Photon Transfer Curves

[ngc2403]

This is a copy of the the ptc for my camera at iso 800 and after a quick search on the internet you will see that the it is not that good

(click to enlarge)

ally

[ngc2403]

just to give you all a hand to grasp the graph it is a diagram of the signal to noise responce of the camera,

From this grasp you can find out how much read noise there is, how much the gain is and how long before you reach the max snr of any image

ally

[Sam]

mmmm.... what does that actually mean?

[John]

mmmm.... what does that actually mean?

Yep that's my reaction as well - thought maybe I'm a bit dim :scratch:

John

[beamer3.6m]

Does it mean the maximum exposure you can take before your image degrades - if not I have no idea either.

[dph1nm]

Actually this plot doesn't make much sense to me. If the units on the axis are electrons then your noise is much too low - a signal of 10**4 should give a noise of 10**2 or greater. However the read noise would then be about 15e-, which is sort of sensible, and full well 60000e-, also sensible.

If the numbers are DN then the gain would have to be about 11 (i.e. 11 e- per DN) for 10**4 DN to give noise of ~30DN. But then the read noise would be about 160e- and the full well 660000e-, neither of which makes any sense.

It should look something like this one for the D70.

NigelM

[ngc2403]

Hello,

i am not sure of the meaning of it to be honest but to possibibly answer Nigels question i think i may have used natural log by mistake hence the values don't add up.

i am a going to be doing this as part of my astro project early next year and i am awaiting narrowbandpaul's help on sorting it out, however in the mean time i can say this.

the camera graph should be smooth! and this is not,

it should have a it should start out flat then curve up across the graph from left to right before falling off.

much like the plot Nigel has in his post

the middle part of that (Nigel's plot) plot, is the point were you can find out how good an image you can get. i.e. the best SNR you can get. if the camera was perfect you could just expose for a very long time to get a better image however this graph can show you the limit as to how good your signal to noise can be if you don't flat field.

Paul did this for his camera and i try to follow what he had do but it's not right, however he was able to tell how long before the camera was at it best possible image SNR and it was very low ~10

Am not sure what you would get out of this appart from the gain max SNR and read noise but as a technical exercise i think it will be interesting.

ally

P.S I hope to use this and other things to completely examine a H16 CCD to find everthing their is to know about the camera including it's Quantum Efficiency to within 1% across it full range of wavelengths

[dph1nm]

To quote from Kodak on the meaning of Photon Transfer Curves:

Each point in a photon transfer curve represents a portion of a flat field image (e.g. 100 x 100 pixels) taken with a different exposure time. Generally the integration time is fixed, but the amount of time the light source is on varies for each image. This way the dark current noise is constant throughout the curve. For each 100 x 100 pixel square, the noise, i.e. the standard deviation in the pixel values, is plotted against the average signal. The plot is traditionally done on a log-log scale. The resulting curve has three sections. At the lowest signal levels, the curve is flat. This portion of the curve can be extended to the noise axis to give the camera s read noise floor. The middle portion of the curve has a slope of ½ and represents the part of the camera s dynamic range over which its operation is shot-noise limited. The last third of the curve has a slope of 1 and corresponds to the range in which the camera s operation is pattern-noise limited.

To add to this, if you plot this in DN rather than electrons, then the plot will be the same shape, but will be offset, and the y-axis value where the extrapolation of the 1/2 slope section crosses the noise axis (at log(signal)=0) gives you 1/2 log of the gain (or the inverse, I can never remember the definition of gain).

NigelM

[narrowbandpaul]

the x intercept of the slope 1/2 line is the gain...

the y interept is the read noise

the constant multiplying the unity gradient is the PRNU

what value did you get ally...

dont worry about the apparent shot noise not squaring to the signal....it will only do this when converted to electrons...the gain is S/noise^2 for all points showing shot noise limited...

SNR max= 1/PRNU

easily proved mathematically and even easier in Excel/Matlab

just plot S/N versus S...if it plateaus there is FPN domination...if it goes as S^1/2 all is well.

Might post mine later...

But that curve does not look good.

Hence my like for CCD's which for the most part behave themselves...

what gain did you get and what full well does that correspond to....in e-

cheers

[ngc2403]

please tell me where the 1/2 slope is on that graph ? it's not clear to me where it is on the graph.

however the solution to this could be to redo the images and take more using sun light and not 52Hz lighting

as for PRNU ???

i will try this all again no monday and try using log to the base 10 might be a good start

ally

[narrowbandpaul]

on log-log axes a gradient of 1/2 corresponds to a curve of general equation y= Cx^1/2...ie vary C until ypu have the best fit...then find x when y=1...ie the intercept...

same for unity slope (PRNU=C in this case) y=PRNU*x^1

same for read noise...(C=read noise) y=C*x^0

perhaps the frequency of the mains is affecting the brightness measurement...but its only the signal that counts...so an exposure starting in the minimum just produces a low signal...which is fine...its bad for photometric/linearity measurement, but not for PTC

make sure you have a stack of the shortest possible darks for bias measurement...and use base10 logs....

and use two or four images per exposure length to remove FPN, for accurate gain measurement...

if you use two images then subtract them and divide by sqrt2 because of the increase in shot noise...

