«

A few weeks ago we had a long discussion on a new
announcement of an ISO 20,000 sensor followed by
discussion of intensified sensors. Some of us
pointed out that digital cameras already had quantum
efficiencies in the tens of percent with very low read
noise. Kennedy didn't think low noise DSLRs could
detect very low light levels amounting to less than
one photon per pixel per frame.

I've done an extensive analysis of my Canon 1D Mark II
and performed tests to see what could be done. I determined
the quantum efficiency of the 1D Mark II sensor is
about 28% and the system read noise is only 3.9
electrons. (The 20D has similar low read noise, and
probably similar QE.)

This performance results in great low signal detections and
I show on a new web page detecting image detail with
only 0.1 photon per pixel per frame in low light conditions
such as 0.00016 lux, and providing credible signal detection
when combining only 64 frames.
See:

Night and Low Light Photography with Digital Cameras
http://www.clarkvision.com/photoinfo...ht.photography

A detailed evaluation of the Canon 1D Mark II read noise,
full well capacity, and dynamic range as a function of ISO
is at:

Procedures for Evaluating Digital Camera Noise and Full
Well Capacities; Canon 1D Mark II Analysis
http://www.clarkvision.com/imagedetail/evaluation-1d2

The determination of quantum efficiency is at:

Digital Cameras: Counting Photons, Photometry, and Quantum Efficiency
http://www.clarkvision.com/imagedeta...photons.and.qe

The bottom line is that if one removed the Bayer filter from
current DSLRs and used the full range of the CMOS/CCD sensor,
the performance we now have in color at ISO 1,600 we would have
at ISO 16,000 in black and white (but with images like infrared film).
So claims of a new higher sensitivity sensor that is orders of
magnitude more sensitive than current DSLRs is bogus from all
the data I've seen as well as investigated on the above pages.

Roger

On Sun, 26 Feb 2006 22:46:46 -0800, Roger N. Clark (change username
to rnclark) wrote:

> A few weeks ago we had a long discussion on a new
> announcement of an ISO 20,000 sensor followed by
> discussion of intensified sensors.
>
> . . .
>
> The bottom line is that if one removed the Bayer filter from
> current DSLRs and used the full range of the CMOS/CCD sensor,
> the performance we now have in color at ISO 1,600 we would have
> at ISO 16,000 in black and white (but with images like infrared film).
> So claims of a new higher sensitivity sensor that is orders of
> magnitude more sensitive than current DSLRs is bogus from all
> the data I've seen as well as investigated on the above pages.

An ISO of 20,000 isn't an order of magnitude greater than 16,000,
but did you really mean "20,000"? The thread you mentioned referred
to an ISO an order of magnitude higher than that. So if the
respective sensors are 16,000 and 200,000, that would be slightly
more than an order of magnitude difference, but less than "orders",
which would need a super sensor of at least 1,600,000 ISO.

There's another point I'm not clear on. The OP from that thread
says that the sensor's sensitivity is 2000X. (2000 x 100 ==
200,000). That would seem to imply that it was based on a
comparison with the more or less universally accepted value of ISO
100 for the native sensitivity of today's typical camera. I
understand why you chose to compare ISO 16,000 with ISO 20,000 (or
200,000) but is that fair? I didn't read the news item posted in
the OP. Did it claim that the sensor had an effective ISO only for
B&W? If it did, your results may be more or less valid. But even
if so, from reading your messages it seems that you might not be a
disinterested observer, and instead of trying to determine what the
Korean sensor is actually capable of producing, seem to be looking
mainly for ways to back your earlier assertions that it can't be as
good as claimed. And if you accidentally used ISO 20,000 instead of
the claimed ISO 200,000, then the Korean sensor's ISO more than
meets the 10x improvement which you previously said is bogus.

