Digital camera and film signal-to-noise

Hi
This weekend I completed a study of the
signal-to-noise of a Canon 1D Mark II camera, and
compared results to scanned Fujichrome Velvia and
the maximum theoretical possible based on photon
statistics. The results are at:

http://clarkvision.com/imagedetail/digital.signal.to.noise

The ISO 50 film has worse signal-to-noise than the
1D Mark II does at ISO 800. The 1D Mark II is essentially
photon noise limited at ISO 100, with a signal-to-noise
3 to 6 times greater than Velvia.

I have searched a long time for full well
capacity of digital camera sensors, and while I can find
a lot of information on the scientific sensors, there is
very little information on the sensors in consumer
digital cameras. If anyone knows of any web sites with
this information, please let me know.

The 1D Mark II has a full well of about 52,000 electrons
(really at 4095 camera DN). If these new 8 megapixel P&S
small sensors have a much lower full well, it could
explain how they could be operating near the photon
limit and have greater noise. So I am particularly
interested in finding the full well capacities of
these smaller sensors. If anyone wants to measure their
camera's gain, read noise, and full well capacity, there
are links on the above page which will help you do it.
It is not hard; it just takes some time.

Roger

 

Nice! Great work, as always.

Thanks for doing the work. You've made yet another major contribution.

David J. Littleboy
Tokyo, Japan

 

Well done. Interesting conclusions.

Bart

 

There's a problem I see here. How do you know what noise reduction
and/or dynamic range scaling was done before you got to do the
measurement?

Perhaps you used dcraw and selected linear gamma, in which case my
worries are unfounded.

Andrew.

 

Question: isn't the signal to noise ratio of scanned film, in principle
inversely proportional to the resolution of the scan? Presumably, if one
applied a bit of noise reduction to the scan and then downsampled (from, say
5400 dpi to 2700 dpi), the resulting 2700 dpi image should have a much
better (about 2 bits better, I'd guess) SNR than the 5400 dpi???

There's of course the problem of grain aliasing, but in principle, the
larger the area of film used to create each pixel, the lower the noise
should be. (And this game can be played with the digital camera as well.)

Anyway, I think you should Neatimage and downsample the scan to the same
pixel count as the 8MP camera before presenting the results. Or at least
present multiple noise curves for the film at different degrees of
downsampling.

The 1DII is 2336 x 3504 pixels, a 4000 dpi scan is 3780 x 5670 pixels, and a
5400 dpi scan is 5100 x 7650 pixels. So a 4000 dpi scan is 2.5 times as many
pixels, and a 5400 dpi scan is 4.7 times as many pixels.

David J. Littleboy
Tokyo, Japan

 

David,
I disagree on making the images the same number of pixels.
The fundamental reason is that the digital camera is directly
sampling the image, while the scan of the film is a
second order sampling. The second order sampling must be
at a higher sampling density in order to preserve detail.
After all it is well established that a digital camera images
of lower megapixels have the spatial information of
much higher resolution scans of film.

For my tests, I used 4000 dpi scans. I'll go back and look at
some higher resolution scans to see if it makes much difference.

Roger

 

I use a program called ImagesPlus and the Canon 1DII raw converter,
both in linear node, and both give the same results in all
tests I have done. ImagesPlus is used in the astronomy
community for adding multiple (linear) images together to
increase signal to noise with astrophotos. I did find a bug
in Canon's software that comes with the 1D Mark II: the batch
converter does not use the settings you select for an image,
it does some default. Perhaps I set it up wrong. But if one
converts one image at a time and make each linear, it works
fine. There was no noise reduction turned on and no dynamic range
scaling (proof by multiple exposures at different exposure
times are have levels as one would expect in a linear
system).

Roger

 

Agreed. Completely.

But once you've captured the detail, you should be able to downsample the
image, removing some of the noise. Careful downsampling should preserve
image detail up to just under the Nyquist frequency at the downsampled dpi.
(Care is required to prevent the aliasing any lost information, although I
personally don't think there's much information in the upper half of the
range of spatial frequencies.)

Thanks.

