Does it make sense? An $1800 Full frame DSLR from Canon

[Jon Ruyle]

Thanks for taking the time to explain this, Daniel.

For an example, take a comparison between pixel sizes of 4 microns and 2 microns:

* If the 2um pixel has read noise that is twice as bad, then the final image has the same read noise.
* If the 2um pixel has read noise that is the same, then the final image has much less read noise.
* If the 2um pixel has more than twice as much read noise, then the final image has more read noise.

I agree with all of this, assuming by “twice as bad” you mean snr is twice as bad, not absolute noise is twice as big (ie, you don't mean sd of a Gaussian distribution of number of electrons of read noise is twice as big for the smaller pixel)

So smaller pixels can be noisier per pixel (by an amount equal to the square root of the difference in size) to achieve the same amount of noise as larger pixels. My position is that, all other things being equal, smaller pixels do tend to be just about that much noisier, and so the final image is just about the same.

Okay, so you're asserting that other things being equal, read noise is inversely proportional to the square root of the size of the pixel? That this is true for photon noise is clearly true, and requires very little knowledge of how the camera works. That it is true for read noise seems less obvious to me (as I said, I would have thought reading a pixel would have caused the same amount of noise no matter how big the pixel is) but I believe you []

[Daniel Browning]

You're welcome, Jon.

I agree with all of this, assuming by “twice as bad” you mean snr is twice as bad, not absolute noise is twice as big

Yes, that's what I meant; I'm glad you saw through the obtuse writing.


One would naturally think that read noise must shrink in linear proportion to the area decrease in order to keep the same read noise. But the fact that noise adds in quadrature is what makes it possible for read noise to shrink at a slower rate while retaining the same "image level" noise (i.e. noise power for any given spatial frequency).

Okay, so you're asserting that other things being equal, read noise is inversely proportional to the square root of the size of the pixel?

How about "read noise is directly proportional to the square area"? (e- / um^2.) I think that works.

That this is true for photon noise is clearly true, and requires very little knowledge of how the camera works. That it is true for read noise seems less obvious to me, but I believe you

Read noise is indeed the much more complicated, inconsistant, and difficult to generalize. However, I think the rule is a pretty close fit to the cameras at hand. It's more often true for cameras of a similar sensor size and budget from the same manufacturer (e.g. 40D and 1D3), but not always. Unlike photon shot noise, read noise varies significantly from model to model, manufacturer to manufacturer, and even unit to unit of the same model. Generally, though, one of the exceptions seems to be that when comparing tiny inexpensive digicams to large, pricey DSLR, the DSLR has less read noise at high ISO, while the very tiny pixels in the cheap digicam have less read noise at low ISO. And yet high ISO read noise has dropped with each new and smaller pixel, and other DSLR seem to also indicate that it isn't something inherent to pixel size specifically, but another factor (design tradeoff?).

If we look at the example of the LX3 vs 5D2 at low ISO again:

5D2 6.4 microns vs LX3 2 microns using signal of 1 * N.
6.4um S/N = 23.5:23.5 (1:1)
2um scale factor = (6.4/2.0)^2 = 10.24
2um S = 23.5/10.24 = 2.2949
2um N = 5.6
2um S/N = 2.2949:5.6
2um resampled S = 2.30*10.24 = 23.5
2um resampled N = sqrt(5.6^2 * 10.24) = 17.92
2um resampled S/N = 23.5:17.92 = 1.31:1
31% better S/N.

Here's a visual example of a base ISO comparison that contains quite a bit of read noise (pushed from ISO 100 to ISO 13,000 in post).

http://forums.dpreview.com/forums/read.asp?forum=1018&message=28607494

Pattern noise is another important factor to consider. The 5D2, for example, has very strong and objectionable pattern noise. Yet even the inexpensive Canon digicams have no pattern noise. So while the RMS read noise number may indicate the 5D2 has a certain amount of read noise, it's actually worse than the indicated number because the pattern noise makes it objectionable to the eye.

It's also worth mentioning that the OLPF means that when the small pixels are resampled to the same pixel size as the large pixels, they will have more detail. In other words, for comparisons near the Nyquist frequency of the large pixels, small pixels have a big advantage in resolution and contrast. But if the desired spatial frequency is below the strongest effect of the AA filter, this difference is not relevant.

We have a lot to look forward to in the future.

[dmckinny]

Daniel,


Thank you for all the excellent information. I think you hinted at a couple of impacts in your last post, things that I was thinking about while reading through the entire thread. I thought one of the 'hidden' assumptions was that all pixels were designed the same way. By this I mean the circuitry of the pixel itself. If that is true then you can compare the relative performance of pixels of different size. However, I think that both the circuit design and the manufacturing processes are constantly improving. With these improvements we are seeing smaller pixels, and higher densities on the sensors, while at the same time seeing either no degradation or sometimes actual improvement in performance. Throw in micolens design changes and it becomes very difficult to compare various sensors and/or their pixels to each other.


