Actually, we need even more resolution.

[Theodoros]

Hi,

This paper is a good read: http://www-isl.stanford.edu/~abbas/group/papers_and_pub/pixelsize.pdf

The impression I have is that the papers says that 4-5 microns is optimal with 0.18 micron technology, but I also feel that the pixel model in the paper is a bit dated. Modern pixel designs use shared transistors and utilise the pixel area better than older designs.

Best regards
Erik

    Actually the conclusion suggests a 6.5μm pixel as perfect, here is the conclusion only...

                                                                                                              CONCLUSION
    We proposed a methodology using a camera simulator, synthetic CSF scenes, and S-CIELAB for selecting the optimal pixel size for an image sensor given process technology parameters, imaging optics parameters, and imaging constraints. We applied the methodology to photodiode APS implemented in CMOS technologies down to 0.18μ and demonstrated the tradeoff between DR and SNR on one hand and spatial resolution and MTF on the other hand due to the selection of pixel size. Using mean ∆E as an image quality metric, we found that indeed an optimal pixel size exists, which represents the optimal tradeoff. For a 0.35μ process we found that a pixel size of around 6.5μm with fill factor 30% under certain imaging optics, illumination range, and integration time constraints achieves the lowest mean ∆E. We found that the optimal pixel size scales with technology, albeit at slower rate than the technology.

Given the last phrase (on technology dependance), but also considering the photon's reception angle for the edges/corners of a huge MF sensor, I wouldn't trust anything less than 6μm...

[ErikKaffehr]

Yes,

With 0.35 micron technology, with 0.18 microns it would be 4-5 microns.  Please keep in mind that the article is based on a couple of years old pixel designs.

Technology sort of evolves over time…

6.8 microns	3.8 microns

Two images, same camera position and focal length. Which one do you prefer?

Best regards
Erik

    Actually the conclusion suggests a 6.5μm pixel as perfect, here is the conclusion only...

                                                                                                              CONCLUSION
    We proposed a methodology using a camera simulator, synthetic CSF scenes, and S-CIELAB for selecting the optimal pixel size for an image sensor given process technology parameters, imaging optics parameters, and imaging constraints. We applied the methodology to photodiode APS implemented in CMOS technologies down to 0.18μ and demonstrated the tradeoff between DR and SNR on one hand and spatial resolution and MTF on the other hand due to the selection of pixel size. Using mean ∆E as an image quality metric, we found that indeed an optimal pixel size exists, which represents the optimal tradeoff. For a 0.35μ process we found that a pixel size of around 6.5μm with fill factor 30% under certain imaging optics, illumination range, and integration time constraints achieves the lowest mean ∆E. We found that the optimal pixel size scales with technology, albeit at slower rate than the technology.

Given the last phrase (on technology dependance), but also considering the photon's reception angle for the edges/corners of a huge MF sensor, I wouldn't trust anything less than 6μm...

[Theodoros]

Yes,

With 0.35 micron technology, with 0.18 microns it would be 4-5 microns.  Please keep in mind that the article is based on a couple of years old pixel designs.

Technology sort of evolves over time…

Best regards
Erik

You should take in mind the last phrase Erik... "We found that the optimal pixel size scales with technology, albeit at slower rate than the technology." 4-5 microns is huge difference (about 55-70% reduction) than the 6.5microns conclusion... It's no where near a sensible difference.

Edit: (on edit).... Out of the two images, I would prefer none... It's not the kind of image I care to shoot, ....I prefer to make my choices by my needs of capturing (as subject).

[ErikKaffehr]

See post above.

Best regards
Erik

You should take in mind the last phrase Erik... "We found that the optimal pixel size scales with technology, albeit at slower rate than the technology." 4-5 microns is huge difference (about 55-70% reduction) than the 6.5microns conclusion... It's no where near a sensible difference.

[Theodoros]

See post above.

Best regards
Erik

Erik, it's YOU that posted the article (Great article, thanks for posting!) ...and now you disagree with the conclusion of the article you posted!

[Jim Kasson]

As we have seen in the discussion, it is very hard to utilize 60" wide paper at 360 PPI fully, with high transferred modulation at the 1/360" pixel level. So quite obviously, the prints normally don't fully utilize the resolution of the printer.

Some media make it impossible to use the full resolution of the printer. I'm thinking of matte finishes, especially on uncoated paper, and canvas. That's not a bad thing, if you plan for it. I recently had a 60x60 canvas print made of this image for an exhibition (I love it when they want large prints):



The original file was 6000x6000, with very high-quality pixels, since it was res'd down from 6000x1,000,000 pixels. (16 stitched Betterlight captures). That gives only about 100 ppi. We res'd the image up to 360 ppi before we sent it to the printer, but I have the impression that it wouldn't have looked much sharper on canvas even if I'd had that much real resolution. Certainly the 15x15 proofs on canvas looked a lot softer than 15x15 proofs on Exhibition Fiber.

That's not to say that the final result was bad in any way; I was delighted. For another subject, the story might have been different, and canvas would have been inappropriate.

