Read noise: get it 30x down?

I looked through papers on sensor read noise, and it looks like all
the major contributors to it (kT and 1/f) are random. AND, reading is
non-destructive.

And I can't combine these two statements together. It LOOKS LIKE if
one reads the sensel voltage 100 times, then averages the value, then
the noise should go down 10 times. So: does it?

One example: the new Sony APS CMOS sensor has 4000 ADCs; this is about
1000x more than the previous generation of sensors. In particular (at
least, if one forgets power consumption issue), they could spend 1000x
time to read each sensel.

Assume they read each sensel 100 times, and assume they get the same
read noise as Canon did in its CMOS sensors (about 4 electrons). Then
the noise goes down to 0.4 electrons.

Sounds too good to be true... Puzzled,
Ilya

[A complimentary Cc of this posting was NOT [per weedlist] sent to
Ilya Zakharevich
<nospam-abuse@ilyaz.org>], who wrote in article <fdmlkr$1fld$1@agate.berkeley.edu>:
> I looked through papers on sensor read noise, and it looks like all
> the major contributors to it (kT and 1/f) are random. AND, reading is
> non-destructive.
>
> And I can't combine these two statements together. It LOOKS LIKE if
> one reads the sensel voltage 100 times, then averages the value, then
> the noise should go down 10 times. So: does it?

It was a temporary shutdown of mind. :-( ;-)

The fact that the noise is "colored" (not white) implies that the
measurements made at different times are not independent. Thus the
average noise won't scale down as 1/sqrt(N).

> One example: the new Sony APS CMOS sensor has 4000 ADCs; this is about
> 1000x more than the previous generation of sensors. In particular (at
> least, if one forgets power consumption issue), they could spend 1000x
> time to read each sensel.
>
> Assume they read each sensel 100 times, and assume they get the same
> read noise as Canon did in its CMOS sensors (about 4 electrons). Then
> the noise goes down to 0.4 electrons.
>
> Sounds too good to be true...

Now the REAL question becomes: what is the LOWER cut-off frequency of
the 1/f noise? ("Pure" 1/f noise is physically impossible, since it
carries infinite power; so the density should not increase that quick
on very low frequencies. My question is about this frequency. Yet in
other words: actual low-frequency power of actual noise coincides with
power of 1/f noise filtered above some frequency; what is it?)

If, for example, we know that this frequency is about 20KHz, then
spending more than 50 usec reading a pixel would significantly
decrease noise; while spending 5 usec vs 50 usec would not influence
the noise very much.

Thanks,
Ilya

Ilya Zakharevich wrote:
> [A complimentary Cc of this posting was NOT [per weedlist] sent to
> Ilya Zakharevich
> <nospam-abuse@ilyaz.org>], who wrote in article <fdmlkr$1fld$1@agate.berkeley.edu>:
>> I looked through papers on sensor read noise, and it looks like all
>> the major contributors to it (kT and 1/f) are random. AND, reading is
>> non-destructive.
>>
>> And I can't combine these two statements together. It LOOKS LIKE if
>> one reads the sensel voltage 100 times, then averages the value, then
>> the noise should go down 10 times. So: does it?
>
> It was a temporary shutdown of mind. :-( ;-)
>
> The fact that the noise is "colored" (not white) implies that the
> measurements made at different times are not independent. Thus the
> average noise won't scale down as 1/sqrt(N).
>
>> One example: the new Sony APS CMOS sensor has 4000 ADCs; this is about
>> 1000x more than the previous generation of sensors. In particular (at
>> least, if one forgets power consumption issue), they could spend 1000x
>> time to read each sensel.
>>
>> Assume they read each sensel 100 times, and assume they get the same
>> read noise as Canon did in its CMOS sensors (about 4 electrons). Then
>> the noise goes down to 0.4 electrons.
>>
>> Sounds too good to be true...
>
> Now the REAL question becomes: what is the LOWER cut-off frequency of
> the 1/f noise? ("Pure" 1/f noise is physically impossible, since it
> carries infinite power; so the density should not increase that quick
> on very low frequencies. My question is about this frequency. Yet in
> other words: actual low-frequency power of actual noise coincides with
> power of 1/f noise filtered above some frequency; what is it?)
>
> If, for example, we know that this frequency is about 20KHz, then
> spending more than 50 usec reading a pixel would significantly
> decrease noise; while spending 5 usec vs 50 usec would not influence
> the noise very much.
>
> Thanks,
> Ilya
>

As Roger Clark has pointed out the read noise is pretty low in most
cameras and quantizing noise is going to dominate, in which case there
would be little benefit in reducing it by a factor of 10.

The other part is that part of the noise will be how evenly the quantum
well are filled at the beginning of an exposure, if there in noise here
multiple reads will not help. In other words can you will a 50,000
electron well within an electron or 2 from cell to cell?

