Megapixels vs Signal-to-Noise

[MarkS]

Here's an interesting question.  If you keep a CCD the same size but increase the number of megapixels (from 6 to 8 to 10 to 12) then, all other things being equal, the signal-to-noise ratio of each pixel is compromised because each pixel receives fewer photons.

There are two viewpoints on this.
Argument 1 says the signal-to-noise ratio of each pixel is paramount, so increasing the number of pixels is a bad thing.
Argument 2 says that the same number of electrons per unit area are being captured so the signal-to-noise measured per unit area is unchanged by increasing the number of pixels.  So increasing the number of pixels is a good thing because it increases the resolution of the image (even though each individual pixel is noiser).

Argument 1 taken to its extreme would have a couple of hundred super noise-free pixels on a CCD.
Argument 2 taken to it's extreme would have billions of really noisy pixels on a CCD.

So what is the truth?  Does the extra resolution outweigh the extra noise?  It must do up to a point - a picture containing a few hundred pixels would be a waste of time.  I think the answer probably lies in human perception - if you were asked to examine the a whole series of (identical size) astro-photos produced by sensors with differing numbers of megapixels you would probably come up with a favourite. 

What do you think?

Mark

[mickw]

I think I'll have another beer 

Yes, yes, I know it was predictable 

[mickw]

Or............
The same number of photons hit the same surface area, so I think it would come down to the efficiency of an individual pixel's handling of the photon.  Assuming not all pixels (CCDs/CMOSs) are created equal

A crap CCD will still be a crap CCD no matter what its size.

Don't be afraid to ask more questions if you don't understand 

[Tony G]

Bloody 'ell Mick,

If that is the response after the 'predictable beer' you should have a case of Stella, and give us a talk at the next members evening on 'Relationship of Quantum mechanics to classical electromagnetism and classical relativistic mechanics in relation to Astro Physics'  -

Tony G

[mickw]

 

OK, now I'm scared   
 

[Ian]

wot Mick said.

Or, given human image processing I would think that the maths actually gives a bit of bum steer. I would suggest that resolution wins overall, as while the overall noise in the image increases, perception of that noise will decrease as the resolution increases, to the point where background noise merely becomes a uniform background level. Being uniform it's likely to be much less objectionable.

However, upscale the image and all bets are off again.

On a related point, by how much does noise increase as resolution increases? The relationship between the two may give an understanding of where noise perception is minimised even if absolute noise levels are higher.

And to muddy the water further, we're only considering the fact that the noise floor is raised due to lower absolute signal being generated by a pixel. As pixels get smaller, diffraction may start to have a role increasing the noise floor itself.

Mick, gizza beer mate. I'm as dry as a wallaby's *&^*&

[Tony G]

 

As pixels get smaller, diffraction may start to have a role increasing the noise floor itself.

Even the police are using that forbidden word. should know better. 

Tony G

[Rick]

On a related point, by how much does noise increase as resolution increases? The relationship between the two may give an understanding of where noise perception is minimised even if absolute noise levels are higher.

I suspect the read noise may skew things a bit, depending on how it's affected by pixel area. It probably doesn't decrease linearly with pixel area. I'd expect smaller pixels to have proportionally higher read noise. There's also the problem that more cells means more circuitry to allow them to be read, and that may reduce the effective collecting area of the CCD.

[Rick]

Even the police are using that forbidden word. should know better. 

Nah. "Diffraction" is a perfectly honest and allowable word.
The forbidden phrase is [fake] spikes spaghetti hoops.
_______________________________ ________________________ 

[mickw]

gizza beer mate

Sorry Ian, how rude of me

Here ya go mate, sorry it's late (or early) 

[MarkS]

I suspect the read noise may skew things a bit, depending on how it's affected by pixel area. It probably doesn't decrease linearly with pixel area. I'd expect smaller pixels to have proportionally higher read noise. There's also the problem that more cells means more circuitry to allow them to be read, and that may reduce the effective collecting area of the CCD.

Rick is on the ball here!  The people in camp 2 conveniently overlook the fact that read noise measured in electrons tends to remain constant as pixel size reduces and therefore read noise becomes larger and larger per unit area as pixel size reduces (I don't know what happens to dark current - maybe it reduces with pixel size).  This doesn't totally invalidate their viewpoint but means that the "sweet spot" occurs sooner than expected as pixel size is reduced.


So where is the "sweet spot" - as you may have guessed, I had a reason for asking the question.

I have a Canon 300D (6 Mpixels) and a 350D (8 Mpixels).  Doing a detailed comparison of the two, under identical conditions, the 350D collects as many photons per pixel as the 300D in spite of their smaller size.  Also the read noise and the dark current is smaller.

So both camps would agree that the the 350D is better than the 300D.

What about the 400D (10 Mpixels)?  Christian Buil's tests (http://www.astrosurf.com/~buil/400d/400d.htm) indicate that each pixel collects fewer photons per pixel but this is in proportion to the reduced size of the pixel so the CCD collects the same number of pixels per unit area.  The read noise is slightly better than the 350D and the dark current slightly worse.

So camp 1 would say that the 400D is a step backwards (because of reduced signal to noise) but camp 2 would say it is a step forwards because of resolution.

What about the 450D (12 Mpixels)?  It appears that this collects fewer electrons per pixel and fewer electrons per unit area.  The only figure I've seen published implies that each pixel collects only 50% of the electrons of the 350D even though they are 66% of the area.  Both the read noise and dark current are slightly better than the 350D.

So camp 1 would say that the 450D steps even further backwards than the 400D.  Even camp 2 might begin to have their doubts about the 450D.

What about the 1000D (10 Mpixels)?  I just don't know.

The frustrating thing is that no-one seems to be producing the figures needed for either camp to make an informed choice i.e. read noise, dark current, relative sensitivity (or QE).

Don't get me wrong, all the above cameras are competent performers for astrophotography but the 350D or the 400D probably represent the "sweet spot" in terms of entry level Canon DSLRs.

Mark

[mickw]

the 350D collects as many photons per pixel as the 300D in spite of their smaller size

How/why ?

(400D) each pixel collects fewer photons per pixel but this is in proportion to the reduced size of the pixel

[MarkS]

How/why ?

I don't know the exact reason but clearly a design improvement took place in the 350D CMOS sensor.

[RobertM]

It's probably due to conversion efficiency improvements such as microlens design, electronic layout allowing a larger (by proportion) photosite and quantum efficiency.  I believe the Canon sensors have read, A/D converters and noise reduction at each pixel so better design and miniturisation would yeald a larger (in proportion) collection area.  The 350d/400d sensors certainly look like the sweet spot in the design and in part explain the slightly poorer 450d performance.

Robert

[Ian]

cheers Mick. *buurrrrpppp*