The title should reflect that. Obviously this is misleading otherwise, there are a million ways of representing grayscale in data.
Sure, you can argue just picking a color channel and making it the grayscale source is not gonna work as expected, but what if we're talking about the data off a pan sensor vs an RGB array? What about a pan sensor created by taking an RGB sensor and removing the filter so it's producing an image with 3x the resolution but in pan?
This should end up being true for any presentation image format (the one you're going to actually show people, not necessarily the one you might do editing work on) optimized for human vision as well.
It's more a property of our eyes than a quirk of JPEG. So I don't find it misleading personally.
I'd argue compressed file formats make up a lot less of the broader problem space than most people realize. If you're doing anything high performance (where the difference matters) but not immediately consumer facing you're spending a lot of time working with lossless/uncompressed imagery in ram/disk. Stuff like MIPI->CUDA for classification, this question can go a number of different ways very quickly.
Keep in mind, at least in my experience, if you're even remotely thinking about doing anything with grayscale and this question comes up you're probably doing something...interesting (read: multi band remote sensing shenanigans).
> What about a pan sensor created by taking an RGB sensor and removing the filter so it's producing an image with 3x the resolution but in pan?
This got me thinking about how subtle and slippery the concept of "resolution" is.
The relationship is not strictly 3x; an ideally demosaiced Bayer filtered image still contains some information at the original capture resolution, because R, G, and B are correlated in nontrivial ways. Nyquist is a nice and simple result for the special case where nothing can be predicted about the original function, measurements are assumed to be perfect, and the criterion for reconstruction is "exact" - but in the general case of "recovering information, more is better" even a simple interpolation procedure with tame priors like "luminance exists" might perform so much better than the naive approach that we'd be silly not to use it.
Even the Nyquist frequency itself is less fundamental than it first appears; a less well known result is that a function can also be perfectly recovered if sampled at half the Nyquist frequency, if the slope is captured as well as the value. In other words you can losslessly convert a 2MP image into a 1M value+slope data matrix and back. What resolution is it? Depends on your point of view.
The same idea is used when doing complex IQ sampling of radio frequency data. I don't know if it's as obscure as you say, it's just not commonly considered when dealing with plain cameras. It's a bigger deal when working in the RF domain, or SAR imagery...
This would be the case for any lossy format with decent compression. We are much less sensitive to color than we are to brightness so we can compress color much more than we can compress brightness for the same perceived quality.
Not only that but color and brightness are often correlated, and advanced compression techniques use that property, see "chroma from luma" in AV1.
Without compression, indeed, it can be 3x, or even 4x (for alignment).
This post jumps into using "4:4:4" / "4:2:2" / "4:2:0" notation without seeming to explain what it means. Would be very helpful to add a short explanation!
And that notation is not really intuitive even when you do understand what it means. For example the initial 4 is technically a denominator of other two numbers, but it is not intended to be normalized! A consistent notation would use the reciprocal instead, say, "1-1" for 4:4:4, "2-2" for 4:2:2 and "1-∞" or "1-x" for 4:4:0.
I don't understand the point of the article, it seems to me that the point is that non compressed grayscale images don't take a third of the space of a compressed image? If we look at the last example the fist image 4:4:4 is 150KB the one in gray is 50KB (it seems exactly 1/3), the compressed images in-between take less space, but also the quality is much worse, I would say as expected. What am I missing?
I appreciate the article, but isn't the example image of a grey bird on cobblestone looking on a dark pool of water an odd choice to illustrate color subsampling?
The Y channel of the 4:4:4 "sub"sampled image looks sharper than the other comparison images. Look at the ground next to the bird's feet. I perceive it as sharper in the 4:4:4 image even though it's just about a single colour and the luma was supposedly not subsampled.
Yeah the "without visual impact" bit is a lie. People working in image / video compression are in denial about the noticeable quality drop with 4:2:0. 4:2:0 is extremely noticeable if your computer monitor gets into that mode.
There are some major failure cases involving alternating single pixel width red / black lines. That caused issues with DVDs because the format also used interlaced frames, meaning things like red tail lights at night could end up looking awful.
Luckily AFAIK they've given up on interlacing.
I really think with modern tech they'd be better off trying to compress 4:4:4 instead of immediately throwing out a lot of colour information.
I remember white/black-dithering (at even twice the dpi) being a lot smaller than the greyscale jpeg, when I was scanning documents. Greyscale was a bit smaller than full color, but dithered was a big step up (er, down).
When I scanned a lot of documents about ten years ago, I empirically found that JBIG2 is so much better than anything else including JPEG. That's specially optimized for scanning text documents.
