Look guys, I do not want to insult anyone but there has been a lot of misinformation passed back and forth on this subject. But don't feel too bad, I read a lot of nonsense about the aa filter, what it is for and how it works. My credentials include being a scientist doing image processing research for many years.
Here is a mini tutorial:
A digital photo sensor measures descrete samples from the image projected upon it by the lens. All digital images processing is based on what is called the sampling theorm. And the sampling theorem guarantees that the image can be reconstucted from a finite number of samples. It is a remarkable result that suggests how the image can be reconstructed at any point in the image plane, not just at the samples. But there is a catch. One can reconstruct perfectly only if the image has no information above some spatial frequency (the Nyquist frequency) and the number of samples in the sensor is sufficient (the sampling frequency must be at least twice the Nyquist frequency). On the other hand, if the lens creates an image with too high a frequency content, the image becomes distorted. I want to emphasize, that this distortion mixes the (too) high frequency information into the recosntructred image, distorting it; and no process can rigorously remove the distortion. Just because one does not see the tell tale aspects of alaising, does not mean that the image is fine. The information above the Nyquist frequency will appear somewhere in the image, limiting image quality. The idea of the aa filter is to remove the high frequency information above the Nyquist frequency before it influences image formation. There are some additional assumptions underlying the theory, like the data are nosieless, but the basic ideas are accepted throughout the image processing community. When you remove the aa filter, you are hoping that the lens does not pass too much high frequency information. The end results will depend on both the lens and the scene photographed. A too sharp lens could actually be a problem!
There is no free lunch. No one will be able to design a system that outperforms one with a properly designed aa filter.
nma wrote:
Look guys, I do not want to insult anyone but there has been a lot of misinformation passed back and forth on this subject. But don't feel too bad, I read a lot of nonsense about the aa filter, what it is for and how it works. My credentials include being a scientist doing image processing research for many years.
Here is a mini tutorial:
A digital photo sensor measures descrete samples from the image projected upon it by the lens. All digital images processing is based on what is called the sampling theorm. And the sampling theorem guarantees that the image can be reconstucted from a finite number of samples. It is a remarkable result that suggests how the image can be reconstructed at any point in the image plane, not just at the samples. But there is a catch. One can reconstruct perfectly only if the image has no information above some spatial frequency (the Nyquist frequency) and the number of samples in the sensor is sufficient (the sampling frequency must be at least twice the Nyquist frequency). On the other hand, if the lens creates an image with too high a frequency content, the image becomes distorted. I want to emphasize, that this distortion mixes the (too) high frequency information into the recosntructred image, distorting it; and no process can rigorously remove the distortion. Just because one does not see the tell tale aspects of alaising, does not mean that the image is fine. The information above the Nyquist frequency will appear somewhere in the image, limiting image quality. The idea of the aa filter is to remove the high frequency information above the Nyquist frequency before it influences image formation. There are some additional assumptions underlying the theory, like the data are nosieless, but the basic ideas are accepted throughout the image processing community. When you remove the aa filter, you are hoping that the lens does not pass too much high frequency information. The end results will depend on both the lens and the scene photographed. A too sharp lens could actually be a problem!
There is no free lunch. No one will be able to design a system that outperforms one with a properly designed aa filter.
Thank you for the explanation. Could you please clarify how Leica gets the results that they do from their AAless sensor and supposedly sharper lenses? Are they near their limit in terms of IQ with that design?
Thank you for the explanation. Could you please clarify how Leica gets the results that they do from their AAless sensor and supposedly sharper lenses? Are they near their limit in terms of IQ with that design?
I am not able to comment on Leica and their cameras. However, I am certain that their results cannot violate the sampling theorem without there being some important cost. I have read some breathless comments from devotees, but whether any pleasing results obtained actually come from removing the aa filter is not clear. You should remain skeptical that Leica or anyone else knows something that has escaped Canon or the other makers. Digital cameras may be relatively new, but their designs are all based on published and well known principles.
By your hypothesis (or science) quoted above, it would seem odd that Leica - having some of the sharpest lenses around - would not have run into the issues that you bring up. Almost seems counterintuitive.
