In most cases, anti-aliasing means removing data at too high a frequency to represent. When sampling is performed without removing this part of the signal, it causes undesirable artifacts such as the black-and-white noise near the top of figure 1-a.
See the articles on signal processing and aliasing for more information about the theoretical justifications for anti-aliasing; the remainder of this article is dedicated to anti-aliasing methods in computer graphics.
Figure 1-a illustrates that visual distortion which occurs when anti-aliasing is not used. Notice that near the top of the image, where the checkerboard is very distant, the image is impossible to recognize, and is displeasing to the eye. By contrast, figure 1-b is anti-aliased. The checkerboard near the top blends into gray, which is usually the desired effect when the resolution is insufficient to show the detail. Even near the bottom of the image, the edges appear much smoother in the anti-aliased image. Figure 1-c shows another anti-aliasing algorithm, based on the sinc filter, which is considered better than the algorithm used in 1-b. Figure 2 shows magnified portions of Figure 1 for comparison. The left half of the image is taken from Figure 1-a, and the right half of the image is taken from Figure 1-c. Observe that the gray pixels, which aren't very attractive at that size, help make 1-c much smoother than 1-a.
|File:Anti-aliased diamond.png||File:Anti-aliased diamond enlarged.png|
| Compare the |
diamond on the
left with the
on the right
Fig 3 shows the effect of anti-aliasing on a small black and white image. The enlarged image shows how anti-aliasing adds gray pixels around the border between black and white, which visually smooths the outline. Text is affected in just the same way.
First-principles approach to anti-aliasingEdit
There are infinitely many such functions. However, the computer screen is capable of displaying only finitely many different images. Indeed, an ordinary computer screen has no more than a few million pixels, and each pixel only has a finite number of colors it can display.
Hence, to convert an image f(x,y) into something that the screen can display, we must simplify it. By the pigeonhole principle, sometimes two different ideal images f(x,y) and g(x,y) will be converted to the same picture on the screen. This cannot be avoided. The question is how to choose the reduced image so that it looks better.
An example of a poor choice is illustrated in Figure 1-a. The most direct way to simplify an image for display is to use a sample of the image at f(i,j) for each pixel (i,j). That is how figure 1-a was generated. At the top of the checkerboard, multiple black and white tiles may be represented by a single pixel. But since only black and white points occur in the ideal image, an area containing both colors in similar proportion will be represented with a strange pattern of black and white. This type of aliasing is called a Moiré effect.
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. To see why this works better, it helps to look at the problem from a signal-processing perspective.
Signal processing approach to anti-aliasingEdit
In this approach, the ideal image is regarded as a signal, the image displayed on the screen is viewed as a certain filtering of the signal. Ideally, we would understand how the human brain would process the original signal, and provide the image on the screen which most resemble the original signal according to the human brain.
The most widely accepted method is to use the Fourier transform. The Fourier transform decomposes our signal into basic waves, or harmonics, and gives us the amplitude of each wave in our signal. The waves are of the form:
where j and k are arbitrary non-negative integers. (In fact, there are also waves involving the sine, but for the purpose of this discussion, the cosine will suffice; see Fourier transform for technical details.)
The numbers j and k together are the frequency of the wave: j is the frequency in the x direction, and k is the frequency in the y direction.
It has been observed (by Harry Nyquist) that to measure a signal of frequency n, you need at least 2n sample points, and they need to be well-placed. If your sample points occur near the zeros of the signal, you will be led to the belief that the signal is in fact zero.
And so, in signal processing, we choose to eliminate all high frequencies from the signal, keeping only the frequencies that are low enough to be sampled correctly by our sample rate.
Although there is no known justification why this form of aliasing is less troublesome than some other form of averaging, such as the uniform averaging algorithm, experimentation with humans has suggested that the Fourier approach is superior and somehow matches what the brain would expect to see.
Figure 1-c was generated with this approach. It was not possible to do an exact Fourier series truncation; however, an approximation was used which we hope comes close to the correct image. To highlight the differences between 1-b and 1-c, observe that 1-c manages to be a bit clearer further up on the image than 1-b does. We are able to distinguish some texture other than uniform gray higher up on the image in 1-c than in 1-b.
