P/S Camera challenging 4x5" ......

[Christoph C. Feldhaim]

Okay, okay, okay ...... we all know a P/S camera is no match for a 4x5" concerning IQ.

But:
How far could one boost IQ by stitching and multi exposure for stills ?
I'm pondering about doing an experiment with architecture, but wanted to ask
about the opinion of the more experienced and technical people here.

My thoughts on the experiment:
--------------------------------------
If I'd do some mad stitching, exposure bracketing, maybe even focus bracketing -
how far would I have to go to match a 4x5" film camera?

If I want huge pixels, lets say 12 Micron, I'd need 6x6 Images, since my G11 has about
2 Microns pixel size. I could stop down to f 8 (the cameras max f-stop) safely without diffraction
penalty assuming I'd downsize later.

People say 39 Megapixels is about a 4x5" drum scan.
So - I'd have to make a 39 Megapixel Image.

The G11 produces 10 MP. I need 36 Images to get a huge pixel pitch for stopping down.
This done 4 times would equal 40 MP.
So I'd have to shoot 144 frames.

If I want more DR I'd exposure bracket from 2 images 2-4 stops apart.
An HDR Merge would also decrease noise.
So I'm at 288 double exposed frames with a 12x12 stitch layout, downsized to 40 MP later.

Adding I need overlap of maybe 33%, the 288 frames grow to a 2*18*18=648 frames .....


Maybe someone has already done this?

What do you think?

Any suggestions for a change of experiment layout?


Cheers
Chris

[jjj]

Maybe someone has already done this?

Yup.
Here's where to go to buy equipment to do just as you wanted. The original version was for compact cameras.
Gigapan

[Christoph C. Feldhaim]

So - how far would this monster machine help to compensate for DR ? (I believe bracketing at a location would solve it).

... and for lack of lens optical quality? As far as I understand the stitching would allow for very low magnification (I'm interested in raising image quality for prints, not producing Gigapans), so if I'm not mistaken, the method would exploit the best part of the MTF curve.
Right ?

And how far could the recording of color get improved by multi exposures and HDR merge?

And what about noise reduction?

After all this system together with a P/S Camera seems to be much less bulk than a MF or LF Camera,
especially when combined with a lightweight carbon fiber tripod.



(I myself wouldn't buy such a monster, since I seriously believe it was constructed by some demon from hell and thus it will mess with your mind and drive you insane after some time ... however ...  )

[NikoJorj]

Interesting theoretical question...  

First, I just don't get how you make "12 microns pixels" by combining images taken with a 2 micron pixel sensor. My take on this : as pixel size is all about DR (ie nise), the HDR bracketing should already take care of that.
Ditto for focus stacking : unless you're doing macros, depth of field of a 2µm-pitch sensor is generally enough. Moreover, I'd think that if you stitch many frames, the ability to change focus between frames gives a bit of a poorman's tilt (but I didn't test it, you'd need a good stitcher).

For resolution, yes stitching is a solution (unless anything moves in the frame). To equate 39MP with 10MP frames, given that you'll need a bit of overlap, 2x3 frames may be more conservative that simply 2x2.
With 3 shots 3 EV apart to get a huge #12EV photographic DR, you're about to do 6*3=18 frames to equate one 39MP.

[Christoph C. Feldhaim]

.....First, I just don't get how you make "12 microns pixels" by combining images taken with a 2 micron pixel sensor. My take on this : as pixel size is all about DR (ie nise), the HDR bracketing should already take care of that. .....

The idea behind that was, that if I downscale the image in the end by combining 6*6=36 images into one
I'd have the information of a 12 square micron area in one pixel in the final image,
since 12 by 2 microns is that factor of 6.
This should equal out some noise and fix the diffraction penalty.
This also assumes, that I have a chance to zoom in enough.

[NikoJorj]

This should equal out some noise and fix the diffraction penalty.

I still don't fully get the thing...

Are you talking about an averaging (stacking) of the 36 images?
In this case, you cure the noise by that sqr(36)=6 factor, but that wouldn't do anything to diffraction (which, due to the small sensor size and its big depth of field, is no problem here?).
Moreover I'd say that bracketing and HDR blending (see Zero Noise) is much more efficient to remove noise and increase DR.

[Christoph C. Feldhaim]

Seems I wasn't exact enough.

If I take a wide angle image with the G11 i have 2 square Microns per pixel.
If I zoom in by factor 6 and take 36 images (a 6x6 grid) and afterwards downscale it by factor 6. I'd have the information of 12 square Microns per pixel.
The zooming and using the telephoto is the trick to use more sensor area as if using the wide angle and shoot only one image or stack HDR with the wide angle.
If I'd stop down to f8 I'd not see the blurriness caused by diffraction as I would see when using the wide angle for the same FOV.
The reduction after first zooming in would allow for higher f-stops.

[NikoJorj]

Ok, now I get it ('tchuldigung, ich bin aber blond).

This kind of massive oversampling should do the same to noise as stacking, I'd say... But it's only a guess.
And for diffraction, I'd say a tad more surely that it doesn't do anything : what you gain in ability to stop down, you lose in depth of field because of the longer focal length you use. Netto result = not that much.