Have you read the entire 'Leica DMR Bible'? I'm sure there's some critical insight contained in there, somewhere...
Jeff wrote:
By your hypothesis (or science) quoted above, it would seem odd that Leica - having some of the sharpest lenses around - would not have run into the issues that you bring up. Almost seems counterintuitive. ???
Have you read the entire 'Leica DMR Bible'? I'm sure there's some critical insight contained in there, somewhere... :D
Jeff,
I understand your skepticism, after all you can perceive better image quality with the Leica lens. But why? Is it resolution or some other property, such as contrast? One useful way to think about such things is to think of a more extreme case. Let's take the example of a textile shot, with a high frequency woven pattern and tiny threads. Imagine shooting with the Leica lens, no aa filter and a 3 megapixel sensor. This would be grossly undersampled and one expects to see distortions, alaising, beacuse the sampling frequency of the sensor is not high enough. All the high frequency information above the Nyquist frequency is mixed in a lower frequencies. Now, imagine gradually increasing the sampling frequency of the sensor. The amount of aliasing will decrease and at some point disappear. But in the intermediate region a system with a properly designed aa filter and the Leica lens will perform better.
There really is no free lunch. In situations where the very high frequency content of the image is low, you may prefer the Leica without an aa filter. But the same lens mounted on the Canon 1Ds II, would work even better. I should also point out that building an appropriate aa filter may also be difficult, I don't really know how it is done. We are talking about real world issues. The idea is to roll off the the high frequency portion of the MTF of "any" lens to zero for frequencies above Nyquist for the sensor on a given camera. Implementation of such an optical filter will have its own imperfections.
One of the problems with making an effective AA filter is that it's apparently impossible to make an optical filter with a sharp cutoff frequency. That is, to blur the high frequencies you must also blur the lower frequencies to some extent. A similar problem exists in audio where, although sharp cutoff filters can be made, they are expensive to produce. The audio engineers solve the problem by oversampling to a great degreee so as to push the Nyquist frequency much higher to a point where a gradual rolloff filter can be used without affecting the audio frequencies. In optical terms this would mean making a sensor with a huge number of pixels. But then you have all sorts of other problems such as noise and diffraction which can't be overcome. In other words your DSLR becomes basically a p&s but with more pixels.
Some people might like a DSLR with 40 mp that begins diffraction limiting at f/5.6 and can't be used at higher than ISO 400 but most DSLR users would not.
George Deliz
gdeliz2 wrote:
One of the problems with making an effective AA filter is that it's apparently impossible to make an optical filter with a sharp cutoff frequency. That is, to blur the high frequencies you must also blur the lower frequencies to some extent. A similar problem exists in audio where, although sharp cutoff filters can be made, they are expensive to produce. The audio engineers solve the problem by oversampling to a great degreee so as to push the Nyquist frequency much higher to a point where a gradual rolloff filter can be used without affecting the audio frequencies. In optical terms this would mean making a sensor with a huge number of pixels. But then you have all sorts of other problems such as noise and diffraction which can't be overcome. In other words your DSLR becomes basically a p&s but with more pixels.
Some people might like a DSLR with 40 mp that begins diffraction limiting at f/5.6 and can't be used at higher than ISO 400 but most DSLR users would not.
George Deliz
I actually think your audio example could be used effectively for dSLRs. One could use a very high-sampling sensor and a filter that is at relaitvely high frequency. Then down sampling the result to a more modest sampling frequency would restore most of the lost SNR. In fact, this is exacly how the audio systems work.
NMA -- thanks for the really interesting and useful information!
I have a related question that you can hopefully answer regarding lens versus sensor resolution.
It's often stated that the very high res digital sensors like the 5D and 1DsII are 'merciless' on poor glass, implying that their resolution is so high that substandard glass really limits their performance.
But that is really inconsistent with experience in the medium format and large format worlds, where by virtue of the negative size even the best lenses in the world are outresolved by the negative -- take an extreme example like an ultralarge format 11x14 or 20x24, for instance.
In these cases, the resolution is always lens-limited because of the resolution of the capture medium. But no one in the LF or MF communities complain about the format being merciless on lower end glass. In fact it's the opposite -- people get away with 100 year old uncoated lenses and stop down to f/64 (way into the diffraction world) without any decrement in image quality, specifically because the capture medium is of such high resolution that lens errors are trivial next to the absolute amount of detail that gets captured.
