Talk:Anti-aliasing

From Wikipedia, the free encyclopedia

Contents

[edit] Request for article

This needs a proper article about anti-aliasing with sampling theory etc. Unfortunately I haven't got time at the moment.


We also need to establish in the opening of the article that anti-a only applies to computer displays and printouts -- digital stuff. "real" typaces don't need it. -- Tarquin

[edit] Improving the article

Anti-aliasing is far more general than discussed here: the general outline is:

  • sampling wraps high frequencies into low
  • therefore high frequencies should be removed before sampling
  • a low-pass filter used for this purpose is called an anti-aliasing filter
  • (make mention of reconstruction filters on the reconstruction side of the system)
  • applies to digital imaging, digital audio, etc. etc.

-- Anon.

I dispute this engineer-centric recommendation. "Wrapping high frequencies into low" isn't a very accurate description at best, and simply misleading at worst. Aliasing is based on the pigeonhole principle, and the sinc/signal processing based approaches comes from blind faith in the Fourier transform. Don't take me wrong, I'm a mathematician with some specialization in the Fourier transform, and the Fourier transform is an extremely powerful tool. However, to use a Fourier approach without justification doesn't mean anything. In particular, I oppose "therefore high frequencies should be removed before sampling." It's mostly right, but there's no proof of any kind that's the best thing to do. Loisel 22:32 Jan 23, 2003 (UTC)
The article is hugely improved, in any case. I agree, the sketch above is (cough) non-rigorous handwaving, but even that was better than the really dodgy article which prompted the comments. Hooray for the Wiki-process, and thanks for your contribution! -- Anon.


[edit] Etymology

BTW, what's the exact etymology of the word anti-aliasing ? The fact that two images can be mapped to the same image looks irrelevant to me, as that cannot be avoided (pigeonhole principle tells), and "anti-aliasing" does not specifically try to avoid that either. --FvdP

All sources seem to indicate that the first practical solutions to anti-aliasing were invented Frank Crow, computer graphics pioneer (who incidentally works for nVIDIA these days), in 1977. Earliest reference I could find was: The Aliasing Problem in Computer-Generated Shaded Images, Communications of the ACM, vol. 20(11), pp 799-805. Nov. 1997. Of course I've also seen mentions of work done at the Architecture Machine Group at MIT in 1972, but I have not seen any references to published papers. Chadloder 23:06 Jan 24, 2003 (UTC)
PS here is a link to the article. You must be an ACM member to view the text (I am a member so I will try to distill anything I learn from it). [1] Chadloder 23:09 Jan 24, 2003 (UTC)
I think the term aliasing was used in its radio-engineering sense, related to carrier mixing in a superheterodyne radio: you can hear the same program at two positions on the dial if you don't use prefiltering. If this is the case, the other mathematical/computer science use of the term is inappropriate here (although the pigeonhole info is still relevant). The Anome
I thought the same, and modified the article accordingly. The article at aliasing is nice. --FvdP 00:19 Jan 25, 2003 (UTC)
Right now, I'm having a discussion in Talk:Nyquist-Shannon sampling theorem about the aliasing and Nyquist-Shannon sampling theorem and articles, because I believe that they can be improved. My slant is that I'm a mathematician with some speciality in Fourier series, and I don't entirely agree with the blind-faith-Fourier-series-Frequencies-are-a-God-Given-Thing approach.


[edit] More work needed

This article is getting better all the time: but this is a difficult topic which even confuses experts. More work is needed. The Anome 00:11 Jan 25, 2003 (UTC)

[edit] Gamma

I don't think this passage is pertinent to this article. I am moving it here, it can be restored later. If we were to discuss Gamma correction in here, there's no reason why we wouldn't discuss pixel formats, color theory, interlacing and all these other topics, which do not help at all with anti-aliasing.

(Note that the above assumes that the numerical value of a pixel is proportional to its intensity: for systems where this is not the case, gamma correction must be performed to preserve the linearity of the system, for this approach to hold true).

Loisel 05:31 Jan 25, 2003 (UTC)

Please put it back. Gamma correction problems are the #1 cause of poor anti-aliasing in computer graphics -- doing the A-A computation at the wrong gamma, or displaying the A-A'd image at the wrong gamma, will screw everthing up. <Michael Caine>Not a lot of people know that.</Michael Caine> The Anome

Having worked at SGI and NVidia, I'd say that almost all graphics hardware get the gamma wrong. The only place I've ever seen people complain about it is in film or print, and usually not exactly in those terms because they don't dump the images to the screen, and hence don't have a "gamma" in the final product (although they have to deal with more complicated response curves.) The problem of getting gamma right isn't specific to antialiasing, in fact I would say it affects lighting a whole lot more than antialiasing.

