PDA

View Full Version : Query: High pixel density and diffraction limit (for Roger Clark)



Sinuhe Hahn
10-23-2012, 05:23 AM
Hi Roger

I wonder if you can help with this, as you are an advocate of high pixel density sensors.

A criticism that has been raised at high density pixel cameras such as the D800 (with 36mp on a FX chip), is that they have a noticeable diffraction limit.

ie even the Nikon brochure does not advocate using apertures greater than f8.0 / f11.0, as this can lead to a loss in resolution and perceived sharpness.

If one is mainly a landscape photographer, this becomes a real hindrance.

What is your view / experience regarding this issue?

Thanks

Sinuhe

Jon Rista
10-23-2012, 05:12 PM
I'd like to hear what Roger has to say as well. In his absence, though:

When it comes to diffraction, there is a lot of general misunderstanding of what it means when a sensor becomes "diffraction limited". To start off:


All other things being equal, in no way can a higher density sensor produce worse quality images than a lower density sensor because of diffraction.

The Diffraction Limited Aperture, or DLA, is simply the point at which diffraction begins to affect IQ for a sensor of a given spatial resolution, but not the point at which it makes IQ unusable. Let's assume we have one sensor that is diffraction-limited at f/8, and another that is diffraction limited at f/5.6. When the sensor with an f/5.6 DLA becomes diffraction limited, I believe the general (mis) understanding of that is that the f/5.6 camera is now producing "worse" IQ than the f/8 camera. The contrary is actually quite true. The f/5.6 sensor may be experiencing overlapping airy discs, but it is still extracting more detail than the f/8 sensor. When diffraction increases such that it begins to affect the f/8 sensor, and at subsequent apertures, the IQ of both sensors will begin to normalize...that is, the f/5.6 and f/8 sensors will start to produce images that have increasingly similar IQ.

You experience diminishing returns for the benefit of having a sensor with smaller pixels...but at no point will the sensor with smaller pixels ever produce worse image quality (all else being equal...that is).

In reality, a sensor with smaller pixels will have a lower maximum saturation point, which reduces S/N. That leads to more pronounced photon noise, and in many cases the electronic/read noise floor will be a grater portion of the signal, which at lower ISO can lead to even further problems with noise. Noise is the great unequalizer in sensors of differing pixel sizes. It is more likely that you'll experience a greater loss in IQ from noise than you will from diffraction, relative to a sensor with larger pixels.


The D800 has 4.6 micron pixels. Those are actually slightly larger than the 4.3 micron pixels of the 7D. I use the 7D myself, and often at apertures like f/7.1, f/8, f/9. I have several images posted in the Avian forum at such apertures. In my opinion, the IQ is stellar, despite diffraction...and my crops tend to be fairly significant (50% or so), so I am taking an area of my sensor that is very small and producing an image from it. I wouldn't be able to do that with a larger sensor with larger pixels...the detail simply wouldn't exist. I would expect the D800 to actually fair a little better simply because of its slightly larger pixel size, however on top of that, it also has a very low-noise sensor (unlike the 7D), which should give it a further edge.

One thing about high density sensors is that they are extremely demanding on your optics. When pixel-peeping, assuming you don't address the high demand on your lenses by such a camera, you'll likely see a greater degree of 100% crop "softness" than you would with a lower density sensor. That doesn't necessarily mean the D800 or 7D are not "as sharp" as a lower resolution camera. It just means you have to work harder to fully realize their potential. To truly realize the full benefit of the kind of spatial resolution offered by the 7D or D800, you'll want to make sure you are using glass that can produce a comparable real image at the sensor. You will also want to make sure you stabilize your camera as much as you can before taking a shot, either with a ground pod or a sturdy tripod, to minimize blur from camera shake.

