The DOF vs Diffraction challenge: FF or crop?

[fike]

Two topics: focus stacking and tilt shift to increase DOF in panos.

FOCUS STACKING
In landscape panoramic photography, I find that focus stacking is very difficult to do well. Typically you are oustide and there is some air movement. Leaves move.  That's it.  Any movement is too much.  It is the same attribute that tends to make HDR problematic with woodland or wilderness photography. If your subject are hoodoos in the southwest, it may work because they don't move, ever.  With all the ways to mess up a panoramic (each photo is another opportunity for movement, lens flare, or an earthquake) I have given up on panoramic focus stacking--too many variables.

TILT SHIFT PANOS
I experimented rather extensively with the new 24mm TS-E II.  That is a fantastically sharp lens.  It can do flat stitches, though there is no gain in DOF with a flat stitch.  You can do some small amount of tilt in the first row of a multi-row pano and then stitch it all together, but it is tricky and error prone.  Stitching a pano containing images with different tilts tends to create problems because the shape of the projection changes.  I was unable to get reliable results...though sometimes it was stunning. 

The most reliable method I managed to create with a TS is the least obvious.  It involved mounting the camera in the portrait orientation on a pano head and tilting the TS-E lens sideways fairly aggressively. Effectively what I did was make an in focus plane running vertically from the back of the image area to very close to the camera.  This plane runs roughly perpendicular to the camera, but actually it angles away slightly to the left or right.  The in focus region would be about the middle fourth of the frame.  This wasn't really a problem because I was stitching. Now here is where it is annoyingly time-consuming, instead of making a new image every 20 degrees, or so, I made a new image every 4-5 degrees, and then in stitching I only used that narrow-in-focus slice of the image.  This was complicated, but the results were quite good.  A sample can be seen here: http://www.flickr.com/photos/trailpixie/5124820432/in/photostream/lightbox/

[deejjjaaaa]

Diffraction depends only on aperture, there are no lenses more tolerant of diffraction. It's a property of light. The form of the aperture may matter, but I presume that all lenses we discuss have circular or near circular apertures.

sure, however it is a fact that different lenses deliver quite different results (same focal length, same stopped down to big F numbers aperture) on the same body... so in some cases when you have a subpar lens you get much worse results that can be explained by diffraction alone.

[deejjjaaaa]

Two topics: focus stacking and tilt shift to increase DOF in panos.

you forgot superresolution to fight w/ loss of resultion (while increasing DOF by stopping down further... albeit it might allow you to stop just 1-1.5 stops further down... not much, but still) from diffraction using programs like from http://photoacute.com (it can do focus stacking as well)

[hjulenissen]

you forgot superresolution to fight w/ loss of resultion (while increasing DOF by stopping down further... albeit it might allow you to stop just 1-1.5 stops further down... not much, but still) from diffraction using programs like from http://photoacute.com (it can do focus stacking as well)

But superresolution will mainly combat limited sensor resolution, not severly limited optical resolution due to diffraction. If diffraction is the main factor limiting the system resolution, I would not expect SR to fix it.

-h

[deejjjaaaa]

But superresolution will mainly combat limited sensor resolution, not severly limited optical resolution due to diffraction. If diffraction is the main factor limiting the system resolution, I would not expect SR to fix it.

-h

why again we are talking about "severely limited optical resolution" ? I wrote that it might help you to stop down a 1-1.5 stops more... not 10 stops... if you still can get a good resolution @ certain aperture F(n), then superresolution might help you @ F(n+1 stop)... not a magical bullet but less work than focus stacking...

[Wayne Fox]

The fact that you act surprised suggests that you don't see a difference between an actual image taken at e.g. f/8 and f/22. I'm puzzled by that.

I think it is image dependent ... depends on the micro detail and how critical it is to the overall image.  Gaining sharpness in important detail but perhaps losing some of the ultra fine detail can still result in a very good image. Certainly you can't completely overcome the loss of sharpness and restores all the detail once you start stopping down.

So curious what others think, which image would be sharper and have better detail printed at 40x60, one taken with a good lens at f/22 on a 60-80mp sensor (no AA filter) or a dSLR taken at an optimum of f/8 (with no diffraction but with blurring from an AA filter).

Focus stacking is great but can't always be used (things moving, things in front of other things can leave fringe or halo of "softness" since the background around the edge of front objects can't be captured sharp).  Scheimpflug is great, but can't always be applied well (if at all).  Sometimes your only hope of any depth of field is just stop the thing down and take what you can get, and with a high resolution back what you get is usually really good despite diffraction.

[ErikKaffehr]

Hi,

I essentially tried this with one of the images from the great 2006 MFDB shootout. There was an aperture series with a P45+ back. I "reproduced" the subject (a one dollar bill) and shot it with my 24.5 MP DSLR. Large prints were made from both. At f/8 a blind man could see that P45+ was wastly superior. With the lens on the P45+ stopped down to f/22 the difference was essentially gone.