With small pixels, edge antialiasing such as FSAA should only be visible on a small fraction of the pixels (those that are on the edge of a triangle), and for in-triangle AA such as mip-mapping, the adjacent texels are of such a similar color that linear interpolation with or without gamma correction gives almost the same result. Hence, I would argue that being careful with gamma is a whole lot more important for lighting (which affects ALL pixels) than it is with anti-aliasing. And you don't really hear people complaining about how the gamma correction screws up their lighting.

To discuss "frequent problems" with antialiasing, we might need to restructure the article somewhat. Perhaps a section near the end summarizing some related issues. Gamma correction only applies to computer monitors in 8 bits or so per components; perhaps a reference to color theory or some comment about the response curve of the display would be more general. The upcoming graphics hardware should have a flat gamma curve when using floating point framebuffers. Loisel 05:31 Jan 27, 2003 (UTC)

The Anome, do you agree with the placement and wording of the Gamma Correction note now? Loisel 03:25 Jan 28, 2003 (UTC)

[edit] About images

Removed "This article certainly needs images!" from Antialiasing article. Created images. Tried to add images. Edit conflict - the new Antialiasing article someone made was merged with this Anti-aliasing article...

What I was going to add:

This image is not antialiased.   This image is antialiased.
Image:Nonantialiased line and ellipse.png   Image:Antialiased line and ellipse.png

Unless someone can think of a good use for them, I guess they can be deleted by whoever can delete uploaded files... Cyp 19:08 Feb 12, 2003 (UTC)


[edit] Merge

This was merged in. I think all the information in this stub article is contained in more detail in the current article, so I am moving it to talk. If certain details should be included in the current article, please feel free to integrate them. However, I feel that the language level in the article below is not adequate, and definetly lower than, the language level of the anti-aliasing article. Of course, I am biased, since I wrote most of the anti-aliasing article. Loisel 02:56 Feb 18, 2003 (UTC)

Antialiasing is a graphical effect used to reduce the effect of aliasing (better known as jaggies) in computer graphics.

On a computer lines are drawn in pixels (square dots) the result can look very raw (stairstep-like), depending on the resolution (i.e. the size of the pixels). This stairstep-effect is reduced by surrounding the line with grayscale pixels (black line on white background). You can compare this with a stone stair on which you throw mud and sand as long till you can't see the edges of the steps anymore.

Modern graphic chips like GeForce do this by calculating the gray pixels on their own. See: http://www.nvidia.com/view.asp?IO=feature_hraa

Most modern operating system GUIs (e.g. Windows 2000/XP, Xfree86 on UNIX / GNU) support antialiasing. This is very important for fonts and TFT-Displays. By the way, on TFT-Displays antialiasing is more important since the pixels have a precise square shape. Therefore on TFT subpixel-antialiasing is also used. To increase the optical resolution of a display the color of the neighbour pixels is altered depending on the oriantation of the red, green and blue cells.

[edit] Aliasing and radios

I'm coming very late to this discussion, but I just noticed a glaring error/misconception. The statement " ..the term aliasing was used in its radio-engineering sense, related to carrier mixing in a superheterodyne radio: you can hear the same program at two positions on the dial if you don't use prefiltering" is plain wrong. For one thing, this phenomenon in radio is called imaging, not aliasing - I wrote a small article about this under Image response; yes it arises because of mixing within a superheterodyne receiver, but that is not the same thing as sampling/quantization. Whoever wrote that statement either doesn't understand radio receivers, or antialiasing - one of the two! Secondly, the aliasing in graphics is precisely related to the Nyquist sampling theorem - it is in fact a direct example of it, albeit in two dimensions. It arises because the mathematical description of an object - a line, say, has to be rasterised into a pixel grid array, which is effectively sampling it at a lower rate. If the original object contains frequency components higher than the twice the sampling rate, they will be aliased to new lower frequency positions in the frequency spectrum - this manifests itself as jaggies. GRAHAMUK 03:52 10 Jul 2003 (UTC)


