Michael, many thanks for this very straightforward, clear and sensible discussion. And all the more so after what you have just been through.

I have only one observation to contribute. Toward the bottom you mention that one probably won't see much difference in image quality between a 16 and 22 MP sensor. Based on the underlying logic and assuming all else equal, no argument.

However, as your article is intended to provide factual guidance to those contemplating equipment options, it may be useful to caution that this isn't necessarily a reason not to buy a 22 MP camera (rumoured to be the next generation DSLR forthcoming some time soon) - because most of the time not all else is equal and there are other considerations.

For one thing, as we saw between the 1Ds and the 1Ds Mk2, when they bring out a new model, not only the pixel count changes, but so do other things which add-up to producing cleaner images - i.e. ancillary technical change in a number of variables at the same time all contributing to improved image quality.

For another, with 22MP, one has considerable added luxury for cropping images and still getting excellent resolution on large prints. For example a 1Ds Mk 2 at 16MP provides about 4900 pixels on the large dimension, for a print dimension of 16.3 inches at 300PPI. Bump up the sensor to 22 MP with the same aspect ratio and the 4900 PPI dimension changes to about 5800 PPI, or 19.3 inches printed at 300 DPI. This is about an 18% increase in "croppable real estate" on the large dimension, while roughly maintaining 1Ds Mk2 resolution.

The only additional comment that I would make is that the article, like many other discussions, confuses pixel pitch with pixel size. Early sensors may have had a larger pixel pitch than current sensors, though due to design and production limitations, the fill factor was lower than current sensors and, therefore, the amount of light capturing area on the chip was less. Where we are seeing increasing pixel density on the chip then there is probably no reduction in the light capturing area compared with previous generations - i.e. the pixel pitch is decreasing but the pixel size (light capturing area) remains constant as the fill factor increases. We will probably see another couple of generations of improvement in pixel density as manufacturers move production to more advanced production processes with finer resolution in the lithography and improvements in fill factor.

David,

It's even more complicated that that, as you likely know. Fill factor is an issue, but so is the use of microlenses, which some chips use and others don't, and which helps conversion efficiency.

I deliberately refrained from going into too much esoteric detail in an essay of just a few paragraphs, with a very specific intent.

Michael

Just a few thoughts. When a new Canon model comes out, I usually check the dpreview noise and resolution charts, after they get around to publishing their in-depth review. It seems to be the case so far that Canon never introduces a new model that has higher pixel noise than a previous model. On the noise front, it seems Canon refuses to go backwards (with regards to DSLRs anyway).

Many people were disappointed the 30D had no more pixels than the 20D. I guess it's because Canon were simply unable at that stage to provide smaller pixels that would not also be noisier. But I'm confident they are working on it and that we'll see more pixels on the next generation of cropped format and full frame sensors and those individual pixels will be no noisier (and possibly even less noisy at high ISOs) than the current 30D and 1Ds2 pixels.

Another issue is the effect of noise on a given size print. Most of us do not restrict ourselves to making prints of a size that does not require either upsampling or downsampling of the image file. If we did, we'd be restricting ourselves to a 12x18" print for the 5D and a 9.5x14' print for the 30D. When we upsample (interpolate) an image to make a large print, we interpolate the noise also. The reverse is also true.

The consequence of this is, for any equal size prints, big or small, the sensor with the greater number of pixels will produce less visible noise, provided the smaller, more numerous pixels have, individually, at least equal noise to the larger, less numerous pixels on the other sensor.

I'm also of course confusing pixel size with pixel pitch, which I think is forgivable because Canon never seems to make public the actual pixel (photodiode) size. But, as DiaAzul mentions, there's probably scope for improvement here by reducing the space on the sensor taken up with on-chip processing, whilst keeping the actual size of the photodiode the same, even though the pixel pitch has increased. However, it's not clear to me to what extent there might be benefit here. As I understand it, the microlens directs the incoming light onto the photodiode, whatever its size, so the photons are not wasted by falling unproductively on processing transistors. Nevertheless, a physically larger photodiode should translate to a greater full-well capacity and greater dynamic range. It's worth noting that, according to some facts in a Canon brochure on the D30 (not 30D), that Michael referred to in a thread a few years ago, the actual pixel size of the 10 micron-pixel-pitch D30 is only 5.25 microns. (I know! I've got a good memory, for some things ).

I was sort of coming to the same conclusion. As the inter-site walls become thin, the microlense geometry overlaps and increases purple fringing. That's my theory.

One point that Michael made in his article, a point which is often made in the interests of realism and downplaying of unrealistic expectations, is the fact that the laws of Physics are 'inviolate'. In other words, that's where we come to a dead end.

This stance has always struck me as curious because the laws of Physcics are also the laws of man, and what man creates he can often break or modify, according to his imagination and ability.