Now I'm going to lay down for a while and get ready to shoot some soccer this weekend. []

[Jon Ruyle]

Hi Daniel,


Thanks again. Things are clearing up for me a bit.

Okay, so you're asserting that other things being equal, read noise is inversely proportional to the square root of the size of the pixel?

How about "read noise is directly proportional to the square area"? (e- / um^2.) I think that works.

How embarrassing. I meant "read noise (in absolute terms) is directly proportional to square root of area". If read noise was directly proportional to area, high pixel density would be better (because noise adds "in quadrature", as you say... so 4 times as many pixels 1/4 the size and 1/4 the noise gives half the noise when you add it all up... 4 times as many pixels 1/4 the size but with half the noise, though, is the same.) I *think* that's right []

Read noise is indeed the much more complicated, inconsistant, and difficult to generalize. However, I think the rule is a pretty close fit to the cameras at hand.

This makes sense. I was a little baffled by the idea that there may be some reasoning with which one could come up with "read noise is proportional to square root of area".

Generally, though, one of the exceptions seems to be that when comparing tiny inexpensive digicams to large, pricey DSLR, the DSLR has less read noise at high ISO, while the very tiny pixels in the cheap digicam have less read noise at low ISO.

I'll take low read noise at high iso any day


I measured the read noise of my 5DII (not numerically, just by taking a picure and looking at it) and, as far as I can tell, it is extremely low anyway until iso is high.


The reason I measured it was that for a long time I thought there was no point in very long exposure photography: I reasoned it is better to take many images and stack (equivalent signal and noise, but far easier for a number of reasons). But then I realized read noise is worse when you stack. But after taking my dark pictures, I concluded that at iso3200 and below, read noise is pretty much negligible unless your signal size is minute.

The 5D2, for example, has very strong and objectionable pattern noise.

True, though only regularly visible by me above iso6400. At iso3200, I can't see it (in real pictures, that is). But IMO pattern noise is the worst part of very high iso images on the 5d2.

We have a lot to look forward to in the future.

Agreed. I'm very happy with my 5D2, but I think I should start saving for the 1DsX now. []

[Daniel Browning]

I'll take low read noise at high iso any day

Me too.

I measured the read noise of my 5DII (not numerically, just by taking a picure and looking at it) and, as far as I can tell, it is extremely low anyway until iso is high.

For a typical tone curve, that's normal. When 50,000 photoelectrons are flooding the camera, such as the bright exposure zones in a low ISO shot, the 23 electrons of read noise seem like nothing.


It's only when you try to maximize the dynamic range of the camera that you run into the limitations of the read noise, because it's in the extreme shadows where the number of photons gets closer to the number of electrons.


ISO 1600 has 10 times less read noise than ISO 100. If Canon could get ISO 100 down to the same level as ISO 1600, dynamic range would increase greatly and 14-bits would no longer be sufficient.

True, though [pattern noise is] only regularly visible by me above iso6400. At iso3200, I can't see it (in real pictures, that is

This is another low ISO dynamic range thing. Typical tone curves don't show it because they crush the blacks and don't exploit the full dynamic range of the camera.


If it weren't for pattern noise, I would be able to use another 2 stops of dynamic range (say, 12 instead of 10) at ISO 100. But I can't because the pattern noise ruins it. I can see into the promised land, but I can't enter.

[Jon Ruyle]

If it weren't for pattern noise, I would be able to use another 2 stops of dynamic range (say, 12 instead of 10) at ISO 100. But I can't because the pattern noise ruins it.

Pattern noise at iso 100? The darks must be *dark*. I mean, I can't see anything if I take a total black picture at iso 100 (by total black I mean lens cap on, shutter speed 1/8000 and f/22.) If I then adjust the exposure upward, I see a few spots, but I don't recall anything I would call a pattern.

I can see into the promised land, but I can't enter.

I laughed when I read this. Did someone promise you a land of 12 stops of dynamic range? []

[Daniel Browning]

Pattern noise at iso 100? The darks must be *dark*. I mean, I can't see anything if I take a total black picture at iso 100 (by total black I mean lens cap on, shutter speed 1/8000 and f/22.) If I then adjust the exposure upward, I see a few spots, but I don't recall anything I would call a pattern.

Your raw converter is not set for high dynamic range. There is no such thing as black in a raw file. It just goes from white to noise. The point which is chosen as "black" is entirely a creative choice, though some raw conversion software does not allow the photographer to choose the true raw black point.

