Readout noise in sCMOS cameras

[Sripad Ram-2]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hello Gerhard,
This is a slightly off-topic question in connection to your recent response
to the thread on the choice of sCMOS cameras.

You made a comment that the distribution of noise in sCMOS is not Gaussian.

Can you clarify whether you meant noise during the readout process (charge
to voltage conversion step)? If it is not Gaussian, what is the underlying
noise process? Is there any literature that you can point me to?

Thanks.

Sripad

[Michael Giacomelli]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hi Sripad,

The read noise (error when measuring number of photoelectrons in a
pixel) distribution of (s)CMOS sensors is definitely not Gaussian.  It
is an asymmetric distribution where a few pixels are essentially
noiseless, most have read noise a little less than the median value,
and then there are a long tail of increasingly bad pixels with much
higher read noise:

http://camera.hamamatsu.com/jp/en/technical_guides/read_noise/index.html

This noise reflects in part (I believe) the extent to which charge
becomes trapped in the insulating oxide layer around each pixel,
resulting in random fluctuations in electron counts.  Since many CMOS
sensors have median read noise in the 1-5 electron range, it doesn't
take much trapped charge to have a big influence on photoelectron
counts. There is a lot of literature on making low noise CMOS pixels,
but I couldn't suggest any specific paper.

Mike

On Mon, Mar 5, 2018 at 12:54 PM, S Ram <[hidden email]> wrote:

> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
> Hello Gerhard,
> This is a slightly off-topic question in connection to your recent response
> to the thread on the choice of sCMOS cameras.
>
> You made a comment that the distribution of noise in sCMOS is not Gaussian.
>
> Can you clarify whether you meant noise during the readout process (charge
> to voltage conversion step)? If it is not Gaussian, what is the underlying
> noise process? Is there any literature that you can point me to?
>
> Thanks.
>
> Sripad

[BROWNE Mark]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Here is a good paper on noise in emccd and sCMOS

arxiv 1506.07929

Mark
Sent from my iPhone

> On Mar 5, 2018, at 7:02 PM, Michael Giacomelli <[hidden email]> wrote:
>
> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
> Hi Sripad,
>
> The read noise (error when measuring number of photoelectrons in a
> pixel) distribution of (s)CMOS sensors is definitely not Gaussian.  It
> is an asymmetric distribution where a few pixels are essentially
> noiseless, most have read noise a little less than the median value,
> and then there are a long tail of increasingly bad pixels with much
> higher read noise:
>
> http://camera.hamamatsu.com/jp/en/technical_guides/read_noise/index.html
>
> This noise reflects in part (I believe) the extent to which charge
> becomes trapped in the insulating oxide layer around each pixel,
> resulting in random fluctuations in electron counts.  Since many CMOS
> sensors have median read noise in the 1-5 electron range, it doesn't
> take much trapped charge to have a big influence on photoelectron
> counts. There is a lot of literature on making low noise CMOS pixels,
> but I couldn't suggest any specific paper.
>
> Mike
>
>> On Mon, Mar 5, 2018 at 12:54 PM, S Ram <[hidden email]> wrote:
>> *****
>> To join, leave or search the confocal microscopy listserv, go to:
>> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
>> Post images on http://www.imgur.com and include the link in your posting.
>> *****
>>
>> Hello Gerhard,
>> This is a slightly off-topic question in connection to your recent response
>> to the thread on the choice of sCMOS cameras.
>>
>> You made a comment that the distribution of noise in sCMOS is not Gaussian.
>>
>> Can you clarify whether you meant noise during the readout process (charge
>> to voltage conversion step)? If it is not Gaussian, what is the underlying
>> noise process? Is there any literature that you can point me to?
>>
>> Thanks.
>>
>> Sripad

___________________________________________________________________________
This e-mail is confidential and is for the addressee only.   Please refer to
www.oxinst.com/email-statement<http://www.oxinst.com/email-statement> for regulatory information.

[James Pawley]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Dear Sripad,

In a CMOS chip, every pixel has its own read amp. All of these vary slightly in gain and DC-offset. So the raw output from a a black (no light) image would have a noise term related to how much the offsets of the pixel amps varied and a uniform white image would have Poisson noise on the photo charge plus a noise term mostly related to the variation in the gains of the pixel amplifiers.
In an sCMOS chip to these uncertainties must be added variations in the gain and offer of the 4,000-plus separate ADCs mounted at the edges of the chip. An effort is made to correct for the multi-amplifier and multi-digitizer noise by “flat fielding” the raw data from the chip using data from previous “black” and “white” images, The system works quite well but as the white image always involves a lot of charge, its Poisson noise (sqrt of n) is large and this can skew the results. So can using the chip at a different temperature, dwell time or pixel clock than was used for the “black” and “white” images. Other sources on non-“Gaussian”  noise include “hot pixels” (perhaps leaky photodiodes that are sometimes flagged and removed by the camera system software).

