more pixels are better!

John Carson wrote:

The new Fuji sensor design relies on pixel binning in low light. The
basic objection to pixel binning in general is that same-coloured
pixels aren't adjacent. The new Fuji design changes that so that
same-coloured pixels are adjacent.

http://www.dpreview.com/news/0809/08092210fujifilmexr.asp

Loks like it will have some good uses, but remember, as far as Bayer resolution is concerned, this is still 6MP.

I didn't notice if the two sets can operate at the same sensitivity, which is necessary for traditional "binning". You don't want to add a low sensitivity photosite to a high-sensitivity one for shadow areas, as the low-sensitivity photosite will add far too much noise.

Graystar wrote:

So it seems that this new wonderful sensor only performs AS well as
previous 12MP sensors...it's just learned a few new tricks. Current
cameras don't bin, so this could be an advantage in terms of feature
set. But I think it questionable as to how advantageous binning is
to the kind of photography we generally discuss. This may matter to
celestial photographers...don't know how much it will matter to
others.

The way I see it, hardware binning is of limited value. It can reduce read noise, but I appreciate the finer grain of the full capture, despite higher noise. I'd have to have very high oversampling before I'd appreciate it, I think.

Statistically, of course, hardware 2-> 1 binning can reduce image-level read noise by up to 29%. That assumes 2 pixels with the same sensitivity; I'm not clear on whether Fuji does this.

Gentlemen! There is no fighting in the war room!

Okay, I concede Graystar's point; I intended to do that earlier when I said there were other factors besides pixelization that affected a comparison between DSLR's and DP&S's. but as Les understands, I'm trying to isolate the practical and theoretical limits of increasing pixel count.

I'm not trying to compare digital compacts with SLR's. Rather I'm suggesting that in some situations - good light, huge blow-up - a high pixel count P&S can rival a smaller count SLR. One could do this with faster lens like the LX3's 2.0, but that is more expensive, and the SLR can always slap on a 1.7, whereas we do not (yet) have interchangable back end sensor modules... Perhaps putting 12Mp in, instead of 6Mp is simply a cheap and easy way to extend the P&S's envelop of performance.

Anyway, I'm guessing there are 3 liimits on pixel count (density) - noise, optical resolution, and pixel precision (how finely it can register light values before it is saturated). Or is noise a product of the latter two? I am hoping somewhere theres a discussion of what these limits are and how close we are getting to them...

Iforgetwhat8was4 wrote:

I'm not trying to compare digital compacts with SLR's.

Okay, got it.

Rather I'm
suggesting that in some situations - good light, huge blow-up - a
high pixel count P&S can rival a smaller count SLR.

I understand your suggestion. The answer is no. With the right content it is possible that it may look more detailed, as I've mentioned earlier. However there are two points to make. First, of course, it that it will always have more noise. You can't get around that. Second, the difference in detail isn't as great as the MP difference might suggest. DSLR images respond better to resizing than compact images. A 6MP image from a compact, no matter how optimal the conditions, will not exhibit the same apparent detail that a 6MP image from a DSLR will. So, back to your example, the likely result is that you probably won't gain as much detail as you think, but the extra noise of the small sensor will always be there. Under those condition my wager would be on the P&S not “rivaling” the DSLR.

Not to be overlooked, however, is John Sheehy's comment that he likes the detail even with the extra noise. It appears that, along with beauty, IQ is also in the eye of the beholder.

One could do
this with faster lens like the LX3's 2.0, but that is more expensive,
and the SLR can always slap on a 1.7...

Before you look at lens speed...what ISO are we discussing? Are we comparing performance at base ISO (which, I believe, loosely means the point at which gain = 1 for the sensor)? If you're matching ISO speed then the lens speed doesn't matter. Fast lens/fast shutter vs. slow lens/slow shutter...it doesn't make a difference because the same amount of light is captured. Of course, the LX3's fast lens makes a difference at the kid's birthday party when compared to a camera with a slower lens. But that has nothing to do with image quality.

whereas we do not (yet) have
interchangable back end sensor modules... Perhaps putting 12Mp in,
instead of 6Mp is simply a cheap and easy way to extend the P&S's
envelop of performance.