Thanks for the input. I can't say anything about the radio language, because I don't understand it (and that's why I was asking, way above in the talk page, how it worked.) I would appreciate if this matter were elucidated.
However, I did write the mathish/compsci-ish portion of the article. I am a Fourier analyst, and so I have strong opinions about Fourier series and such, and I've written that article very carefully in that respect. The one reference to the "Nyquist sampling theorem" is hidden in this phrase: "By the pigeonhole principle, sometimes two ideal images f(x,y) and g(x,y) will be converted to the same picture on the screen." And yes, I do think that's at the core of the Nyquist sampling theorem.
I could've written more, but there was also an article aliasing which I mostly rewrote. It has a much more detailed explanation of the Fourier version of aliasing. Please see if that corresponds more to your proposition about the Nyquist sampling theorem.
I had a long winded rant about the Nyquist theorem but I decided to remove it before saving, because it's not important. What's more important to me is your remark regarding radio stuff, can we get that straightened out, please?
-- Loisel 08:37 10 Jul 2003 (UTC)
Actually, I found the rest of the article accurate and informative, definitely no problem there, not that my own maths skills is up to criticising it anyway! It was only the paragraph that I quoted I have a problem with, as a (one time) radio engineer. The aliasing article is likewise fine, except that it quotes the same bit about likening the effect to a superhet receiver, so that needs changing too.
To clarify, in a superhet receiver, the input is mixed with a signal from a local oscillator to generate both the sum and difference frequencies, one of which, the intermediate frequency (IF), is selected and amplified by the rest of the receiver. A poorly designed receiver can suffer from the problem of an image of an input signal appearing at another place on the dial by reflection about the IF, and yes, this is a function of the filtering ahead of the mixer stage. However, as far as I can see that's where the similarity ends - the mixing process is not analogous to sampling, and the image is not analogous to the phenomenon of aliasing. I can see why the confusion arises - it sort of looks like the same thing going on - but it isn't. The bottom line is that the statement about radio is wrong, and as such adds only confusion to the article. GRAHAMUK 11:42 10 Jul 2003 (UTC)
Here's the analogy. An RF mixer is essentially a multiplier, multiplying the input signal with the local oscillator. Now consider sampling in the continuous domain, as the multiplication of the input signal with a train of delta functions. There is a clear resemblance here, with an input signal being multiplied with a fixed-frequency signal train. The function of the pre-mixer filter is also directly analogous with the anti-aliasing filter, both of which operate prior to conversion. -- Anon.
I don't understand what that meant, so I can't start throwing my weight around one way or the other. Keep in mind I'm just a mathematician. So an RF mixer is a multiplier -- does that mean that an RF mixer takes in an input signal f(t) and maps it to g(t)f(t) where g is the oscillator, or does it map it to g*f(t), the convolution of g and f (i.e., it's a multiplier in the Fourier analysis lingo?) When you say "in the continuous domain", is that the time domain as opposed to the frequency domain? When you say "train of delta functions" do you mean that g is actually the delta function and the RF mixer operates on f by convolving it with the delta function? Now if that is so, convolution with the point mass is just the identity map, so that wouldn't say much about your RF mixer (it would just say that your RF mixer does nothing at all.) So I'm guessing I've got it wrong. I also don't know what a pre-mixer is or what it does. Loisel 07:46 11 Jul 2003 (UTC)
An RF mixer is a multiplier, true enough. I still don't think that in this context the explanation is a helpful analogy for the phenomenon of antialiasing. I'm still not convinced that it is mathematically equivalent, but even if it is, I don't think the analogy is helpful. It confuses more than enlightens. For one thing, most radio receivers that people will be familiar with don't exhibit the image problem because they have been designed to have a reasonable image response rejection ratio - if an image is detected at all it will usually be put down to the same signal being broadcast more weakly on another frequency. Secondly, without a deep understanding of the maths of the two situations (even if analogous, which I repeat I'm not convinced of yet), most people will not see these two things as being the manifestation of the same phenomenon - jagged graphics sure doesn't LOOK like a phantom radio station! And if one has the maths to see the similarity, then presumably one wouldn't be looking up the explanation of antialiasing in a general purpose encyclopaedia? In other words, even if the analogy makes sense, it will only do so for the type of person for whom this article is NOT written. GRAHAMUK 10:10 13 Jul 2003 (UTC)

I can't fix it, so I'm dropping it here for further discussion.

The term aliasing is used here in its radio-engineering sense, originally used to describe to the phenomenon causing the same transmission to be heard at multiple dial positions on a superheterodyne radio if pre-filtering is not used.

-- Loisel 19:37 14 Jul 2003 (UTC)

[edit] Difference between the anti-aliasing figures

What's the difference between figures 1(b) and 1(c)? The article doesn't even hint at the algorithm of figure 1(b), yet the algorithm of figure 1(c) is "considered better". Isidore 21:51, 29 Aug 2004 (UTC)

The article does explain it, you just have to read further. The algorithm for 1b is explained thus: A better approach is, for each pixel, use the average intensity of a rectangular area in the scene corresponding to the surface area of said pixel. This gives a better, but not yet ideal, "anti-aliased" appearance; figure 1-b was generated this way. The alogorithm for 1c is explained a bit further by making a reference to the sinc filter and the aliasing article. Loisel 11:25, 29 Oct 2004 (UTC)

[edit] Gamma-corrected anti-aliased images?