Indeed, the noise spectrum in these low-light systems is almost never “Gaussian”. Even if the electronic noise (that signal variation which becomes evident when reading the same pixel with no light signal) seems Gaussian, it is usually caused by Poisson Noise (Or Johnson noise) affecting the small number of electrons that constitute the (fairly table currents passing through the elements of the charge amplifier. And of course, at signal levels of more than a few dozen photoelectrons, Poisson Noise on the PE number soon dominates most other noise sources (not hot pixels).

Gaussian noise is just easier to think about, and easier to model. We should remember that in low-light photodetectors, it is almost never appropriate. (Poisson Noise get bigger as the signal increases!).

Best,,

JP
James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC, Canada, V0N3A0 [hidden email]<mailto:[hidden email]>, Phone [phone], cell 1-604-989-6146



On Mar 5, 2018, at 9:54 AM, S Ram <[hidden email]<mailto:[hidden email]>> wrote:

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hello Gerhard,
This is a slightly off-topic question in connection to your recent response
to the thread on the choice of sCMOS cameras.

You made a comment that the distribution of noise in sCMOS is not Gaussian.

Can you clarify whether you meant noise during the readout process (charge
to voltage conversion step)? If it is not Gaussian, what is the underlying
noise process? Is there any literature that you can point me to?

Thanks.

Sripad

[Kyle Michael Douglass]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hi Sripad,


On 03/05/2018 06:54 PM, S Ram wrote:
>
> You made a comment that the distribution of noise in sCMOS is not Gaussian.
>
> Can you clarify whether you meant noise during the readout process (charge
> to voltage conversion step)? If it is not Gaussian, what is the underlying
> noise process? Is there any literature that you can point me to?

I would just like to highlight that a common source of confusion on this topic is the failure to distinguish between temporal noise and spatial non-uniformities.

The temporal noise refers to the noise in the ADU count value from a single pixel over time. Spatial non-uniformities, however, come in two flavors: a variation in the statistical moments of the temporal noise between pixels (dark signal non-uniformity, which includes read noise) and a variation in gain between pixels (photoresponse non-uniformity). As already mentioned, CMOS sensors suffer from both photoresponse and dark signal non-uniformities, and these are usually what are referred to when I see people talk about "CMOS noise."

Making the distinction between the two types of noise is important because I would think that a Gaussian is a pretty good model for *temporal* CMOS read noise for a *single* pixel. I would justify this statement by noting that typical measurements of the read noise produce the mean and variance of the pixel's dark signal ADU values. Unless higher order moments of the temporal noise are incorporated into the modeling, then it's difficult to justify a more complex distribution than a Gaussian.

I highly recommend the EMVA 1288 standard for more reading on these topics:http://www.emva.org/standards-technology/emva-1288/emva-standard-1288-downloads/

Best,
Kyle

--
Kyle M. Douglass, PhD
Post-doctoral researcher
The Laboratory of Experimental Biophysics
EPFL, Lausanne, Switzerland
http://kmdouglass.github.io
http://leb.epfl.ch

[Jeremy Adler-5]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****



Dear James,
a minor point from your interesting post.

"Poisson Noise get bigger as the signal increases!"

This is true as an absolute measure of the size of the variation around the mean but your statement could be misinterpreted as a preference for detecting fewer PEs.
Perhaps adding -      but is smaller as a fraction of the signal
An average of 16 PEs has a standard deviation of 25% while for  256 PEs the SD drops to around 6% and to 3% for 1024.


Jeremy Adler
IGP, Uppsala U, Sweden
====================================




[phone]

http://www.biovis.uu.se







-----Original Message-----
From: Confocal Microscopy List [mailto:[hidden email]] On Behalf Of JAMES B PAWLEY
Sent: den 6 mars 2018 07:22
To: [hidden email]
Subject: Re: Readout noise in sCMOS cameras

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Dear Sripad,

In a CMOS chip, every pixel has its own read amp. All of these vary slightly in gain and DC-offset. So the raw output from a a black (no light) image would have a noise term related to how much the offsets of the pixel amps varied and a uniform white image would have Poisson noise on the photo charge plus a noise term mostly related to the variation in the gains of the pixel amplifiers.
In an sCMOS chip to these uncertainties must be added variations in the gain and offer of the 4,000-plus separate ADCs mounted at the edges of the chip. An effort is made to correct for the multi-amplifier and multi-digitizer noise by “flat fielding” the raw data from the chip using data from previous “black” and “white” images, The system works quite well but as the white image always involves a lot of charge, its Poisson noise (sqrt of n) is large and this can skew the results. So can using the chip at a different temperature, dwell time or pixel clock than was used for the “black” and “white” images. Other sources on non-“Gaussian”  noise include “hot pixels” (perhaps leaky photodiodes that are sometimes flagged and removed by the camera system software).