I don't think it makes a difference. The latest sensors perform better than older sensors because technology marches on. But when comparing sensor of equal manufacture, where the only difference is MP count, in that case the print from the higher MP count will have more detail but the same noise level. Depending on the size of the print, the detail may or may not be noticeable...the reverse of the low MP image where pixelization may or may not be noticeable.

Anyway, I'm guessing there are 3 liimits on pixel count (density) -
noise, optical resolution, and pixel precision (how finely it can
register light values before it is saturated). Or is noise a product
of the latter two? I am hoping somewhere theres a discussion of what
these limits are and how close we are getting to them...

There's lots of good info available on these areas. Personally, the final print is all the info I need

boels069 wrote:

Is the noise independent of shutter time?
(ISO 6400 @ 1/8000s has the same noise as ISO 3200 @ 1/4000s ?)

As with all these kinds of technical nuances, the true answer is long and complex. However, I'll try to keep it simple.

First, a change in ISO changes everything because you're changing the actual exposure to light. So I'm rewriting your example to match what I had written...f/2.0 @ 1/8000s vs. f/2.8 @ 1/4000s.

Simply stated, even with different shutter speeds, shot noise (noise from the light itself) will be the same as long as exposure is the same. Read noise (noise added by the electrical going-ons in the sensor) will be different. That difference in read noise is practically negligible when comparing two such fast times. It gets worse as the shutter speed gets slower. At 1.3 seconds, Canon compacts are already performing a dark-frame subtraction to get rid of the extra sensor noise. I haven't seen any tests of a sensor's noise across the more typical shutter speeds (1/30s to 1/1000s.) If anyone knows of any, post a link. Should be interesting.

My own take, based on not much real knowledge and extrapolating what little I do know with some rough science, to beyond to where it is any way robust...

Two aspects of sensor quality, resolution and signal to noise (S/N) are antagonistic for a given size of sensor.

The bigger the individual pixels the greater the maximum output voltage, and therefore the signal to noise ratio of the individual pixel, so say for a little pixel with certain illumination, 10 mv may be the output,but for a big one, say a little over 3 times the diameter, would give 100mv, A better quality signal for a given illumination

However if you have less pixels you have lower resolution, therefore for a fixed sensor size, resolution and S/N are antagonistic, I would say that why Nikon have the D3 and D3x one for highest resolution with good lighting and one for highest dynamic range and lowest noise with lower light. In one case the light/dark information is best preserved and in the other the spatial resolution,

Of course there are all sorts of complications, like how good does it need to be?, our eyesight has its own limitations so there is no point in engineering a system with subtly beyond what we can perceive, ( and even if you print very large you stand further away to view it right?) and limitations of the optical system that may also limit resolution, no point in having 40 MP if your lens is only good for 10, and the photo-voltaic stuff isn't always linear like i am assuming, and there is amplification and conversion to digital all subject to there own limitations and optimizations that may impose a limit on performance.

The process of "pixel binning" that is of adding together the signal from a number of different pixels, satisfies the need to have higher headline pixel counts to impress Joe the punter, and still having good sensitivity in moderate light, you will of course loose some resolution, but on screen or printed the perceived picture quality may be better because we are not Kestrels when it comes to resolution and like with the movies, that are a sequence of frozen stills. your brain fills in and conveniently ignores many quality issues, Unless it is connected to an internet forum like this one.

Atto

John Sheehy wrote:

Statistically, of course, hardware 2-> 1 binning can reduce
image-level read noise by up to 29%. That assumes 2 pixels with the
same sensitivity; I'm not clear on whether Fuji does this.

I don't understand where the 29% comes from. I would have expected 19%

(2^(1/4)). I say thanks in advance for the explanation (not to waste a post by a "thanks [nt]").

Just my two oere
Erik from Sweden

John Carson wrote:

Yes, and that is what the features are being compared to. The claim
is obvously that in challenging light conditions the Fuji gives
better performance than you would get by downsizing to 6 MP in
post-processing.

I'll believe it when I see it (printed, of course.)

But really...if you want a 6MP camera you don't have to wait...just
buy a Nikon D40 now.