[BobFisher]

I'm also not entirely sure I follow how you get 12 microns out of a 2 micron pixel but either way the math isn't quite right, I don't believe.  You need to quadruple to get the right math.  It's the square area, not the linear area you're looking at.

Consider it this way.  How many 4x5 slides can you fit in an 8x10 slide?  2 or 4?  4.

[Christoph C. Feldhaim]

I'm also not entirely sure I follow how you get 12 microns out of a 2 micron pixel but either way the math isn't quite right, I don't believe.  You need to quadruple to get the right math.  It's the square area, not the linear area you're looking at.

Consider it this way.  How many 4x5 slides can you fit in an 8x10 slide?  2 or 4?  4.

2 microns is the G11.
12 micron was an arbitrary value I chose to be surely far away from the diffraction penalty at f8.

12/2=6. 6x6 = 36.

So - I considered 6*6=36 frames combined into one 10 MP image to be equivalent as if my G11 had a pixel pitch of 12 Micron instead of 2 Micron with still 10 Megapixels.

So - I'd shrink a stitch of 6*6 frames (Okay - overlapping would add) into one frame of one sixth the size.

So - for one 10 Megapixel frame I'd need (assuming 1/3 overlap) 9*9=81 frames, which get stitched together into one frame 6 times bigger than one original frame and that shrunk by 6 so, that it had the size of an original frame = 10 Megapixels.

This 4 times would give me the 40 Megapixel image.
81*4 = 324 frames. Added HDR*2 = 648 frames.

[BobFisher]

You're still using a linear equation and it shouldn't be.

If the film example didn't do it, try a more modern one.

If you've got a 10MP sensor, what's the pixel count need to double the resolution?

[Christoph C. Feldhaim]

You're still using a linear equation and it shouldn't be.

If the film example didn't do it, try a more modern one.

If you've got a 10MP sensor, what's the pixel count need to double the resolution?

Sorry, I just don't understand where you see the error.

40  MP with a 10 MP camera = 2x2 frames, with overlap = 3*3=9 frames.

Each of these frames, made from a downsized 6*6 "superframe"  (=36 frames to combine) for better detail (the better part of the MTF curve), downsizing of the Airy disks (2 Micron=> 12 Micron - the diffraction issue) should solve some IQ issues.
A 6x6 frame with overlap becomes a 9x9 frame (=81 frames to combine)  

So - the edges of that stitch are 2*6 = 12. 12*12 frames, with overlap 18*18 = 324 frames.


So its 324 frames = an 18x18 stitch where at f 1:8 no diffraction should be visible.
Together with a HDR Merge for better DR I'd be at 2 superframes of downsized 18*18 frames to get a 40 MPixel image.

Thats how I come to 324*2 frames = 648 frames to shoot with a "virtual" 12 Micron photosite.

[BobFisher]

OK, maybe I'm wrong.  Either way, I give up.

[EricV]

I think the basic premise of basing this calculation on pixel size comparisons is wrong.  How about considering it from an optical resolution or image blur perspective instead?  After all, the only purpose served by pixels is to adequately sample the intrinsic resolution of the optics.

Setup #1) 4x5 camera (100mm x 125mm sensor) with 180mm lens at f/16.  Assuming the lens is diffraction limited, the point spread blur diameter at the sensor is 20um.  The sensor then captures (100K)(125K)/(20)(20) = 30 million resolution "spots".  You assume this setup has 12um pixels, which samples the 20um blur function moderately well, but apart from that observation, pixel size does not enter the calculation.

Setup #2) Canon G11 camera (5.7mm x 7.6mm sensor) with 180mm lens at f/16.  Yes, that's right, lets use exactly the same lens, just switch sensors.  Then to capture the same image, with the same composition and the same number of resolution "spots", you need the same effective sensor area, which requires (100)(125)/(5.7)(7.6) = 288 frames.  This setup will clearly give equivalent image quality, since every important variable is the same.  Of course the image in this case will have vastly more pixels, since each pixel is smaller, but that is irrelevant to the calculation.  You could downsample quite a bit to reduce the pixel count, without any loss of image quality, since sampling a 20um blur function with 2um pixels is overkill.

Setup #3) Canon G11 camera (5.7mm x 7.6mm sensor) with 90mm lens at f/8.  This time let's try something different -- use a shorter lens at a wider aperture.  To get the same image composition with a lens of 1/2 the focal length, you now need to capture only 1/4 the image area as before, so you will need to stitch 72 frames.  By opening the lens up two stops, diffraction is reduced by a factor of two.  This means that you will still capture the same number of blur "spots" on the reduced sensor area, giving the same image resolution (still assuming diffraction limited optics).  Depth of field will also be similar.  Capturing the 10um blur spots with 2um pixels is still overkill.

Setup #4) Canon G11 camera (5.7mm x 7.6mm sensor) with 45mm lens at f/4.  The logic is the same as the previous example.  Now you need only 18 frames to capture the same image composition, and you will again capture the same number of resolution "spots".  This still assumes diffraction limited optics, which may be pretty difficult to achieve at f/4, so there is probably no point in extrapolating further than this.  Digitizing 5um blur spots with 2um pixels is still good sampling.