In the mid-nineties, there was a printer at a service bureau in Las Vegas that had been designed and constructed to print billboards. It printed with acrylic automobile paints. The halftoning software wanted the image at 18 ppi. I made several really nice prints with that printer. 

Jim

[Theodoros]

I would try three things before I res'it up to 360ppi... 1.Down sample it to 90ppi and then up-sample it to 360ppi... 2.Up-sample it to 120ppi and then re-upsample it up to 360... 3. Down sample it to 90ppi ...and print it!

From sense, I think of "1"... but I would try all of them on my monitor anyway. But still... is it a "perfect" capture?

[Jim Kasson]

But still... is it a "perfect" capture?

I'm not sure what you mean. I suppose, from one perspective, nothing ever is. From another, it's pretty great, with about 160 source pixels being averaged into each pixel in the working copy of the final edit. In the orthogonal direction to the time one, I could have used some more resolution, since I needed f/45 on the 120mm AM-ED Nikkor for the DOF, and the photodiode array is maybe 3 inches high. There was some CA that I had to process out.

Jim

[Theodoros]

I'm not sure what you mean. I suppose, from one perspective, nothing ever is. From another, it's pretty great, with about 160 source pixels being averaged into each pixel in the working copy of the final edit. In the orthogonal direction to the time one, I could have used some more resolution, since I needed f/45 on the 120mm AM-ED Nikkor for the DOF, and the photodiode array is maybe 3 inches high. There was some CA that I had to process out.

Jim

Sounds good... but since this was obviously a still capture, using a Betterlight for so long as 16 captures require can be an issue... I just wonder, I don't say you did anything wrong by any means...   My quote was on my opinion on the printing process... One may ignore the "capture" part... it's only there to underline my opinion that printing big requires a "perfect" capture.

[Jim Kasson]

Sounds good... but since this was obviously a still capture, using a Betterlight for so long as 16 captures require can be an issue...

Well, it was a 14-hour series of exposures, and I did have to write some Matlab code to find and deal with instances of subject motion as well as doing one-dimensional downresing and median filtering, so in that sense the capture wasn't perfect.

Jim

[Theodoros]

Well, it was a 14-hour series of exposures, and I did have to write some Matlab code to find and deal with instances of subject motion as well as doing one-dimensional downresing and median filtering, so in that sense the capture wasn't perfect.

Jim

I will stay with your comment: "That's not to say that the final result was bad in any way; I was delighted."... A creation belongs to the creator! Well done.    What do you think on my printing proposals?

[Jim Kasson]

What do you think on my printing proposals?

Sending the image at 90 ppi to the printer driver wouldn't be my first choice. The Epson driver resamples the image it is fed to either 360 or 720 ppi (depending on the state of the "fine detail" check box, using nearest neighbor. With a 90 ppi image, that means that every pixel in the input image becomes a 4x4 square of equal-value pixels in the image the halftoner works from. The canvas will probably smooth that out enough so that you won't notice it, but I prefer not to take chances, and upsample using a more sophisticated algorithm such as Lanczos 3 or the proprietary algorithms in QImage.

Your other suggestion was to do the upsampling in two stages. That has proven to be a successful regime for some, and, if I remember right, was the basis for the Genuine Fractals plug-in many years ago. My own personal experience is that doing the resolution changing in one shot is at least as good, and I'd rather spen my time making sure I'm using an algorithm suited to the image than trying various staged possibilities. But to each his own, and I'm sure that, done with care, the staged approach can yield successful results. In any event, I think there is more leverage in the choice of algorithm than in the choice of ratio, but that's just a personal opinion.

We have some experts on this board, Bart van der Wolf and Nicolas Robidoux, who have forgotten more than I'll ever know about resampling.

Jim

[Theodoros]

Sending the image at 90 ppi to the printer driver wouldn't be my first choice. The Epson driver resamples the image it is fed to either 360 or 720 ppi (depending on the state of the "fine detail" check box, using nearest neighbor. With a 90 ppi image, that means that every pixel in the input image becomes a 4x4 square of equal-value pixels in the image the halftoner works from. The canvas will probably smooth that out enough so that you won't notice it, but I prefer not to take chances, and upsample using a more sophisticated algorithm such as Lanczos 3 or the proprietary algorithms in QImage.

Your other suggestion was to do the upsampling in two stages. That has proven to be a successful regime for some, and, if I remember right, was the basis for the Genuine Fractals plug-in many years ago. My own personal experience is that doing the resolution changing in one shot is at least as good, and I'd rather spen my time making sure I'm using an algorithm suited to the image than trying various staged possibilities. But to each his own, and I'm sure that, done with care, the staged approach can yield successful results. In any event, I think there is more leverage in the choice of algorithm than in the choice of ratio, but that's just a personal opinion.

We have some experts on this board, Bart van der Wolf and Nicolas Robidoux, who have forgotten more than I'll ever know about resampling.

Jim

OK... But, when downsampling to 90 or up-sampling to 120 (in first stage), there is still a choice of algorithm involved... it's just kept to the minimum possible influence... If one aims to affect the capture as little as possible it seems a logical approach, ...doesn't it?