The basic waves need not be cosine waves. See, for instance, wavelets. If one uses basic waves which are not cosine waves, one obtains a slightly different image. Some basic waves yield anti-aliasing algorithms which are not so good (for instance, the Haar wavelet gives the uniform averaging algorithm.) However, some wavelets are good, and it is possible that some wavelets are better at approximating the functioning of the human brain than the cosine basis. Sinc functions, with the usual spacing, are orthogonal and have the property of being zero at each-other's centers. This minimizes the loss of data, on resampling.
Two dimensional considerations Edit
The above assumes that the rectangular mesh sampling is the dominant part of the problem. It should seem odd that the filter usually considered optimal is not rotationally symmetrical, as shown in this first figure. 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. This must be the justification for doing signal processing, along each axis, as it is traditionally done on one dimensional data.
If the resolution is not limited by the rectangular sampling rate of either the source or the target image, then one should ideally use rotationally symmetrical filter or interpolation functions, as though the data were a two dimensional function of continuous x and y. The sinc function of the radius, in the second figure, has too long a tail to make a good filter (not square integrable). One might consider a Gaussian plus enough of its second derivative to flatten the top. This is shown also. Functions based on the Gaussian function are natural choices, because convolution with a Gaussian gives the same result, whether applied to x and y or to the radius. Another of its properties is that it (similarly to wavelets) is half way between being localized in the configuration (x and y) and in the spectral (j and k) representation. As an interpolation function, a Gaussian alone seems too spread out to preserve the maximum possible detail.
As an example, when printing a photographic negative, with plentiful processing capability, on a printer with a hexagonal pattern, there is no reason to use sinc function interpolation. This would treat diagonal lines differently from horizontal and vertical lines, which is like a weak form of aliasing.
Practical real-time anti-aliasing approximationsEdit
There are only a handful of primitives used at the lowest level in a real-time rendering engine (either software or hardware accelerated.) These include "points", "lines" and "triangles". If one is to draw such a primitive in white against a black background, it is possible to design such a primitive to have fuzzy edges, achieving some sort of anti-aliasing. However, this approach has difficulty dealing with adjacent primitives (such as triangles that share an edge.)
To approximate the uniform averaging algorithm, one may use an extra buffer for sub-pixel data. The initial, and least memory-hungry approach, used 16 extra bits per pixel, in a 4×4 grid. If one renders the primitives in a careful order, for instance front-to-back, it is possible to create a reasonable image.
Since this requires that the primitives be in some order, and hence interacts poorly with an application programming interface such as OpenGL, the latest attempts simply have two or more full sub-pixels per pixel, including full color information for each sub-pixel. Some information may be shared between the sub-pixels (such as the Z-buffer.)
Main article: Mipmap
There is also an older approach specialized for texture mapping called mipmapping, which does not require any sub-pixel samples, and in fact can improve the speed of the rendering by improving locality of reference, hence improving the performance of any cache system.
Mipmapping works by creating lower resolution, prefiltered versions of the texture map. When rendering the image, the appropriate resolution mip-map is chosen and hence the texture pixels (texels) are already filtered when they arrive on the screen.
Typically, if a texture is of resolution 256×256 (for instance) then mipmaps will be created with the following resolutions: 128×128, 64×64, 32×32, 16×16, 8×8, 4×4, 2×2 and 1×1. The advantage here is that having such a hierarchy of texture maps requires only 1/3 more memory than having simply the base 256×256 texture map. However, in many instances, the filtering should not be uniform in each direction (it should be anisotropic, as opposed to isotropic), and a compromise resolution is used. If a higher resolution is used, the cache coherence goes down, and the aliasing is increased in one direction, but the image tends to be clearer. If a lower resolution is used, the cache coherence is improved, but the image is overly blurry, to the point where it becomes difficult to identify.
To help with this problem, nonuniform mipmaps (also known as rip-maps) are sometimes used. With a 16×16 base texture map, the rip-map resolutions would be 16×8, 16×4, 16×2, 16×1, 8×16, 8×8, 8×4, 8×2, 8×1, 4×16, 4×8, 4×4, 4×2, 4×1, 2×16, 2×8, 2×4, 2×2, 2×1, 1×16, 1×8, 1×4, 1×2 and 1×1.