Best regards
Erik

I think it is image dependent ... depends on the micro detail and how critical it is to the overall image.  Gaining sharpness in important detail but perhaps losing some of the ultra fine detail can still result in a very good image. Certainly you can't completely overcome the loss of sharpness and restores all the detail once you start stopping down.

So curious what others think, which image would be sharper and have better detail printed at 40x60, one taken with a good lens at f/22 on a 60-80mp sensor (no AA filter) or a dSLR taken at an optimum of f/8 (with no diffraction but with blurring from an AA filter).

Focus stacking is great but can't always be used (things moving, things in front of other things can leave fringe or halo of "softness" since the background around the edge of front objects can't be captured sharp).  Scheimpflug is great, but can't always be applied well (if at all).  Sometimes your only hope of any depth of field is just stop the thing down and take what you can get, and with a high resolution back what you get is usually really good despite diffraction.

[torger]

For landscape when you strive after "sharpness in whole frame" (maximum DOF), sensor size actually has no meaning. It is about resolution.

If your sensor is small, you can use shorter focal length for the same view (larger DOF), but if your resolution on the sensor is the same as on a larger sensor you'll have to use larger aperture to minimize diffraction, and it exactly cancels out what you gain on focal length. In other words, 20 megapixel on a small sensor and 20 megapixel on a large sensor have the same ability to produce large DOF - the larger diffraction problem on the small sensor is compensated with the ability to use shorter focal lengths. The two properties cancels out precisely so maximum DOF ability becomes exactly the same regardless of format. That smaller formats may appear to have larger DOF is only because they usually have lower resolution and thus accept larger blurs, but if resolution is the same (and the lens can resolve it) there is no difference.

However, in practice a larger sensor is preferable, due to better noise properties (can gather more light) and that current lenses can produce sharper results. Larger sensors requires larger image circles from the lenses, and it is harder to make a lens that resolves say 5 microns on a medium format image cirle that 5 microns on for example APS-C. Actually the sharpest lenses in terms of microns on the sensor are in cheap compact cameras. However, unlike diffraction vs aperture this is not a you-gain-as-much-as-you-lose game, with increasing lens image circle size you gain more in total pixel count than you lose in individual pixel size, so while it is possible to resolve say 100 megapixel with the best medium format lenses today, you can only do say 20 megapixel with the best APS-C lenses.

Another aspect is that today the full frame 35mm (and even APS-C) produce as high resolution that you need to start thinking more like a large format photographer. In the film days and early digital, tilt lenses were rarely used to optimize DOF on 35mm, since the resolution of the format was so low that you did not gain much from doing so. With a modern digital 35 mm camera resolution is comparable to larger film formats, and thus you can indeed make use of tilt in landscape photography to optimize DOF, just as large format view camera photographers always have. In the Canon system, the TS-E 24mm version II is a fantastic landscape lens. Of course a view camera is more flexible in terms of tilt, but the price-performance equation and the flexibility to use DSLRs for other things than still life photography makes them the most practical alternative to most of us.

Also note that we will probably see further developments in deconvolution technique in the future that will to some extent be able to restore blur caused by diffraction and lens resolution limitations.

[erpman]

Hi,

I essentially tried this with one of the images from the great 2006 MFDB shootout. There was an aperture series with a P45+ back. I "reproduced" the subject (a one dollar bill) and shot it with my 24.5 MP DSLR. Large prints were made from both. At f/8 a blind man could see that P45+ was wastly superior. With the lens on the P45+ stopped down to f/22 the difference was essentially gone.

Best regards
Erik

Hey, Erik, that would be interesting to see, can you post the results for us?

BTW, comparing f/8 and f/22 is a bit unfair to the p45, since f/8 won´t give sufficient dof on the FF anyway. How was the difference between f/16 on the FF and f/22 on the p45?

[Bart_van_der_Wolf]

So curious what others think, which image would be sharper and have better detail printed at 40x60, one taken with a good lens at f/22 on a 60-80mp sensor (no AA filter) or a dSLR taken at an optimum of f/8 (with no diffraction but with blurring from an AA filter).

Hi Wayne,

Let's assume both cameras can resolve detail up to the Nyquist frequency. In my experience many cameras can have more than 10% luminance response at Nyquist, even if an OLPF (AA-filter) is used and despite demosaicing.

The Nyquist frequency of a Canon EOS-1Ds Mark III (6.4 micron sensel pitch) is 78.1 cycles/mm.
The Nyquist frequency of a PhaseOne IQ180 digital back (5.2 micron sensel pitch) is 96.2 cycles/mm.