[Jim Kasson]

OK... But, when downsampling to 90 or up-sampling to 120 (in first stage), there is still a choice of algorithm involved... it's just kept to the minimum possible influence... If one aims to affect the capture as little as possible it seems a logical approach, ...doesn't it?

When upsampling in two stages, there are two opportunities to select and tune the algorithms. When upsampling in one stage, there is only one such set of choices -- and there are many choices within each algorithm: window sizes, parameters, post-resampling deconvolution filtering, etc. I don't that two stages is any less likely to affect the capture than one (and I'm not sure precisely what that means). Actually, we want to affect the capture; otherwise we'd leave it at the original resolution. We just want the effects to be whatever we decide are good ones, and to minimize whatever we decide are bad effects. But that's not so easy to define.

Jim

[Torbjörn Tapani]

What does quality pixel even mean. It's a dimensionless entity with a RGB value as it's only property. I'll take more of them before I pay for a bigger sensor. The cost of a chip is closely related to size. With improved technology more pixels in a given size basically comes for free. If I can keep dynamic range I'll take more. When we are diffraction limited wide open we have enough.

But speaking of DR. What is the limit of DR that could be obtained in a single capture?

[Jim Kasson]

What does quality pixel even mean. It's a dimensionless entity with a RGB value as it's only property. I'll take more of them before I pay for a bigger sensor. The cost of a chip is closely related to size. With improved technology more pixels in a given size basically comes for free. If I can keep dynamic range I'll take more. When we are diffraction limited wide open we have enough.

But speaking of DR. What is the limit of DR that could be obtained in a single capture?

One measure of quality is the signal-to-noise ratio. So, more light, higher quantum efficiency, larger fill factors, bigger chips, all help. But you want to measure the quality as sent to the printer, not necessarily with the same number of pixels produces by the camera.

The limits of single capture DR depend on the definition you are using. There are those who like "engineering" DR, which is th max signal over the read noise or the max signal over the signal that produces an SNR of 1, depending on who you're talking to. I prefer the ratio of the max signal to that signal which produces a photographically acceptable SNR, which for me is around 10.

Once you've agreed on a definition, then the result is a function of the resolution of the output image.

And, in the real world, a function of lens flare, etc.

It's not easy to answer that question.

Jim

[RobertJ]

Question:
Do you judge how a print will look like on screen by viewing at 50% (I don't know if this is true)?

[Jim Kasson]

Question:
Do you judge how a print will look like on screen by viewing at 50% (I don't know if this is true)?

The only way that I know of to judge what a print will look like is to make a proof print. You can get close on-screen, but you can't really tell for sure.

Jim

[Torbjörn Tapani]

Jim, I'm too dumb to be asking these kinds of questions as I don't understand all of your reply. But it's that last part I'm curious about. There is some practical limit of what a lens can deliver. Are we even close to being able to record it? Is flare the biggest limiting factor?

[hjulenissen]

Its noise per sensor area that is lower with larger sensors Erik... at least that's what Physics suggest.

This can be approached by introducing arguments and references to relevant theory, or introducing real-world side-by-side tests that confirm or fail to confirm your view. Simply claiming "because... physics" is not very fruitful.

Back to the artefacts matter though... I see that you only mean moire when you refer to artefacts, isn't it?  We all have to understand that if Bayer pattern was absent, there would have been no moire problem... Detail would also improve considerably since what the human eye understands as resolution is the colour difference between pixels and thus interpolated colour mistakes affects resolution too and DR would improve considerably...

There is aliasing in any sampling system as long as the input signal has a bandwidth >= fs/2. Man-made sampling systems might include some kind of "anti-aliasing" prefiltering (explicit or implicit), but like any man-made filters they won't have infinite attenuation in the stop-band. What might be realistic (for some systems/cases) is for this aliasing to be sufficiently attenuated to levels where they are no longer a big problem. There is no need for the signal to be periodic, by the way. A single impulse (think single-sensel-size star against a black sky) contains a wide range of frequencies and may cause aliasing.

With the CFA used in "Bayer" sensors, the 3 sampling frequencies are simply shifted downwards. If you prefilter at this new minimum fs/2, I believe that there will be no Bayer artifacts (and Debayering will be a very simple linear operation). One reason that we don't do such prefiltering is that the sensel density is simply not high enough. The resulting (luminance) resolution would simply be too low, and the artifacts from the CFA pre/post processing are often surprisingly small anyways. The current level of prefiltering and postprocessing may be a compromise that works well for current sensel densities and print sizes/viewing distances.

Therefore, I am eagerly awaiting the time when sensor manufacturers can give us a (near) photon-noise-limited Bayer sensor of e.g. 180MP with good color selectivity and angular light response, where natural and/or introduced filtering limits the degree of "detail" to a lot less than 10x what can be had with my current 18MP camera. The 100% pixel view will be a blurry mess, but the A2 images that I frame ought to be better looking with less processing effort.

-h