Re: ISO 200000 ?

eawckyegcy@xxxxxxxxx wrote:
Roger N. Clark (change username to rnclark) wrote:


Back to your desired test of imaging under a cloudy night sky, can you
post a link to an image that shows what an image from the intensified
devices looks like? I also assume that these devices are used with no
filters in order to capture the most photons possible. Remember a DSLR
has a colored Bayer filter to do color imaging, which reduces the
possible photon count to the sensor by 5 to 10x compared to that with
no filters. That has nothing to do with sensor performance.


You assume right. The L3Vision gizmo McEwen has referenced at

http://e2vtechnologies.com/datasheets/publications/brochures/l3c95.pdf

is fed a raw 400-1060nm stream of photons. Furthermore, the pixels are
gigantic: 15 x 35.5 micrometres (EIA model), or about 8x larger than
Canon pixels (assuming the same fill-factor, probably not a safe
assumption given the CCD nature and the monster size of the pixel in
the L3V). If we accept a factor of 5-10 for the lack of IR, and a
factor of 8 for the pixel, a fair test would be to bin Canon images in
squares anywhere from 6 to 9 on a side, or increase the exposure time
to 40-80x. The CFA is also a complication; perhaps only the "G"
pixels should be considered (though the multi-spectral nature can
probably be used to some advantage)? What exactly is the spectrum of
"overcast starlight" anyways? All I can find is the relatively useless
lux level.


Yes, I agree with your assessment. The bandwidth of the CCD with
no filters is on the order of 0.7 and the DSLR bayer filter on the
order of 0.07 microns in full-width-half-maximum, so a factor of
10 for bandwidth, and another 0.8 to 0.9 factor for transmission
of the filter for a total factor of ~ 10/.85 ~ 11 to 12. Your point
about pixel size is another important factor too. We can add pixels
in the DSLR to bring the two systems into equality.

I think there has been a number of misunderstandings in this
thread, and some unfortunate name calling. I have registered
at the e2v.com, downloaded and read many of the articles, and
watched the videos. I too find it frustrating that Kennedy waved
his hands and said the information is there. He should have
referenced real papers and could have cleared up a lot of
confusion early on.

Example: Kennedy said to see the videos, and implied there would
be data/images/videos taken under overcast night conditions. I could
not find any. The lowest light level I could find was video number
4 (I think that is the right number) which was driving under
starlight. One thing to note about that video was you see (car) lights
in the distance at the beginning.

What is the most important paper, in my view is the one you
referenced that is on the e2v website after registration but which
you found a direct link to:

Sub-Electron Read Noise at MHz Pixel Rates
http://www.arcetri.astro.it/~rtubbs/papers/lllccd/sern_main.html
SPIE 4306 Conference Proceedings, 289-298, January 2001.

This paper has some key statements, many of which Kennedy alluded
to as secret, yet here they are in the open (SPIE) literature.
The key statements about performance of these devices are,
quoting from the paper:

Mode "1. The conventional CCD mode, with no gain in the multiplication
register, and the signal-to-noise set by the photon shot noise
added in quadrature with the readout noise of the CCD output amplifier."

This means that in a high signal environment the device works exactly
like conventional CCDs. From data in the paper, this regime is
a few hundred photons, exactly what both of us were saying.

Mode 2. "The CCD operated with a gain in the multiplication register
that substantially overcomes the readout noise of the output
amplifier. In this case the signal-to-noise is worse than would
be expected from the number of photons detected by a factor of
root 2. Another way to think of this degradation is to calculate
on the assumption that the signal-to-noise is set by the photon
shot noise but that the detector has half the detective quantum
efficiency that it has been for mode 1 above."

I pointed this out earlier and Kennedy said there was a way around it.
The way around it is only in single photon counting mode. That mode
only works when there is a maximum of one photon per pixel per frame.
So mode 2 works at signal levels less than a few hundred photons
per pixel per frame (this is very dark conditions), but the QE is reduced.
At the upper end, the S/N would be worse than conventional CCDs,
but better as one approaches a few photons where read noise in
conventional CCD becomes dominant. THIS IS EXACTLY WHAT BOTH OF
US WERE SAYING AND MODELED and which Kennedy criticized us saying
we were wrong.

Mode 3. "The CCD is operated with high gain in the multiplication
register so that the readout noise of the CCD output amplifier is
completely negligible for each multiplied electron. If each event
is then thresholded and treated as a single event on equal weight
without making any attempt to consider its amplitude then the
quantum efficiency that is lost by operating in mode 2 above is
restored, giving the same quantum efficiency essentially as that
in mode 1 above. There are, however, the major limitations that the
maximum photon rate used must be kept extremely low in order to
avoid coincidence losses which will give rise to non-linearities
in the response curve of the detector system, and the corresponding
need for deep cooling to maintain correspondingly low levels of
dark current."

