Theory Help: How to fake digital back resolution from current DSLRs

[ikrima]

I've been toying with this idea in my head but wanted to get some other's perspective to see if my thought process is correct.

I have a DSLR (5Dmkii) which has a 21 megapixel sensor with a bayer filter on top. Thus, there's a loss of image resolution, due to the interpolation/demosaicing process, especially when compared to digital backs (which record three color values at each pixel site).

So my thought process was to cut the image resolution in half and use the raw data from the sensor to directly assign the RGB values. Would that give me full resolution? And, does anyone know how I would go about converting the recorded "green", "red", "blue" to the correct colorspace? There's also the problem of each 4 pixels would contain 2x green channel information so I'm not sure how to handle that data.

It would also be interesting to try this out with high f/stops to make the lens diffraction limited. I figured if I can stop it down to ~f/16-f/22, that would cause the size of the Airy disk to encompass 4 pixel sites on the sensor which would in turn fake 4 different CCDs at one pixel site. Does that logic pan out?

Cheers,
Ikrima

[ikrima]

I should also add: would it be worth it to go through all of this instead of just downsampling the image by 50% with a sinc or bicubic filter?
a
Cheers,
Ikrima

[Colin Southern]

I'd suggest taking a step back and asking "does current technology really limit image resolution"? I've seen photos taken with an iPhone side by side with a digital back ... And there's nothing in it (for small prints). So unless you're producing really large prints, you won't see any difference between prints from a high resolution SLR and a higher resolution back.

[ikrima]

Yes actually. Quite a bit for my uses. I do fashion photography and want the resolution and tonal gradation you get with what I've seen come out of the medium format digital backs.

I feel like I can fake it with spending inordinate amounts of time in post through sharpening but it's too time intensive.

[Colin Southern]

Yes actually. Quite a bit for my uses. I do fashion photography and want the resolution and tonal gradation you get with what I've seen come out of the medium format digital backs.

I feel like I can fake it with spending inordinate amounts of time in post through sharpening but it's too time intensive.

Hi Ikrima,

I hear what your saying - but the limiting factor (as I see it) usually isn't the resolution; assuming that one frames the shot correctly so that excessive cropping isn't required then even a camera like the 5D2 or 1Ds3 will give you 12 x 18 inch prints at a ridiculous 300dpi - if you're not going any larger than that then the eye just won't be able to resolve the extra detail from the higher resolution of a MF camera. For landscape photography the sensor MP wars became a moot point once the counts went past about 8MP - for fashion photography past about 12MP.

With regards to better tonal graduation ... I've heard this said may times too, but I'm not too sure I buy it; the maths just doesn't stack up. I suspect that it's probably more the case that people who have high-end medium backs probably also have people who know how to properly retouch the images that they're producing.

Just my 10c anyway

[ikrima]

Ah but you have to remember that digital sensors on DSLRs fake their resolution. They have one a bayer filter on the sensor and go through a demosaicing process. So each pixel sensor measures only one color and then it's interpolated to provide color at each location. So technically, a 21 mp sensor = 8 mp.

Digital backs on the other hand measure r,g,b at each location, so you get actual resolution. That's why if you zoom in at 100% on a picture from a digital back, it looks good. On a DSLR, not so much. As for tonal gradations...i'm not really sure. That's a mystery to me too I couldn't find too much info on how the digital backs work. Maybe they scan at 16 bits which would give a higher dynamic range than the 10-12 DAC coming out of the DSLRs

[Colin Southern]

That's why if you zoom in at 100% on a picture from a digital back, it looks good. On a DSLR, not so much.

That's my point though; for sure, a digital back WILL look better at high magnifications - but - nobody examines a real-world print with a microscope. If you're going to check your carpets, do you stand at the door and look, or do you go through the pile with microscope?

Same with cameras; if you're producing a 12 x 18" print, you won't be able to see additional detail from the MP camera as it'll be too small for the eye to resolve.

[arith]

Is it just me or did this just get complicated. Do medium format camera's come in three bits, a lens, body and back; because I've seen prices of 17000 for body and 40000 for back. But that isn't the point, I can't see how anything could record three colours at a point unless it was a three pixel triangular point and I can't see the difference, a little mathematics is used for the bayer array which won't be far out.
Or does a medium format work in a completely different and mysterious way? Just interested

[Colin Southern]

Ikrima, you might find this interesting ...

http://news.cnet.com/8301-13580_3-10154386-39.html

[ikrima]

@Colin Southern, well I think we're in agreement! If I can paraphrase, you're saying "it doesn't matter because you won't ever view the image at 100%. Only a 12"x18" @ 300 dpi would require 20 megapixels. Every other size is smaller than a 100%". Which means, at some point, you are going to resize the image.....which is exactly what I am referring to:

"I should also add: would it be worth it to go through all of this instead of just downsampling the image by 50% with a sinc or bicubic filter?"

But, this simple approach doesn't seem theoretically correct...but this is where the math gets fuzzy in my head. So the question I was asking was what's the right way to do it and will that even provide a noticeably improved image compared to simple downsampling?