For text documents with line drawings? For sure, absolutely -- especially when it's JBIG compression like fax machines use (or even Group 4 before that).
If it's a full scale photograph full of grayscales that's been dithered, then it's a harder comparison to make. Because obviously a ton of useful detail is lost in the dithering, in a way that isn't with lines and letterforms. So the comparison would really have to be with a terrible, blocky JPEG super-compressed file. In theory the JPEG should win.
I was honestly expecting this to be an article about grayscale versus dithered black and white, where the answer to "do grayscale images take less space" could go either way. Instead, it's an article about grayscale versus sRGB.
Somewhat related, a few years ago I noticed that an image with gaussian blur applied to it takes up significantly less space in JPEG than does the original image. I suppose it works out because the blur diffuses information throughout a particular area such that each JPEG block has less detail to encode internally?
A gaussian blur is essentially a low-pass filter so all the high frequency components are gone. The resulting image can be represented and stored using fewer DCT coefficients.
It converts the image into frequency domain, and reduces precision of the frequency data, with a special case when the data rounds down to zero.
However, whole-image blur crosses the 8x8 block boundaries, so it isn't perfectly aligned with the blur that JPEG uses. Lowering quality setting in JPEG (or using custom quantisation tables) will be more effective.
There's also a fundamental information-theoretic foundation for this — lower frequencies carry less information.
Naively i would assume that even if you have the unnessary the C_b C_r channels in a greyscale image, they are going to compress really well since they have very little information in a greyscale image.
Readit News
The title should reflect that. Obviously this is misleading otherwise, there are a million ways of representing grayscale in data.
Sure, you can argue just picking a color channel and making it the grayscale source is not gonna work as expected, but what if we're talking about the data off a pan sensor vs an RGB array? What about a pan sensor created by taking an RGB sensor and removing the filter so it's producing an image with 3x the resolution but in pan?
This is a complicated question.
This should end up being true for any presentation image format (the one you're going to actually show people, not necessarily the one you might do editing work on) optimized for human vision as well.
It's more a property of our eyes than a quirk of JPEG. So I don't find it misleading personally.
Keep in mind, at least in my experience, if you're even remotely thinking about doing anything with grayscale and this question comes up you're probably doing something...interesting (read: multi band remote sensing shenanigans).
This got me thinking about how subtle and slippery the concept of "resolution" is.
The relationship is not strictly 3x; an ideally demosaiced Bayer filtered image still contains some information at the original capture resolution, because R, G, and B are correlated in nontrivial ways. Nyquist is a nice and simple result for the special case where nothing can be predicted about the original function, measurements are assumed to be perfect, and the criterion for reconstruction is "exact" - but in the general case of "recovering information, more is better" even a simple interpolation procedure with tame priors like "luminance exists" might perform so much better than the naive approach that we'd be silly not to use it.
Even the Nyquist frequency itself is less fundamental than it first appears; a less well known result is that a function can also be perfectly recovered if sampled at half the Nyquist frequency, if the slope is captured as well as the value. In other words you can losslessly convert a 2MP image into a 1M value+slope data matrix and back. What resolution is it? Depends on your point of view.
Not only that but color and brightness are often correlated, and advanced compression techniques use that property, see "chroma from luma" in AV1.
Without compression, indeed, it can be 3x, or even 4x (for alignment).
https://en.wikipedia.org/wiki/Chroma_subsampling
There are some major failure cases involving alternating single pixel width red / black lines. That caused issues with DVDs because the format also used interlaced frames, meaning things like red tail lights at night could end up looking awful.
Luckily AFAIK they've given up on interlacing.
I really think with modern tech they'd be better off trying to compress 4:4:4 instead of immediately throwing out a lot of colour information.
If it's a full scale photograph full of grayscales that's been dithered, then it's a harder comparison to make. Because obviously a ton of useful detail is lost in the dithering, in a way that isn't with lines and letterforms. So the comparison would really have to be with a terrible, blocky JPEG super-compressed file. In theory the JPEG should win.
This page has a few visual examples at the bottom: https://www.cs.unm.edu/~brayer/vision/fourier.html
It converts the image into frequency domain, and reduces precision of the frequency data, with a special case when the data rounds down to zero.
However, whole-image blur crosses the 8x8 block boundaries, so it isn't perfectly aligned with the blur that JPEG uses. Lowering quality setting in JPEG (or using custom quantisation tables) will be more effective.
There's also a fundamental information-theoretic foundation for this — lower frequencies carry less information.
Naively i would assume that even if you have the unnessary the C_b C_r channels in a greyscale image, they are going to compress really well since they have very little information in a greyscale image.