Diffraction puts an absolute (unrecoverable by deconvolution sharpening) limit (MTF=0%) on resolution at 81.9 cycles/mm for f/22, and at 225.2 cycles/mm for f/8, when we look at 555 nanometer wavelength. It is clear that the IQ180's resolution suffers much more from the f/22 aperture than the 1Ds3 resolution does from f/8. Here is a graph showing the diffraction limited MTF:

If we assume perfect (only diffraction limited) optics on the IQ180, and want a minimum MTF response of 10%, we get a maximum resolution at f/22 of 66 cycles/mm. We will assume a maximum resolution of the 1Ds3, not limited by diffraction at f/8, at it's Nyquist frequency of 78.1 cycles/mm.

What remains is the output magnification of the on-sensor resolution. The IQ180 sensor (40.4 mm in the vertical dimension) compared to the 24mm of the 1Ds3, gives a factor of 1.68 benefit to the larger sensor which needs less magnification. That would scale the 1Ds3 maximum resolution to a 46.4 cy/mm relative to the IQ180's for output.

If our goal would be 5 lp/mm output resolution (~ 254 PPI), then the 1Ds3 could stand magnification by 9.28x, thus 223 mm for the short dimension, and the IQ180 could stand 13.2x magnification, thus 533mm for the short dimension while maintaining equal resolution. Of course this is when both systems use quality optics, perfect focus, and a rock steady shooting setup. The defocused parts of the DOF zones would make the MTF suffer gradually more at narrow apertures, but the exact effect depends on the specific MTFs (lens aberration and AA-filter effects).  

Cheers,
Bart

[ErikKaffehr]

Hi!

Check Bart's explanation above, a very good one.

I'm on vacation now, but will try to post some images from the comparison.

The "test" I have done was described here, at the end of the article: http://echophoto.dnsalias.net/ekr/index.php/photoarticles/24-how-many-megapixels-do-we-need

Regarding the difference between f/8 and f/22 on the P45 it was huge at actual pixels.

I have done some studies on DoF and diffraction here: http://echophoto.dnsalias.net/ekr/index.php/photoarticles/49-dof-in-digital-pictures?start=1

Regarding comparing f/22 on P45 and f/8 on full frame it was not really about being fair, more about seeing how much of the MF advantage would be lost by indiscriminate stopping down. I'm not in doubt about the advantages of MFD just that I want to understand why they are at advantage.

Best regards
Erik

Hey, Erik, that would be interesting to see, can you post the results for us?

BTW, comparing f/8 and f/22 is a bit unfair to the p45, since f/8 won´t give sufficient dof on the FF anyway. How was the difference between f/16 on the FF and f/22 on the p45?

[aman74]

Subscribed.

[erpman]

Of course, the difference is huge, I do agree.

BTW, the comparison that I posted had f/11 for the 5d and f/18 for the 645d

I looked into photoacute, very interesting program. Maybe a way to go could be to use a 645d, wider lens (35mm), and wider apertures. Then use superresolution to do part of the upscaling, and avoid stitching altogether, making the whole shooting ten times easier. Then I guess you could blow it up to gigantic proportions with fairly good quality, and a lot less hassle when shooting.

[torger]

A comment on the test when combining CoC and diffraction blur. Adding blurs works in the way that if you add a small blur to a large blur the total blur becomes a bit larger.

That is, if CoC is 4 um and diffraction is 5 um the result will be less sharp than if CoC is 0 um and diffraction 5 um.

In other words if you improve something in the chain of blur factors that is not the worst factor you may still see an improvement.

So the assumption that as long as diffraction is less than CoC it does not have any effect on sharpness is actually incorrect. However, at such a small scale and when working with DOF that assumption is perfectly ok concerning the lack of preciseness in the general photographic workflow (one rarely measure distances to objects in millimeters etc).

[Bart_van_der_Wolf]

In other words if you improve something in the chain of blur factors that is not the worst factor you may still see an improvement.

Indeed, however the effect on image quality can be complex and hard to predict. Diffraction blur has quite a different effect on resolution loss than defocus (COC) has. In an attempt to demonstrate that, I've prepared 2 graphs that show the combined effect of diffraction and defocus on the MTF.

First the situation with f/8 diffraction (10.8µm diameter), and different amounts of defocus blur added:

As expected, adding more defocus blur (a larger COC limit) to the diffraction blur will further reduce the limiting resolution progressively. The overall MTF is also reduced, resulting in a lower contrast image also for the spatial frequencies that are still resolved. In general we can conclude that at wider apertures image quality will suffer significantly from defocus blur. That is nice for portraits, but very unwelcome for deep DOF landscapes. There is a significant difference as we transition from the plane of best focus to the COC zone limit, and the quality at the plane of best focus is still high, compared to the theoretical 5.2 micron sensel pitch sensor potential (Nyquist frequency is 96.2 cy/mm).