Here's a clip from one of your replies in that thread:

> Message-ID: <>
> Date: Sun, 29 Jan 2006 11:03:01 -0800
> From: "Roger N. Clark (change username to rnclark)"
> Newsgroups: rec.photo.digital
> Subject: Re: ISO 200000 ?
> . . .
>
>> Even if the claimed
>> "2000X light sensitivity" can't be attained, most people would be
>> giddy if only a 500X improvement could be achieved. The first
>> result would be that cameras would no longer need to include
>> flashes, so only external units would be sold. Bye bye red-eye,
>> unless as the OP asked, it turns out to be too good to be true.
>>
>
> But the point is that even 500x is bogus, so is 10x. The quantum
> efficiency is already 20 to 30% in consumer systems. You can up it
> 3 to 5x if you could get to ~100% quantum efficiency. That is for
> the high signal case (more than about 50 to 100 photons). For the
> low signal case, read noise dominates. Best systems are already in
> the 3 to 4 electron read noise range, so you see big gains in photon
> detection below a few photons. That is a regime that doesn't
> make photographs. That was the point of the tables I posted.

ASAAR wrote:

> On Sun, 26 Feb 2006 22:46:46 -0800, Roger N. Clark (change username
> to rnclark) wrote:

>>A few weeks ago we had a long discussion on a new
>>announcement of an ISO 20,000 sensor followed by
>>discussion of intensified sensors.

>>
>>The bottom line is that if one removed the Bayer filter from
>>current DSLRs and used the full range of the CMOS/CCD sensor,
>>the performance we now have in color at ISO 1,600 we would have
>>at ISO 16,000 in black and white (but with images like infrared film).
>>So claims of a new higher sensitivity sensor that is orders of
>>magnitude more sensitive than current DSLRs is bogus from all
>>the data I've seen as well as investigated on the above pages.
>
> An ISO of 20,000 isn't an order of magnitude greater than 16,000,
> but did you really mean "20,000"? The thread you mentioned referred
> to an ISO an order of magnitude higher than that. So if the
> respective sensors are 16,000 and 200,000, that would be slightly
> more than an order of magnitude difference, but less than "orders",
> which would need a super sensor of at least 1,600,000 ISO.

Your are right. I mean to write 200,000.

> There's another point I'm not clear on. The OP from that thread
> says that the sensor's sensitivity is 2000X. (2000 x 100 ==
> 200,000). That would seem to imply that it was based on a
> comparison with the more or less universally accepted value of ISO
> 100 for the native sensitivity of today's typical camera. I
> understand why you chose to compare ISO 16,000 with ISO 20,000 (or
> 200,000) but is that fair?

I think you are missing the point. Sensitivity is expressed as
photon conversion efficiency, called the quantum efficiency (QE).
Important too is low electronics noise (low read noise).
Current QE of DSLRs are in the 20 to 35+% range and backside illuminated
CCDs are in the 90+% range. So sensor sensitivity of current
DSLRs can only be improved 3 to 4x to be perfect. So any claim
of "orders of magnitude" improvement was bogus from the start for
a high signals tens of photons and greater.

One person argued that it was not the high level signal that is
the improvement but the low signal regime, including less than
a photon per pixel per frame. I posted astronomical images
that I new went deep, recording low numbers of photons.
But I had no information on how low that really was compared
to the E2V sensors whose web sites showed terrestrial night
scenes and lab images where the photons recorded were less than
one photon per pixel per frame. So no one really knew how low
a DSLR could go.

From my experience with other sensors, and with DSLRs in astronomical
imaging, I knew one could record signals less than the system noise,
but I did not have any hard data at the below one photon per pixel
level. So I was challenged to show what could be done with DSLRs.

After these new tests, it proves that DSLRs can record signals less
than 1 photon per pixel per frame. In fact the performance is
significantly better than I thought. In a terrestrial night,
fraction of a second image, DSLRs record less than a photon
per pixel per frame per fraction of a second. In astronomical
imaging, that level gets pushed lower: less than a photon per
pixel per frame per minute(s).

I didn't read the news item posted in
> the OP. Did it claim that the sensor had an effective ISO only for
> B&W? If it did, your results may be more or less valid. But even
> if so, from reading your messages it seems that you might not be a
> disinterested observer, and instead of trying to determine what the
> Korean sensor is actually capable of producing, seem to be looking
> mainly for ways to back your earlier assertions that it can't be as
> good as claimed. And if you accidentally used ISO 20,000 instead of
> the claimed ISO 200,000, then the Korean sensor's ISO more than
> meets the 10x improvement which you previously said is bogus.

