Super Resolution?

[Hening Bettermann]

My former Nikon D200 had about 10 MP.  An exposure stack of 6 exposures in PhotoAcute is displayed as having about 35 MP! At 300 dpi, it would print about 60 cm on the long side without any uprezzing...? and 120 cm with a factor of 2? When I set resolution to 96 dpi, scale to 120 cm and view on screen at 100 %, the image looks indeed gorgeous. Dare I believe this as the base for planned printing in a print shop??  PhotoAcute does in fact call it super resolution... Too good to be true??

Good light! Hening.

[feppe]

Send them a representative A4-sized crop of the final image if you want to ensure it indeed looks good full size before splurging on the big print.

[Hening Bettermann]

Thank you. Yes, of course I will run a small test, also to determine the amount of print sharpening needed. The same would go for a "normal" image, I guess. I am of course aware of that the screen image is never the "same" as a print.  In other words: The 35 MP figure is a realistic figure akin to the 10 MP for the single slice, as a point of departure? Sounds like that.

[feppe]

In other words: The 35 MP figure is a realistic figure akin to the 10 MP for the single slice, as a point of departure? Sounds like that.

It very well might be. I have no experience with super-resolution, but the theory is sound AFAICT, and some form of it is used by Hasselblad in some of their backs.

[Hening Bettermann]

Unfortunately it turns out that the stacks suffer from artifacts, probably due to moving foliage. Whereas Helicon Focus renders such objects as unsharp, which looks much more "natural", PhotoAcute shows double contours. So this time no.

[feppe]

Unfortunately it turns out that the stacks suffer from artifacts, probably due to moving foliage. Whereas Helicon Focus renders such objects as unsharp, which looks much more "natural", PhotoAcute shows double contours. So this time no.

You could take one un-processed shot and mask the offending sections of the image to reveal the un-processed shot. Those areas will be at lower res, but I bet it's less offending than artifacts like you describe and possibly indistinguishable to anyone but you in print. Of course blow up the image to similar dimensions first.

[Hening Bettermann]

I don't quite see the procedure before me. Also, these offending areas are many, small, and scattered over the image.
For now, I have settled with just one slice. I think it will print OK at 300 dpi and 200%, that is 64 cm on the long side, that has to do.

Good light - Hening.

[Hening Bettermann]

Uff - that fool I am! I forgot that Photo Acute has an option "Take moving options from 1st image"! With that enabled, it works! The result will print 55x82 cm @ 240 dpi, that will be fine.

[milt]

I have done some experimenting with the "super-resolution" mode of PhotoAcute, and I think its a very useful technique.  For an old duffer like me, it lets me carry an SLR up into the mountains and yet produce a near-MF result.  In my experiments, I can see improved noise in the result up to about 15 shots.

Except for edge effects (where you lose just a few pixels) you should get exactly 4X pixels with super-resolution, so I'm not sure I understand your 10 -> 35 results.

Its ironic to think that this is one case where don't want a really super steady tripod/technique.  Its the slight camera jiggle that leads to capturing additional information.

--Milt--

[hjulenissen]

Except for edge effects (where you lose just a few pixels) you should get exactly 4X pixels with super-resolution, so I'm not sure I understand your 10 -> 35 results.

Why exactly 4x? If you had flawless lenses, no OLPF/AA-filter and each sensel was sensitive to a small percentage of its total area, then only the amount of images in the stack should limit uprezzing capabilities (i.e. infinite). In practice, all lenses have flaws and most sensors have OLPF.

Its ironic to think that this is one case where don't want a really super steady tripod/technique.  Its the slight camera jiggle that leads to capturing additional information.

I think that movement on the order of one sensel is going to happen almost "no matter what". So good technique is still a good thing to avoid camera shake during capture, or loss of side-pixels?

-h

[Hening Bettermann]

Except for edge effects (where you lose just a few pixels) you should get exactly 4X pixels with super-resolution, so I'm not sure I understand your 10 -> 35 results.

It was my sloppy math done in my head - the exact figure is indeed 40.144.896  pixels!

[Bart_van_der_Wolf]

Why exactly 4x?

Because that's how the super resolution feature is implemented in PhotoAcute. It doesn't necessarily mean that spatial resolution is also doubled, but it does improve with the correct (= undersampled with very slight displacements) images as input. On a regular 2x upsampling on would expect the resulting spatial resolution to reduce to 50%, with super resolution it usually is significantly higher.