Indeed, the noise spectrum in these low-light systems is almost never “Gaussian”. Even if the electronic noise (that signal variation which becomes evident when reading the same pixel with no light signal) seems Gaussian, it is usually caused by Poisson Noise (Or Johnson noise) affecting the small number of electrons that constitute the (fairly table currents passing through the elements of the charge amplifier. And of course, at signal levels of more than a few dozen photoelectrons, Poisson Noise on the PE number soon dominates most other noise sources (not hot pixels).

Gaussian noise is just easier to think about, and easier to model. We should remember that in low-light photodetectors, it is almost never appropriate. (Poisson Noise get bigger as the signal increases!).

Best,,

JP
James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC, Canada, V0N3A0 [hidden email]<mailto:[hidden email]>, Phone [phone], cell 1-604-989-6146



On Mar 5, 2018, at 9:54 AM, S Ram <[hidden email]<mailto:[hidden email]>> wrote:

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hello Gerhard,
This is a slightly off-topic question in connection to your recent response to the thread on the choice of sCMOS cameras.

You made a comment that the distribution of noise in sCMOS is not Gaussian.

Can you clarify whether you meant noise during the readout process (charge to voltage conversion step)? If it is not Gaussian, what is the underlying noise process? Is there any literature that you can point me to?

Thanks.

Sripad

[James Pawley]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Dear Jeremy,

Of course you correct.

I stated it as I did because I find that people who correctly understand that the signal-to-Poisson Noise ratio increases with the sqrt of the signal level forget that the absolute value of the uncertainty in the signal level increases with the sqrt of the signal level, i.e., they are surprised to see that the “grass” on the signal of apparently uniform bright areas is much greater than that on dark areas.

While it may be convenient to estimate that  the noise in low contrast widefield image from a CCD can be approximated as being Gaussian, this really is not true when applied to images from, for instance, an slow-scan CCD image from a confocal. In the latter, many pixels will be essentially black and in these pixels, the noise will be Gaussian and represent the read noise of the charge-to-voltage converter (much of which is Johnson noise related to the electron statistics of small, fairly constant currents). However, it would be a mistake to think that this same Gaussian accurately defines the uncertainty of signals from the brighter parts of the image (>100 photoelectrons?).

sCMOS has a whole raft of other noise terms and as noted the noise in their output is poorly estimated by a Gaussian, particularly on images having high-contrast.

Great sensors though!

JP

James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC, Canada, V0N3A0 [hidden email]<mailto:[hidden email]>, Phone [phone], cell 1-604-989-6146



On Mar 6, 2018, at 2:47 AM, Jeremy Adler <[hidden email]<mailto:[hidden email]>> wrote:

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****



Dear James,
a minor point from your interesting post.

"Poisson Noise get bigger as the signal increases!"

This is true as an absolute measure of the size of the variation around the mean but your statement could be misinterpreted as a preference for detecting fewer PEs.
Perhaps adding -      but is smaller as a fraction of the signal
An average of 16 PEs has a standard deviation of 25% while for  256 PEs the SD drops to around 6% and to 3% for 1024.


Jeremy Adler
IGP, Uppsala U, Sweden
====================================




[phone]

http://www.biovis.uu.se







-----Original Message-----
From: Confocal Microscopy List [mailto:[hidden email]] On Behalf Of JAMES B PAWLEY
Sent: den 6 mars 2018 07:22
To: [hidden email]
Subject: Re: Readout noise in sCMOS cameras

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Dear Sripad,

In a CMOS chip, every pixel has its own read amp. All of these vary slightly in gain and DC-offset. So the raw output from a a black (no light) image would have a noise term related to how much the offsets of the pixel amps varied and a uniform white image would have Poisson noise on the photo charge plus a noise term mostly related to the variation in the gains of the pixel amplifiers.
In an sCMOS chip to these uncertainties must be added variations in the gain and offer of the 4,000-plus separate ADCs mounted at the edges of the chip. An effort is made to correct for the multi-amplifier and multi-digitizer noise by “flat fielding” the raw data from the chip using data from previous “black” and “white” images, The system works quite well but as the white image always involves a lot of charge, its Poisson noise (sqrt of n) is large and this can skew the results. So can using the chip at a different temperature, dwell time or pixel clock than was used for the “black” and “white” images. Other sources on non-“Gaussian”  noise include “hot pixels” (perhaps leaky photodiodes that are sometimes flagged and removed by the camera system software).

