Talk:Dither

From Wikipedia, the free encyclopedia

Contents

[edit] Noise

Some scientists/engineers define dither broadly enough to include forms that have little or no noise content. In the examples given, the WWII bomb trajectory calculator, tapping on the cookie sheet, etc., the dithering signals are not particularly noisy. As long as the signal is large enough to overcome the slip-stick friction of the calculator's moving parts, or of the tater tots, but not so large as to introduce significant error, or send the tots across the kitchen, it will be an effective dither source. The "colored" and "psychoacoustic" dither used in digital audio is a step away from randomness, and indeed, some audio manufacturers (e.g. Lexicon) experimented with narrowband dither (at the Nyquist rate) which cannot really be described as noise.

So more broadly, dither can be defined as a signal which is injected into a system to overcome low-level non-linearity, be it quantization error or slip-stick friction. With this broad definition, it becomes practically synonymous with "stochastic resonance", which is the term the scientific -- especially the neurophysiological -- community uses to describe, well, dither. Not surprisingly, the transfer function of a neuron is not particularly linear. Therefore, not surprisingly, injecting a random (stochastic) signal into a neurological system increases its sensitivity. -Jim MacArthur, Chief Engineer, Harvard Electronic Instrument Design Lab Squinto 20:37, 6 September 2005 (UTC)

[edit] 4 bit example pictures

Would be interesting to see not dithered and dithered pictures with an unoptimized 16 colour vga standard palette. Sure, the first one will look terrible and the second on not really good, but they'll serve as good comparisons to the other pictures and show what's possible with very few colours.

[edit] Noise in image dithering?

The explanations of the uses of dithering in both audio och image processing are excellent. However, I fail to see how the techniques are related. How does the "noise" fit into the picture when dithering an image? It would be great if someone could elaborate on this. MEMark 11:54, 27 October 2005 (UTC)

Audio is 1-dimensional, images are 2-dimensional. But otherwise what is happening is identical - noise is added to the signal to spread the quantization error over a number of samples. If you examine a dithered image closely, you'll see it looks noisy - the noise is actually visible as pixels of slightly variable colour which from further away (averaged by the eye) look like a single solid colour. Graham 10:34, 21 December 2005 (UTC)

[edit] Quantization error

"Another plausible solution would be to take 4.8 and round it so that four times out of five it rounded up to 5, and the other time it rounded to 4. This would average out to exactly 4.8 over the long term. Unfortunately, however, it still results in repeatable and determinable errors, and those errors still manifest themselves as distortion to the ear."

Is this only in practice? Why do these errors occur? Does oversamplng realy average this?

203.214.75.127

4, 4, 4, 4, 5, 4, 4, 4, 4, 5 does not create the same waveform as 4.8, 4.8, 4.8, 4.8, 4.8, 4.8, 4.8... even though mathematically speaking they average to the same value over the long run. The ear will pick this up quite easily as a nasty sort of low-level scratchiness (distortion). By adding dither the error is spread randomly which is far more acceptable to the ear - it sounds like a quiet hiss instead of distortion. The brain is wired to detect patterns in the sounds it hears, so a systematic attempt to eliminate the rounding error will create a detectable pattern (called correlation) whereas simply randomising it makes it 'invisible' to the brain (uncorrelated). Graham 10:30, 21 December 2005 (UTC)

[edit] Frequency domain

I deleted the sentence about the ear being more sensitive in the frequency domain. I'm an electrical engineer and this statement makes no sense. Frequency and time domain are two ways of representing the SAME signal. — (unsigned comment by 195.243.189.189, 2005-12-21 01:41:21)

[edit] Split into multiple articles?

