WCAG 2.
• This allows Bridge PCA to calculate for Dark Mode which WCAG 2 can not do.
• Bridge PCA is polarity aware, so provides different (more accurate) results depending on if text is light or dark mode.
• Bridge PCA shows a modified PCA Contrast in Lc, but ALSO shows a "WCAG Compatible Ratio"
  • The WCAG 2 compatible ratio display will meet or exceed the results from contrast calculators using the old WCAG 2 contrast math.
  • In other words, WCAG 2 math says #688cf7 against black is 6.7:1, which is grossly too high—BridgePCA rates that as 3:1 for light text on a black background.
• In order to achieve full backwards compatibility to WCAG 2 contrast SCs, Bridge PCA give up some flexibility in terms of white text on a colored background, and has an offset in that area to force backwards compatibility.
  • This means that there will be the same false fails as WCAG 2, for compatibility reasons only.

The Paper Reading Experience and More

The scope of the research into readability has expanded, and a number of exciting things ahead as the work progresses. One is a set of guidelines to "The Paper Reading Experience" aimed to create highly readable content with minimal fatigue.

But also, the APCA was designed to be both adjustible, and extendable, to be able to flexibly address new technologies, such WGD and HDR. At present, APCA is setup to accommodate sRGB, Display-P3, and AdobeRGB. It is fairly straightforward to create an input module to address any display color space.

TOOLS

APCA Demo Tool

The canonical demo tool for APCA is at http://www.myndex.com/APCA/ this tool takes a pair of colors, and provides a recommended minimum font size and weight along with the contrast results.

This tool was just updated this week for the three year anniversy, with a new look, and new features!

Bridge PCA Demo Tool

The Bridge PCA tool is at http://www.myndex.com/BPCA/ this tool takes a pair of colors, and returns a WCAG 2 style ratio, which is intended for use with WCAG 2 SCs, and is a drop-in, backwards compatible replacement for the old WCAG 2 math.

This tool also freshly updated this week, with a better "transition to APCA" method, and of course, is still bakwards compatible with WCAG 2 math.

Third Party Tools

There is a growing list of third party contrast tools using APCA technology, and you can see them at Myndex/SAPC-APCA#51 — and if you have a tool you are developing, please list it here!

Color Vision Deficiency Simulator

The Myndex CVD sim is at https://www.myndex.com/CVD/ and provides clinically accurate simulation (Brettel model) of various forms of color insensitive vision (inaccurately called "colorblind"). Here you can process a screenshot of your content to see how it is perceived by color insensitive vision.

RESOURCES — Third Party Review/Discussion

Independent, third party criticism of WCAG 2 contrast.

The inherent problems with the WCAG 2 contrast math have been known for some time and widely critisized. Including studies by others showing that color insensitive types are not well served. The WCAG 2 contrast specs often cause enough problems for designers that it is ignored.

Independent, third party reviews of APCA.

Several independent reviews that include demonstrations of APCA as a significant improvement over WCAG 2.

• It’s Time For A More Sophisticated Color Contrast Check A look at APCA in the world of DataViz.
• Improving Accessibility with APCA
• Adv. Perceptual Contrast Algorithm A comparitive review of APCA.
• Color Contrast Is Not As Black And White As It Seems A comparison of the math and perceptual functioning of WCAG 2 and APCA, including custom infographics.

RESOURCES — APCA Education

Why APCA

A plain language article explaining the need to replace WCAG 2 contrast math, and the benefits of APCA.

APCA in a Nutshell

A plain language synopsis of the APCA

Draft White Paper on APCA

Discusses the underlying methods of the technology, with bibliography.

Let's Flip for Color!

If you want your text to be either black or white if the user selects some random color, just where is that inflection point?
Hint: It's NOT 18% Y. This is not exactly APCA, but is related. Included are code snippets and a CodePen demo.