Indeed, the noise spectrum in these low-light systems is almost never “Gaussian”. Even if the electronic noise (that signal variation which becomes evident when reading the same pixel with no light signal) seems Gaussian, it is usually caused by Poisson Noise (Or Johnson noise) affecting the small number of electrons that constitute the (fairly table currents passing through the elements of the charge amplifier. And of course, at signal levels of more than a few dozen photoelectrons, Poisson Noise on the PE number soon dominates most other noise sources (not hot pixels).

Gaussian noise is just easier to think about, and easier to model. We should remember that in low-light photodetectors, it is almost never appropriate. (Poisson Noise get bigger as the signal increases!).

Best,,

JP
James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC, Canada, V0N3A0 [hidden email]<mailto:[hidden email]>, Phone [phone], cell 1-604-989-6146



On Mar 5, 2018, at 9:54 AM, S Ram <[hidden email]<mailto:[hidden email]>> wrote:

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hello Gerhard,
This is a slightly off-topic question in connection to your recent response to the thread on the choice of sCMOS cameras.

You made a comment that the distribution of noise in sCMOS is not Gaussian.

Can you clarify whether you meant noise during the readout process (charge to voltage conversion step)? If it is not Gaussian, what is the underlying noise process? Is there any literature that you can point me to?

Thanks.

Sripad

[Sripad Ram-2]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Thanks everyone for the insightful remarks, and Kyle for the comment about
temporal versus spatial non-uniformity.

Regards,
Sripad


On Tue, Mar 6, 2018 at 7:24 AM, JAMES B PAWLEY <[hidden email]> wrote:

> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
> Dear Jeremy,
>
> Of course you correct.
>
> I stated it as I did because I find that people who correctly understand
> that the signal-to-Poisson Noise ratio increases with the sqrt of the
> signal level forget that the absolute value of the uncertainty in the
> signal level increases with the sqrt of the signal level, i.e., they are
> surprised to see that the “grass” on the signal of apparently uniform
> bright areas is much greater than that on dark areas.
>
> While it may be convenient to estimate that  the noise in low contrast
> widefield image from a CCD can be approximated as being Gaussian, this
> really is not true when applied to images from, for instance, an slow-scan
> CCD image from a confocal. In the latter, many pixels will be essentially
> black and in these pixels, the noise will be Gaussian and represent the
> read noise of the charge-to-voltage converter (much of which is Johnson
> noise related to the electron statistics of small, fairly constant
> currents). However, it would be a mistake to think that this same Gaussian
> accurately defines the uncertainty of signals from the brighter parts of
> the image (>100 photoelectrons?).
>
> sCMOS has a whole raft of other noise terms and as noted the noise in
> their output is poorly estimated by a Gaussian, particularly on images
> having high-contrast.
>
> Great sensors though!
>
> JP
>
> James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC,
> Canada, V0N3A0 [hidden email]<mailto:[hidden email]>, Phone
> 1-604-885-0840, cell 1-604-989-6146
>
>
>
> On Mar 6, 2018, at 2:47 AM, Jeremy Adler <[hidden email]
> <mailto:[hidden email]>> wrote:
>
> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
>
>
> Dear James,
> a minor point from your interesting post.
>
> "Poisson Noise get bigger as the signal increases!"
>
> This is true as an absolute measure of the size of the variation around
> the mean but your statement could be misinterpreted as a preference for
> detecting fewer PEs.
> Perhaps adding -      but is smaller as a fraction of the signal
> An average of 16 PEs has a standard deviation of 25% while for  256 PEs
> the SD drops to around 6% and to 3% for 1024.
>
>
> Jeremy Adler
> IGP, Uppsala U, Sweden
> ====================================
>
>
>
>
> [phone]
>
> http://www.biovis.uu.se
>
>
>
>
>
>
>
> -----Original Message-----
> From: Confocal Microscopy List [mailto:[hidden email]]
> On Behalf Of JAMES B PAWLEY
> Sent: den 6 mars 2018 07:22
> To: [hidden email]
> Subject: Re: Readout noise in sCMOS cameras
>
> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
> Dear Sripad,
>
> In a CMOS chip, every pixel has its own read amp. All of these vary
> slightly in gain and DC-offset. So the raw output from a a black (no light)
> image would have a noise term related to how much the offsets of the pixel
> amps varied and a uniform white image would have Poisson noise on the photo
> charge plus a noise term mostly related to the variation in the gains of
> the pixel amplifiers.
> In an sCMOS chip to these uncertainties must be added variations in the
> gain and offer of the 4,000-plus separate ADCs mounted at the edges of the
> chip. An effort is made to correct for the multi-amplifier and
> multi-digitizer noise by “flat fielding” the raw data from the chip using
> data from previous “black” and “white” images, The system works quite well
> but as the white image always involves a lot of charge, its Poisson noise
> (sqrt of n) is large and this can skew the results. So can using the chip
> at a different temperature, dwell time or pixel clock than was used for the
> “black” and “white” images. Other sources on non-“Gaussian”  noise include
> “hot pixels” (perhaps leaky photodiodes that are sometimes flagged and
> removed by the camera system software).
>
> Indeed, the noise spectrum in these low-light systems is almost never
> “Gaussian”. Even if the electronic noise (that signal variation which
> becomes evident when reading the same pixel with no light signal) seems
> Gaussian, it is usually caused by Poisson Noise (Or Johnson noise)
> affecting the small number of electrons that constitute the (fairly table
> currents passing through the elements of the charge amplifier. And of
> course, at signal levels of more than a few dozen photoelectrons, Poisson
> Noise on the PE number soon dominates most other noise sources (not hot
> pixels).
>
> Gaussian noise is just easier to think about, and easier to model. We
> should remember that in low-light photodetectors, it is almost never
> appropriate. (Poisson Noise get bigger as the signal increases!).
>
> Best,,
>
> JP
> James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC,
> Canada, V0N3A0 [hidden email]<mailto:[hidden email]>, Phone
> 1-604-885-0840, cell 1-604-989-6146
>
>
>
> On Mar 5, 2018, at 9:54 AM, S Ram <[hidden email]<mailto:s
> [hidden email]>> wrote:
>
> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
> Hello Gerhard,
> This is a slightly off-topic question in connection to your recent
> response to the thread on the choice of sCMOS cameras.
>
> You made a comment that the distribution of noise in sCMOS is not Gaussian.
>
> Can you clarify whether you meant noise during the readout process (charge
> to voltage conversion step)? If it is not Gaussian, what is the underlying
> noise process? Is there any literature that you can point me to?
>
> Thanks.
>
> Sripad
>
>
>