On the other end of the spectrum, I'd expect that a 6 megapixel 300D would actually be less forgiving of cheap glass than a 17 megapixel 1DsII. The rationale is that the 6 MP sensor is much more of a limiting factor in detail capture, and any further loss of detail from cheap glass would have a much greater effect on the image quality.
I actually think your audio example could be used effectively for dSLRs. One could use a very high-sampling sensor and a filter that is at relaitvely high frequency. Then down sampling the result to a more modest sampling frequency would restore most of the lost SNR. In fact, this is exacly how the audio systems work. "
But there is no audio analog for diffraction. Down sampling will not restore losses due to diffraction.
nma wrote:
I actually think your audio example could be used effectively for dSLRs. One could use a very high-sampling sensor and a filter that is at relaitvely high frequency. Then down sampling the result to a more modest sampling frequency would restore most of the lost SNR. In fact, this is exacly how the audio systems work.
Precisely. And I think that's where DSLRs are headed.
gdeliz2 wrote:
But there is no audio analog for diffraction. Down sampling will not restore losses due to diffraction.
George Deliz
I'm afraid that this is a misunderstanding of diffraction. Amount of diffraction doesn't depend on the pixel size because diffraction happens on the lens not on the sensor. For a given lens and aperture it would be the same regardless of the media used - film, sensor with 10 or 1000 megapixels. If (or better when) it is possible to build a sensor exceeding the lens resolution we can loose the AA filter for good.
If you guys do check out the link above, make sure you check out Sean's photography, it's fantastic! In fact, this is one of my favorite sites, by virtue of the fact that the technical information is of such high quality, and is so well presented.
[/tangent]
Diffraction limits the resolution of the lens and soften the image. If you stop your lens down it may happend that the lens will not be able to resolve details as small as pixels. This doesn't affect how the sensor or idividual pixels work. You can get to this point with whatever camera you use right now. Just get yourself a pin hole. It is a diffraction limited "lens", but you sensor will work just fine and will be giving you an oversampled image. In this case you could take off the AA filter (if t was removable) without any risk of moire. For real lenses we will need much smaller pixels to get an oversampled image, but if it is technicaly possible we will not need an AA filter.
But you're less likely to run into that diffraction limit with full frame than with APS-C, and again with larger formats. By the time you're shooting 4x5 and above the amount of diffraction relative to the film size is miniscule. Many of Ansel Adams shots were at f/64 -- hence the 'f/64 club'.
Re: gdeliz2/Alex's comments, they're both right (if added together). There is an analog to diffraction-in-audio: if an optic element was able to bend all frequencies of light to the same angle, it would be free of diffraction in a pathological sense. Amplifiers generally cannot amplify all frequencies exactly the same, hence distortion.
The trick is downsampling...like they said. In Audio and in Signals Processing its assumed that 2x is the absolute minimum for reliable sampling. We're going to see by the end of the decade, down-sampling in sensors as a way of getting image quality.
I have lived for 2yrs with a camera with a removable AA filter, and, yeah, I've only had 1 truly ruined picture, on very predictable scenario: A shot of a big house (big local news) under construction, in framing with diagonal lumber giving zebra stripes.
Panic, find the AA filter, put it back in (looks like a metal-framed slide with one screw) and life is good.
losloslos wrote:
Hey guys. I'm here with another obtuse point.
Re: gdeliz2/Alex's comments, they're both right (if added together). There is an analog to diffraction-in-audio: if an optic element was able to bend all frequencies of light to the same angle, it would be free of diffraction in a pathological sense. Amplifiers generally cannot amplify all frequencies exactly the same, hence distortion.
The trick is downsampling...like they said. In Audio and in Signals Processing its assumed that 2x is the absolute minimum for reliable sampling. We're going to see by the end of the decade, down-sampling in sensors as a way of getting image quality.
I have lived for 2yrs with a camera with a removable AA filter, and, yeah, I've only had 1 truly ruined picture, on very predictable scenario: A shot of a big house (big local news) under construction, in framing with diagonal lumber giving zebra stripes.