The unfortunate problem with this approach is that rip-maps require four times as much memory as the base texture map, and so rip-maps have been very unpopular.
To reduce the memory requirement, and simultaneously give more resolutions to work with, summed-area tables were conceived. Given a texture , we can build a summed area table as follows. The summed area table has the same number of entries as there are texels in the texture map. Then, define
Then, the average of the texels in the rectangle (a1,b1] × (a2,b2] is given by
However, this approach tends to exhibit poor cache behavior. Also, a summed area table needs to have wider types to store the partial sums than the word size used to store . For these reasons, there isn't any hardware that implements summed-area tables today.
A compromise has been reached today, called anisotropic mip-mapping. In the case where an anisotropic filter is needed, a higher resolution mipmap is used, and several texels are averaged in one direction to get more filtering in that direction. This has a somewhat detrimental effect on the cache, but greatly improves image quality.
An example of an image with extreme pseudo-random aliasing Edit
Because fractals have unlimited detail and no noise other than arithmetic roundoff error, they illustrate aliasing more clearly than do photographs or other measured data. The dwells, which are converted to colors at the exact centers of the pixels, go to infinity at the border of the set, so colors from centers near borders are unpredictable, due to aliasing. This example has edge in about half of its pixels, so it shows much aliasing. The first image is uploaded at its original sampling rate. Since most modern software anti-aliases, one may have to download the full size version to see all of the aliasing. The second image is calculated at five times the sampling rate and down-sampled with anti-aliasing. Assuming that we would really like something like the average color over each pixel, this one is getting closer. It is clearly more orderly than the first.
It happens that, in this case, there is additional information that can be used. By re-calculating with the distance estimator, points were identified that are very close to the edge of the set, so that unusually fine detail is aliased in from the rapidly changing dwell values near the edge of the set. The colors derived from these calculated points have been identified as unusually unrepresentative of their pixels. Those points were replaced, in the third image, by interpolating the points around them. This reduces the noisiness of the image but has the side effect of brightening the colors. So this image is not exactly the same that would be obtained with an even larger set of calculated points.
To show what was discarded, the rejected points, bled into a grey background, are shown in the fourth image.
Finally, "Budding Turbines" is so regular that systematic (Moiré) aliasing can clearly be seen near the main "turbine axis" when it is downsized by taking the nearest pixel. The aliasing in the first image appears random because it comes from all levels of detail, below the pixel size. When the lower level aliasing is suppressed, to make the third image and then that is down-sampled once more, without anti-aliasing, to make the fifth image, the order on the scale of the third image appears as systematic aliasing in the fifth image.
The best anti-aliasing and down-sampling method here depends on one's point of view. When fitting the most data into a limited array of pixels, as in the fifth image, sinc function anti-aliasing would seem appropriate. In obtaining the second and third images, the main objective is to filter out aliasing "noise", so a rotationally symmetrical function may be more appropriate.
Important early works in the history of anti-aliasing include:
- Freeman, H.. "Computer processing of line drawing images", ACM Computing Surveys vol. 6(1), March 1974, pp. 57–97.
- Crow, Franklin C.. "The aliasing problem in computer-generated shaded images", Communications of the ACM, vol. 20(11), November 1977, pp. 799–805.
- Catmull, Edwin. "A hidden-surface algorithm with anti-aliasing", Proceedings of the 5th annual conference on Computer graphics and interactive techniques, p.6–11, August 23–25, 1978.
- Dithered ray tracing and anti-aliasing
- Statistical sampling and anti-aliasing
- Temporal anti-aliasing or motion blur
- Measure theory and anti-aliasing
- Font rasterization and anti-aliasing
- Interpolation and Gamma Correction In most real-world systems, gamma correction is required to linearise the response curve of the sensor and display systems. If this is not taken into account, the resultant non-linear distortion will defeat the purpose of anti-aliasing calculations based on the assumption of a linear system response.
- Color theory for certain physical details pertinent to images which are not grayscale.
- reconstruction filter
- quincunx (pattern used for anti-aliasing)
- Full Scene Anti-aliasing
- Anti-Aliasing: Interactive diagram and visual explanation