Let's get real about resolution vs. noise and NR...

[Steve Deutsch]

Remember, they are turning on NR and if you look at Phil's samples in the review, you'll see clean 'gray' but fairly 'soft' edges in the queen.

I think the 7-8 um is the current sweet spot of pixel pitch.

It is clear (to my understanding of the subject) that
signal-to-noise ratio decreases as sensor pixel pitch decreases.
Although there are those on these fora who seem to think this is
not a significant issue. The bulk of this noise on higher end
cameras comes from photon (shot) noise, which you don't really
cover in your essay.

Measurements by Roger Clark suggest that in the 1DII at least, SNR
is pretty close to the theoretical limit imposed by photon noise:

http://clarkvision.com/imagedetail/does.pixel.size.matter/

There is little room for futher improvement in the hardware in
terms of signal to noise ratio. Smaller pixels mean more noise, and
although there are ways to attempt to improve this using
processing, what can be achieved seems to be limited. This is what
we see in the digital compacts.

The best of the current crop of compacts for noise would seem to be
the Fuji F10, which seems pretty good by comparison to other
compacts, but of course nowhere near what is produced by the DSLRs.
The F10 has a pixel pitch of approx 2.5 microns.

Of the current DSLRs, the 1DII (and 5D) have a pitch of 8.2
microns. 20D = 6.4, D2X = 5.5 and E300 = 5.6.

The question is how small pixel pitch can become and still produce
good signal to noise ratio (of a quality consistent with that
expected of DSLRs rather than comsumer compact).

For full frame sensors, this is not too much of a problem (1DsII
pitch is 7.2 microns giving a 17MP sensor). 13MP on an APS-C sized
sensor requires a pixel pitch of approx. 5 microns. A 4/3 sensor
with 13MP requires a pixel pitch of 4.3 microns. Nikon have
demonstrated that it is possible to have good noise control at 5.5
microns. the question is will it be possible at 4.3? this obviosuly
remains to be seen.

As I understand it, there are some possible improvements in sensor
design that could improve noise for a given pixel pitch, because
with current designs only part of the "available" area of each
pixel is used for the light sesing component, the rest being taken
up by processing components. It seems clear that new sensor designs
will be required to take resolutions above the 12MP mark,
particularly for APS-C and four thirds sensors and anything smaller.

 

[Shaun Williamson]

Photography is peculiar because it involves both art and science. A
photographer without an artists eye won't be able to take a good
photograph and a photographer who doesn't understand photographic
technology and the limitations of his equipment will never know how to
make a good photograph.

 

[GordonBGood]

That leaves:

"Blind application of averaging type NR:

1. Does NOT improve image quality for general image use.
2. In fact, it may actually reduce image quality.
3. Should not be applied as a matter of course, as it can not be
removed."

I think #3 is your "point"...

On a more technical level, how do we know how much internal NR a
selected camera applies at any ISO setting? How do we find out what
type of NR a camera uses? Does Phil give us any clues? If not, is
there a better source of DC P&S cameras?

Charlie, thanks again. I have to watch the too long sentences.

1) We can only guess at how much NR is being applied by looking at how much resolution is diminished with increasing ISO sensitivity. This assumes that NR is not being applied at the lowest sensitivities, or at least not of a type that would reduce resolution, as there would not then be much point in increasing the resolution by packing photosites.

We might also guess at whether NR is being applied as without NR, noise tends to increase with ISO sensitivity with a recognizably shaped curve.

2) Camera manufacturers aren't going to tell us how they implement NR, as that will be regarded to be proprietary. We can guess that they likely apply some sort of Gaussian filtering using a convolution engine, since convolution is the way that almost all of the other in-camera processing is done.

3) No, Phil doesn't give us any clues because a) he can't find out any more than we can and b) he relies an other technical experts as he is more of a photographer than a technical guru.

4) I'm watching Fuji's developements as far as P&S sensors, as hopefully their success in this area will force the other manufacturers to compete.