Additional articles and discussions:

• Better reading on the web published in UX Collective, this article discusses and demonstrates the problems with automated testing and WCAG 2 contrast.
• Please Stop Using Grey Text on Tangled Web, debunking one of the worst myths about design contrast.
• What’s Red & Black & Also Not Read? on Tangled Web, which examines the nature of color insensitivity and readability.
• A Contrast of Errors on Tangled Web, which looks at the history and the current international readability crisis.
• Part I: Orange You Wondering About Contrast? Answering some contrast questions, and demonstrating a real solution to the infamous orange conundrum.
• Part II: The Lighter Side of Dark Backgrounds An article comparing some parts of APCA with the old WCAG 2 contrast methods, demonstrating how WCAG 2 contrast can cause poor results.
• Part III: WCAG 2 vs APCA Contrast Shootout Answering some recent questions regarding APCA, with comparisons and examples of the old (WCAG 2 1.4.3) and the future WCAG 3 / APCA.

EXAMPLES

WCAG 2 does not calculate appropriate contrast values for "dark mode".

WCAG 2 does not provide consistent contrast results over the visual range,

RESOURCES — Code and Repositories

• SAPC-APCA The main repository for all things related to SAPC and APCA, and the discussions tab is the ideal place for comments and discussion.
• APCA W3 This is the specific version of the APCA licensed to the W3 for the use in web accessibility guidelines such as the WCAG_3 guidelines.
• Bridge PCA BPCA is the drop-in replacement for WCAG_2 contrast that is backwards compatible, but using APCA technology, to "bridge" toward the future of readability contrast.
• Fancy Font Flipping Fancy Font Flipping is a demo also up in CodePen to illustrate the issues with flipping the text color from black to white based on a given estimated background luminance.
• Color Parsley lightweight and versatile color parsing functions, is a dependency of some of the above releases.

NPM Releases

The key releases are available as npm packages:

APCA W3:

npm i apca-w3

Bridge PCA:

npm i bridge-pca

Color Parsley:

npm i colorparsley

SPECIAL THANKS

For their direct and constructive help, guidance, and continued support, and without whom this project would not have developed as successfully as it has, I want to thank:

Visual Contrast Subgroup members, and AGWG chairs, without whom this project would not have been possible:

Bruce Bailey @bruce-usab
Alastair Campbell @alastc
Chris Loiselle
Chuck Adams
Rachael Montgomery
Cybele Sack
Sam Waller
Todd Libby

Members of the Accessibility, Design, & Development communities (in no particular order) your feedback, discussions, and support have been invaluable:

Charles Hall
Lisa Charlotte Muth
Roger Attrill
Dan Hollick
Jonathan Avila
Gareth Ford Williams
Nate Baldwin
Alex Krolick
Kilian Valkhof
Heather S. Neff
Alexey Ardov
Colin Fulton
James O'Leary
Tom Keenoy
Eugene Fedorenko
Munawwar Firoz
C. Liam Brown
Simon Bertrand
Protean Studio
Marco Vettorello
Joshua Krämer
Jamie Kyle
Steve O'Connor
Elizabeth Buie
James Daniel
Daniel Eden
Shanti
Lonny O'Brien
Alberto Calvo
Dustin Wilson
Taiga Yamada
Michael Cooper
Shawn Lauriat
Jeanne Spellman
Chris Lilly
Frank Elavsky
Clar Fon
Dan Hendersonede
(Sorry if I missed anyone!!)

And the Research and Writings of these noted visual perception professionals, whose work provided the foundation for APCA, SAPC, SARCAM and the other developments we are working on:

Mark Fairchild (RIT)
Robert W.G. Hunt (Kodak)
Charles Poynton
Maureen Stone (PARC, NIST)
Larry Ahrens (NASA)
Ian Bailey
Jan E. Lovie-Kitchin
Gordon Legge
Aris Arditti
Peter Barten
(And countless more, see bibliography)

In Closing

When I think back to the start of this project, I was thinking a solution could be had in about three months. LOL!!! While there was a basic solution in three months, it became clear that this is an under-researched area, and the rabbit hole was not only deep but endlessly fascinating. Today, we have a variety of solutions including a drop-in replacement for WCAG 2 math to "bridge the gap" toward a more readable future.