Scott

Ilya Zakharevich wrote:
> [A complimentary Cc of this posting was NOT [per weedlist] sent to
> Ilya Zakharevich
> <nospam-abuse@ilyaz.org>], who wrote in article <fdmlkr$1fld$1@agate.berkeley.edu>:
>> I looked through papers on sensor read noise, and it looks like all
>> the major contributors to it (kT and 1/f) are random. AND, reading is
>> non-destructive.
>>
>> And I can't combine these two statements together. It LOOKS LIKE if
>> one reads the sensel voltage 100 times, then averages the value, then
>> the noise should go down 10 times. So: does it?
>
> It was a temporary shutdown of mind. :-( ;-)
>
> The fact that the noise is "colored" (not white) implies that the
> measurements made at different times are not independent. Thus the
> average noise won't scale down as 1/sqrt(N).
>

What you want to do is read each pixel many times real fast ....
but also interleaved with these read a source known to be actually
constant. This allows subtracting out the slower 1/f noise.
It is the standard way of dealing with 1/f noise ("chopping").

On the other hand, if you get the thermal noise down enough,
and the actual quantization noise low enough, you can, at least at low
light levels, figure out the number of electrons exactly. This
gets you to the ultimate "perfect" situation.

I have no idea whether this is feasible on a CCD or CMOS sensor
running at camera speeds.

Doug McDonald

[A complimentary Cc of this posting was sent to
Scott W
<biphoto@hotmail.com>], who wrote in article <46fefd97$0$26341$4c368faf@roadrunner.com>:
> As Roger Clark has pointed out the read noise is pretty low in most
> cameras and quantizing noise is going to dominate, in which case there
> would be little benefit in reducing it by a factor of 10.

??? Roger has his blind spots, but on this topic he is quite kosher.
He would not say anything as wrong as this.

Just look at his tables: quantization dominates the low-ISO region.
Read noise dominates high-ISO region.

And adding another bit or two to quantizer is not a big deal.
Decreasing readout noise may be quite tricky...

> The other part is that part of the noise will be how evenly the quantum
> well are filled at the beginning of an exposure, if there in noise here
> multiple reads will not help. In other words can you will a 50,000
> electron well within an electron or 2 from cell to cell?

Readout noise of 4 electrons matters only on the individual cells
where the exposure is below about 25 electrons. So full well of 50Ke
is not relevant.

When low readout noise is very important is the situations when people
long for cameras of yesterday, those with 6MP. Given low enough
readout noise, you can combine pixels of 24MP camera so that it
performes as 6MP one - has lower noise, and lower resolution. (And
you can make it selectively - at low zones only.)

Hope this helps,
Ilya

[A complimentary Cc of this posting was sent to
Doug McDonald
<NOmcdonald@SPscs.AMuiuc.edu>], who wrote in article <fdn4p3$9m6$1@news.ks.uiuc.edu>:
> > The fact that the noise is "colored" (not white) implies that the
> > measurements made at different times are not independent. Thus the
> > average noise won't scale down as 1/sqrt(N).

> What you want to do is read each pixel many times real fast ....
> but also interleaved with these read a source known to be actually
> constant. This allows subtracting out the slower 1/f noise.
> It is the standard way of dealing with 1/f noise ("chopping").

Thanks; looks like google may be enough to self-educate myself. [The
problem of having no mental model of 1/f-noise is that I have no clue
what happens when you switch between sources...]

> On the other hand, if you get the thermal noise down enough,
> and the actual quantization noise low enough, you can, at least at low
> light levels, figure out the number of electrons exactly. This
> gets you to the ultimate "perfect" situation.

Are there cases where astronomers do this? Two years ago the best
reference I have seen was for 2e noise. (Starting from about 0.2e,
the noise becomes just a ditherer for the ADC. Seems like a long way
to go.)

Yours,
Ilya

Ilya Zakharevich wrote:
> [A complimentary Cc of this posting was sent to
> Scott W
> <biphoto@hotmail.com>], who wrote in article <46fefd97$0$26341$4c368faf@roadrunner.com>:
>> As Roger Clark has pointed out the read noise is pretty low in most
>> cameras and quantizing noise is going to dominate, in which case there
>> would be little benefit in reducing it by a factor of 10.
>
> ??? Roger has his blind spots, but on this topic he is quite kosher.
> He would not say anything as wrong as this.
>
> Just look at his tables: quantization dominates the low-ISO region.
> Read noise dominates high-ISO region.
>
> And adding another bit or two to quantizer is not a big deal.
> Decreasing readout noise may be quite tricky...
>
>> The other part is that part of the noise will be how evenly the quantum
>> well are filled at the beginning of an exposure, if there in noise here
>> multiple reads will not help. In other words can you will a 50,000
>> electron well within an electron or 2 from cell to cell?
>
> Readout noise of 4 electrons matters only on the individual cells
> where the exposure is below about 25 electrons. So full well of 50Ke
> is not relevant.
>

In a CCD the wells start full and discharge, so a 25 electron exposure
with a 50,000 electron well will take you from 50,000 to 49,975. In
which case you better be very careful to get 50,000 in your well be for
exposure and not say 50,010. This error in filling the wells will show
up as read noise, and doing multiple reads on a cell will not lower the
this component of noise.