I'm not suggesting that Maxwell's equations, Einsteins theories of Relativity or the current theories of Quantum Mechanics are necessarily wrong for our current purposes. They are probably inaccurate, as we'll probably eventually find out.

As technology progresses, new ways of 'circumventing' rather than 'violating' the so-called laws of physics are discovered, which is rather exciting, don't you think?

There's even promising research underway to circumvent the 'laws of diffraction' by producing materials with a negative refractive index. Refer to the following link here

Ray,

Good point.

My intention though is not to work from a base assumption that any inherent limitation can be overcome, but rather to suggest that people not have unrealistic expectations.

Remember, man can not fly.

Michael

There may be hope with quantum efficiency, readout noise, etc, but digital cameras are counting photons, and without increasing subject lighting, there is a finite limit to how many you can capture per unit of focal plane area. Shot noise is not some foreign disturbance to digital photography; it is a facet of its very substance.

However, shot noise is not the most immediate obstacle to greater DR with current sensors; sloppy readout noise at low ISOs is. Shot noise affects mainly the highlights at low ISOs, and highlights and midtones at high ISOs.

At the risk of adding one more variable into the mix, I'd be interested in some thoughts on how print size impacts perception of micro detail.

Michael indicates that I should expect to see the difference between a 1d2 and 1ds2 (8 vs 16 mpx) and consequently, obviously a difference between a 1d2 and a 22mpx sensor.

And equally clearly, I won't see any difference printed 5x7 or 8x12 or (probably?) 12x18. My prints from the mpx dvd don't demonstrate clear differences at 17x25 even between the 5d and p45 (at least to my aging eyes). I'm wondering, based on the best interpolation techniques available today, at what point up-rezzing begins destroy detail apparent in print to a skilled observer?

That may be true. Supposing a sensor manufacturer discovers a way of cutting read-out noise by 75%. We could have 4 small pixels occupying the space of one large pixel, say 4x3 micron pixels in place of one 6 micron pixel. The total read noise from the 4x3 micron pixels would be the same as the read noise from the one 6 micron pixel. The shot noise would be the same because the area the light falls on is the same. However, the resolution would be higher because of the greater pixel count.

Add to that a few other noise cancellation breaks-through, cancellation of thermal noise and perhaps a little photon multiplying and we could have something really special. I'm getting excited already .

Possibly not much difference between 16mp and 22mp in 35mm format, but it will be intriguing to see if there are differences, and if so, how much, between 22mp and 22mp in 35mm and 645 format. That will help tell us whether the smaller format has really reached its limits.

Here's a question for those of you who are more technically apt:

It seems that sensor size is being squeezed between the image that a lens can relay to the sensor, and the limit (on the small end) of how small pixels can be made. In other words, couldn't really have, say, a 36mp sensor using Canon 35mm lenses because the pixels would have to be so small that you'd get unacceeptable noise, or the sensor qould have to be so large that you wouldn't get complete coverage. So if 5 micron pixels, say, wind up being the smallest acceptable, how big a chip could you squeeze into a camera that uses 35mm lenses? Also, do 35mm lenses have design aspects that means that they are most effective (or only effective) when they are locked into one (horizontal) format? Or would it be possible to build a squarish chip (say, 6x7 format) into a 35mm camera, using existing lenses?

JC

The biggest difference between the full frame 35mm chips and the 645 chips can be illustrated by my image at:

http://www.photo.net/photodb/photo?photo_id=4251879

I have two cameras. One is a Canon 5D and the other is a Hasselbad with a Leaf Valeo 22 digital back. For this picture I had shot a number of hand held test shots with the 5D in the morning light. After studying the results of the hand held shots, I chose the composition that I liked. The next morning I then set up my tripod with the Hasselbald SWC 38mm Biogon with the Leaf Valeo 22 and made this shot.

Notice the detail in the bright, directly lit highlights outside and then the shadow detail in the walls of the room in deep shadow. The Canon 5D could not handle this dynamic range. The MF back captures 11 to 12 f/ stops of dynamic range. The chips with smaller pixels cannot do this.

While it is possible to do multiple shots with the 5D and use the HDR feature in Photoshop, it is not possible to capture this dynamic range in one click.

Ray M.

Bravo, Michael.

Nuff sed.

Wow, that paragraph got really mangled; I must have deleted part of the sentence. Let me try again:

However, shot noise is not the most immediate obstacle to greater DR with current sensors; sloppy readout noise at low ISOs is. Shot noise affects mainly the highlights at low ISOs, and highlights and midtones at high ISOs.

I've read some articles from people who have done astrophotography with IR-modified cameras. What they suggest, is that a cold night can reduce apparent noise in their long exposures. I really don't know the science behind it, but the rule of thumb is that the colder the camera, the more controlled the noise is... the hotter the camera, the more noise.

Maybe the next important development in noise reduction is finding a way to produce circuitry which makes less heat, or devise additional systems which cool the camera's internals.