A doubly silly comment:
1. The Fuji is both a 6 and a 12 megapixel camera depending on
circumstances.

2. The D40 is doubtless a fine choice for many purposes, but people
buy compact cameras for a reason (just as people buy DSLRs for a
reason); a D40 is not a substitute for a compact for some purposes.

I see no statements in the press release that this new technology will be limited to compact sized sensors. In the press release, references are made to both compact and DSLR cameras, as well as technologies that are in the DSLR sensors. I believe this new sensor design is to be used across their entire range.

Graystar wrote:

boels069 wrote:

Is the noise independent of shutter time?
(ISO 6400 @ 1/8000s has the same noise as ISO 3200 @ 1/4000s ?)

As with all these kinds of technical nuances, the true answer is long
and complex. However, I'll try to keep it simple.

First, a change in ISO changes everything because you're changing the
actual exposure to light. So I'm rewriting your example to match
what I had written...f/2.0 @ 1/8000s vs. f/2.8 @ 1/4000s.

Simply stated, even with different shutter speeds, shot noise (noise
from the light itself) will be the same as long as exposure is the
same. Read noise (noise added by the electrical going-ons in the
sensor) will be different. That difference in read noise is
practically negligible when comparing two such fast times. It gets
worse as the shutter speed gets slower. At 1.3 seconds, Canon
compacts are already performing a dark-frame subtraction to get rid
of the extra sensor noise. I haven't seen any tests of a sensor's
noise across the more typical shutter speeds (1/30s to 1/1000s.) If
anyone knows of any, post a link. Should be interesting.

Read noise ought to be the same regardless of exposure duration since the read is an event occuring at the end of the exposure. Thermal noise will become more significant as exposure length increases and for constant temperature is linear but increases by a factor of 2 every 6-7 deg C (in CCDs, not certain if CMOS follows the same rules). Unless you are dealing with exposures in multiple seconds or minutes, thermal noise shouldn't be an issue.

Read noise for CCDs also increases with temperature and increases as readout clock frequency is raised to reduce read times. For this reason, low noise scientific cameras often have very slow readouts that take many seconds to read each image from the chip.

Andrew dB wrote:

Graystar wrote:

As with all these kinds of technical nuances, the true answer is long
and complex.
...

Read noise ought to be the same regardless of exposure duration since
the read is an event occuring at the end of the exposure. Thermal
noise will become more significant as exposure length increases and
for constant temperature is linear but increases by a factor of 2
every 6-7 deg C (in CCDs, not certain if CMOS follows the same
rules). Unless you are dealing with exposures in multiple seconds or
minutes, thermal noise shouldn't be an issue.

Read noise for CCDs also increases with temperature and increases as
readout clock frequency is raised to reduce read times. For this
reason, low noise scientific cameras often have very slow readouts
that take many seconds to read each image from the chip.

See...long and complex

Sorry, meant to change "read noise" into the all-encompassing "sensor noise" like I did later in the post, but I let a couple slipped.

Ehrik wrote:

John Sheehy wrote:

Statistically, of course, hardware 2-> 1 binning can reduce
image-level read noise by up to 29%. That assumes 2 pixels with the
same sensitivity; I'm not clear on whether Fuji does this.

I don't understand where the 29% comes from. I would have expected 19%
(2^(1/4)).

Say there is 4 electrons of read noise.

If you do the hardware bin of two pixels into one, the total charge has a read noise of 4 electrons. If you bin them in software, the read noise is the square root of 4 squared plus 4 squared, or the square root of 32, or 5.66. 4 is 29% less than 5.66.

I say thanks in advance for the explanation (not to waste
a post by a "thanks [nt]").

Don't you just love it when a hot thread is approaching 150, and the "thanks" and "Bump"s start flying around!

John Sheehy wrote:

Ehrik wrote:

John Sheehy wrote:

Statistically, of course, hardware 2-> 1 binning can reduce
image-level read noise by up to 29%. That assumes 2 pixels with the
same sensitivity; I'm not clear on whether Fuji does this.

I don't understand where the 29% comes from. I would have expected 19%
(2^(1/4)).

Say there is 4 electrons of read noise.

