I am a bit surprised that Roger was a bit surprised that a change of camera only has a small effect on the optimal aperture, as if he has no faith in Imatest actually measuring lens properties. :-)
The effect of diffraction on MTF50 (or DOF) is just one aspect of image quality, and too shallow a view if you are interested in the overall picture. I agree with mpmendenhall that the drop in macro contrast is more harmful, and there are no sharpening techniques that will restore this contrast.
A spinning iris or otherwise time-varying aperture is not a substitute for a soft-edged aperture, because the mean diffraction pattern of a varying aperture is not the same as the diffraction pattern of the mean aperture. Furthermore, there is no free lunch. Softening the edges of an aperture improves macro contrast, but reduces micro contrast. (However, for instance, f/11 with tapering may yield the same micro contrast as f/22 without tapering, and a better macro contrast.)
The importance of the aperture shape should not be exaggerated. Sure, the effect on a bright point source is clearly different, but the effect on an image with extended elements is not. Here is an MTF-comparison between a round and a square aperture at F/22, where the f-number is defined in the usual way for the round aperture, and where the square aperture has the same area as the round aperture. The square-aperture MTF is measured in the direction of the diffraction spikes that would occur with a point source.
Toothwalker wrote:
I am a bit surprised that Roger was a bit surprised that a change of camera only has a small effect on the optimal aperture, as if he has no faith in Imatest actually measuring lens properties. :-)
Paul,
Given my ongoing discovery of the limitations of both Imatest and Roger's knowledge, my faith has become, at least, lessened
Lessened faith seems to increase my ability to be surprised - I suppose there's some inverse relationship between the two.
RCicala wrote:
Given my ongoing discovery of the limitations of both Imatest and Roger's knowledge, my faith has become, at least, lessened
Lessened faith seems to increase my ability to be surprised - I suppose there's some inverse relationship between the two.
I expect (well, hope really) that once the sharpness crack addiction winds down a little, people come to realise the role of low spatial frequencies in image quality. Erwin Puts and Mike Johnston emphasise this facet, and Puts always points out the drop-off in 5 lp/mm data as a sign of diffraction's early affects on Leica lenses.
Toothwalker wrote:
. Furthermore, there is no free lunch. Softening the edges of an aperture improves macro contrast, but reduces micro contrast. (However, for instance, f/11 with tapering may yield the same micro contrast as f/22 without tapering, and a better macro contrast.)
If we use a window function depressing all sidelobes much better that a rectangular window, it is practically impossible to get a widening of the main lobe any less than about 2x as you suggest. However I am not sure myself if it is the very high first and second sidelobes that is the main problem, or if it is that the sidelobes decrease slowly even after that. If it is the higher order sidelobes that are most harmful to macro contrast, we could use a window function optimized for narrow main lobe and high suppression of higher order sidelobes, like a Tukey window.
Anyway, f/11 with tapering would also increase the DOF compared to straight cut f/11, and since we often stop down for DOF, this balances the trade-off again.
For the purpose of using diffraction as the only AA filter in a system, a tapered aperture would certainly be a preferrable solution.
alundeb wrote:
If we use a window function depressing all sidelobes much better that a rectangular window, it is practically impossible to get a widening of the main lobe any less than about 2x as you suggest.
This depends of course on how you define "much better". A Hann window would serve my purposes, but anyhow my example was hand-waving and not the result of a detailed investigation.
However I am not sure myself if it is the very high first and second sidelobes that is the main problem, or if it is that the sidelobes decrease slowly even after that. If it is the higher order sidelobes that are most harmful to macro contrast, we could use a window function optimized for narrow main lobe and high suppression of higher order sidelobes, like a Tukey window.
Any reduction of sidelobes is welcome, but since the sidelobes of an untapered aperture decay very slowly I am inclined to say that it is particularly important to suppress the higher-order sidelobes.
Anyway, f/11 with tapering would also increase the DOF compared to straight cut f/11, and since we often stop down for DOF, this balances the trade-off again.
For the purpose of using diffraction as the only AA filter in a system, a tapered aperture would certainly be a preferrable solution.
Agreed. The technical solution to implement the optimal (in some sense) tapering is not obvious for a variable-aperture lens, but manufacturers reading this thread have a clear business case.
alundeb wrote:
For the purpose of using diffraction as the only AA filter in a system, a tapered aperture would certainly be a preferrable solution.
Toothwalker wrote:
Agreed. The technical solution to implement the optimal (in some sense) tapering is not obvious for a variable-aperture lens, but manufacturers reading this thread have a clear business case.
What are the odds
Brian Caldwell had som recent cooperation with a somehow innovative manufacturer recently, maybe he could pick up on this.
Thinking about it, a starting point could be a fixed cosine tapering on the iris blades. This would form a Tukey window with the alpha coefficient varying with the aperture. For smaller apertures, it would approach a Hann window, and for larger apertures it would approach a rectangular window. This behaviour would act positively if the goal is to maximize macro contrast at the smallest apertures.
If the goal is to reduce the variation in the main lobe width with the aperture (constant AA filter approximation), it would be a bad thing though.
SoulNibbler wrote:
This long thread demonstrated to me that it was possible to recover quite a bit of resolution from diffraction smeared images using careful sharpening.