Speaking of gamma correction, it looks very much like the anti-aliased example images are not gamma-corrected... Anders Kaseorg 03:30, 28 Oct 2004 (UTC)

[edit] Mip-mapping article

At present this article has far more information about mip-mapping than does the mip-mapping article itself. I believe the text on mip-mapping should be moved to the mip-mapping article and a summary of it should be kept in this article. What do other people think? MIT Trekkie 04:07, 14 Dec 2004 (UTC)

The mipmapping section here is now just a duplicate of the entire mipmapping article. I am removing it and putting in a summary. Valarauka 10:22, 17 September 2006 (UTC)

[edit] "Nyquisty" paragraph

I might be reading it wrong, but the sentence "It has been observed that to measure a signal of frequency n, you need at least n sample points, and they need to be well-placed." seems to be slightly misleading. It doesn't specify any sort of time within which the n sample points have to be in. Rspanton 00:03, 24 Dec 2005 (UTC)

[edit] pronounciation

how do you pronouce anti aliasing?


Its not rocket science... ăn'tī (like antimatter, or antidisestablishmentarianism) ā'lē-əs or āl'yəs (like the tv show Alias (TV series)) -ing (exactly what it looks like, if you have any experience with the english language at all, you know how to pronounce this suffix).

[edit] Pictures...

The layout of the pictures really needs to be sorted - at the moment the page looks untidy and some pictures are nowhere near the text that references them. Romansanders 17:31, 8 April 2006 (UTC)


[edit] Turning anti aliasing off?

I heard theres a ay to turn anti aliasing off in MS paint, can anyone shed some light on that?

[edit] New article for Antialiasing (Computer Graphics) ?

The 'real time approximations' section and its mipmapping subcomponent don't really fit with the tone of the rest of the article, which is more mathematical/theoretical (as it should be). I was thinking I'd create a new article for AA as used specifically in graphics; mipmapping fits there naturally, as does information about supersampling/multisampling, different sample patterns, coverage masks etc. which again shouldn't really go here, and I haven't seen anywhere else. There's already an article at FSAA but it's pretty scanty. Valarauka 20:49, 17 September 2006 (UTC)

I don't think we need to split the article. Every technique applicable to 2D signals (graphics) can be (and is) used for 1D signals. Mip mapping is not just a graphics technique, it's just most commonly a graphics technique. The same goes the other way for the other techniques. The slower filters are used less in real-time graphics (i.e. games), but they're used plenty in imaging software, etc. A decent article on anti aliasing in "computer graphics" is going to overlap with 95% of what goes in this article. A split is not appropriate. - Rainwarrior 21:53, 17 September 2006 (UTC)
On second thought, you're probably right. I'm still itchy about the article though... needs improvement. Let's see if I get time. Valarauka 00:48, 18 September 2006 (UTC)
A lot of articles on wikipedia need improved organization. This one could probably use some as well. - Rainwarrior 06:04, 18 September 2006 (UTC)

[edit] merge with fsaa to here

I think this article is better than the (already mentioned as small) FSAA (full screen antialiasing) article, and the content should be merged. They are relatively the same thing. I don't think it would hurt this article to include a mention of the methods used in graphics cards, and fsaa could be one of those subheadings. --gatoatigrado 05:48, 4 October 2006 (UTC)

Isn't mipmapping for textures? --gatoatigrado 05:52, 4 October 2006 (UTC)

Agreed on both counts. I don't really see that mipmapping has a place in the general article on anti-aliasing - it's a very specific technique to avoid having to do antialiasing in real-time rendering, and it's already covered sufficiently in articles on texturing and texture filtering. - Valarauka(T/C) 05:55, 4 October 2006 (UTC)

Mip-mapping has a few applications to audio signals, and even in non-realtime graphics there are uses. It is rightly an anti-aliasing technique, not an anti-aliasing "avoidance" technique. - Rainwarrior 15:30, 18 October 2006 (UTC)
As far as FSAA, it is an application of oversampling, which is an often used anti-aliasing technique. I don't see it being expanded much, so I would support a merge into either this article, or oversampling. - Rainwarrior 15:36, 18 October 2006 (UTC)

[edit] Whats with this sentence?

Since eyes can rotate in their sockets, this must have to do with the fact that we are dealing with data sampled on a square lattice and not with a continuous image.

What does that have to do with anything? I think this sentence should go. User A1 00:14, 15 December 2006 (UTC)

Yes, clearly nonsense. Please do get rid of it. Dicklyon 00:36, 15 December 2006 (UTC)
Retrieved from "http://en.wikipedia.org../../../a/n/t/Talk%7EAnti-aliasing_8346.html"