One important point not mentioned in the article is the relationship between pixel size and lens resolution. Increasing the pixel count of a sensor of a given size by making smaller pixels provides little improvement in image quality once the pixels become smaller than the lens resolution. At this point, image quality is dominated by the lens rather than the sensor. This consideration has nothing to do with noise.

In this situation, a sensor with larger pixels will provide a higher resolution image at the same pixel count. (Of course the sensor will be larger, so a different lens will be required to give the same angular coverage, but let's assume both lenses have similar resolution.)

Even a perfect lens is limited by diffraction, which is a function of the lens aperture. A lens stopped down to f/4 for example will have a diffraction-limited blur of around 5 microns at the image plane. It makes a significant difference whether this image is digitized with pixels of size 2 microns or 5 microns or 10 microns, even if sensor noise is negligible.

Noise due to heat is not applicable to short exposures, at least with some cameras. Neither of my Canon DSLRs show any difference in noise in temperature extremes until the exposures get to be multiple seconds in length. The noise from long exposures is mostly stray, bright pixels, which repeat from frame to frame.

Even at short exposures, I have found that a 120 deg F camera will always have much more image noise (mainly chromatic) in the shadows than the same camera at 20 deg F. It is noticable.

That's not quite true - noise is directly related to the temperature of the sensor. This is thermally generated noise associated with the random re-combination/splitting of electrons from the material. The effect that you mention with respect to longer exposures and repeatability from frame to frame is more associated with variations in the actual pixel construction during fabrication. One of the hardest parts of sensor production is ensuring that all pixels are fabricated within a certain given tolerance (i.e. +/- 5% of required specification). Higher quality and larger sensors will typically have greater uniformity across the sensor and greater immunity from variations (i.e. there is an argument that MF sensors have perhaps +/-2% variance from pixel to pixel in terms of electrical characteristics whereas a point and shoot may be +/-20%). This variance across the sensor will render an image of lower perceived quality even though there is the same (big assumption) thermally generated noise and will be particularly evident in longer exposures.

You can also add in that all of the electronics will be calibrated for a given temperature (range) and operating outside of that range will give imprecise results - e.g. if you have a colour profile for room temperature than it will be slightly off at extremes of hot and cold.

We're a long way from DSLR pixels being smaller than good 35mm lenses can resolve. These Airy Disk diameters and diffraction spot sizes can be misleading. The question should be, 'Is the pixel sensitive enough, efficient enough and sufficiently protected from extraneous noise, to record fine detail that has lost most of its contrast?'; because that's what happens as you try to record finer and finer detail. The detail is there, it's resolved by the lens, but it loses a lot of its original contrast as a result of lens diffraction, opacity of the glass, reflections and other aberrations.

A lens that is truly diffraction limited at F4 has awesome resolving power. Even at 50% MTF it would resolve way beyond the capabilities of current sensors.

To get things in perspective, old-fashioned B&W film such as T-Max 100 is (was) capable of recording 50 lp/mm without any loss of contrast. (Ie, what the lens delivered, the film recorded at 100% MTF, up to 50 lp/mm). Above 50 lp/mm the film began to lose contrast, but at 100 lp/mm it could still record detail at 60% MTF.

Consider the result of using a lens diffraction limited at f4 with T-Max 100. At the Rayleigh's limit of 9% MTF, lens resolution is about 800 lp/mm; at 50% MTF it's about 200 lp/mm; at 75% MTF it's about 100 lp/mm (figures off the top of my head).

100 lp/mm at 75% MTF from the lens gets recorded at (60% x 75%=45%) on the film. That's quite a good result in my view.

Now, I don't want to get into a debate of film versus digital. We know that these results are going to be somewhat spoiled by clumps of grain and subjected to further degradation in the scanning process. The point I make here is, a sensor that could record 100 lp/mm with a loss of only 40% in contrast would finally be on a par with the best B&W film in terms of absolute resolution (but not in terms of over all image quality, obviously. Resolution isn't everything.)

No you didn't. You just missed a full stop. I understood you perfectly because this is not a computer program .

I tend to agree that lowering read noise of small pixels is a major obstacle to further imrovement in digital image quality. We appear to be stuck with photonic shot noise, as we are with lens diffraction (although this is something that nanotechnology is already addressing in the laboratory). Thermal noise seems to be more of a problem with the larger pixels and/or the longer exposure, which, for equal DoF the larger format also needs. Extremes of temperature difference might also produce a noticeable difference with short exposures, as Boku has found, so there's probably some scope for thermal noise reduction technology even in the small camera.

The quantum efficiency factor is also an avenue for further improvement. We can never reach 100% efficiency, but technological progress seems to be directed at improving the efficiency of just about everything. The 2 litre engine in my Daewoo wagon is both more powerful and less gas hungry than the 2.4 litre engine in my previous car.