Scott

Doug McDonald <NOmcdonald@SPscs.AMuiuc.edu> wrote in news:fdn4p3$9m6$1
@news.ks.uiuc.edu:

> On the other hand, if you get the thermal noise down enough,
> and the actual quantization noise low enough, you can, at least at low
> light levels, figure out the number of electrons exactly. This
> gets you to the ultimate "perfect" situation.
>
> I have no idea whether this is feasible on a CCD or CMOS sensor
> running at camera speeds.

Then the problem becomes, "how do you do ISO 100 on the same camera?".

Maybe we'll see dedicated low-light bodies in the future?

I believe that shot noise, in isolation, without any read noise, is much
less distracting than (current levels of) read noise, in very low photon
captures. I've simulated pure shot noise from well-exposed ISO 100 images,
and the shot noise looks quite nice compared to the garbage we call read
noise and total blackframe noise. Read noises always have a chance to be
patterned; shot noise can not be, not with any reasonable chance. Shot
noise looks like a texture in the subject; read noise looks like a film of
streaks, smears and fireworks super-imposed upon the image.

--

<>>< ><<> ><<> <>>< ><<> <>>< <>>< ><<>
John P Sheehy <JPS@no.komm>
><<> <>>< <>>< ><<> <>>< ><<> ><<> <>><

John Sheehy wrote:
> Doug McDonald <NOmcdonald@SPscs.AMuiuc.edu> wrote in news:fdn4p3$9m6$1
> @news.ks.uiuc.edu:
>
>> On the other hand, if you get the thermal noise down enough,
>> and the actual quantization noise low enough, you can, at least at low
>> light levels, figure out the number of electrons exactly. This
>> gets you to the ultimate "perfect" situation.
>>
>> I have no idea whether this is feasible on a CCD or CMOS sensor
>> running at camera speeds.
>
> Then the problem becomes, "how do you do ISO 100 on the same camera?".
>
> Maybe we'll see dedicated low-light bodies in the future?
>
> I believe that shot noise, in isolation, without any read noise, is much
> less distracting than (current levels of) read noise, in very low photon
> captures. I've simulated pure shot noise from well-exposed ISO 100 images,
> and the shot noise looks quite nice compared to the garbage we call read
> noise and total blackframe noise. Read noises always have a chance to be
> patterned; shot noise can not be, not with any reasonable chance. Shot
> noise looks like a texture in the subject; read noise looks like a film of
> streaks, smears and fireworks super-imposed upon the image.
>
John,
You're confusing fixed pattern noise sources with read noise.

But you do bring up a major limiting factor to image quality
in digital cameras. Contributing to fixed pattern noise are
sources like reset noise. Already done in some cameras is
double correlated sampling to try and reduce the fixed pattern
noise. But even if fixed pattern noise were zero, and
the read noise reduced below the current typical 4 electrons
in the best cameras, it really wouldn't help image quality
for most images. Having a signal of 4 (photons) gives a
signal-to-noise ratio of only 2 (very noisy). It will help
astronomers and photographers who co-add images.

Here is a good summary by a sensor manufacturer of
CCD noise sources:

http://learn.hamamatsu.com/articles/ccdsnr.html

Before read noise is further reduced, we need higher
quality A/D converters (more bits and higher S/N from the
A/D converters), and reduction of fixed pattern noise.

To answer Doug, thermal noise is insignificant in digital
cameras for short exposure times (less than a few tens of
seconds). E.g., the thermal dark current in the 1D mark II is
1/4 electron/second at room temperature, and noise is the
square root of the dark current. Newer cameras have even lower
dark current. Dark current of 0.25 electron/sec needs
a 64 second exposure to generate thermal noise = 4 electron
read noise.

Roger

On Sep 30, 8:12 pm, John Sheehy <J...@no.komm> wrote:
> Read noises always have a chance to be
> patterned; shot noise can not be, not with any reasonable chance. Shot
> noise looks like a texture in the subject; read noise looks like a film of
> streaks, smears and fireworks super-imposed upon the image.

Is the patterned read noise consistent from frame to frame? That is,
if I take two blackframe images and subtract them, does the patterned
noise cancel, or is it simply patterned, with the precise pattern
varying from image to image (so that the subtracted images would have
sqrt(2) as much pattern noise)?