Regardless of what number they claim, 10x, 100x, ISO 20,000, ISO 200,000,
there is no actual sensitivity gain beyond improvement in quantum
efficiency of 3 to 4x. More than that is like digital zoom to magnify an
image! Call digital ISO zoom. The sensitivity is the Quantum efficiency.
The read noise is low enough in the better DSLRs such that
photon noise is still a significant factor in the lowest signals.

If the Korean sensor claims ISO 200,000, it must be using different
specifications than is used in DSLRs, and correspondingly
different noise specs. Digital ISO zoom. You can
do that now in photoshop.

The main improvement in the sub one photon per pixel case of the
intensified sensors is lowering read noise, but you still have
photon noise and in the case I show on the web page of a moonlit
scene, the zero read noise sensor would show only about a 2x improvement
in image signal-to-noise ratio over current DSLRs.

Yes, the 200,000 ISO claim is bogus, and current DSLRs are performing
very well indeed.

Roger

On Mon, 27 Feb 2006 08:00:37 -0700, Roger N. Clark (change username
to rnclark) wrote:

> The main improvement in the sub one photon per pixel case of the
> intensified sensors is lowering read noise, but you still have
> photon noise and in the case I show on the web page of a moonlit
> scene, the zero read noise sensor would show only about a 2x improvement
> in image signal-to-noise ratio over current DSLRs.
>
> Yes, the 200,000 ISO claim is bogus, and current DSLRs are performing
> very well indeed.

If everything you claim is correct, you raise an interesting point
nearly as significant as the claims made for the Korean sensor. What
is involved with using "the full range of the CMOS/CCD sensor"? You
say that when that is combined with removing the Bayer filter, the
color ISO 1,600 performance could be boosted to b&w ISO 16,000
performance (with images resembling infrared film). As a fairly
large number of people would love to be able to have an infrared
sensitive camera having performance comparable to ISO 16, why aren't
cameras given this capability? Security concerns? If the only
change was to remove the Bayer filter, where do you estimate the
increased ISO would fall between 1,600 and 16,000?

In article <>, "Roger N. Clark (change
username to rnclark)" <> writes
>A few weeks ago we had a long discussion on a new
>announcement of an ISO 20,000 sensor followed by
>discussion of intensified sensors. Some of us
>pointed out that digital cameras already had quantum
>efficiencies in the tens of percent with very low read
>noise. Kennedy didn't think low noise DSLRs could
>detect very low light levels amounting to less than
>one photon per pixel per frame.
>
Actually Roger, as I suspect you are aware, that is not what I said. In
fact it was another correspondent in that thread who claimed that less
than one photon per pixel per frame was impossible to achieve. In fact,
not only did I state that this was possible but provided evidence of
intensifiers which did achieve this.

My comment was that digital SLRs were unable to achieve a useful image
in 10^-4 lux (overcast starlight) illumination in an exposure of 1/50th
of a second with any amount of post processing - which is what the
intensified camera achieved and which is also what high ISO would
achieve. None of your examples on your page come close to that - the
1/20 sec moonlit single frames appear to be about a factor of almost 700
wide of the target, based on the illumination level and the exposure you
used.

Close, but no cheese this time.
--
Kennedy
Yes, Socrates himself is particularly missed;
A lovely little thinker, but a bugger when he's ****ed.
Python Philosophers (replace 'nospam' with 'kennedym' when replying)

In article <>, "Roger N. Clark (change
username to rnclark)" <> writes
>
>I think you are missing the point. Sensitivity is expressed as
>photon conversion efficiency, called the quantum efficiency (QE).
>
Actually Roger, I think *you* are missing the point, still. Sensitivity
isn't *just* quantum efficiency! For example, those digital cameras
which achieve ISO3200 today don't have a QE that is 4x better than the
cameras that have only ISO800. In fact, both detectors have pretty much
the same QE.