Ray,

It is true that MTFs multiply, but not so simply as you show above. For details, see Norman Koren's website. Before you can perform the multiplication, you need to do a Fourier transform to convert from the spatial to frequency domain. You then multiply the components and then you must apply a reverse Fourier transform via a complex convolution algorithm to go back from the frequency to spatial domain.

http://www.normankoren.com/Tutorials/MTF.html

For MTF at the Rayleigh limit, the equation 1/r = 1/r1 + 1/r2 is often used, where r = system resolution and r1 and r2 are system components (lens and film, for example). If the lens resloved at 800 lp/mm and the sensor at 100 lp/mm at Rayliegh, the system resolution would be about 89 lp/mm.

Bill,
Did I write 800 lp/mm as the Rayleigh's resolution limit of a lens diffraction limited at f4? You should have corrected me . It is of course 400 lp/mm, but maybe still 200 lp/mm at MTF 50% and possibly 100 lp/mm at 75% MTF.



I'm aware of this formula but don't know how to apply it for digital photography. Sensor manufacturers don't give us resolution and MTF details as film manufacturers do or used to. But generally, it seems to me that using that formula to generatet a lower sytem resolution than the component resolutions at the same MTF is roughly equivalent to multiplying the MTF responses to generate the same system resolution as the component resolutions (if they are both equal) but at a lower MTF.

In other words, if lens and film independently both produce 100 lp/mm at 50% MTF, then the system resolution is 50 lp/mm at 50% MTF (1/100+1/100=1/50).

Multiply the MTFs and you get 100 lp/mm at (50%x50%)=25% MTF, which might be a good approximation if one is considering high MTFs and provided the MTF fall-off for both lens and sensor is fairly even. The straight line MTF response for the 10D system resolution, on Norman Koren's site, would suggest to me that multiplying MTFs of equal component resolutions is a good approximation.

Ray, you mentioned the figure 800 lp/mm and I assumed incorrectly that you were referring to Rayleigh. Using the correct figure of 400, the system resolution becomes 80 lp/mm rather than 88. At such high lens resolutions, one is reaching the point of dimishing returns and the system is primarily limited by the sensor.



As I understand things, the simplified formula only works for MTF around 10% or less, and it is using spatial resolution in lp/mm not the MTF. For MTF of 50%, you would have to use the Fourier transform and convolution in the case of unequal component resolutions, but if the resolutions are similar perhaps you can multiply the MTFs as you suggest. This is beyond my expertise, and perhaps an expert can reply.

I am quite worried about this article because when reading it, one will believe that a 20MP 35mm sensor will have inherently more noise than a 20MP 645 sensor... which is not true, of course (this is the usual 'Sensor centric' or 'Same ISO comparison' bias).
A better picture is that Signal/Noise Ratio implies the sensor AND THE LENS... and the consequence is that for a given Field Of View, it is the lens' diameter that set the Signal/Noise Ratio, not the Sensor's size (ie: where are the 35mm f/1.4; 85mm f/1.2; 200mm f/1.8; 600mm f/4.0... equivalents in the 645 world???).

Bottom line: "The larger the sensor, the better" is probably worse than "The more MP, the better"!

Olivier
PS: I am not saying that MF backs don't make sense. On the contrary, I believe they deliver because they are optimized for quality: CCD Full Frame, Sensor design, no microlenses, Dynamic Range, etc...

I've never seen a satisfactory explanation as to why large format digital backs perform relatively poorly at high ISOs, compared with FF 35mm sensors just half the size. Join 2x1Ds2 sensors together and you have a 32MP digital back which would be very usable at ISO 800. At ISO 1600, noise is slightly greater (in the 1Ds2) than the 5D, but resolution is better. Higher ISO does slightly degrade resolution and I imagine this also applies to the large sensor digital MF backs.

Given the same manufacturing process technology, this is most likely true.

If you have evidence that suggests otherwise, you should definitely produce documentation for it, because your claim goes against generally accepted knowledge in the field.


This has nothing to do with signal/noise ratio in the sensor.

This has something to do with signal/noise ratio in the system.

If you want to compare systems, you have to take that into account, yes, but if you want to compare sensors, you have to try to eliminate the system dependent issues.

Fortunately, there are ways of comparing different sensors in the same sensor generation on systems that otherwise are very similar, because camera manufacturers occasionally release models within the same model line at the same time.


I've never seen a comparison, even, because medium format digital backs don't seem to offer ISO 800, 1600 or 3200.

ISO 400 comparisons are possible, however, and they do show a relatively poor performance vs. CMOS based 135 format sensors.

Then again, Nikon's CCD sensors also show relatively poor high ISO performance compared to Canon's CMOS sensors.

It appears that it's CCD vs. CMOS all over again, doesn't it?