Regards, GordonBGood

 

[GordonBGood]

i have heard that Canon even processes their "raw" images. Which
would make it even harder to get an honest assessment of their
sensors ... and to subsequently apply future "better" algorithms
for NR. Is this the case? Do any other Mfrs also fiddle with "raw"
data? In particular the Fuji S3 seems to have two types of "raw"
output. Is either really raw? A standard assessment of raw output
would seem the only way to get an honest assessment of the ultimate
capabilities of a camera. i wish some review site would make that a
regular feature.

I have also heard that Canon do some NR manipulations in their Raw files. I can't really comment about Canon's DSLR Raw, as I have never had the oportunity to analyse them, but I find no evidence of NR tampering in the Raw files for the Powershot series up to at least the G5. The files I have looked at don't even do compensations for dark current drift, but preserve all the the pixels used to determine and apply such corrections.

I would expect most if not all manufacturers to not apply NR to their Raw file in their Raw formats, although some appear to do dark current compensation. Since it can not be undone, applying NR to a Raw file would seem to undermine the advantage of such files as being as close as possible to a "digital negative" without manipulation.

Regards, GordonBGood

 

[Ron Parr]

I have also heard that Canon do some NR manipulations in their Raw
files. I can't really comment about Canon's DSLR Raw, as I have
never had the oportunity to analyse them, but I find no evidence of
NR tampering in the Raw files for the Powershot series up to at
least the G5. The files I have looked at don't even do
compensations for dark current drift, but preserve all the the
pixels used to determine and apply such corrections.

Actually, Canon has said explicitly that they do not do any NR.

--
Ron Parr
Digital Photography FAQ: http://www.cs.duke.edu/~parr/photography/faq.html
Gallery: http://www.pbase.com/parr/

 

[Tristan Cope]

I'm not so sure about that. I agree technology marches on, and things we considered difficult or impossible a few years ago are on the shelves a few years later.

But, if the current high end DSLRs are really have outputs that are close to the theoretical limits imposed by photon counting, then its difficult to see how a new material will make any difference to that. The lower limit of noise is governed by quantum uncertainty (dead cats etc). Presumably any new material is also going to be subject to quantum effects (dilithium crystals not withstanding). So no matter how efficient you make the sensor itsself, and the rest of the electronics to go with it (amplifier, porcessor etc) there will eventually reach a point where the only way to reduce noise any further is to reduce the statistical effect of photon noise - and the way to do that is to make pixel size larger, not smaller.

It is difficult (or impossible) to see how software (or for that matter hardware) processing can reduce noise any further when this level is reached, because the photon noise is inherently indistinguishable from the signal.

 

[aukhawk]

no text

 

[aukhawk]

since for a given pitch, it is still theoretically possible (in fact likely, I'd say) that photosite size can be increased - isn't this part of what Fuji have done?

Francis

 

[Ron Parr]

It is clear (to my understanding of the subject) that
signal-to-noise ratio decreases as sensor pixel pitch decreases.
Although there are those on these fora who seem to think this is
not a significant issue. The bulk of this noise on higher end
cameras comes from photon (shot) noise, which you don't really
cover in your essay.

Measurements by Roger Clark suggest that in the 1DII at least, SNR
is pretty close to the theoretical limit imposed by photon noise:

http://clarkvision.com/imagedetail/does.pixel.size.matter/

I don't think Roger has an airtight argument that current DSLR noise is limited by photon shot noise.

He never fully explains his method for computing the well capacity of the sensor. If you follow the links, he seems to be computing well capacity as a function of gain, and he provides links to a method for computing gain.

However, the method for computing gain only works if you assume no other noise sources. This might be a reasonable assumption with scientific grade CCDs, which typically have very low dark current, but it should not be assumed with a consumer CMOS sensor.

I'd need to look at it a little more carefully, but his argument may , in fact, be circular. If he assumes no other noise sources when he computes the gain, then uses this to compute the well capacity, and then uses this to prove that SNR is limited by photon shot noise, he has built an argument where his conclusion is actually one of his premises.