This is the mode where the device outshines other devices: an imaging
photon counting device. But photon rates must be <= 1 photon per
pixel per frame. Very dark and will not work at all at high
photon imaging rates because if you got 2, 3, 5, 12,000, ... photons in
one pixel, it would only be counted as one.

Bottom line: there is NO "several orders of magnitude" improvement in
performance that Kennedy claimed. While there are several orders
of magnitude gain to the system, that gain is solely to reduce
effective read noise. There is a reduction in effective read noise
to the system by several orders of magnitude, but that does not
translate to several orders of magnitude sensitivity gain as
Kennedy claimed. The reason for this is at these low photon count
rates, noise is dominated by Poisson statistics of the photons,
not sensor or system electronics noise. These papers, the data, the
text make no such claims of orders of magnitude SENSITIVITY gains, nor
do the data support such claims. The papers, however, provide excellent
data and the text clearly shows what the devices do, how they are better
than conventional CCDs for certain applications, and what the limitations
are.

It is clear that: they are NOT several orders of magnitude more
sensitive at any level in terms of photons detected/noise, the
only reasonable definition, and the one you find in these papers.

Quote from the above paper:
"There are two main regimes experienced when operating CCD cameras
for scientific applications. In the high light level case the photon
shot noise is significantly greater than the system readout noise.
Here there is no advantage at all in using the L3Vision CCD technology."

"In the other low light level regime, the photon shot noise in the image
is comparable to or less than the readout noise. The effect is to make
the overall system noise significantly poorer than would be expected
purely from photon shot noise statistics, something that effectively
reduces the overall system detective quantum efficiency. It is in
this regime that the L3Vision CCD technology has a great deal to offer."

I showed in a previous thread how that gave better S/N in certain
low signal levels of a few photons that a perfect system would have
factors of several over systems with 3 electron read noise. After
reading the e2v papers, that assessment is unchanged. These devices
produce at best S/N a few times better at the bottom end (a few
photons/pixel/frame). That few can be very important for some
applications (military and scientific), so I am not trying to
minimize the sensor.

Note the papers on the L3Vision technology date from 2001. Since
then, Canon has developed CMOS with lower read noise than the ~15
electrons used in the L3Vision CCD technology. That is what I said in
earlier threads. It is clear that as read noise becomes lower
the advantage of the L3Vision technology is less because the
L3Vision technology is designed for the purpose of reducing
effective read noise.

Kennedy also criticized me when I said the high gain reduced the
dynamic range of the devices. Quote from the above paper:
"In order to preserve dynamic range as much as possible it is
sensible to minimize the multiplication register gain since
that gain value is the same factor by which the dynamic range
of the device is reduced."

Finally, if people want to test their cameras at low light levels,
I have added info to my page to derive lux from a digital camera:

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

See equations 3 and 4, which gives the calibration of a scene
measured in exposure time, f/ratio, and iso, to lux.

Coming this weekend will be first quarter moon. The intensity
from the moon at 1st quarter is about 0.027 lux. It will be
interesting to see what can be imaged in an outdoor scene
with that light level, which is about 270 times brighter
than a moonless overcast night away from city lights.

The moonless overcast night away from city lights problem
is really difficult. The formula for lux (this would be
0.0001 lux) says you need a 970s exposure at f/1 ISO3200
to properly expose an 18% gray card (see table 2 on my web
page above). A Canon 10D needs 250 photons at iso 3200
(see table 1 on the above web page) so 0.0001 lux would be
970/250 = 0.3 detected photons/second (take off the bayer
filter and that would rise to about 3.4 photons/second).
The e2v sensor has larger pixels, so if one adds about 9
10D pixels together, to equal the area of the e2v
sensor (one model), that would raise the photons/second to
30.6 photons/second (which at video rates is around 1 photon
per pixel per frame).

Previously, I said I did a 30-minute exposure (1800 seconds)
and got an upper limit of 178 photons at f/1.8 iso3200.
With the exposure calibration to lux, that level corresponds
to an upper limit of 0.00002 lux, or at least 5 times darker
than a cloudy night away from the city. Adding hundreds
of pixels together still detects no signal, pushing the
upper limit to < 0.000001 lux (< 1 microlux).

Bottom line, the ISO 20,000 claim of the Korean device
is bogus. So are claims of orders of magnitude more
sensitive of e2v sensors.
This is not to denigrate the e2v sensors,
which are impressive for certain situations, and work
as expected and no wild claims are made in their
literature or research papers, in contrast to claims in
this thread.

I still plan to do some low light examples of what a DSLR can
do. That has already been done and well demonstrated on
astronomical objects, but it would be interesting to see
similar results for night scene conditions.

Roger



.

Loading