[Gerhard Holst]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hi Sripad,

I just have read the all the contributions, because I returned to my office today and no access to my mailbox the last two days.

I was referring to the fact, that the dark noise distribution in sCMOS can be approximated by a Gaussian, but by the shape it is visible, that it is not, because of that "tail" with the high noise pixels. Reasons for that are manifold, some related to the semiconductor manufacturing process, to the chosen technology (surface channels vs. buried channels), and some of them simply to the high number of parallel readouts, because it doesn't matter how precise you handle the process, the gains of the all the transistors, the resistor an capacitor values, the A/D converters all of them might have some slight even small differences, which might contribute to the noise distribution.
This in fact was a big advantage of CCDs, because they had a serial readout, all charge packages generated by each individual pixel, had to pass the same readout circuit, therefore they all received more or less the same noise contribution by the readout circuit, and this resulted in a noise distribution which could be approximated by a Gaussian distribution.

Concerning the different types of noise, I could nothing add to what James B Pawley and Kyle Douglass have said, and supporting that everybody who is interested in these topics should have a look to the EMVA1288 standard. If you are more interested in the general noise sources you might want to have alook at the webinars and publications by Prof. Albert Theuwissen or the papers of Jim Janesick.

with best regards,

Gerhard


Dr. Gerhard Holst
Head of Science & Research
[phone]
[phone]

PCO AG, Donaupark 11, 93309 Kelheim, Germany, www.pco.de
USt. ID-Nr. / VAT: DE128590843, Registergericht / Register court: Regensburg HRB 9157
Sitz der Gesellschaft / Registered office: Kelheim, Vorstand / Chairman: Dr. Emil Ott
Vorsitzender des Aufsichtsrats / Chairman of the supervisory board: Johann Plöb

-----Ursprüngliche Nachricht-----
Von: Confocal Microscopy List [mailto:[hidden email]] Im Auftrag von S Ram
Gesendet: Dienstag, 6. März 2018 19:52
An: [hidden email]
Betreff: Re: Readout noise in sCMOS cameras

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Thanks everyone for the insightful remarks, and Kyle for the comment about temporal versus spatial non-uniformity.

Regards,
Sripad


On Tue, Mar 6, 2018 at 7:24 AM, JAMES B PAWLEY <[hidden email]> wrote:

> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
> Dear Jeremy,
>
> Of course you correct.
>
> I stated it as I did because I find that people who correctly
> understand that the signal-to-Poisson Noise ratio increases with the
> sqrt of the signal level forget that the absolute value of the
> uncertainty in the signal level increases with the sqrt of the signal
> level, i.e., they are surprised to see that the “grass” on the signal
> of apparently uniform bright areas is much greater than that on dark areas.
>
> While it may be convenient to estimate that  the noise in low contrast
> widefield image from a CCD can be approximated as being Gaussian, this
> really is not true when applied to images from, for instance, an
> slow-scan CCD image from a confocal. In the latter, many pixels will
> be essentially black and in these pixels, the noise will be Gaussian
> and represent the read noise of the charge-to-voltage converter (much
> of which is Johnson noise related to the electron statistics of small,
> fairly constant currents). However, it would be a mistake to think
> that this same Gaussian accurately defines the uncertainty of signals
> from the brighter parts of the image (>100 photoelectrons?).
>
> sCMOS has a whole raft of other noise terms and as noted the noise in
> their output is poorly estimated by a Gaussian, particularly on images
> having high-contrast.
>
> Great sensors though!
>
> JP
>
> James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC,
> Canada, V0N3A0 [hidden email]<mailto:[hidden email]>, Phone
> 1-604-885-0840, cell 1-604-989-6146
>
>
>
> On Mar 6, 2018, at 2:47 AM, Jeremy Adler
> <[hidden email] <mailto:[hidden email]>> wrote:
>
> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
>
>
> Dear James,
> a minor point from your interesting post.
>
> "Poisson Noise get bigger as the signal increases!"
>
> This is true as an absolute measure of the size of the variation
> around the mean but your statement could be misinterpreted as a
> preference for detecting fewer PEs.
> Perhaps adding -      but is smaller as a fraction of the signal
> An average of 16 PEs has a standard deviation of 25% while for  256
> PEs the SD drops to around 6% and to 3% for 1024.
>
>
> Jeremy Adler
> IGP, Uppsala U, Sweden
> ====================================
>
>
>
>
> [phone]
>
> http://www.biovis.uu.se
>
>
>
>
>
>
>
> -----Original Message-----
> From: Confocal Microscopy List
> [mailto:[hidden email]]
> On Behalf Of JAMES B PAWLEY
> Sent: den 6 mars 2018 07:22
> To: [hidden email]
> Subject: Re: Readout noise in sCMOS cameras
>
> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
> Dear Sripad,
>
> In a CMOS chip, every pixel has its own read amp. All of these vary
> slightly in gain and DC-offset. So the raw output from a a black (no
> light) image would have a noise term related to how much the offsets
> of the pixel amps varied and a uniform white image would have Poisson
> noise on the photo charge plus a noise term mostly related to the
> variation in the gains of the pixel amplifiers.
> In an sCMOS chip to these uncertainties must be added variations in
> the gain and offer of the 4,000-plus separate ADCs mounted at the
> edges of the chip. An effort is made to correct for the
> multi-amplifier and multi-digitizer noise by “flat fielding” the raw
> data from the chip using data from previous “black” and “white”
> images, The system works quite well but as the white image always
> involves a lot of charge, its Poisson noise (sqrt of n) is large and
> this can skew the results. So can using the chip at a different
> temperature, dwell time or pixel clock than was used for the “black”
> and “white” images. Other sources on non-“Gaussian”  noise include
> “hot pixels” (perhaps leaky photodiodes that are sometimes flagged and removed by the camera system software).
>
> Indeed, the noise spectrum in these low-light systems is almost never
> “Gaussian”. Even if the electronic noise (that signal variation which
> becomes evident when reading the same pixel with no light signal)
> seems Gaussian, it is usually caused by Poisson Noise (Or Johnson
> noise) affecting the small number of electrons that constitute the
> (fairly table currents passing through the elements of the charge
> amplifier. And of course, at signal levels of more than a few dozen
> photoelectrons, Poisson Noise on the PE number soon dominates most
> other noise sources (not hot pixels).
>
> Gaussian noise is just easier to think about, and easier to model. We
> should remember that in low-light photodetectors, it is almost never
> appropriate. (Poisson Noise get bigger as the signal increases!).
>
> Best,,
>
> JP
> James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC,
> Canada, V0N3A0 [hidden email]<mailto:[hidden email]>, Phone
> 1-604-885-0840, cell 1-604-989-6146
>
>
>
> On Mar 5, 2018, at 9:54 AM, S Ram <[hidden email]<mailto:s
> [hidden email]>> wrote:
>
> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> Post images on http://www.imgur.com and include the link in your posting.
> *****
>
> Hello Gerhard,
> This is a slightly off-topic question in connection to your recent
> response to the thread on the choice of sCMOS cameras.
>
> You made a comment that the distribution of noise in sCMOS is not Gaussian.
>
> Can you clarify whether you meant noise during the readout process
> (charge to voltage conversion step)? If it is not Gaussian, what is
> the underlying noise process? Is there any literature that you can point me to?
>
> Thanks.
>
> Sripad
>
>
>

[Zdenek Svindrych-2]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hi Listers,
I have a related question: since every pixel of a CMOS sensor behaves a bit
differently, some of them are so much off the specs that they can be
regarded as bad pixels (hot pixels, dead pixels, etc). For example one of
our 4-megapixel front-illuminated sCMOS cameras has over 11000 bad pixels
(they are easy to spot, as they are 'corrected' by averaging neighboring
pixels, so their standard deviation is much lower that what is dictated by
Poisson statistics).