>Yes, the 200,000 ISO claim is bogus, and current DSLRs are performing
>very well indeed.
>
Your 1/20sec exposures in quarter moonlight, are still more than two
orders of magnitude away from the objective, using the data on your
page, an objective that intensifiers and LLTV sensors achieve every day,
well, night actually.
--
Kennedy
Yes, Socrates himself is particularly missed;
A lovely little thinker, but a bugger when he's ****ed.
Python Philosophers (replace 'nospam' with 'kennedym' when replying)

ASAAR wrote:

> If everything you claim is correct, you raise an interesting point
> nearly as significant as the claims made for the Korean sensor. What
> is involved with using "the full range of the CMOS/CCD sensor"?

Stripping out the CFA, back-illumination, and upsizing the pixel. Read
noise reduction via L3CCD techniques (start writing those checks to
TI), super-fast pixel rates (for DR), etc.

> You
> say that when that is combined with removing the Bayer filter, the
> color ISO 1,600 performance could be boosted to b&w ISO 16,000
> performance (with images resembling infrared film).
> As a fairly
> large number of people would love to be able to have an infrared
> sensitive camera having performance comparable to ISO 16, why aren't
> cameras given this capability?

N people want an IR camera. C*N people want a color camera. C is a
very large constant (probably over a thousand, probably even higher).
Special fab run for a few photographic kooks? Maybe some day. Not
even the 20Da got a CFA-less sensor, since it was likely too big a hit
to the production costs ("Why buy this $4k un-cooled camera when I can
get a fancy-ass cooled one from SBIG for a bit more?")

> Security concerns?

Despite the claims of some here, it's already far too late for that.
Not that this will keep El Presidente from withdrawing the Poisson
distribution from the public sphere....

> If the only change was to remove the Bayer filter, where do you estimate the
> increased ISO would fall between 1,600 and 16,000?

I can't find a definitive spectrum of "overcast starlight". However:

http://arxiv.org/PS_cache/astro-ph/pdf/9909/9909153.pdf

(and others of the genre)

may be representative (ignore some of the spikes in the visual due to
so-called civilization). Note the large increase in intensity at the
near infrared, and also recall that as the wavelength increases, energy
per photon decreases, so the number of photons available scales even
more. Now recall that the Bayer CFA puts essentially tight (but lossy)
passbands in the visual [[look at the spectrum again: UBVRI passbands
are shown: B=blue, V=green, R=red]] and there is a deep IR filter
beforehand (this IR filter is needed because the dye filters in the
typical photographic cameras are can be transparent to longer IR).
The exact number needs an integration, but we can expect at least a
factor of 2 (just for bandwidth), but clearly there is much more to be
gained.

In the end, the game is converting photons. This is what "quantum
efficiency" is all about: sensitivity. The labels we attach to these
photons -- what photographers call "ISO" -- are beside the point.
Anyone who claims or says otherwise needs to present new physics to
back up their statements, because current, known, physics 'explains'
all existing cameras. Even soooper-dooper secret decoder ring level
cameras.

ASAAR wrote:
> On Mon, 27 Feb 2006 08:00:37 -0700, Roger N. Clark (change username
> to rnclark) wrote:
>
>
>>The main improvement in the sub one photon per pixel case of the
>>intensified sensors is lowering read noise, but you still have
>>photon noise and in the case I show on the web page of a moonlit
>>scene, the zero read noise sensor would show only about a 2x improvement
>>in image signal-to-noise ratio over current DSLRs.
>>
>>Yes, the 200,000 ISO claim is bogus, and current DSLRs are performing
>>very well indeed.
>
>
> If everything you claim is correct, you raise an interesting point
> nearly as significant as the claims made for the Korean sensor. What
> is involved with using "the full range of the CMOS/CCD sensor"? You
> say that when that is combined with removing the Bayer filter, the
> color ISO 1,600 performance could be boosted to b&w ISO 16,000
> performance (with images resembling infrared film). As a fairly
> large number of people would love to be able to have an infrared
> sensitive camera having performance comparable to ISO 16, why aren't
> cameras given this capability? Security concerns? If the only
> change was to remove the Bayer filter, where do you estimate the
> increased ISO would fall between 1,600 and 16,000?
>
Consider that CMOS and CCD sensors have response from less than
4000 angstroms to about 10,000 angstroms, with a full width
at half maximum from about 3800 to 8500 angstroms. Full Width
at Half Maximum (FWHM) is the point at which the response
of a system falls by a factor of two from the peak response.
The FWHM is a good approximation of the throughput of a system.
So the FWHM is ~ 8500 - 3800 = 4700 angstroms.