Jani,
The Leaf Aptus 75 has ISO 800, but it's a pretty degraded image compared to ISO 50 and 100. The CCD sensor must have some qualitative advantage over CMOS at low ISOs, I guess. Two joined 1Ds2 sensors would produce a sensor of the same size as the Dalsa 33mp sensor in the Aptus 75, which I believe actually consists of 2 joined smaller sensors which occasionally reveal the join line vertically down the middle, according to a current thread on this very problem.

You can't talk about Signal/Noise in the sensor independantly from the whole system, as there is no Signal in the sensor considered separately.
You can talk about Noise generated in the sensor, but Noise alone doesn't provide any information about the noise in your picture.

The 'generally accepted knowledge' is that under the same illumination, a larger photosite will receive more photons and have a better Signal/Noise.
I agree: the Signal will be higher while the Noise will have little variation. I would even say that smaller photosites have actually LESS Noise generated in the sensor (more about noise here) and this is exactly why I consider that saying "a bigger pixel will have less noise" is not true... (an oversimplification, at best).

As I said, there is little meaning in comparing different systems without taking into account the optics (current lenses and/or optical limitations implied by the system).
I will make a comparison: people now understand that resolution cannot be determined by Sensor's definition (MP) alone and that the lens has to be factored in. The big difference here is that the theorical resolution based on sensor's definition has a meaning... whereas the Signal/Noise considered independantly for the sensor alone has NO meaning at all.

Now, about that 'Same ISO comparison bias': people tend to compare Signal/Noise under a same illumination and say "everything else being equal". What this really mean is considering an Olympus 150mm f/2.0 on '4/3', a Nikon 200mm f/2.0 on DX, a 300mm f/2.0 on 24x36 (huh? where is this one???) and a 450mm f/2.0 on 645!!!
That "everything else being equal"(ISO/Illumination) is in reality the most absurd way to compare things: weight, price, Depth Of Field and even existence of products is not even comparable due to optics consideration.

Olivier
PS: in addition, here is an excerpt of an email I sent to Michael after reading his article "Industry Push Pull" 1 year ago:

Duh, it was right there in the review I linked to, too.

The quality at ISO 800 isn't bad, just not up to the standards that medium format photographers might like. It seems to be useful at approximately the same level as ISO 800 is on the 20D (where image quality is pretty degraded, too), but it's hard to tell without a side-by-side comparison. Unfortunately, other system differences might get in the way of testing this properly.


It could also be that CCD sensors are easier to produce at such a large size.

That was one of the arguments against CCD "full frame" 135 format sensors, before Canon proved them wrong.


Well, maybe that explains one of the rumours for the 1Ds3 having a "double frame" size.

How many Canon lenses have a 60mm image circle?

None, AFAIK.

But my (subtle) point wasn't that Canon would be making a "double" sensor for the 1Ds3; on the contrary, I was trying to point out that the rumour might have confused a few things.

I beg to differ.

The sensor receives and converts a "signal" regardless of whether you have a lens cap, a lens or an AA/UV/IR filter.


Yes, I -- and most of the regulars here -- are well aware of that.

However, when someone uses the term "noise" without any qualifiers in digital photography, he or she most likely means visible noise, which is an indicator of the signal/noise ratio.

While it's not technically precise, it's the generally accepted term.

I can rant and rave about how stupid the use of e.g. "pathetic" as a derogatory term is in today's English, but it doesn't mean that I'll get any sympathy.


So far, so good.


Well, considering that a pixel only has a physical size on a screen, it's not even an oversimplification, it's a meaningless statement.

However, given the common usage of "noise" I mention above, saying "a bigger photosite will, in the same process technology and generation, have less noise ((compared to signal))" is quite reasonable.


You're missing my point.

Take two cameras with sensors from the same process technology and generation, with the same noise reduction system, but with different photosite sizes, from the same sensor manufacturer.

Use the same lenses, subject matter and conditions.

Compare.

If you don't worry about comparing possibly different noise reduction systems, you should be able to find something. I suggest the following cameras:

- Nikon D1H
- Nikon D1X

Well... when I said "a bigger pixel will have less noise" I used the simpler 'joe average' vocabulary... but it is like "a bigger photosite will have less noise ((compared to signal))" and I agree that it should be compared at the same technology level. Therefore, your statement is equally wrong... (sorry).

Well, I think exactly the same: it seems you focused on use of words, whereas my original message was about what's going on... and not nitpicking about vocabulary, really.

Before using real examples, we should first understand what we want to look for... and I believe the D1X is not exactly a good example of simplicity.

Olivier,
I find your reasoning a little confused and don't quite know where to begin, but I see Jani has done a good job already .

When comparing image quality and image quality alone, getting everything equal without exception is of course absurd. The only things we need to get equal (as far as possible) are the 'direct effects' on image quality, such as quality of lens, choice of f stop for equivalent DoF and choice of focal length for same field of view.

If the physical size of the sensors being compared are different, then it's not meaningful to use the same lens and the same f stop. Even after adjusting focal lengths and f stops to get the same DoF and FoV, it may not be appropriate to use the same ISO. The larger sensor will require use of a larger f stop # for equivalent DoF and consequently a slower shutter speed at the same ISO.