Panic, find the AA filter, put it back in (looks like a metal-framed slide with one screw) and life is good.
I want to emphasize once more the consequences of undersampling. Imagine that you could take the 2-dimensional Fourier transform of the "airiel" image that is projected by the lens onto the digital sensor. If the Fourier transform does NOT become zero for all frequencies above Nyquist, THERE WILL BE DISTORTION. We are only arguing over how much distortion is present and whether it is obtrusive. If the sampling theorem is not satisified, you will loose image quality because the high frequency information goes somewhere in the image. I suspect that in the best case without the aa filter, there is always some gritty haze that we may perceive as noise.
This stuff is taught in great detail to all electrical engineers in signal processing 1.01. So, all the manufacturers know about it and certainly base their designs on the sampling theorem and other basic signal processing principles. If Leica leaves out the aa filter, they are making certain assumptions about the modulation transfer function of the lens and perhaps how the camera is likely to be used.
nma wrote:
I want to emphasize once more the consequences of undersampling. Imagine that you could take the 2-dimensional Fourier transform of the "airiel" image that is projected by the lens onto the digital sensor. If the Fourier transform does NOT become zero for all frequencies above Nyquist, THERE WILL BE DISTORTION. We are only arguing over how much distortion is present and whether it is obtrusive. If the sampling theorem is not satisified, you will loose image quality because the high frequency information goes somewhere in the image. I suspect that in the best case without the aa filter, there is always some gritty haze that we may perceive as noise.
This stuff is taught in great detail to all electrical engineers in signal processing 1.01. So, all the manufacturers know about it and certainly base their designs on the sampling theorem and other basic signal processing principles. If Leica leaves out the aa filter, they are making certain assumptions about the modulation transfer function of the lens and perhaps how the camera is likely to be used.
'Hope this helps. ...Show more →
As an former EE with years of experience collecting data digitally I gotta chime in.
It's an optical filter - electronic sampling theory for the electromagnetic frequency domain has similarities to optics but it can only stretch so far. The pixel spacing and size could loosely be construed as a "sample" of the image being projected on the capture device, but no sampling frequency is involved in the true sense. What is of concern with these samples is the performance with image details that meet or exceed the pixel capabilites to capture that data. The image data (think lines here) could roughly translate to the frequency being sampled but there is no equivalent to the frequency of the sampling device (the imaging sensor). The best you can do is relate the sample window (pixel width) and the sample interval (inter-pixel distance) as related to the actual image resolution being projected on the imaging device. The only relevance Nyquist has is identifying the point in the frequency domain that you need to start rejecting frequencies that can show up as aliases. Optically this would be a spectral filter like the ones that reject or pass IR, or a "fuzzifier" which in effect is a low pass filter that eliminates the "frequency" component past a certain point (less data = fuzzier image). In electronics the transfer function of the filter is by necessity a compromise and determined by the amount and rate of rolloff at the cutoff frequency. These design compromises are what introduce distortion. The design of an optical filter is similar but to my knowledge (admitted slim pickins optically) things like non-linear response, phase shifts and delays are not much of a concern if any. Thus the term distortion as you use it is somewhat misleading in the way that most of these folks are familiar with which is in things like barrel and pincushion. The type of distortion you refer to and that is described by sampling theory is simply loss of information. Optically this would be a degraded image from the theoretically perfect image produced by the imaging device (lens). The lens distortions and image degradation being considered seperately.
Now in plain English here's my take on the AA filter. It fuzzies up the image so the "data" is spread out making it easier to determine what the pixel "should be" in the case of a coin toss where the image detail is uncertain because of the device geometry. Part of this is the expected quality of the image being projected. So if your optics are the sharpest and most distortion-free possible - then you have a higher confidence that the data the sensor reads is "correct". This is my guess why the Leica has no AA filter - the lenses are sharp and error free to the point that the resolved image is mostly correct - in their estimation anyway. I'm quite sure that close examination of the raw data would show errors - but statistically low enough that they can be filtered out in software.
Note to would-be flamers that I'm a Canon guy and have no Leica glass! Don't shoot the messenger.
regards,
Jon
Thanks guys.