Graeme Sheppard
10-23-2012, 05:43 PM
Just want to add my thoughts and hypotheses.
I also believe that a lot of diffraction limit tests are not done fairly - cropping images of two different sensors to 100%, which leads to a different image area being assessed. In theory (yes, I know) the fact that a black line is one pixel wide on a low resolution sensor or two pixels wide on a high resolution sensor will not affect its sharpness against a pure white background.
However, if the line being photographed was actually not a monotone then the high resolution sensor will detect that and, perhaps due to a little diffusion limiting blur the edge.
I'd say this means that objectively the high resolution is no worse for diffusion, but since sharpness and detail have an inverse relationship in real images, we can perceive the reduced sharpness as a reduced image quality when in fact it is technically superior.

Jon Rista
10-23-2012, 09:05 PM
Great point, @Graeme. There is a lot of perception involved in subjectively defining image quality, which does make it rather difficult to evaluate,

Roger Clark
10-23-2012, 10:28 PM
Hi Roger
I wonder if you can help with this, as you are an advocate of high pixel density sensors.
A criticism that has been raised at high density pixel cameras such as the D800 (with 36mp on a FX chip), is that they have a noticeable diffraction limit.
ie even the Nikon brochure does not advocate using apertures greater than f8.0 / f11.0, as this can lead to a loss in resolution and perceived sharpness.
If one is mainly a landscape photographer, this becomes a real hindrance.
What is your view / experience regarding this issue?
Thanks
Sinuhe

Hi Sinuhe,

Jon and Graeme have some good points. The problem with most discussions we see online, and in things like DXOMark is the focus on the sensor. The focus (puns appropriate here) should be on the image. The sensor is the device that samples the image, but it is the image that counts. If the image is diffraction limited, one can not get any more detail, so it is the best that can be done. So diffraction limited is a good thing as long as you have enough detail for your purposes. For example, superzoom P&S cameras have supposedly great equivalent focal length, but they are diffraction limited such that the sensor has such a high pixel density that the diffraction is way oversampled and the image detail is greatly limited by that diffraction (e.g. small lenses and pixels smaller than 2-microns sampling an Airy disk at f/8 that is over 10-microns in diameter).

Important in fine detail image quality is 1) sufficient photons per pixel to give a decent signal-to-noise ratio, and 2) enough pixel-to-pixel contrast to show enough accutance to give the perception of sharpness and detail. Certainly large pixels deliver that because they can hold a lot of photoelectrons, and diffraction is not reducing contrast (or reducing it less) in the fine details. But large pixel under sample the potential detail. I'm attaching an image from my article on nightscapes that will illustrate these points. The article is: http://www.clarkvision.com/articles/nightscapes/
You can also see the images from the 3 cameras not resampled on the above web site.

There are images from 3 cameras: 5D Mark II, 1D Mark IV, and 7D. The images are not diffraction limited, but they are worse: a 20 mm f/2.8 lens at f/2.8 is aberration limited. Further the scene is a star field and by reading the net, one would think small pixels are less sensitive. But the 7D comes out on top, showing fainter stars and better resolution. The 1D4 comes in a close second and the larger pixel 5D2 comes in last.

So the bottom line is more pixels are better for image detail as long as they are not too small. In my opinion, around 5 microns is optimum for current silicon sensor technology (cmos and ccd). The D800 is very close to the optimum pixel size. The 7D is still pretty nice, but might be just a hair on the too small side, but still close to the optimum. The 5D4 is slightly on the too large side of the optimum. I hope the 1D5 has 5 micron pixels.

Roger

Roger Clark
10-23-2012, 10:56 PM
The Diffraction Limited Aperture, or DLA, is simply the point at which diffraction begins to affect IQ

Hi Jon,
Actually, if it is diffraction limited, it can't get any better than that (for that focal length and f/ratio), so it is not where diffraction begins to have an effect, it is where all aberrations are much smaller than diffraction. It is not limited if it is only beginning to have an effect. The main effect of diffraction is reduction in contrast in the fine details.