In situations where a fast shutter speed is required, one might then make the observation that the smaller sensor provided better image quality with less noise.

Jani,
Below is a comparison of the leaf Aptus 75 at ISO 50 and ISO 800. The crops are very small crops from a much larger image at http://www.alpa.ch/en/focus/focus.html



The question arises, would a 1Ds2 shot of the same scene at ISO 400 provide better detail, bearing in mind that noise is reduced as downsampling takes place?

Ah... this is the 'Equivalence' I would like to see used (Equivalent Focal length, f-stop and ISO based on the crop factor => identical FOV, DOF, Shutter speed, Signal/Noise at the same level of technology and comparable lens weight/price).
The "new" thing I emphasize here is with that "fair" comparison, you get the same Signal/Noise Ratio whether you use a 20MP 24x36 sensor or a 20MP 645 sensor.

I am trying to communicate the message that a small format system is not inherently doomed with noise (cf: my "Noise is lower in a smaller photosite") and that it all depends on the Optics (available lenses or limitations). As an example, the 24x36 is probably the best performer regarding Signal/Noise due to the availability of very fast lenses.

Unfortunately, the flawed 'sensor centric'/'same ISO' point of view is widely spread and leads to wrong conclusions (like: superiority of large formats regarding noise and the evolutionary dead-end of smaller formats). Therefore I am trying to foster discussion and understanding about this very issue.

Olivier

It depends on how small you want to compare with how big. For me there's a sense of deja vu here because some time ago I irritated a few people with my views on big versus small and got into trouble. I'm not sure I want to revisit this issue.

Anyway, here's what I still think, whether right or wrong.

1. A lens is most efficient when it's used at a diffraction limited aperture because at that aperture no further improvement is possible.

2. For equal FoV and DoF of the same scene under the same lighting conditions, the smaller sensor receives less light. Ie. fewer photons to describe the scene being captured.

3. There are many sources and types of noise that affect sensors, but the one type we can do nothing about, apparently, is photonic shot noise, the random arrival of the photons at a particular location (the photodiode), which apparently has a Poisson distribution. (I just threw in that to give the impression I know what I'm talking about ).

4. The amount of photonic noise in a given exposure is calculated as the square root of the total number of photons impinging upon the sensor.

5. The image that is 'comprised of, results from' the smaller number of photons (the image from the smaller sensor) contains a greater proportion of the one type of noise we can do nothing about.

Conclusion: If we were able to eventually reduce all types of noise to virtually zero, except photonic shot noise which seems to be in the same category as diffraction in lenses, we would find that the smaller sensor has an inherent disadvantage with regard to noise and image quality.

There are also other issues of dynamic range and resolution where we can come to an impasse earlier (determined by the laws of Physics) with the smaller sensor. The size of photodiodes is limited by the wave length of light. You can always fit more pixels on a larger sensor.

I'll focus on this one...

Let's consider formats A and B where B has a 2.0 cropping factor (24x36 and '4/3' are quite close, except aspect ratio):
- To get the same Field of View, you need to use a 2x longer focal on A compared to B.
- To get the same Depth of Field, you need to use a f-number x2 (close 2 stops) on A compared to B (you can use DOFMaster to check that. It is based on the fact that DoF formula is proportional to f-number, Circle of Confusion (which is proportional to the format for identical critera on output) and 1/(focal length)^2 => coef x coef / coef^2).
- We consider exactly the same scene/lightning and we use the same shutter speed.

So I put a 300mm f/4.0 on A, used at f/4.0, 1/500s and 400 ISO and a 150mm f/2.0 on B, used at f/2.0, 1/500s and 100 ISO.
=> FoV, DoF, Shutter speed are the same and I use the same Exposure Value as (1/500s & f/4.0 & 400 ISO) = (1/500s & f/2.0 & 100 ISO).

With the same number of pixels, each photosite of A is 4x larger compared to photosites of B. But to keep the same DoF, the lens on B is opened more by 2 stops (4x more photons per area) => photosites of B gets exactly the same amount of photons as photosites of A.
Bottom-line: settings used to get the same photo (FoV, DoF, Speed) imply that Signal will be identical in both formats.

About Noise: Read Noise and Dark Current Noise will be very similar for both formats (the previously linked document hints that the smaller photosites will have less of those Noises... this is actually supported by the Kodak CCD FFT sensors specs and by a friend Physicist/Researcher working in the energy/matter field). Shot Noise will be identical, as we have the same Signal for A and B.

=> If you compare different formats based on the same photographic criteria (FoV, DoF, Speed...) you end up with the same Signal/Noise ratio, and therefore the same noise in your picture, provided the same technology level and settings.
(This doesn't mean that all systems are equals: you need to find out their limitations. This is what I did in this excel sheet. You then have to include existing specifics: current level of technology, choices/orientations taken, etc...)