Strong contrast and little noise is the most pleasing conversion for most photographs and matches the taste of the most photographers, so the default settings on most raw conversion software are tuned to provide that by setting the black point far above the noise floor and using a tone curve that crushes the shadows.


In other words, the most common conversion is tuned for high contrast and low dynamic range.

If one is interested in utilizing more dynamic range than the typical "high contrast/crushed blacks" conversion, and the photographer chooses their own black point, gamma, and tone curve, it becomes possible to see down to the point where there is nothing but random noise. On some cameras, such as Canon DSLR, the pattern noise becomes a problem long before random noise is reached. Pattern tends to get magnified even further in most raw converters, perhaps because of how it interacts with the edge detection algorithms.


In Lightroom, for example, you can set a "linear" tone curve, "0" blacks, then set shadows to "100" and you should be able to see down into the pattern noise.


IRIS is an exmaple of a raw converter that allows the photographer to have complete control over the black point and true linear exposure compensation. 5.57 is the latest version and it can read 5D2 files. Here's how to do a simple conversion:

File->Open Raw
On the toolbar, find the little icon of the camera, and set it to your model (e.g. 5D2).
Processing->Subtract (enter 1024 for 14-bit Canon cameras such as 5D2, 128 for the 12-bit cameras).
Digital photo->convert a CFA image
To white balance, find a portion of the image that is white, drag a square in it,
Open the console (also an icon on the tool bar) and enter "white"
View->Logarithmic
File->Save (Tiff)

Then open the tiff in Photoshop and analyze the deep, deep shadows. You should see noise (and pattern noise): nothing should be black.

Most people think high ISO has more noise than low ISO, but in the most literal sense that's untrue. It only seems true because photographers tend to reduce light for high ISO and increase light for low ISO. The noise is not caused by the increase in ISO, but the *decrease* in the amount of light. Read noise is reduced at high ISO.

For example, one chooses the f-number for a certain depth of field and shutter speed based on the desired amount of motion blur. The next setting to choose is ISO. Lower ISO for more highlight headroom and more shadow noise. High ISO for less shadow noise and less highlight headroom. In Canon DSLR cameras, that shadow noise includes pattern noise. If, at ISO 100, important highlights are *almost* about to clip, it would be unwise to increase ISO because they would be lost. But if the shadows were more important, then ISO 1600 might be chosen: sacrificing four stops of highlight detail in order to get less noise in the shadows.

Did someone promise you a land of 12 stops of dynamic range?

We hold these truths to be self-evident, that all photographs are created equal, that they are endowed by their Photographer with certain unalienable Qualities, that among these are Color, Contrast and 12 stops of dynamic range.

That to secure these rights, Raw Converters are instituted among Men, deriving their just powers from the consent of the users, That whenever any Form of Software becomes destructive of these ends, it is the Right of the People to alter or to abolish it, and to institute new raw conversion, laying its foundation on such principles and organizing its powers in such form, as to them shall seem most likely to effect their Dynamic Range and Contrast. Prudence, indeed, will dictate that Raw Conversions long established should not be changed for light and transient causes; and accordingly all experience hath shewn, that mankind are more disposed to suffer, while evils are sufferable, than to right themselves by abolishing the software to which they are accustomed.

[]

[Keith B]

Seriously, who is this Daniel Browning?


Is no one else frightened.


Like Doc Brown from Back to the Future.

[ShutterbugJohan]

We hold these truths to be self-evident, that all photographs are created equal, that they are endowed by their Photographer with certain unalienable Qualities, that among these are Color, Contrast and 12 stops of dynamic range.

That to secure these rights, Raw Converters are instituted among Men, deriving their just powers from the consent of the users, That whenever any Form of Software becomes destructive of these ends, it is the Right of the People to alter or to abolish it, and to institute new raw conversion, laying its foundation on such principles and organizing its powers in such form, as to them shall seem most likely to effect their Dynamic Range and Contrast. Prudence, indeed, will dictate that Raw Conversions long established should not be changed for light and transient causes; and accordingly all experience hath shewn, that mankind are more disposed to suffer, while evils are sufferable, than to right themselves by abolishing the software to which they are accustomed.

LOL! :-) I currently shoot JPEG (because I have a very slow computer and lousy software; when I get a faster computer and better software, I will switch to RAW).


A question: Most of this thread has been a discussion of APS-C vs. FF sensor size noise and signal-to-noise ratio. (At least, that's how I understood it.) :-) How about the same thing with CMOS vs. CCD?

[ShutterbugJohan]

ISO 1600 has 10 times less read noise than ISO 100. If Canon could get ISO 100 down to the same level as ISO 1600, dynamic range would increase greatly and 14-bits would no longer be sufficient.

What? I don't understand that... Wouldn't it be the other way around? :-)