I think I originally advocated combining the general dither article and the image dither article (I think it used to be dithering), but now this article is very big and deals with some rather specialized stuff, so maybe we should split out the audio and image content, so that we have three articles:

  • a general dither article, with two short summaries and links to the two major types:
  • dither (audio)
  • dither (images) — Omegatron 21:51, 13 January 2006 (UTC)
Hmmm, I'm not so sure. The processes are the same - not just analogous, but identical. It's just that one type of data is 1-dimensional audio, the other 2-dimensional images. I think having them in one article makes this much clearer, and I don't think the article's THAT long. Graham 11:39, 23 January 2006 (UTC)
Ok. It is pretty long, but it would probably be fine to just rearrange it. It currently looks like two articles glued together, which, actually, I think I'm responsible for.  :-) — Omegatron 14:32, 23 January 2006 (UTC)
My concern is that audio dither is given much more coverage than digital imagry. For example, the article only goes into (a few sentences of) Floyd-Steinberg dither and neglects other forms of dithering, like Bayer ordered dithers or halftoning. --Trixter 22:07, 9 February 2006 (UTC)
I agree, except that I would call the articles audio dithering and image dithering. I think this split is a good idea because many readers will only care about one particular type of dithering (for example, a web graphic artist would only really care about image dithering, while a sound engineer would only care about audio dithering). Also, this will give the articles more room for much-needed expansion and lots of illustrative images. Deco 22:12, 28 April 2006 (UTC)

[edit] Photography

So I guess the random placement of grains in regular photographic film is a form of dither? — Omegatron 15:43, 24 January 2006 (UTC)

I guess you're right... since a photographic emulsion molecule can really only be in one of two states - either "on", changed to silver, or "off", as silver halide - then smooth tones must be made from particles having a random distribution but a probability of being in one or the other state proportional to the grayscale of the image at that point. This is very similar to a dithered 1-bit image which are common on computers (especially older ones such as 80s-era Macs), though the "grain size" there is relatively enormous.Graham 22:04, 24 January 2006 (UTC)
Now that I rethink about it, I was thinking about anti-aliasing, not dither. But I guess it does both. — Omegatron 00:44, 13 February 2006 (UTC)

[edit] Why minimize quantization error?

Can somebody explain how dither minimizes the quantization error compared to simply rounding to the nearest quantization level?

It doesn't, really. It removes systematic error, such as if you're repeatedly rounding by the same amount, which can cause "new" signals to appear. When dithering, the quatization error manifests itself as noise rather than a signal. europrobe 19:56, 29 May 2006 (UTC)
I should have made myself clearer. I was actually addressing the very first sentence of the page: "Dither is a form of noise, or 'erroneous' signal or data which is added to sample data for the purpose of minimizing quantization error.". This sentence is not correct since quantization error is not minimized ( at least by any reasonable measure that I can think of ).
So change it. europrobe 17:26, 30 May 2006 (UTC)

[edit] Good article

Interesting stuff. Has anyone considered nominating this for Good Article status? Tpth 04:32, 2 August 2006 (UTC) I agree. This is a really informative and well-written article.216.107.192.114 18:00, 5 November 2006 (UTC)

[edit] Optimised Palette

With regard to dithering images, what is the technique used to determine an optimised palette? Is it part of the dithering process, or can it be done independently? More specifically, would the optimised palette be the same regardless of which dithering process is to be used? --Malcohol 12:57, 29 August 2006 (UTC)

Aha! The Color quantization article discusses this issue. A quote: "Color quantization is frequently combined with dithering, which can eliminate unpleasant artifacts such as banding that appear when quantizing smooth gradients and give the appearance of a larger number of colors. Some modern schemes for color quantization attempt to combine palette selection with dithering in one stage, rather than perform them independently."--Malcohol 10:24, 31 August 2006 (UTC)

[edit] Dither & Bombsights

Norden bombsight currently states:

In combat, this accuracy was never achieved - because the Norden had been tested under "artificial conditions" at the US proving grounds, for example in the absence of anti-aircraft fire and/or adverse weather.