Also, when one needs more depth of field, sometimes closing down to diffraction limits is still better the the out-of-focus blur from limited depth of field. Even if a lens becomes truly diffraction limited, sometimes an image at f/22 is still better due to depth of field. Image quality is more than just diffraction limited.




The f/5.6 sensor .... f/8 sensor.

I'm not sure what you mean by f/5.6 sensor and f/8 sensor.



In reality, a sensor with smaller pixels will have a lower maximum saturation point, which reduces S/N. That leads to more pronounced photon noise,


I've written extensively on this here on BPN and on my web site. See, for example, etendue:
http://www.clarkvision.com/articles/telephoto.system.performance/

The reason for your statement is not equalizing etendue. A camera with smaller pixels is chopping up the focal plane into smaller pieces, each seeing a different amount of the subject, thus more pixels on the subject. If one equalizes the pixels on subject, and uses the same aperture diameter lens with the same exposure time, the camera with the small pixels with have the same signal-to-noise ratio as the large pixel camera, show the same detail on the subject, and show the same depth of field.



...and in many cases the electronic/read noise floor will be a grater portion of the signal, which at lower ISO can lead to even further problems with noise. Noise is the great unequalizer in sensors of differing pixel sizes. It is more likely that you'll experience a greater loss in IQ from noise than you will from diffraction, relative to a sensor with larger pixels.

I disagree. As stated above, the noise will be the same when etendue is equalized. Usually what people do is change the pixels on subject between tests. Or they change clear aperture. It is aperture (diameter of the lens aperture) and angular area of the pixel that determines 1) signal-to-noise ratio and 2) detail on subject (for a given light level and exposure time). Note two things are missing from the fundamental rule: f/ratio and linear pixel size, the very two things that seem to dominate photographic discussions in the digital era.




One thing about high density sensors is that they are extremely demanding on your optics....

Yes, I agree.

Roger

Sinuhe Hahn
10-24-2012, 12:11 AM
Thanks for the detailed reply gents - very helpful - cleared up quite a few misconceptions.

Jon Rista
10-25-2012, 09:23 PM
Hi Jon,
Actually, if it is diffraction limited, it can't get any better than that (for that focal length and f/ratio), so it is not where diffraction begins to have an effect, it is where all aberrations are much smaller than diffraction. It is not limited if it is only beginning to have an effect. The main effect of diffraction is reduction in contrast in the fine details.


Sorry, I was referring to the sensor's diffraction limited aperture. I understand that once a lens is "diffraction limited", spatial resolution of the real image projected by the lena at the sensor focus plane is limited, and as you stop down, spatial resolution drops. My comment was referring to the DLA of sensors, and that once the aperture of the lens is stopped down to just below the DLA of a sensor with small pixels, it is still resolving more detail than a sensor with a larger DLA. If we use a 200mm lens as an example, at f/6.3 on a sensor with a DLA of f/5.6, that sensor, despite now being "diffraction limited" itself, is still resolving more detail (from a spatial resolution standpoint) than a sensor with a DLA of f/8. When that 200mm lens is stopped down to f/8, the sensor with smaller pixels will start approaching the detail of the sensor with larger pixels, and as you continue to stop down, the differences in IQ as resolved by those two sensors will eventually disappear. I guess I might have been postulating in a focal-length limited scenario, would preclude the normalization of pixels on subject, as in your arguments below. Sorry for the lack of proper context.



Also, when one needs more depth of field, sometimes closing down to diffraction limits is still better the the out-of-focus blur from limited depth of field. Even if a lens becomes truly diffraction limited, sometimes an image at f/22 is still better due to depth of field. Image quality is more than just diffraction limited.


I don't disagree at all. Artistic and practical concerns should always come first, and if they necessitate a narrower aperture, you should use one. Actually, that is exactly my argument. The fact that diffraction affects a sensor with smaller pixels at wider apertures than a sensor with larger pixels should not be something of concern if you need to use a narrow aperture, as both sensors, regardless of pixel size, will ultimately record the same thing, just at different levels of granularity. I was solely referring to the DLA of sensors and how a sensor with smaller pixels does not produce worse image quality than a sensor with larger pixels (all other things being equal).