Next the situation with f/22 diffraction (29.8µm diameter), and the same amounts of defocus blur as above added:

Perhaps somewhat less expected, the added defocus has hardly any effect on the limiting resolution, but the medium spatial frequencies (for which human vision is most sensitive) still suffer a bit. Diffraction is clearly the limiting factor, and the blur is also seriously affecting the plane of best focus. There is relatively little difference between the COC zone limit and the plane of best focus.

An added consideration in the trade-off between the effects at different apertures is that diffraction blur potentially responds better to deconvolution sharpening than defocus blur (especially for smaller sensel pitch sensors), but that deconvolution will have no effect when the MTF becomes zero, there is no signal to deconvolve. It will help to boost the remaining medium spatial frequencies though, so there is still some benefit in using it.

Cheers,
Bart

[hjulenissen]

Diffraction puts an absolute (unrecoverable by deconvolution sharpening) limit (MTF=0%) on resolution at 81.9 cycles/mm for f/22, and at 225.2 cycles/mm for f/8, when we look at 555 nanometer wavelength.

Is there such a thing as a brick-wall diffraction limit above which mft == 0? Or are you talking about the limit above which you have experienced practical deconvolution algorithms to fail due to falling SNR or whatever?

-h

[Bart_van_der_Wolf]

Is there such a thing as a brick-wall diffraction limit above which mft == 0? Or are you talking about the limit above which you have experienced practical deconvolution algorithms to fail due to falling SNR or whatever?

It's not really a sudden diffraction brick wall, but rather a gradual spatial frequency slope of reduced contrast that ends in zero modulation. Even with only 10% modulation response, it would render a subject contrast of 10:1 to a barely perceptible 1% response. Raw converters performance becomes very important. Diffraction does set an upper limit to what can be resolved/restored, as does defocus.

Cheers,
Bart

[hjulenissen]

It's not really a sudden diffraction brick wall, but rather a gradual spatial frequency slope of reduced contrast that ends in zero modulation. Even with only 10% modulation response, it would render a subject contrast of 10:1 to a barely perceptible 1% response. Raw converters performance becomes very important. Diffraction does set an upper limit to what can be resolved/restored, as does defocus.

Cheers,
Bart

So what you are saying is that in your experience, deconvolution algoritms are unable to recover 1% contrast into something meaningful, not that the contrast is identical to zero.

The reason that I am asking, is that there have been lots of discussions about the "diffraction limit" at dpreview. A lot of people seem to think that it is a theoretically perfect brick-wall, but I am sceptical as theoretically perfect brick-walls are very seldomly seen in nature. Nature seems to dislike 100000 tap sin(x)/x filters, rather going for low-order ones.

-h

[ErikKaffehr]

Hi,

My understanding is that diffraction is not a disc (like defocus for an ideal thin lens) but more like a "bell curve". For peak shapes similar to "bell curves" most often FWHM (Full With Half Maximum) is used, but the effect of diffraction will be broader than FWHM. But some detail may also be resolved within the FWHM diameter as we still have some gradient.

My article here: http://echophoto.dnsalias.net/ekr/index.php/photoarticles/49-dof-in-digital-pictures?start=1 demonstrates this with real world samples. Diffraction is red circles and defocus is green circles. For diffraction the conventional value is used. FWHM would be somewhat smaller.

When looking at the above article keep in mind that diffraction is constant for each row. Defocus is increasing from left to right.

Last page of the article shows examples of sharpening using "basic" deconvulution using Smart Sharpen in CS5 and Topaz inFocus.

Best regards
Erik

So what you are saying is that in your experience, deconvolution algoritms are unable to recover 1% contrast into something meaningful, not that the contrast is identical to zero.

The reason that I am asking, is that there have been lots of discussions about the "diffraction limit" at dpreview. A lot of people seem to think that it is a theoretically perfect brick-wall, but I am sceptical as theoretically perfect brick-walls are very seldomly seen in nature. Nature seems to dislike 100000 tap sin(x)/x filters, rather going for low-order ones.

-h

[torger]

Diffraction is often described as a hard limit, which is wrong. It is a soft onset, a kind of a bell curve filter with some ringing (varies with wave length too), and it can be reversed with deconvolution, theoretically all of it. In practice there'll be noise and there'll be too low resolution and too many unknown parameters (such as color filters in the sensor, color of incoming light etc) to make a perfect result. A deconvolved image has typically a bit grainy look when pixel-peeped, but can make very good results when printed.

As with nearly all aspects of image quality, the less you need to fix up in post-processing the better the end result will be, so if you can avoid a large diffraction offset that is better. And of course, diffraction is only relevant in high resolution photography, still at dpreview one can see photographers that shoot hand-held worry about it ;-).