But the green filter has a full width at half maximum from
about 5000 to 5770 angstroms, so a FWHM of only about 770 angstroms.
Then the green filter is somewhat less than 100% transmission.

The FWHM ratio of the sensor and green filter is ~4700/770 ~ 6.
Multiplying in transmission due to the color filter times
the transmission of the IR blocking filter (somewhere
from 80% to less than 25% seem to be numbers in documents (but
that may include detector quantum efficiency). If the transmission
is as high as 80%, the effect of removing the filters would let
6/0.8 = 7.5x more light in. If the transmission is 25%, the
improvement would be 6/.25 = 24! A factor of 10 seems
to be consistent with other data I've seen in scientific papers.

It is a good question why it hasn't been done. Amateur astronomers
have removed the IR blocking filter and replaced them with
filters than block less IR light (google Heutech: they do it for
a fee), then Canon did it with the 20Da. But for someone
wanting a black and white sensor, lenses working over such a large
wavelength range would be an issue, but a more modest design
could replace the IR + color filter that only allowed light
from the UV to about 7000 angstroms, and gaining some 6x in
speed should be easy. You would also need to put in a compensating
piece of glass in place of the color filter so the focal point
doesn't change (so the AF system maintains accuracy).

Example spectral response functions:
http://www.graftek.com/pdf/Manuals/b...600fmanual.pdf
see figures 1-3 and 1-4 which shows QE spectral response and
filter transmission.

If you are interested, you might contact Heutech and see if they
would do a modification for you. (I have no affiliation with
them nor have I bought anything from them).

Roger

Kennedy McEwen wrote:

> In article <>, "Roger N. Clark (change
> username to rnclark)" <> writes
>
>> I think you are missing the point. Sensitivity is expressed as
>> photon conversion efficiency, called the quantum efficiency (QE).
>>
> Actually Roger, I think *you* are missing the point, still. Sensitivity
> isn't *just* quantum efficiency! For example, those digital cameras
> which achieve ISO3200 today don't have a QE that is 4x better than the
> cameras that have only ISO800. In fact, both detectors have pretty much
> the same QE.

Ahah! Finally an admission. I've been saying that for quite
a while. Let's normalize the test conditions.

>> Yes, the 200,000 ISO claim is bogus, and current DSLRs are performing
>> very well indeed.
>>
> Your 1/20sec exposures in quarter moonlight, are still more than two
> orders of magnitude away from the objective, using the data on your
> page, an objective that intensifiers and LLTV sensors achieve every day,
> well, night actually.

Let's normalize the conditions between sensors to see what the
true difference in the sensors really is.

The LLTV intensified sensors in your test are

1) Black and white, with no color filters, and no IR blocking filter.
This is a huge factor, about 10x between the bandwidth difference
and the transmission loss of the filters.
2) They are back side illuminated with QE >~ 90%, so about
3.4x higher sensitivity than DSLRs.
3) The videos on the E2V site use f/1.4 lenses, I only have
and f/1.8 lens, so another factor of 1.7x difference.
4) Like you admit above with different digital cameras achieve
different ISOs, which is due to pixel size, the E2V sensors
are much larger than DSLR pixels. If I remember right, it
was something like 25x35 microns compared to 8.2 microns
for the Canon 1D Mark II. The area ratio is then about
13x for those numbers.

So equalizing the two systems, there is 10*3.4*1.7*13 ~ 750x
difference. Now of that 750x, only 3.4x is actual sensitivity
differences between the two systems. 750/3.4 = 220 in optical
differences.

With the DSLR results showing it the sensor can detect
and provide image information at the <1 photon per pixel per
frame and normalizing the test conditions, proves there are
not "orders of magnitude" improvement by using the intensified
sensors.