Olivier

Olivier,
Good attempt! But I can summarise your argument as follows. A large bucket that is half full holds as much water as a small bucket that is completely full.

A large sensor will always receive the same light as a small sensor if you allow the same amount of light through the lens. That's obvious but needs mentioning.

Your argument could also be used to claim that 35mm film delivers approximately the same image quality as 8x10" format. DoF at F8 on 35mm film is equivalent to f64 on 8x10" format. To get the best, finest grain, lowest noise, highest resolution result with 35mm color film, I'd use something like Royal Gold ISO 25 or Ektar 25.

If I were to use the same type of film with 8x10", there would be no question that the large format would produce vastly superior results. However, if I insist on letting the same amount of light through the lens at f64 (ie. same shutter speed), I'd either have to push process the film by a ridiculous amount or use an ISO 1600 film. Either way, results would be considerably degraded.

Ray,

Well...
I used your explanation for comparing with same FoV and DoF and just added your own suggestion on using the same shutter speed in order to slightly correct that statement...

Thanks. Actually, I am pretty sure 99% of photographers don't consider my previous message as obvious... and just mix everything (Signal, noise, Sensor's size, lenses...) in that flawed and simple credo: "A bigger photosite is better" I can read everywhere.

By the way, I propose a much better one: "A larger lens is better"! (and really: it's quite accurate as the lens diameter for a given FoV will be rather well related to the amount of light reaching the sensor... whatever the sensor's size)


About your film comparison (who won? 25 ISO on 35mm or 1600 ISO on 8x10"? ):
My message was on the noise that plagues 95% of photographers: when you don't have enough photons to get a "good" picture. This is also the issue raised by Michael in 'Sensor & Sensibility' and other articles. And as I explained, this is related to the lenses you can use with the system... not to the sensor's size.

If you consider the optimal situation where the photographer has enough light (ie: tripod, etc...), the issue is quite different and is related to the limited Dynamic Range and impact in the shadows. I agree that sensor's size is quite important here...
But this represent 5% of photographers (ok: they are all on this board... ) and, more important, unlimited Dynamic Range may become available soon, whatever the sensor's size... => that *real* advantage of larger sensors regarding Noise will vanish in the mid/long term.

Olivier

To nitpick: no, it's not better, because it depends on other factors, too.

"All generalisations are bad, including this one."

BTW, I see that you didn't answer my point regarding the Nikon D1H and D1X.

...Ok: it's not better... it's less bad...
Well, I read that the D1X has actually a tricky 11MP sensor... ...not exactly the best situation to compare things.

Olivier

Wow, that was new information for me, too.

But that's the D1, not the D1H or D1X.

Too bad he doesn't mention what they did with the D1H and D1X.

Olivier,
From a purely technical point of view, we could say that image quality generally improves as the amount of light available to describe the scene increases. There are exceptions for esthetic reasons. For example, a very long exposure on a moonlit night (with a low-noise Canon DSLR) will produce a remarkably high quality image, almost as though it was taken in full daylight, but might destroy the mood of the scene.

The larger format camera (and it is assumed the size of the lens is going to match the size of the sensor) allows one to use more light when taking the shot. Of course, one can choose not to let more light through the lens for the sake of a faster shutter speed, whatever the size of the sensor, assuming one has manual control.

My impression is, when you compare a small format digital camera with a larger format under conditions of same DoF, same FoV, same shutter speed, the camera with the better or most technological innovation wins. The larger camera is generally more expensive, bigger, heavier and can therefore 'hold' more technological wizadry.

For example, an 8mp P&S such as the KM A200 with a sensor diagonal of 11mm has about a 2.5 stop advantage compared with a Canon 20D. To get the same DoF and shutter speed with both cameras, you would need to use ISO 600 on the 20D (if that were possible) in place of ISO 100, or ISO 300 in place of ISO 50 on the KM A200. My guess is, the 20D would still produce the better image in these circumstances, especially when you take into consideration the 'real' ISO values. I believe ISO on the 20D is conservative. ie. the ISO 400 setting is actually ISO 500.

It would be interesting to compare the Leaf Aptus 75 at ISO 800 with the 1Ds2 at ISO 400, after adjustments to get the real ISO values comparable. Since the Aptus 75 has double the number of pixels and higher resolution, the differences might not be as great as one might imagine. Image quality might be very close. If image quality were still worse from the Aptus, one could lay the blame on a lack of on-chip processing, which is much easier with the CMOS type sensor and a major advantage.



I can't see any logic here. Whatever the size of the lens, equal DoF and equal FoV means equal aperture diameter. The amount of light the lens lets through is then governed only by the shutter speed in these circumstances. For example, the diameter of a 50mm lens at f8 is 50/8=6.25mm. The diameter of an equivalent lens at f64 for 8x10 format is 400/64=6.25mm).