If you do the hardware bin of two pixels into one, the total charge
has a read noise of 4 electrons. If you bin them in software, the
read noise is the square root of 4 squared plus 4 squared, or the
square root of 32, or 5.66. 4 is 29% less than 5.66.

I say thanks in advance for the explanation (not to waste
a post by a "thanks [nt]").

Don't you just love it when a hot thread is approaching 150, and the
"thanks" and "Bump"s start flying around!

This one's got just 117 to go
--
Bob

114 now
--
Bob

I apologise if I have misunderstood what folk meant or quoted anything out of context.

Here is a link to the relationship between SNR and exposure: http://www.microscopyu.com/tutorials/java/digitalimaging/signaltonoise/index.html

Nikon allergic folk can find the same information on the Olympus site or on the original Florida State U site.

The key point the graphs make is that you need a lot of light to make sensors photon-noise limited, ie, to make it true that SNR is related to total sensor area, rather than to the size of individual photosites.

Improvements in sensor design are a factor, but DxO's data show that they have not been sufficient to outweigh the effects of increased pixel density: SNRs have been falling steadily over the last few years. From their graphs "normalised" to constant pixel pitch it is clear that SNRs have gone up 3-6 dB because of improved sensor design. That is nice - but it means that if the manufacturers had not increased pixel numbers we could have had 1-2 stops more ISO head room, instead of a nett effect of nothing which is what we got!

It should, however, be noted that DxO's graphs of SNR vs date are all at 18% reflectance, which is quite a lot of light - roughly, the centre of the histogram. Noise is not such a big deal at the centre of the histogram, and the question is what has happened down the left hand end. If the weather here stays miserable I will re-plot the SNR vs date graphs using SNRs for low light and post the results.

Les Olson wrote:

I apologise if I have misunderstood what folk meant or quoted
anything out of context.

Here is a link to the relationship between SNR and exposure:

http://www.microscopyu.com/tutorials/java/digitalimaging/signaltonoise/index.html
Nikon allergic folk can find the same information on the Olympus site
or on the original Florida State U site.

The key point the graphs make is that you need a lot of light to make
sensors photon-noise limited, ie, to make it true that SNR is related
to total sensor area, rather than to the size of individual
photosites.

Improvements in sensor design are a factor, but DxO's data show that
they have not been sufficient to outweigh the effects of increased
pixel density: SNRs have been falling steadily over the last few
years. From their graphs "normalised" to constant pixel pitch it is
clear that SNRs have gone up 3-6 dB because of improved sensor
design. That is nice - but it means that if the manufacturers had
not increased pixel numbers we could have had 1-2 stops more ISO
head room, instead of a nett effect of nothing which is what we got!

It should, however, be noted that DxO's graphs of SNR vs date are all
at 18% reflectance, which is quite a lot of light - roughly, the
centre of the histogram. Noise is not such a big deal at the centre
of the histogram, and the question is what has happened down the left
hand end. If the weather here stays miserable I will re-plot the SNR
vs date graphs using SNRs for low light and post the results.

I think the point is that in photography, we always are dealing with situations where there's a 'lot of light'. You set the xposure so that the highlights pretty much fill the pixel well at base ISO, and even with a typicl 16x exposure push for high ISO, we're working with wells 1/16 full. Many of the CCD tutorials that are cited are primarily aimed at astronomy/science users, where very often the aim is to catch some phenomenon in a single or very few pixels - that's a completely different regime. In photography the design aim is to get as many pixels as possible over any individual piece of detail, in which case 'image level' noise is the thing we're interested in.

Sure, but noise, as a problem that spoils a print or the image on the screen, is not a highlight problem, it is a shadow problem, and even where there is a lot of light in a scene overall, the shadows are still dark. Black is black (I want my detail back) and grey is grey (I want it here today), what can I do (except turn up the ISOO)?

Plus, we usually take our photographs with a hand-held camera and often have moving subjects, so we must use short exposures and cannot, as a scientific or astronomical photographer can, wait for the wells to fill and therefore use low gain so the black sky stays black, not speckled.
--
2 November 1975.

'... Ma come io possiedo la storia,
essa mi possiede; ne sono illuminato:
ma a che serve la luce?'