Hopefully this provides a good overview/resource for all things related to APCA and readability contrast. To be even more concise, here's a link-tree of the most important links:

https://linktr.ee/Myndex

Thank you all,

Andy

Andrew Somers
_{W3C AGWG Invited Expert
Creator of the APCA
Senior Color Science Researcher
Myndex Technologies
West Hollywood, Ca.}

APCA • THE REVOLUTION WILL BE READABLE

APCA QUICKSTART

The thread above is pretty massive — 127+ comments and 44,000 words. And elsewhere there are many other threads, articles, white papers, documentation... All told, this project has spawned a mass of material, making understanding challenging. It's a lot to weed through to fine the important material.

This is a case where "less is probably more." So in the interest of making this learning curve as short as possible, I've created a brief plain language overview, and organized a link catalog of further resources.

Get Shorty

This is a short introduction to the importance of perceptually uniform contrast for readability, simply titled: "Why APCA".

Tool Time

This is the simple demo tool of the technology. On this page there are also a few "twists" with additional documentation, including how to use it and other design guidance. The URL is easy to remember: APCAcontrast.com

A Brief History of Contrast

For those wanting the deeper dive, this is a curated page of resources relating to color and contrast. It includes third party and peer reviews of APCA, supporting articles, documentation, and more tools, code and other resources, all organized in sections. git.myndex.com

Thank You For Reading!

The following peer-reviewed article published by Smashing Magazine provides a good overview of the essential concepts involved in visual readability and accessibility:

The Realities And Myths Of Contrast And Color

The Realities And Myths Of Contrast And Color

I regret missing the GitHub ping back in January. @Myndex, great article!

Hi @bruce-usab Thank you very much!

Consider the APCA values for this VGA-like palette:

For cyan-on-green [fg=#0aa,bg=#0a0], Lc is 0. For brightred-on-green [fg=#f00,bg=#0a0], APAC Lc is also 0, despite brightred-on-green having ever-so-slightly better contrast. The cause for 0 is that many contrast functions operate only on the lightness, and this includes APCAcontrast. Is ignoring the hue a conscious decision to account for eye deficiencies,
and/or would it make sense to define another function that can be used to compute some Lc values for capable eyes?

Hi Jan @jengelh

Let's Go Clamping

"...The cause for 0 is that many contrast functions operate only on the lightness, and this includes APCAcontrast..."

No, that is not the cause for zero. In the case of APCA, the guideline version is clamped to zero for contrasts that are less than about Lc ±8. This is intentional and by design, for a number of reasons.

Among other things, APCA and the APC-Readability Criterion sets a lower limit for non-specific non text at Lc ±15, and Ancillary text at Lc ±30, so as far as guidelines are concerned, the very low values are not useful.

Zero to do...

This has nothing to do with perceptual lightness contrast. There are APCA versions that are full range. And APCA full range and SACAM are certainly capable of discriminating small contrasts. The full range APCA is perceptually uniform over the entire visual range all the way down to invisibility, if that's what you are asking.

Using the SAPC research tool for your examples, g=#0aa,bg=#0a0 is about Lc 1 or -2 (depending on polarity), and in the case of the red on green, it would be Lc 5 or -6 if reversed,

. . . . . . . . . . . .
EDIT:
however, you have a dark line around the zero in your illustration. This outline makes a substantial increase in contrast. So taking the line into consideration, it is more like Lc 22

You have a lot of figures with outlines in your example, so those are not representative of APCA contrast values.

EDIT NOTE: I misinterpreted the artifacts in the image, which I discuss in a later post, below. Apology to Jan @jengelh for the confusion.

Also, APCA is tuned for a range of text sizes, and it is incorrect to demonstrate APCA with a single size and weight.