I'm not sure what you mean by f/5.6 sensor and f/8 sensor.


I was referring to the diffraction limited aperture of two hypothetical sensors. In this case, a hypothetical sensor that has a DLA of f/5.6, and another hypothetical sensor that has a DLA of f/8.




I've written extensively on this here on BPN and on my web site. See, for example, etendue:
http://www.clarkvision.com/articles/telephoto.system.performance/

The reason for your statement is not equalizing etendue. A camera with smaller pixels is chopping up the focal plane into smaller pieces, each seeing a different amount of the subject, thus more pixels on the subject. If one equalizes the pixels on subject, and uses the same aperture diameter lens with the same exposure time, the camera with the small pixels with have the same signal-to-noise ratio as the large pixel camera, show the same detail on the subject, and show the same depth of field.


Hmm, I guess I'll have to re-read your article on etendue... I don't think I disagree when you say "same pixels on subject" with the same lens/same aperture, it does seem logical the two sensors would have the same S/N (since the same amount of light for the same part of the subject would reach each pixel, large or small...any benefit the larger pixel has in terms of maximum saturation would never be realized.... although could one not argue that the smaller pixels could potentially saturate sooner than the larger pixels, resulting in clipped highlights? Would that necessitate the use of a higher ISO setting on the sensor with larger pixels?)



I disagree. As stated above, the noise will be the same when etendue is equalized. Usually what people do is change the pixels on subject between tests. Or they change clear aperture. It is aperture (diameter of the lens aperture) and angular area of the pixel that determines 1) signal-to-noise ratio and 2) detail on subject (for a given light level and exposure time). Note two things are missing from the fundamental rule: f/ratio and linear pixel size, the very two things that seem to dominate photographic discussions in the digital era.


Assuming you equalize pixels on subject, I agree. The point I was trying to make is that there is no reason to think diffraction specifically will result in worse image quality in the sensor with smaller pixels. Diffraction limits spatial resolution of what the lens projects onto the sensor (the "real image"), but two sensors with different pixel densities will be resolving the same "real image" if you use the same lens at the same aperture and get the same pixels on subject.



Regarding electronic noise (fixed pattern, horizontal and vertical banding, etc.). In Canon sensors, electronic noise (vs. photon noise) tends to be fairly significant at ISO 100 & 200, and even at ISO's as high as 800, and can pose a problem in the shadows and even in the mid-tones at times (something I believe is a significant failing in Canon sensors these days, and an area where they are seriously failing in comparison to Sony's Exmor sensor.) In a Nikon camera that uses an Exmor sensor, electronic noise is extremely low. There are a number of Nikon cameras using that technology now (D800, D600, D3200), with a variety of sensor and pixel sizes, all of which produce some phenomenal IQ with two stops or more DR than any comparable Canon camera, and they exhibit no visible pattern noise at all unless you lift shadows by around six stops or more. In a Nikon camera with an Exmor sensor, I would say the above statements about etendue hold true implicitly. In a Canon sensor, electronic noise can be over 30 electrons worth (I believe it is over 33e- per pixel in the 5D III, which seems unreasonably high), with pronounced horizontal and vertical banding (hatch patterned noise that can be so pronounced that it affects the low mid-tones to a frustrating degree when pushing exposure around in post). My Canon 7D, even at ISO as high as 400, often exhibits banding, particularly vertical, with even a minimal amount of post-process exposure tuning. That was what I was referring to when I mentioned "Noise is the great unequalizer". I mean electronic noise, not photon noise. The 5D II/III both have similar banding noise problems, but it seems to require more significant pushing around of exposure in post for it to exhibit in a problematic way than it does with my 7D. I presumed the reason for that was the 7D's smaller pixels, which have a much lower maximum saturation point than either the 5D II or III.