What are the specifications for the number of bad pixels? I could not find
any useful info on the websites of the camera vendors, nor the chip
manufacturer. Does anyone know how the new back-illuminated sensors compare
with the mainstream 4 megapixel CIS2020 (or SCI2020?) sensors in terms of
bad pixels?


Thanks!

Best, zdenek
-- 
Zdenek Svindrych, Ph.D.
Research Associate - Imaging Specialist
Department of Biochemistry and Cell Biology
Geisel School of Medicine at Dartmouth
email: [hidden email]


---------- Původní e-mail ----------
Od: JAMES B PAWLEY <[hidden email]>
Komu: [hidden email]
Datum: 6. 3. 2018 2:23:24
Předmět: Re: Readout noise in sCMOS cameras
"*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Dear Sripad,

In a CMOS chip, every pixel has its own read amp. All of these vary slightly
in gain and DC-offset. So the raw output from a a black (no light) image
would have a noise term related to how much the offsets of the pixel amps
varied and a uniform white image would have Poisson noise on the photo
charge plus a noise term mostly related to the variation in the gains of the
pixel amplifiers.
In an sCMOS chip to these uncertainties must be added variations in the gain
and offer of the 4,000-plus separate ADCs mounted at the edges of the chip.
An effort is made to correct for the multi-amplifier and multi-digitizer
noise by “flat fielding” the raw data from the chip using data from previous
“black” and “white” images, The system works quite well but as the white
image always involves a lot of charge, its Poisson noise (sqrt of n) is
large and this can skew the results. So can using the chip at a different
temperature, dwell time or pixel clock than was used for the “black” and “
white” images. Other sources on non-“Gaussian” noise include “hot pixels”
(perhaps leaky photodiodes that are sometimes flagged and removed by the
camera system software).

Indeed, the noise spectrum in these low-light systems is almost never “
Gaussian”. Even if the electronic noise (that signal variation which becomes
evident when reading the same pixel with no light signal) seems Gaussian, it
is usually caused by Poisson Noise (Or Johnson noise) affecting the small
number of electrons that constitute the (fairly table currents passing
through the elements of the charge amplifier. And of course, at signal
levels of more than a few dozen photoelectrons, Poisson Noise on the PE
number soon dominates most other noise sources (not hot pixels).

Gaussian noise is just easier to think about, and easier to model. We should
remember that in low-light photodetectors, it is almost never appropriate.
(Poisson Noise get bigger as the signal increases!).

Best,,

JP
James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC, Canada,
V0N3A0 [hidden email]<mailto:[hidden email]>, Phone [phone],
cell 1-604-989-6146



On Mar 5, 2018, at 9:54 AM, S Ram <[hidden email]<mailto:sripad.ram@
GMAIL.COM>> wrote:

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hello Gerhard,
This is a slightly off-topic question in connection to your recent response
to the thread on the choice of sCMOS cameras.

You made a comment that the distribution of noise in sCMOS is not Gaussian.

Can you clarify whether you meant noise during the readout process (charge
to voltage conversion step)? If it is not Gaussian, what is the underlying
noise process? Is there any literature that you can point me to?

Thanks.

Sripad

"

[Gerhard Holst]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hi Zdenek,

now, I can only speak for our cameras, because I don't know what the others are doing. In general there are two classes of bad, defect or hot pixels.

Class "defect pixels":
they don't react to light or their dynamic is dramatically reduced (for example by having a pretty high start value), these pixels are replaced by a kind of neighborhood average value, always. They are known and stored in a defect or hot pixel list in the camera, and this list can be read out, and in case that the averaging operation has an unwanted impact on applied image processing algorithms, the averaging can be switched off by an SDK (software development kit) command.

Class "blinkers" or "higher noise" pixels (referring to the tail in the histogram:
They have higher pixel values in some images and in some images they behave normal, or some of the pixels show a larger noise. They are locally fixed, such they can be identified (which for example thermally or clock induce charges in emCCDs are not). For these pixels the camera manufacturers have installed a filter, which can remove them dynamically. Usually it is a kind of lowpass which takes the signal into account, which means in case a single pixel has a value that is more than x times sigma of the neighborhood average different, it is assumed not to contain information, and it is replaced by a weighted neighborhood average. But these are software filters, which can be switched off either by SDK or by the controlling camera application software.

Our storage or the list for defect pixel values can contain as max 4096 values, which are usually not filled, therefore I am surprised by your number of 11000 bad pixels, which to me appears to be too high, based on my experience. But maybe you had the software filter on, because as demonstrated by the noise histograms, there are still some "higher noise" pixels, which are not considered to be "defect".