So the lab test showed detecting 0.1 photon per pixel per frame
in a 0.00016 lux 1-second exposure. That equals (multiply
by 220 and divide by 1.6 then another 50 (to get to your
1/50 second exposure:

= 0.1 *220/1.6/50

= 0.27 photon, per pixel per frame per 1/50 second,
which is above the detection threshold already
established.

In this low light condition, the E2V back side illuminated sensor
would get 0.27 * 3.4 ~ 0.9 photon per pixel per frame. With
its lower read noise, photon noise would dominate any image
made, thus would be another ~2x lower noise over the bare DSLR
sensor. So between effectively zero read noise, and the QE,
the sensor would perform about 7x better than the DSLR,
not "orders of magnitude." The no course as light levels
increase to a few tens of photons, then performance is only
the QE difference.

QED

Roger

Kennedy McEwen wrote:
> In article <>, "Roger N. Clark (change
> username to rnclark)" <> writes
>
>> A few weeks ago we had a long discussion on a new
>> announcement of an ISO 20,000 sensor followed by
>> discussion of intensified sensors. Some of us
>> pointed out that digital cameras already had quantum
>> efficiencies in the tens of percent with very low read
>> noise. Kennedy didn't think low noise DSLRs could
>> detect very low light levels amounting to less than
>> one photon per pixel per frame.
>>
> Actually Roger, as I suspect you are aware, that is not what I said. In
> fact it was another correspondent in that thread who claimed that less
> than one photon per pixel per frame was impossible to achieve.


I guess I misunderstood when you said:
Kennedy McEwen wrote (Re: ISO 200000 ?), 1/24/2006 3:04 PM:
> However that 0.5 phot/pix/frame is now substantially
> lower than the read noise which is applied to *every* pixel. Hence my
> insistence that the read noise you quote for the Canon device of 3
> photons is much higher than a useful signal level.

Kennedy McEwen wrote (Re: ISO 200000 ?), 1/26/2006 4:01 PM:
> What I am not convinced of is that
> the example you referenced is a tougher challenge than the one I gave -
> NOT the issue about detecting signals in SNRs of less than unity.

Kennedy McEwen wrote (Re: ISO 200000 ?), 1/23/2006 2:31 PM:
> How is the integration of several frames with a noise
> floor of >1 phot/pix/frame equal to <1 phot/pix/frame?

Kennedy McEwen wrote (Re: ISO 200000 ?), 1/28/2006 7:49 PM:
> Limitations which are still orders of magnitude
> beyond what can be achieved with a DSLR sensor.

Kennedy McEwen wrote (Re: ISO 200000 ?), 1/29/2006 1:55 PM
> the Korean press release made it quite clear why they are
> interested in such sensor improvements, and they are only
> claiming useful images in 0.1lux.

(Note the moonlit scene test with a DSLR got images at 0.06 lux,
and the lab test 0.00016 lux).

In fact,
> not only did I state that this was possible but provided evidence of
> intensifiers which did achieve this.
>
> My comment was that digital SLRs were unable to achieve a useful image
> in 10^-4 lux (overcast starlight) illumination in an exposure of 1/50th
> of a second with any amount of post processing - which is what the
> intensified camera achieved and which is also what high ISO would
> achieve. None of your examples on your page come close to that - the
> 1/20 sec moonlit single frames appear to be about a factor of almost 700
> wide of the target, based on the illumination level and the exposure you
> used.

See my other post. The optics difference (color filters on the
DSLR, bare sensor on your intensified sensor, very large pixel size)
means an optical difference of about 220. The .06 lux scene
would then be equivalent to .00027 lux scene when the sensor
sensitivities are directly compared. That makes the difference
to 0.0001 lux and 1/50 second only a factor of 6.7, 3.4 of that
is the QE difference (>90% for the intensified sensor, versus
about 28% for the DSLR). So now we have a remaining 2x which
is the real improvement in this test case: the lower read noise
of the intensified sensor.
>
> Close, but no cheese this time.

Yep, for you.

Roger