if you use 4 images average two, so you have two sets of two averaged images then subtract...it negates the use of sqrt 2.

then plot....

find gain and read noise...

plot again using a single image exposure set...find PRNU

also signal vs time for linearity, and SNR vs time for laughs!

hope that made sense...

me

[ngc2403]

I know how a gradient of a half looks on a log log plot it's just that there is not best fit to the data because it is poor.

however i will try again tomorrow when am home

[narrowbandpaul]

i didnt think you were struggling for gradient 1/2...i said it anyway...

try again...remember two exposures (or 4) per exposure...

good luck, and post the results

[ngc2403]

Hello,

just though i would update the thread with a new image,

this is the lastest bash at the PTC for my 350D,

it looks better but it's the wrong shape?

(click to enlarge)

still i don't see a bit were the graph have a gradient of a half?

their is read noise then it begins to curve upwards then full well hits????

ally

[ngc2403]

i might add that this is all the results for 4 runs of the camera at iso 800.

i have not averaged or subtracted or added any factors of root 2.

[narrowbandpaul]

dont log10 the data....plot the data on log axes...

the x axis should have tik marks like 10^1, 10^2 10^3 etc...the data is plotted as it is, but with the axes on a log scale

same for the y axis...

i will post mine when i get chance....working now in a bar...4pm till 5am

[dph1nm]

still i don't see a bit were the graph have a gradient of a half?

Between about 3 and 4.5 on the Signal axis the slope is ~0.5 e.g. between 3.5 and 4.5 (a range of 1 in Signal) the Noise increases from about 1.3 to 1.8 (a range of 0.5). This, I think, is what you expect.

However, the noise still doesn't seem right. According to Clarkvision the 350D should have a gain of 1 at ISO 800 (well 1 at 1000 to be precise), so at a log signal of 4 the log noise should be 2. It should also have read noise of only ~4 electrons.

NigelM

[narrowbandpaul]

we are working with DN here, not e-....so it doesnt work like signal=noise^2...signal=K*noise^2 in DN, K is the gain

the axes should be like the ones you had in the first post...i think they are log10...since numbers are expressed as powers of 10.

plot the data from the second go, just above, on the axes on the first graph...

[ngc2403]

Update i took to fitting a line with gradient a half to the data and found it had the equation y=80*x-0.4

i think i can't remember cause i was ill when i did it. but anyway that is 10^-0.4 it i hadn't taken the log of the data. so now i have this how do i convert to electrons from DN?

Next, now that it have the shot noise area i don't see the fixed pattern noise?

the longest exposure was 0.25 seconds and i wonder what would happen if i tried to get a fainter source of light so i had longer exposure lengths?

i will also try iso100 but it takes an age to get the data.

Paul the quickest way to get the data into Matlab, is to get IP to take the images but to leave copies on the camera because it then means i have easier numbers (IMG_1754, IMG_1755, etc...) to work with in the names unlike the IP naming system!

[Sam]

Ally,

If you don't mind, would you explain how you get the data you are plotting. Without equations and in plain language would be appreciated :scratch: . I think I get the gist of what you are doing but would just be interested in how you are doing it.

cheers

Sam

[ngc2403]

Not a problem Sam, i will try my best,

I took an image at each exposure setting between the shortest and the point at which the chipp is fully saturated. It is best done without any optics taked to the camera. the idea is to illuminate the chip with white light evenly. i did this pointing the camera at a white wall with a white plastic bag over the front.

this is just like taking flats for your camera. after taking all of the images you need to take the average of a small 40 X 40 box of pixels in the image and then from the same box find the Standard Devation or the root mean square of the box.

these to things are the signal ( average ) and the noise ( Standard Devation )

Standard Devation is a measure if how mach the value in the box vary from the average.

plotting this for all the images gives this graph and other things like linearity and how the SNR varies with signal.

still not very clear i know but i don't fully understand it all myself

[Sam]

That's great Ally, Now I understand what you are trying to do! Thanks you very much for that - if this rubbish weather keeps going then I might do my 400D as well. One question though, how do you get the average and the SD from the 40 X 40 pixel square? do graphics programs do this or do you somehow import the data into a spreadsheet?

cheers

Sam

[Sam]

That's great Ally, Now I understand what you are trying to do! Thanks you very much for that - if this rubbish weather keeps going then I might do my 400D as well. One question though, how do you get the average and the SD from the 40 X 40 pixel square? do graphics programs do this or do you somehow import the data into a spreadsheet?

cheers

Sam

Ok worked it out now - it's on the histogram window :oops:

[ngc2403]

i have been using some Special maths software to do it and i was not aware of any graphic software that did it?

i am interested to know what software you use that has the SD in the histogram?

it is a slow process mind you took me a few hours to set it up and do it.

ally

[Sam]

I used PixeInsight then crop the image down to 40 X 40 in the middle then select tools/statistics and it gives a table of mean. median, standard deviation, average deviation, variance, minimum and maximum for each of red, green and blue channels. It also gives the option of what data range to use, 12 bit, 16 bit, 32 bit etc.

Which colour channel do you use? and do you use 12bit? or do you have to convert it to something else?

Sam