--
Ron Parr
Digital Photography FAQ: http://www.cs.duke.edu/~parr/photography/faq.html
Gallery: http://www.pbase.com/parr/

 

[Tristan Cope]

since for a given pitch, it is still theoretically possible (in
fact likely, I'd say) that photosite size can be increased - isn't
this part of what Fuji have done?

Agreed. At the moment, the sensor part of a sensor site only takes up part of the available area (?1/2-2/3), the rest being processor, or amplifier or whatever. So as i understand it there is room for improvement in terms of maximising the light sensitive area of each pixel.

Francis

 

[Tristan Cope]

I see what your getting at. I have to say, i was a bit surprised at the idea that the technology might be so "noisless" without any of the paraphernelia that you would usually associate with such - like supercooling etc.

Are you aware if Canon or any of the other manufacturers publish any information about noise levels and limitations? Or for that matter about the sensor well depth so Roger could do his calculations again without the initial assumption? I suspect such information is "commercially sensitive"?

Are you intersted in contacting him to ask him about this?

 

[chuxter]

1) We can only guess at how much NR is being applied...

I was hoping that somebody was set up to reverse engineer the processors. I'm sure the various manufacturers know EXACTLY how their competitors process high ISO pix.

2) Camera manufacturers aren't going to tell us how they implement
NR, as that will be regarded to be proprietary.

That seems silly. Do they think it's a secret from their competitors?

We can guess that
they likely apply some sort of Gaussian filtering using a
convolution engine, since convolution is the way that almost all of
the other in-camera processing is done.

Yes, but...

3) No, Phil doesn't give us any clues because a) he can't find out
any more than we can and b) he relies an other technical experts as
he is more of a photographer than a technical guru.

I suspect Phil knows many details that we don't. He perhaps can't tell us?

I thought he was more of a technical guru (software) than a photographer? Hmmm...seems like he's ideally positioned to help answer this question!

Good thread!

--
Charlie Davis
CATS #25
PAS Scribe @ http://www.here-ugo.com/PAS_List.htm
HomePage: http://www.1derful.info
'I brake for pixels...'

 

[Ron Parr]

I see what your getting at. I have to say, i was a bit surprised at
the idea that the technology might be so "noisless" without any of
the paraphernelia that you would usually associate with such - like
supercooling etc.

Are you aware if Canon or any of the other manufacturers publish
any information about noise levels and limitations? Or for that
matter about the sensor well depth so Roger could do his
calculations again without the initial assumption? I suspect such
information is "commercially sensitive"?

Kodak is very open about their CCD specifications. Here are the specs for the E-300 sensor:

http://www.kodak.com/global/plugins/acrobat/en/digital/ccd/products/fullframe/KAF-8300CELongSpec.pdf

The dark current accumulation rate isn't bad - 200 electrons/s. This is one of the strengths of mature CCD technology.

I'm not aware of any other high quality still camera sensor manufacturers that provide this kind of detail.

Are you intersted in contacting him to ask him about this?

I've got enough distractions for now.

--
Ron Parr
Digital Photography FAQ: http://www.cs.duke.edu/~parr/photography/faq.html
Gallery: http://www.pbase.com/parr/

 

[GordonBGood]

since for a given pitch, it is still theoretically possible (in
fact likely, I'd say) that photosite size can be increased - isn't
this part of what Fuji have done?

It isn't so much that Fuji have increase the size of the photosites as they have made best use of the area that the photosites occupy.

Agreed. At the moment, the sensor part of a sensor site only takes
up part of the available area (?1/2-2/3), the rest being processor,
or amplifier or whatever. So as i understand it there is room for
improvement in terms of maximising the light sensitive area of each
pixel.

But the percentage of the sensor area occupied by the active detection cells is not a factor when an efficient microlense system is used, which redirects most of the available light to the detection areas. This appears to be one of the things that Fuji have done in order to increase overall sensor efficiency with a resultant better SNR for a given sensor area.

Regards, GordonBGood