David J. Littleboy
Tokyo, Japan

 

David,
I agree that sampling in the upper frequencies has diminishing
returns, however, it is those frequencies that contain
the informations that lead to the perception of sharpness
and fine detail. Also, your concept of Nyquist frequency
is mis-applied here. Nyquist applies to sampling in phase
with the signal. If the signal is not in phase, the case
of random detail in an image, Nyquist does not apply;
Images need sampling higher than Nyquist, approximately 3x
Nyquist does a good job. Please see this page:

http://www.clarkvision.com/imagedetail/sampling1.html

Is shows the Nyquist sampling issues for images. But also
shown are real images with sampling at various MTF levels.
The subject is blades of grass in an image. It is clear that
image detail is important to at least 9% MTF (Rayleigh Limit),
and in my opinion, close to the Dawes Limit, (0% MTF).

Roger

 

I suspect that here is where we disagree: trying to use that perveived
sharpeness from that diminishing returns information results in unacceptably
low perceived image quality. I personally don't consider enlargements of 13x
from any film acceptable, and prefer to limit myself to 8x. At 8x, film
looks pretty decent (although the LF types will argue that you can perceive
the difference between 4x and 8x<g>).

Anyway, that's information that's not in the direct digital capture image.
And you are charging film a factor of four in the noise category for that
information. Since the pixel counts are quite different, it seems to me that
you are comparing different things.

I still don't see why careful noise reduction and downsampling should not
result in an image of equal or better quality to a direct digital capture at
the same pixel count.

I'm quite aware that capturing at the Nyquist frequency doesn't capture
square wave patterns.

But direct digital capture doesn't capture anything above the Nyquist
frequency either.

So this point is rather moot.

I still don't see why careful noise reduction and downsampling should not
result in an image of equal or better quality* to a direct digital capture
of
the same pixel count.

In terms of MTF, not noise.

David J. Littleboy
Tokyo, Japan

 

Actually, information theory says just the opposite. Do you understand
Fourier Transforms? For example, Fourier theory says that all
waveforms can be described by an infinite series of sine waves.
Take for example, a square wave, analogous to a traverse
across blades of grass in an image. While a low frequency
sine wave shows the existence of the blades of grass, if the
higher frequencies are there, then the edges of the blades
become sharp.

You probably know I do large format film too. I've
rarely seen an 8x10 inch enlargement from 35mm that can come
close to the sharpness of a 4x5 transparency enlarged to
8x10, EXCEPT fujichrome velvia 35mm slides scanned at
6000 dpi. So I argue I can see a difference between an
8x 35mm enlargement and a 2x 4x5 enlargement. But with 6000
dpi scans that difference is extremely small.
I agree that bigger enlargements there is no
comparison to large format. That's why I've been back to
the "rainbow spot" (the photo on my web page) dozens of times to
try and get that scene on 4x5 (it's a rare event). If you
do things right, and push the limits, it can be rewarding,
and 6000 dpi drum scans are worth the price for those special
images.

I disagree. You need that scan resolution to record the
detail that equals the detail in the digital image.

Try it. Take the image from my sampling page:
http://www.clarkvision.com/imagedetail/sampling1.html

Resample the image
http://www.clarkvision.com/imagedetail/sampling1.gif
(in particular the bottom portion called 3x Nyquist),
and see how much you can drop the sampling without
destroying the basic pattern. Try blurring the image
to reduce the square wave nature (I tried Gaussian blur,
radius 1.5 and a few others). While this helps aliasing,
it reduces apparent sharpness.

Roger

 

Folks,

In all of your measurements, are you applying any sort of weighting curve
when you measure signal-to-noise ratio?

I ask because the eye is more sensitive to some spatial frequencies than
others, so that what measures the same using a flat spatial frequency
response may be perceived quite differently by the eye.

Cheers,
David

 

David,
No I did not apply any weighting curve. I did notice,
quite strikingly, that the *apparent* noise/grain was different
than I measured, meaning mid-level intensities showed the
most apparent noise. If the intensity is real bright,
near white, noise is less apparent, and similarly
near blacks. I think you can see this in
photos with blue sky and clouds. In the brightest
portions of clouds, the noise/grain is less apparent
than the sky or where the clouds are gray. But over
most of the dynamic range, the noise is apparent,
and it is only the extremes where the apparent noise
decreases. Grain/noise is also less apparent where
the scene is spatially more complex.

Roger

 

Roger N. Clark (change username to rnclark) wrote:
[]

Thanks, Roger.

The different noise spectra of digital and film may give an element of
subjectivity to the results, and I had wondered if that could be removed
by using a weighting curve. Where the scene is complex, the larger amount
of higher spatial frequencies in the scene is likely masking the noise
(like the masking in hearing that Dolby compression relies on). Similar
effects of differing noise visibility versus signal level are seen in
analog TV systems - these may have been investigated and formalise
somewhere - I don't know.