Small thin fonts and thin lines need high contrast, and big thick elements work with lower contrasts. Lower contrasts should be demonstrated with larger/bolder fonts.

HK Effect and Contrasts

Next, it's important to recognize the Helmholtz-Kohlrausch effect, where certain highly saturated colors appear brighter than their measured luminance would imply.

However, when it comes to reading and web content, which is specifically what APCA was developed for, color contrast is not per se useful. Luminance is what gets to the visual word form area and lexical processing. Chroma/hue contrasts are not particularly useful for readability.

Take a look here, the top three are very high color/hue contrasts, but nearly zero luminance contrast. The last one is decent luminance contrast, but nearly no hue contrast (some saturation contrast though).

Guess Hue

"...Is ignoring the hue a conscious decision to account for eye deficiencies,
and/or would it make sense to define another function that can be used to compute some Lc values for capable eyes?..."

APCA does not ignore hue. Yes, it is a lightness contrast predictor, but hue is accounted for the degree needed for readability. Moreover, luminance, i.e. Lightness Contrast, is needed by ALL vision types, not just those with color vision deficiencies (aka colorblind). Later this year we will be releasing some enhanced features for CVD and protanopia, to help more with wide gamut displays, but that module is about compensating for the loss in luminance for certain CVD types when red is paired with dark or black.

Color At War

Hering's opponent process describes how the early stages of the human vision system works: the LMS cones (stimulated by RGB) get matrixed into a red/green channel & a yellow/blue channel (chroma/hue), and a light/dark channel (luminance).

The visual cortex processes luminance separately from chroma/hue. Luminance carries all the fine detail, and is three times higher in resolution than chroma/hue. This is why we can literally throw away most of the color information for compressed image encodings like JPEG.

So if we are talking about contrast for readability, it is the luminance aspect that is critical, and this goes for all sighted users.

I hope that answers your questions, feel free to follow up here or at the APCA forum.

Thank you for reading,

Andrew Somers
Director of Research
Inclusive Reading Technologies, Inc.

While clamping is the correct explanation for those zeros (and I was under the impression taking the hue into account was a huge part of the motivation for APCA in the first place), the mention of “dark outlines” is very surprising!
@Myndex There are no outlines in @jengelh's picture. What we see on your upscaled fragment are artefacts of reconstruction filtering introduced by whatever you used to upscale it. These darker regions between red and green won't appear if you resample either in linear or in perceptually uniform colour space (or just use the nearest neighbours instead of introducing any new colours). It has nothing to do with its unscaled appearance.
However, these transitions are as sharp as your display device allows when displayed 1:1, and an illusion similar to Mach bands might be seen on the edges, leading to that “substantial increase in contrast” perceived.
It sounds like something the apparent contrast model itself could take into account. But unless I'm missing something, sharpness or blurriness is treated the same way size and weight are—the value of Lc doesn't depend on it, so it is still zero, and not “more like Lc 22.”

Hi @mwgamera and thank you VERY much for that comment.

"...There are no outlines in jengelh's picture..."

Well, the outlines do exist... when viewed in a browser on MacOS with a retina display. 😳

DOH!

I recently updated to the latest Safari, not sure when this changed, and it may be more related to my recent OS and Safari upgrade, on the retina Macbook.

I had opened the image into its own tab in Safari, and zoomed into that window—at some point in the past, viewing the image by opening it to its own window would have discarded the filtered image processing, essentially displaying the image as if with the CSS tag:

img {
  image-rendering: crisp-edges;
}

Apparently opening an image into its own browser window no longer does that.

Ooopps.

Viewing in Preview or an image editor display the original png correctly. I downloaded the png and indeed it is as:

But when scaling up in Safari, it the image is re-processed. I would have expected that if Safari or the OS was processing it for a given size, it would look more like:

The image I saw, which I posted previously, looked much cleaner, not like a scaled up filtered image, and more like something with intentional outlines on.