It should be noted however that you need to have a clean image as noise should be indistinguishable from detail.
Hope I am still on topic. I recall a talented mathematically inclined alt FMer convoluted a test pattern with diffraction PSF and much data was lost and not recoverable--a bit contrary to the Roger post and LL link. Post #66 in the LL link convoluted an actual image with single diffraction at F32 and was able to deconvolute with quite good success. Of course he had the perfect PSF but I thought MUCH more resolution data would be lost especailly compared to a Gaussian blur. I suppose non circular apertures etc would
also foul things up but makes me wonder how effective blind L-R deconvolution is with diffraction? I remain confused by the contradictory results.
Deconvolution sharpening can (with non-determinate PSF's) only increase detail CONTRAST, you can't increase the maximum detail frequency. At least not in any practical circumstance in the field of normal photography.
Unfortunately the concept of resolution has to be very tightly coupled to "contrast", and often you use MTF50 contrast to get a "resolution" number. This can be confusing, since the point of real detail extinction is a lot higher than that. What deconvolution sharpening (with non-determinate PSF's) can do is to restore detail from low contrast levels back into their object-related contrast levels in a reasonably correct and accurate way.
So if diffraction at a certain aperture value sets a total detail extinction level of "X" MP on a certain format (Rayleigh works well, as long as you have low noise energy in the image), you can use blind R-L to get good detail contrast all the way UP to that maximum frequency/resolution - you can't INCREASE the maximum resolution. Both noise concerns and the Bayer interpolation error put a stop to that, each in their own way.
One reason as to why you often get very good "practical" results from deconvolution sharpening with camera-generated images is that ACCURATE detail frequency in a Bayer-interpolated image never exceeds ~1.5 pixel widths. So when you deconvolve a slightly diffraction-limited image (F16 on a modern 20MP camera) you get quite close to the optimal detail retention even in a theoretically perfect case.
Just as an illustration, here is the modulation transfer versus spatial frequency for a round aperture without apodization, and a round aperture with a Hann intensity falloff. The f-number of the latter is chosen such that the MTF50 frequencies coincide. (I hope I did not make a mistake in the computation.)
It is a classical case of high resolution versus high contrast. The blue scenario reveals more detail to pixel peepers, the red scenario yields an image that looks better from a normal viewing distance.
It is not immediately clear to me how to scale the f-numbers such that the scenarios yield equal depth of field, which is a concept based on a subjective criterion.
One last remark before this thread dies. I tried a few other apodization functions (Hamming, Gaussian, Lanczos, Triangular) and the resulting MTF curves are amazingly similar to the Hann curve when the f-numbers are chosen such that the MTF50 frequencies coincide.
Toothwalker wrote:
One last remark before this thread dies. I tried a few other apodization functions (Hamming, Gaussian, Lanczos, Triangular) and the resulting MTF curves are amazingly similar to the Hann curve when the f-numbers are chosen such that the MTF50 frequencies coincide.
I'm glad this thread got bumped today - I missed the extended discussion as I was on travel. Thanks everyone for the further enlightening discussion.
As to the above, I think this is an expected result. Remember, MTF curves are always shown linear and those of us that obsess about window functions look at everything in dB usually. Once you move away from rectangular to almost any other window function that attempts to suppress side-lobes you have a similar affect on the main lobe. The differences between the various non-rectangular window functions is primarily in the exchange between close and distant side-lobes - but those side-lobes are very low in a linear scale regardless of the function chosen. So I'd suspect that in a linear MTF we'd see all practical non-rectangular window functions looking about the same.
I would also suspect, that as a practical matter, no real lens has macro-contrast at all affected by diffraction side lobes or anything else to do with the aperture. Baffles, optical coatings, etc. etc. are likely a big part of the game in very low spatial frequency contrast.
Oh, one last thing. I came across a link to this article in another forum recently. With regards to what we perceive as "sharp" and the effects of diffraction on MTF it is very interesting. It removes the question of aberrations by only dealing with pinholes but is still a good read. It is a short read despite the page count - giant text and lots of figures.
Just keep in mind that the real implications of diffraction have a totally different effect on passive pinhole-cameras (as in the link) and for lenses with active optical elements.
In a pinhole-camera you have to balance diffraction against the physical size of the aperture (hole-in-the-box) to get maximum resolution, in a very linear way. This is not true for normal lenses.
A normal 85/1.8 at F4.0 as modeled by a pinhole would have an MTF50 point at about ~0.03lp/mm. That's about six pixels per 36x24mm sensor. Not six MegaPixels, six PIXELS. Singular digit. With active optical elements included, we can take that up to 80-90lp/mm, or maybe a few hundred Megapixels in the center of a 36x24mm image..
That's a difference of about 1:3000 in linear resolution, and that's about how relevant the pinhole model is to an active lens model at those image scales. There's principal function difference modifying the diffraction impact on the image by a factor of about 3000 at F4.0.
@theSuede: Yes of course! I thought the link was interesting for its illustrations of what different PSF/MTF result in visually. The point being the higher resolution MTF looks worse than the one with higher contrast at lower spatial frequencies. There are nice illustrations of this in other hard copy resources like Strobel that use lenses but those are available on the web.