Even though I seem to be saying that the answers are likely to have a
subjective element, I certainly think that we should be continuing to seek
a scientific understanding of what our senses tell us, and I value your
contributions towards that end.

Cheers,
David

 

|Hi
|This weekend I completed a study of the
|signal-to-noise of a Canon 1D Mark II camera, and
|compared results to scanned Fujichrome Velvia and
|the maximum theoretical possible based on photon
|statistics. The results are at:
|
|http://clarkvision.com/imagedetail/digital.signal.to.noise
|
|The ISO 50 film has worse signal-to-noise than the
|1D Mark II does at ISO 800. The 1D Mark II is essentially
|photon noise limited at ISO 100, with a signal-to-noise
|3 to 6 times greater than Velvia.
|
|I have searched a long time for full well
|capacity of digital camera sensors, and while I can find
|a lot of information on the scientific sensors, there is
|very little information on the sensors in consumer
|digital cameras. If anyone knows of any web sites with
|this information, please let me know.
|
|The 1D Mark II has a full well of about 52,000 electrons
|(really at 4095 camera DN). If these new 8 megapixel P&S
|small sensors have a much lower full well, it could
|explain how they could be operating near the photon
|limit and have greater noise. So I am particularly
|interested in finding the full well capacities of
|these smaller sensors. If anyone wants to measure their
|camera's gain, read noise, and full well capacity, there
|are links on the above page which will help you do it.
|It is not hard; it just takes some time.

I don't have such a camera, but here's a rough attempt at estimating the
well size. Many of these cameras use the Sony ICX456 CCD, which Sony
specifies as follows
(http://www.sony.net/Products/SC-HP/cx_news/vol36/pdf/icx456aq.pdf):

Sensitivity (G signal): 200 mV at 1/30 sec exposure time, F5.6,
illumination at 3200K, 706 cd/m^2.
Saturation: 420 mV at 60 degrees Celsius.
Pixel size: 2.7 by 2.7 um.

For the sensitivity reference, the exposure works out to 4.297*10^(-12)
lumen-seconds per pixel. To translate this into electrons per well,
we need to know the quantum efficiency of the green channel of the CCD.
Sony unfortunately does not give this, but let's guess the rough
approximation that it is 0.15 from 500 to 550 nm and linearly drops off to
0 at 450 nm and 625 nm. In this case the sensitivity exposure will give 2300
electrons per well, so saturation should correspond to around 4800 electrons
per full well. For comparison, the ICX084, with 7.4 by 7.4 um pixels (7.5
times the area of the ICX456's) is said to have a full well of 40000 electrons.

 

I don't think there is any information out on the web. CCD
manufacturers won't spec well depths--something I discovered when I
tried to get Sony to tell me what the well depth was on a high
sensitivity survailence camera. I had to work out the numbers
experimentally and then call them back and tell them what their well
depth was. At that point they gruding admissioned that I was in the
right ballpark.

Why? At the time I suspected the well depths vary wildly from sensor
to sensor-something I'd suspected ever since II measured a shipment of
cameras straight out of the box and found a factor of 3 variation in
S/N.

I repeated a measurement of well-depth I'd done a few years ago on my
camera- a Oly 3020Z, I hadn't posted my results to the group mainly
because the numbers came out too good. I feared I might have made a
mistake in my logic. Or perhaps I didn't understand how the camera's
firmware worked when it produced a jpg image--the image format I'd
used during my first experiment

At the time I didn't know that the 3020Z had an unoffical raw
mode--something I learned thanks to the folks at the Oly 405080 group
on Yahoo.

First, a few fact on the camera. The sensor and control chip are
identical to the one used in the Nikon Coolpic 990 and the unoffical
RAW mode produces a NEF rather and an OLY image. Moreover instead of
the standard RGB color filter the camera has a CMYG filter. I don't
think either of these facts effect my current measurements but I'm
throwing them out in case some sees something I missed.

My proceedure is different than the one Roger posted on his website. I
have a calibrated Kodak step tablet #2. It is a tranmittance tablet
and has 21 steps where the optical density varies from about 0 to 3, a
1000X difference in transmittance.

To do the experiment I set up an evenly illuminate background,
mounted the tablet in a black box and then took a series of images.
With the illumination I had, an exposure at 1/10 of a second gave me
an image where the step O was just going into saturation. A dark frame
taken at that exposure didn't show much thermal shot noise so I
ignored it,

I did my raw conversion with Photoshop 8 and my data reduction with
ImageJ, an excellent freebe scientific image analysis package
available on the NIH web site. Since I knew the transmission of the
step tablet, the data was easy to linearize dispite the far from
linear output of the PhotoShop raw converter.