"...an illusion similar to Mach bands might be seen on the edges, leading to that “substantial increase in contrast” perceived..."

Good guess, but it is definitely due to reprocessing, I dug into the issue a bit more:

Digging Deeper

Spending a little more time so as not to make such a mistake in the future, I looked into this a bit more. Apparently, for Macs with retina displays, at least on certain newer MacOSes, some web images are processed in this way when upscaled for the retina resolution.

Yuk. Herein lies the caveat that "automated one size fits all" convenience features are not without issues. Clearly an image with hard edged graphics like this should not be filtered/processed for "smoothing".

Regarding Resizing

"...These darker regions between red and green won't appear if you resample either in linear or in perceptually uniform colour space..."

This is not per se relevant to the present issue. The issue is that the Retina display/Safari on MacOS results in some automated processing of some images, where the resampling exhibits this ringing. But there's more:

CLICK to expand discussion on spaces and resizing, etc.

INDEXED COLOR ISSUE

A possible contributing factor is the OP's image is indexed png. In Gimp, trying to scale while still in indexed mode, regardless of the method (linear, cubic, NoHalo, LoHalo), all exhibited this edge ringing. Gimp's "none" I assume is nearest neighbors, and as you pointed out worked for purpose.

That said, if the image is converted to a regular sRGB image, then the various interpolations worked fine, even in sRGB gamma encoded space.

Indexed color, 200% using Cubic

sRGB color, 200% using Cubic

Linear RGB (gamma 1.0), 200% using Cubic

Discussion

In this case, there doesn't appear to be a benefit to converting to linear 32bit floating point (gamma 1.0), and using the cubic scale algorithm in Gimp, the results seem identical. But the results are notably bad (as expected) for indexed color.

"...either in linear or in perceptually uniform colour space..."

Linear (gamma 1.0) RGB spaces are often preferred for blurs, compositing, and may be helpful for scaling and other transforms, depending on the image content.

A perceptually uniform space is not going to be specifically better than sRGB for these specific tasks.

The reason linear can be better for some tasks is that it makes the math simple, and works very much like real light in the real world: if you have 100 photons of light and you double it, you then have 200 photons. In a linear space, applying a blur works more like a lens being out of focus. And merging two images together using "Additive" can have very naturalistic effect.

Perceptually uniform spaces are the opposite of simple, and use spubstantially more complicated maths. But they are useful for creating gradients, picking colors, and converting to a different color model (i.e. going RGB to CMYK).

sRGB does not follow the linear transfer curve, and neither does a perceptually uniform space, as perception is also not linear. But in terms of transfer curve, sRGB is much closer to a perceptually uniform space than it is to linear. While sRGB is not technically "perceptually uniform", the TRC (gamma) is "close" to perception. The use of gamma is not just a hold-over from old CRT based displays. Now that we no longer use CRTs, gamma encoding is still useful because it does follow human perception to a degree, allowing more efficient utilization of the available data space.

. . . . . . .

In short, linear space and sRGB are as different from each other as linear space is from a perceptually uniform space. Different spaces for different purposes—If something is specifically beneficial in one, that typically excludes the other.

A bit more on APCA

"...I was under the impression taking the hue into account was a huge part of the motivation for APCA in the first place..."

No, the reason I started this thread back in 2019, and did the subsequent work on developing APCA and related methods and guidelines was to address the well known problems with WCAG 2 contrast guidelines. As for hue, that is part of it, see "What’s Red & Black & Not Read All Over?"

"...It sounds like something the apparent contrast model itself could take into account. But unless I'm missing something, sharpness or blurriness is treated the same way size and weight are—the value of Lc doesn't depend on it, so it is still zero, and not 'more like Lc 22.'.."

In order to remain simple, the current public APCA calculator calculates only a pair of colors. Sharpness and blurriness are mentioned tangentially in the guideline regarding font smoothing.

Thank you for reading,

Andrew Somers
Director of Research
Inclusive Reading Technologies, Inc.