Cheers,
Bart

[hjulenissen]

Because that's how the super resolution feature is implemented in PhotoAcute. It doesn't necessarily mean that spatial resolution is also doubled, but it does improve with the correct (= undersampled with very slight displacements) images as input.

Ah, I see. I misinterpreted the post. So photoacute is actually locking the output grid to 2x the number of input pixels in each dimension?

On a regular 2x upsampling on would expect the resulting spatial resolution to reduce to 50%, with super resolution it usually is significantly higher.

Not sure I follow you there. The rendition of details should ideally stay constant if you take an 8 megapixel image and scale it to 16 megapixels in e.g. Photoshop. '

Super resolution is a different operation where multiple, aliased images are combined to provide a single, higher resolution image. Ideally, N input pictures should allow for an increase in resolution by N for any N if offsets were carefully chosen/lucky and all kinds of practical limitations did not exist. If one can get an increase of N=2^2 for a number of input images equal to 4-8, I think one should be very pleased.

-h

[Bart_van_der_Wolf]

Ah, I see. I misinterpreted the post. So photoacute is actually locking the output grid to 2x the number of input pixels in each dimension?

That's correct.

Not sure I follow you there. The rendition of details should ideally stay constant if you take an 8 megapixel image and scale it to 16 megapixels in e.g. Photoshop. '

Not the per pixel resolution (doubling the pixel count decreases the per pixel resolution by a factor of sqr(2), resolution is measured in the linear dimension). In order to store the added (per pixel) resolution, more pixels have to be added. One can of course downsample back to the original pixel dimensions, but then in addition to boosting the per pixel resolution one risks the creation of aliasing artifacts. Going from oversampled to potentially undersampled requires low-pass filtering to prevent most of the trouble, but one might also lose real resolution.

Cheers,
Bart

[hjulenissen]

per pixel[/i] resolution (doubling the pixel count decreases the per pixel resolution by a factor of sqr(2), resolution is measured in the linear dimension).

I try to avoid the per pixel/per inch resolution way of thinking as much as possible. It seems to lend itself to a strange way of reasoning ("you need 150 dpi in order to print an image"). The interesting part to me is the total amount of information/resolution contained in an image, no matter how many megapixel, dpi, square inch etc it happens to be contained in.

In order to store the added (per pixel) resolution, more pixels have to be added. One can of course downsample back to the original pixel dimensions, but then in addition to boosting the per pixel resolution one risks the creation of aliasing artifacts. Going from oversampled to potentially undersampled requires low-pass filtering to prevent most of the trouble, but one might also lose real resolution.

Any change of sampling rate should be accompanied by a filter as per Nyquist. In imaging, filters tend to have short kernels, "bad" frequency specifications, but also rely a lot on non-linear phenomena.

If operations are carried out in the Bayer domain, one might use something similar to super-resolution to create an 8 megapixel image where all subpixels are full color without altering the native sampling grid. I believe that at least one MF company does this?

-h

[Hening Bettermann]

While we're talking about Super Resolution: I have always looked upon this as a way of improving the resolution of mediocre lenses. Would it make sense to shoot multiple frames with my Canon 5D2 and Zeiss Contax primes hoping to increase *their* resolution ??

[hjulenissen]

While we're talking about Super Resolution: I have always looked upon this as a way of improving the resolution of mediocre lenses. Would it make sense to shoot multiple frames with my Canon 5D2 and Zeiss Contax primes hoping to increase *their* resolution ??

Actually, I believe that SR is easier for good lenses than for poor lenses. If every sensor sensel had a PSF like an infinitely small point (no reduction of aliasing), doing super-resolution would be trivial - just align the aliased images and watch the amount of detail grow.

Real-world lenses, OLPF/AA, micro-lenses, camera/scene-movement etc blur the PSF, doing SR harder. Some say that you really cannot do SR in practice without doing lense deconvolution at the same time. I guess that is why Photo Acute ask you to calibrate their software with each and every lense.

-h

[Bart_van_der_Wolf]

While we're talking about Super Resolution: I have always looked upon this as a way of improving the resolution of mediocre lenses. Would it make sense to shoot multiple frames with my Canon 5D2 and Zeiss Contax primes hoping to increase *their* resolution ??

Hi Hening,

Yes, it would also help the Zeiss lenses.

Cheers,
Bart

[Hening Bettermann]

Tak, hjulenissen, and thanks, Bart, for your replies. This opens up a great perspective!