From the linearized data, I was able to correct the additional gain
the Photoshop gamma correction gave to the measured noise ( or more
accurately variance) for each transmittace step. With this data I
plotted out what Kodak calls a Photon Transfer curve in their tech
note MTD PS-0233. The peak photo shot noise was as expected at the
maximum transmittancer and it dropped steadily until transmittance had
fallen by a factor of 200X. At that point I hit the readout noise and
the noise stayed essentially constant until the last step (an optical
density of 3.09)

Now for the numbers.

The 3020Z is a three to four year old camera with a sensor area of
12.5 sq mirron

The max S/N was 251 which when squared gives a well depth of 63000 e-.
At the other end the readout noise floor was 15 e-.

So what is going on? As a rule of thumb I've always assumed you could
collect from 500 to a 1000 e- per sq microns of sensor real estate.
How did I end up with 5 times more well depth than I thought was
possible?

My guess is when the process is working perfectly and the piece of
silicon it working on is also close to perfect, you get a sensor
that's way above average. It doesn't happen often but when I bought
the camera I got lucky. It almost makes up for the time, I came within
a pant and gasp of winning a $300 million dollar lottery.

Comments and caveats are welcome.

jpc

 

Yes. (They made us take the EE courses where I did my CS undergrad.) But the
MTF above 60 lp/mm is pretty minimal. 0% for low contrast edges, barely
detectable for high contrast edges. So your whole discussion here strikes me
as off base.

I'm under the impression that there's not a big difference in rendition
between a brickwall filter (digital) and a filter with a long under 10%
tail. That's just not enough information to contribute significantly.

And even worse, the digital image, that we're comparing to, doesn't have any
of that information _AT ALL_, so the idea that that information contributes
to the image seems seriously problematic.

That I understand. But if you attenuate the high-order components, you are
back to just showing the existence. You have to argue that 10% is a lot
better than 0%, and I just don't think so.

Yup. My darkroom experience is that making a lovely 11x14 from 6x6 is like
falling off a log, but that 8x10s from 35mm Panatomic X in Microdol don't
fly.

But being limited to Velvia (slow, can't handle high contrast scenes, dizzy
color reproduction) is painful.

Hmm. It's clearly possible to represent "the detail in the digital image" at
the pixel count of the digital image: the digital image itself proves that.
There are _no_ high-frequency components in the digital image: that's why I
invoked Nyquist in the first place.

So I don't see what the problem with noise-reduction and downsampling is.
Once the detail is captured, the only question is remapping it to a
different resolution. The target resolution is clearly capable of
representing the information.

Downsampling by 1.414x would result in a 2600 x 3900 dpi file (10MP). That's
a bigger file than 1Dm2 image, yet could have 1/2 the noise (maybe less)
than you are seeing.

I do it all the time. 4000 dpi -> 2400 dpi produces lovely, low-noise images
that make lovely prints at 300 dpi. Of course, I'm starting from medium
format, so I still get nice big prints.

This example doesn't represent the case we are concerned with: the magnitude
of the high-frequency components in film images _on the film_ have been
attenuated by 90% or more. The 3x Nyquist image you are starting with there
has the full high-frequency spectrum.

You also appear to be misunderstanding the resultant images I think should
be compared. If the digital capture (your 1Dm2) is capturing at 3x Nyquist,
then I'm talking about downsampling the film scan to about 4x Nyquist _for
the same FOV image_.

I just don't see film at 4000 dpi getting anywhere near capturing that
pattern _at that quality_ if that pattern is projected so that one cycle is
6 pixels. It'll be OK at the fundamental (26 lp/mm), but you only get one
extra term in the fourier expansion (if memory serves, square waves only
have even terms) and that term will be seriously attenuated.

David J. Littleboy
Tokyo, Japan

 

David,
We are not talking a difference in 10% versus 0% MTF.
I think I already said there is very little difference in
such a level and that is what it shows on the sampling page.
You want to down sample a factor of two. That is more like
moving from 10% to 50% MTF, and that is a big difference as
shown on the sampling page,
http://www.clarkvision.com/imagedetail/sampling1.html

NO. In one case (digital sensor) we have a sampling of an image.
In the case of scanning film, we have a sampling of a sampling of an image.
Each generation removed loses information unless you sample
high enough.