In summary, one could say that the inherent advantages of the small format camera are compactness, light weight and low price. But nothing much else.

Why is it assumed that the smaller format must be equipped with a smaller/lesser lens? My point is that we should break those 'rules' to get a better picture.
What are the equivalents of 35mm f/1.4, 85mm f/1.2, 200mm f/1.8, 400mm f/2.8, 600mm f/4 or 1200mm f/5.6 in the Medium Format?

I rather think a manufacturer that has large volumes can put more technological wizardry in its products, as Canon does (its CMOS, noise, processing power, etc... is quite exciting) whereas medium backs are specialized products optimised for high quality photography: they use higher quality components and can avoid degradations required for other types of shooting (handheld low light, speed...) like the use of microlenses, the fast read-out speed, etc...


About different systems comparisons: by using the 'same FoV, DoF, Speed' equivalence (ie: same amount of light, different ISO), I compared samples from a lot of digicams/DSLR and found that the sensor used in the Fuji F30 (6MP 1/1.7") is quite efficient, the latest Sony 1/2.5" and 1/1.8" are close and the Canon CMOS used in APS format are also very good.
I believe the sensors used in the 1D2/1Ds2 (and even 5D) are not quite so light-efficient, Nikon/Sony DX as well as sensors used in '4/3' are behind. Kodak/Dalsa sensors for MF are probably even less light-efficient as it is not their priority (ultimate quality is).
This is not exactly a scientific test, but it gives some interesting trends: light efficiency is not related to sensor's size, but rather to the R&D/Production effort the manufacturer put into it (ie: the level of technology). There is no limit at 5 or 6.8 micron... as the excellent results from the 2.6 micron Fuji F30 shows (OK, This 1.75 micron CMOS sensor is probably pushing the envelope at this time).

Ray:
- Do you agree that same FoV, DoF and shutter speed give the same image (or 'as close as could be') and same Signal/Noise whatever the sensor size?
- Do you agree that for a given FoV, when light is lacking (ie: camera or subject movements), you will be better off with a lens of large aperture diameter on APS rather than with a lens of smaller aperture diameter on 645 to get a picture with little noise/blur?

The logic is to replace the usual "a larger sensor has less noise" which is not true by "a larger lens has less noise" which is really more accurate in those situations.

Olivier

The entrance pupil by itself has nothing to do with how much light is used in an exposure. What is important is the effective aperture, usually given in f-numbers - this is important as it is going to produce the intensity of the signal. That is the factor that the International Standards Organization uses to determine exposure for noise and saturation based ISO. The entrance pupil does not indicate the amount of light used in exposure without the distance to the image plane; what goes in does not matter unless the distribution (magnification) is taken into account. That is why effective aperture is the thing.

Physically big lenses don't mean anything except they are big lenses. And most likely expensive.

I am sorry if I misunderstood the point you folks are trying to make - it is hard to follow what you are talking about.

Anon, I think you've hit the nails on the head: nail #1: indeed, the amount of light reaching the sensor depends on more than just the size of the lens - distance from the lens to the sensor is very critical; nail #2: indeed alot of this discussion is hard to follow, most likely because there may be a large number of non-sequiturs in it, possibly due to the comingling of issues and the elements relevant to one issue but perhaps not the other. One needs to distinguish carefully first of all what one is talking about: noise or resolutiion, or both, or be explicit about how they interact if they do, then sort out clearly in what ways lenses or the sensor is a more determinative factor in the one than in the other. Some such sorting out of the discussion by those who seem to understand it would help the rest of us.

Well, I must battle on .

Olivier,
It's not essential that the smaller format be equipped with a smaller lens. It's just more sensible and more efficient. You wouldn't get far using a 35mm lens on a 2/3rds P&S camera (assuming you could fit it). The lenses used for P&S cameras need to be higher resolving than 35mm lenses because the pixels and sensors are so small.

Likewise, you could use your large format camera and lenses to take MF size images, but you would not get the best of results. It seems to be a fact of lens design that absolute resolution (lp/mm) is compromised in lenses that are optimised for a large image circle. A 400mm lens designed for 35mm format is generally much higher resolving than a 400mm lens designed for LF.

The equivalent MF lenses for 35mm f1.4, 85mm f1.2 and 200mm f1.8 would be approximately 70mm f2.8, 180mm f2.5 and 400mm f3.5, but you are probably not going to get a 400 f3.5, or an 800 f5.6 because they would simply be too heavy, expensive and cumbersome. Even the 1200/5.6 for 35mm has to be manufactured on demand and is ridiculously expensive and heavy.



I don't think so. Are there any laptops that have the same performance as the bigger and heavier workstations? Miniaturisation without quality compromise is terribly expensive.



I've never come across any noise specifications for lenses, so I don't know how to respond. Lens resolution is either limited by various types of aberration, such as coma, spherical aberration, chromatic aberration, etc.... or diffraction.