@Arith, yes, you can have co-located sensor pixels although it's presumably a lot more expensive. All modern HD video cameras have 3 CCDs, not just one. As for "the noticable difference", ken wrote a good summary with pictures. http://www.kenrockwell.com/tech/bayer.htm

[Colin Southern]

Sorry, I'm getting a bit lost here. I suspect that how the down-sampling algorithm treats edges - and this it's effect on the sharpness of the image - is probably more significant than other (source) factors.

[arith]

It's alright I found the info, they come in four bits. When I was at school they had a Hasselblad but apart from the teacher using a Polaroid back I thought the film compartment was integral to the body, and when I was tested loading a rig I thought that was a mock up.

We wasn't allowed to pull this camera to pieces. Still don't get how you record three colours at a point though. Do Sony do something like that?

[arith]

@Arith, yes, you can have co-located sensor pixels although it's presumably a lot more expensive. All modern HD video cameras have 3 CCDs, not just one. As for "the noticeable difference", ken wrote a good summary with pictures. http://www.kenrockwell.com/tech/bayer.htm

Seen it, what a difference. Doing it with three sensors all 60 x 60 will cost a few bob. Actually the interpolation shouldn't be that bad but there is twice the greens, so maybe this image was chosen to highlight faults and is therefore unrepresentative of the general situation.

I like pictures, makes it so easy to understand.

[arith]

I see what you mean now If you diffract enough a 4x4 array can simulate a single pixel colour, is this right? Why not take two images if that is practicable, not in fashion I suppose but I was thinking you could take one at a non diffracted aperture and an extremely diffracted one to compare colour, but I think it still won't work because the interpolation uses more than one 4x4 array and how do you get RAW to display actual pixel colour in photoshop ACR let alone work with it in layers for instance.

[ikrima]

exactly. although after researching this for a day, I realized I was trying to solve what basically amounts to capture sharpening. Looks like the most promising approach is to use an RL-deconvolution filter for capture sharpening and then a sinc filter to downsample. (Richard-Lucy deconvolution http://www.openphotographyforums.com...ad.php?t=3013; sinc downsampling http://www.xs4all.nl/~bvdwolf/main/f...own_sample.htm). Now I understand a little bit more why those 3rd party sharpening plugins work so well

[arith]

Cheers for that; I hardly ever use USM sharpening preferring Topaz Detail, but I don't know the algorithm it uses.

[Steaphany]

Ikrima,

A point that you and everyone else here has missed is all Bayer dSLRs also have an anti aliasing filter in front of the imager. The anti-aliasing filter is an optical element that prevents Moiré pattern and Aliasing stair step image artifacts by slightly blurring the image.

Yup, that 21 MB imager reduced to 8MP by the Bayer mask gets a further reduction by the camera's built in soft focus filter.

So, you need to adjust your airy disk spread by the diffusion imparted by the anti aliasing filter. A problem I realized is that no camera manufacturer ever specifies just how diffusing the anti aliasing filter is for each camera model.

Depending on your budget, you may want to look into the Sigma SD15, or now that the SD15 is out, the Sigma SD14 on the used market. The marketing hype lists both camera models at 14.7 MP, but the Foveon imager stacks the Red, Green, and Blue photodiodes so that each photosite captures the whole spectrum. Neither model use an anti aliasing filter, so the image clarity is excellent. ( Obviously, I have a Sigma SD14 and I chose the camera specifically because the imager technology finally gave me an alternative to Bayer masked imager based cameras. ) An additional capability of both the SD15 and SD14 is the IR blocking filter can be removed for sensor cleaning which allows use for broad spectrum and Infrared photography, but do understand that the Foveon imager operates very differently from other dSLRs when capturing infrared images.

Another option that you may want to explore is to have your camera modified by removing the anti aliasing filter. The web page LDP, LLC High Resolution/Hot Rod Conversion explains the anti aliasing filter and demonstrates what the camera performance advantage is when it's removed. They also provide a conversion service.

[Pictus]

Also about downsampling http://www.topazlabs.com/forum/featu...nsampling.html

Rawtherapee is excellent and the new unstable builds with AmaZE demosaic are even better, you can get them in the forum http://www.rawtherapee.com/forum/vie...asc&highlight=

Besides RawTherapee, the deblur option in Topaz Detail is Richardson-Lucy deconvolution, and this 2 also have Richardson-lucy http://www.raw-photo-processor.com/RPP/Overview.html and http://www.iridientdigital.com/products/
the free GIMP also have plug-in with Richardson-Lucy http://gmic.sourceforge.net/gimp.shtml

BTW, the daylight low ISO shoots from Sigma Foveon are excellent, not only it is sharp, but
the dynamic range is TOP(shame on Canon), I wonder what a full frame Foveon would do...
http://www.rytterfalk.com/
http://www.rytterfalk.com/2010/04/19...-is/#more-1446

[arith]

Besides RawTherapee, the deblur option in Topaz Detail is Richardson-Lucy deconvolution,

I really like that feature, it makes distant foliage sharp without halo's, but is very good when used with small detail. cheers