Right now, I don't have enough statistical data to compare the new sCMOS image sensors to the former ones, but the current data support, that the new image sensors show a better noise behavior, which allows to use them as well without cooling.

with best regards,

Gerhard


Dr. Gerhard Holst
Head of Science & Research
[phone]
[phone]

PCO AG, Donaupark 11, 93309 Kelheim, Germany, www.pco.de
USt. ID-Nr. / VAT: DE128590843, Registergericht / Register court: Regensburg HRB 9157
Sitz der Gesellschaft / Registered office: Kelheim, Vorstand / Chairman: Dr. Emil Ott
Vorsitzender des Aufsichtsrats / Chairman of the supervisory board: Johann Plöb


-----Ursprüngliche Nachricht-----
Von: Confocal Microscopy List [mailto:[hidden email]] Im Auftrag von [hidden email]
Gesendet: Freitag, 9. März 2018 21:44
An: [hidden email]
Betreff: Re: Readout noise in sCMOS cameras

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
Post images on http://www.imgur.com and include the link in your posting.
*****

Hi Listers,
I have a related question: since every pixel of a CMOS sensor behaves a bit differently, some of them are so much off the specs that they can be regarded as bad pixels (hot pixels, dead pixels, etc). For example one of our 4-megapixel front-illuminated sCMOS cameras has over 11000 bad pixels (they are easy to spot, as they are 'corrected' by averaging neighboring pixels, so their standard deviation is much lower that what is dictated by Poisson statistics).

What are the specifications for the number of bad pixels? I could not find any useful info on the websites of the camera vendors, nor the chip manufacturer. Does anyone know how the new back-illuminated sensors compare with the mainstream 4 megapixel CIS2020 (or SCI2020?) sensors in terms of bad pixels?


Thanks!

Best, zdenek
--
Zdenek Svindrych, Ph.D.
Research Associate - Imaging Specialist
Department of Biochemistry and Cell Biology Geisel School of Medicine at Dartmouth
email: [hidden email]


---------- Původní e-mail ----------
Od: JAMES B PAWLEY <[hidden email]>
Komu: [hidden email]
Datum: 6. 3. 2018 2:23:24
Předmět: Re: Readout noise in sCMOS cameras
"*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy 
Post images on http://www.imgur.com and include the link in your posting.
*****

Dear Sripad,

In a CMOS chip, every pixel has its own read amp. All of these vary slightly
in gain and DC-offset. So the raw output from a a black (no light) image
would have a noise term related to how much the offsets of the pixel amps
varied and a uniform white image would have Poisson noise on the photo
charge plus a noise term mostly related to the variation in the gains of the
pixel amplifiers.
In an sCMOS chip to these uncertainties must be added variations in the gain
and offer of the 4,000-plus separate ADCs mounted at the edges of the chip.
An effort is made to correct for the multi-amplifier and multi-digitizer
noise by “flat fielding” the raw data from the chip using data from previous
“black” and “white” images, The system works quite well but as the white
image always involves a lot of charge, its Poisson noise (sqrt of n) is
large and this can skew the results. So can using the chip at a different
temperature, dwell time or pixel clock than was used for the “black” and “
white” images. Other sources on non-“Gaussian” noise include “hot pixels”
(perhaps leaky photodiodes that are sometimes flagged and removed by the
camera system software).

Indeed, the noise spectrum in these low-light systems is almost never “
Gaussian”. Even if the electronic noise (that signal variation which becomes
evident when reading the same pixel with no light signal) seems Gaussian, it
is usually caused by Poisson Noise (Or Johnson noise) affecting the small
number of electrons that constitute the (fairly table currents passing
through the elements of the charge amplifier. And of course, at signal
levels of more than a few dozen photoelectrons, Poisson Noise on the PE
number soon dominates most other noise sources (not hot pixels).

Gaussian noise is just easier to think about, and easier to model. We should
remember that in low-light photodetectors, it is almost never appropriate.
(Poisson Noise get bigger as the signal increases!).

Best,,

JP
James and Christine Pawley, 5446 Burley Place, Box 2348, Sechelt BC, Canada,
V0N3A0 [hidden email]<mailto:[hidden email]>, Phone [phone],
cell 1-604-989-6146



On Mar 5, 2018, at 9:54 AM, S Ram <[hidden email]<mailto:sripad.ram@
GMAIL.COM>> wrote:

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy 
Post images on http://www.imgur.com and include the link in your posting.
*****

Hello Gerhard,
This is a slightly off-topic question in connection to your recent response
to the thread on the choice of sCMOS cameras.

You made a comment that the distribution of noise in sCMOS is not Gaussian.

Can you clarify whether you meant noise during the readout process (charge
to voltage conversion step)? If it is not Gaussian, what is the underlying
noise process? Is there any literature that you can point me to?

Thanks.

Sripad

"