No, with factors of 2 change in samples, it's more like
10% MTF versus 50% MTF.

Velvia has color, panatomic X does not. Velvia is the highest
resolution color film on the market today. That is the
comparison I used.

But such an image would have less detail than the 8 MPixel
original digital image. When you downsample, apparent sharpness
is lost, and one typically has to sharpen to regain that apparent
loss. Sharpening increases noise. I chose to do no processing
to the image. After all, if you invoke noise reduction on the film
scan to make it look better, why not the digital image too?
To me it seems like stacking the test to favor one medium over
the other.

Real images are much more complex and the Fourier transform of an
image shows a continuum of frequencies, so it is not a matter
of losing one specific frequency when downsampling. I used the
square wave to make a simple example.

When I have time, I will put together a web page I've been meaning to
build for several years: apparent grain and signal to noise
as a function of scan ppi. I did drum scans up to 11,000 ppi
for the test. My impression from looking at the images is that
the apparent grain *decreases* at higher ppi above 6000 ppi.

Roger

 

Yes, we are. We're talking about 2800 dpi vs. 4000 dpi. (See below.)

Truth in advertising: I'm changing my tune slightly. In my original note, I
was thinking of a 6MP vs 35mm 4000 dpi Velvia comparison: in that case, you
can downsample to 2400 dpi (my standard workflow for MF scans, resulting in
1/3 the pixel count), but that puts you at under 8MP, so it's not a fair
comparison with the 8MP 1Dm2. Which is why I'm talking about a 1.414x
downsample, which cuts the pixel count in half.

Your argument here assumes that the MTF of the signal in frequencies in the
range that correspond to 2800 dpi to 4000 dpi is over 50%. My argument is
that that assumption is simply wrong. (See below.)

I'm NOT arguing against sampling at a higher frequency.

I'm arguing FOR postprocessing that result to a more sensible dpi.

It still seems that you are failing to see my point.

Yes, for the same frequencies the digital system captures. That's why you
oversample. To capture the lower frequencies.

But in the frequencies between above 2800 dpi and below 4000 dpi, film
certainly doesn't capture 50%. MTF. 2800 dpi is 55 lp/mm, 4000 dpi is 78
lp/mm. Neither Velvia nor 35mm lenses provide 50% MTF in that region. Both
are well under 50% at 55 lp/mm, to say nothing of 78 lp/mm. And that doesn't
include the MTF of the scanner.

To the best that I can tell, there is no significant information above 2800
dpi in a 4000 dpi scan. And given the MTFs of the components (lens, film,
scanner) in the system, there is no reason to believe there could be.

Only if either the original had less detail (which is actually probably the

Not if you bump the threshold to 2 or so. After NeatImage, even that's often
unnecessary.

I'd have no complaint with that. But since you aren't downsampling, you will
lose detail.

But not postprocessing the film image stacks the test in the other
direction. You _oversampled_ the film since that does a better job capturing
the detail that is there.

Again, you are assuming that there's high-MTF information between 55 and 78
lp/mm in the scan. That seems to be wrong. Everything up there is
attenuated.

Yes. That's what lots of people say. Which means that just as 2800 dpi
scanners do hideous things to Kodak Gold 200, 4000 dpi may be a problematic
scan density for Velvia.

Another good reason to postprocess.

(With my scanner, Velvia's pretty noisy, but it cleans up nicely with
NeatImage. Since old NeatImage (the new version is said to be faster) on my
old computer took 25 minutes CPU time per frame, I decided to shoot Provia
100F and Velvia 100F. Maybe I should give Velvia a try again.)

David J. Littleboy
Tokyo, Japan

 

<jpc> wrote in message
SNIP

Some links:
<http://www.fillfactory.com/htm/technology/htm/high_fill.htm>
including the first reference at the bottom of the page mentions some
numbers (you have to almost read between the lines).

<http://www.kodak.com/global/plugins/acrobat/en/digital/ccd/papersArti
cles/bluePlusOverview.pdf> gives a summary for several Kodak sensors,
and more detail per sensor at:
<http://www.kodak.com/global/en/digital/ccd/products/fullframe/fullfra
meFamilyPublications.jhtml?id=0.1.4.12&lc=en> at the top right for
their Blue Plus full frame series.

Well depth levels can sometimes be derived by looking at "Saturation
level" in eV, or by dynamic range numbers in dB combined with noise
levels.

Bart