This seems to partially contradict this page which states:

Airplane bombers used mechanical computers to perform navigation and bomb trajectory calculations. Curiously, these computers (boxes filled with hundreds of gears and cogs) performed more accurately when flying on board the aircraft, and less well on ground. Engineers realized that the vibration from the aircraft reduced the error from sticky moving parts.

Ewlyahoocom 14:28, 6 September 2006 (UTC)

(This issue has been addressed - there is no contradiction. 82.139.85.81 04:28, 16 October 2007 (UTC))

[edit] This article shouldn't require cleanup

I am an electronics engineer and needed to review some theory behind decibel calculations. I found a link to this article on dithering. I was quite surprised to see that someone had placed a 'Cleanup' tag on this article, implying that it needed work. That is ridiculous - the discussion of dithering in this article and the examples of the usage of dithering in making digital photographs clearer, make this article one of the best written I have ever come across in Wikipedia. This subject could not be discussed and illustrated any more clearly, or reduced to anything more elemental without compromising the subject matter.--70.142.9.33 08:45, 3 September 2007 (UTC)

Cleanup tag removed on Sept. 3, 2007--70.142.9.33 09:11, 3 September 2007 (UTC)

I think that audio section is still a bit of a mess; seems to be someone's personal essay, not very well wikified, and completely unsourced. So I added an unsourced section tag. If we get a good source or two that will make it easier to work on. Dicklyon 16:33, 3 September 2007 (UTC)
I am actually doing an investigation on audio mastering and came across this great book by Bob Katz, "Mastering Audio: the art and the science", which explains in great detail the purpose of dithering in digital audio not only when converting between wordlengths, but when recording analong input through an Analog/Digital converter. I am planning to add some quotes and references to this book on the Digital Audio section, since I think it lacks the explanation of some purposes of dithering in digital audio.--Tevex 06:36, 18 September 2007 (UTC)

[edit] Sierra's filter light

Sierra may think his "filter light" is pretty good, but in images with slow gradients and such I think I'm seeing more vertical and diagonal banding than with Floyd-Steinberg. Is anyone else experiencing this issue? Shinobu 10:03, 2 December 2007 (UTC)


[edit] Images

Really nice article. Could we maybe add some more relevant images at the start? The original dictionary definitions are interesting, but could we have something a bit more appealing to the eye? Hate to be a wikifairy...

on the images ... isnt it kinda strange that Dithering_example_dithered_256color.png, contains 607 distinct colors ? (acording to Irfan view) 83.25.83.82 (talk) 23:17, 26 January 2008 (UTC)

Sure enough, someone boned it. I tried to fix, but some of the pixels needed to fix the inset are missing, and the magnification of pixels is by a non-integer, so it's hard to fix right. Dicklyon (talk) 00:02, 27 January 2008 (UTC)

[edit] This is an error, right?

"If we reduce our waveform by, say, 20% then we end up with the following values:"

then

"If any waveform, comprising the original values, were to be processed by multiplying each value by 0.8, the result would contain errors."

note that if you get 1, reduce it by 0.2, you'll have 0.8 . If you multiply 0.8 for 1-(0.2) (he meant, to try to get the original values again) , we get 0.64. The right value should be (0.8)-1 which is 1,25. I've changed so in the page, so if I'm wrong please say and Ill restore to 0.8 . —Preceding unsigned comment added by 189.33.159.46 (talk) 21:40, 7 March 2008 (UTC)

[edit] Dithering and compression

This is my original research so can't go in the article. I did some tests in GIMP of different image file formats (the image used being a screenshot of a web browser), and found that dithering in GIF and (palleted) PNG images substantially increases the compressed file size (by a factor of 2-3) compared to an undithered image, whether using optimized or pre-defined pallets. Presumably this arises as a result of the compression algorithms handling large areas of predictable solid colour far better than the randomness that dithering by definition creates. If anyone can find other studies to back these results up, it would be worth including. 128.232.228.174 (talk) 01:09, 3 April 2008 (UTC)

Yes, it's well known]. Dicklyon (talk) 04:20, 3 April 2008 (UTC)