How much noise is noisy?

[Camelopardalis]

Folks,

I've been thinking about noise, how much and how little we want in our images.

I figured a good place to start is a low noise camera. Then we want to stack a collection to reduce the noise further. Personally, I don't like noise reduction techniques as featured in popular image processing software as I feel there's always a trade-off with detail.

So, how much noise do you get from your subs, and how much do you accept in your pre-processed stack?

Just curious...I'm interested in "benchmarking" my process so I know where I should stop :lol:

All thoughts welcomed...

Cheers,
Dunk

[Merlin66]

Dunk,
In spectroscopy we have basic limits....
Less than 50 SNR is very low and only usable if there's nothing else to use.
Above 100 SNR is acceptable and >200 SNR is really required to submit data to ProAm campaigns.

[RickS]

1. You can overcome read noise with subs that are long enough regardless of camera read noise
2. Shot noise is unavoidable and is sqrt(n) where n is the number of photons you detected (per pixel)

That's really all you need to know, Dunk ;) You use sky limited subs and you stop when the dimmest parts of the image that you are interested in look acceptably clean. Noise reduction is acceptable, IMO, when selectively applied so that data that already has acceptable SNR is not adversely affected.

The only time you get SNR that is consistent over a whole image is when you have a very bright source, like when you take flats. With the objects we normally image there will be areas of good SNR and areas of poor SNR. If you want to see the dimmest areas looking clean you'll need to get a very large amount of integration time under dark skies. A short amount of integration time under very bright skies will also get you high SNR, but this is only useful if you like imaging light pollution gradients :P

[cometcatcher]

I think the answer is subjective. Everyone has different "acceptable parameters" and in time even they change.

[Camelopardalis]

Thanks for your thoughts, chaps... I'm looking through some of my AstroFest data and trying to determine if I need more subs :eyepop:

One of my stacks is claiming noise of 5e-05 in PI. I've never really paid that much attention to the numbers before, more just looked at the image and decided. To be fair, I've usually been time-limited before, so I've probably been of the mentality that more would always be better :shrug:

[sil]

I agree here, i've seen many ways to measure it but they all have drawbacks it seems with consistency and accuracy. you can measure the SNR in one program twice and get two different values.

Chasing a perfect SNR or noise-free image is a nitpicking exercise. There are ways and means to reduce capturing noise in the first place (eg understand the different types of noise and their sources). By stacking you are NOT removing noise (not everyone understands this), just pushing it fainter from the Signal you actually want so its perceptually less noticable. If you set clipping values in your software you can more aggressively remove noise along with some signal too. Noise reduction filters will effect signal too in different ways depending on the method used. Keep in mind your end goal for your image when processing and how noise reduction will impact it. Also many processes actually add noise (unwanted Signal ) to your image too. Personally I capture very noisy subs and my final images are fairly noisy too, I try to compensate for being unable to capture the way I want or should so I roll with the best I can do with my limitations. I keep looking at different ways though to reduce the obvious noise for me AP is a constant cycle of learning and experimenting, so how much noise is too noisy? for me when you see NO signal at all, otherwise ever shot is still in draft for me, i can still work on ways to improve it. I enjoy the process.

[Camelopardalis]

Thanks for your input sil, I think I've been tying thoughts up in knots a little worrying about noise, but I guess it's all a matter of perspective.

Part of the problem is having a 16MP image and feeling compelled that the appearance of noise is minimised even at 1:1...in reality, I don't have a 16MP screen :confused2:

[Atmos]

It all comes down to what you find acceptable, what you can live with and how much noise reduction you're willing to do :P

[Camelopardalis]

See above - not a big fan of noise reduction, as the trade-off bothers me...maybe I'm just using it all wrong :shrug: :lol:

The 1600 is a pretty low noise photon collector, but see above...I think I've got into the bad habit of peeping... :eyepop:

[sil]

Doesn't work for everyone but my approach to most things is examine minutely and ask myself "what's wrong/bad here?" and from there I work on why and how to fix/eliminate it. Work on the broken bits and leave the good alone, its fine already.

Bit like the old adage of a mason and a sculptor making a statue of an elephant. The sculptor will keep adding clay to the pile until its the shape of an elephant, while the mason starts with a block and remove everything that doesn't look like an elephant. different points of view, same end result.

[Camelopardalis]

Also I think one of the traps that easy to fall into is to do a quick stack and apply a STF in PI...always looks bad :sadeyes:

[RickS]

It does require a delicate touch, Dunk ;)



Even if it was zero noise, perfect QE photon hoover you'd still have shot noise to contend with.

Cheers,
Rick.

[Camelopardalis]

Subtlety is a strength, huh Rick :D

[RickS]

Not claiming it is my strength, of course :lol:

[Camelopardalis]

:d

[Astrofriend]

Hi,
I did an Excel sheet about Signal / Noise relation to do some test. I have also implemented how optic vignetting lower the S/N in the corners.

It's very simplyfied but still it give you a good idea how readout, sky background, object strenght, calibration nosie works together.

You can download the Excel sheet on the page:
http://astrofriend.eu/astronomy/tutorials/tutorial-dithering/tutorial-dithering.html#part06

/Lars

[rcheshire]

...as much as you didn't manage to eliminate through acquisition and can't be removed without smudging the image. Unless you like the watercolour / impressionist look.

In an emergency - ImageMagick...

convert img.in -despeckle img.out

...does wonders...

[Merlin66]

For AP images can we call it "cosmetic noise" rather than refer to the scientific SNR - which is not really used......

[Camelopardalis]

Thanks all :thumbsup:

Maybe we should call it "film grain" then Ken? ;)

[Camelopardalis]

Which poses an interesting question/curiosity from me...

How do you calculate the SNR of an image? PI gives an indicator of noise but not signal (unless I'm missing something obvious).

[RickS]

Some of us actually do the calculations to optimise data capture, so perhaps we can apply for a license to use the term SNR? Purely for scientific purposes, of course :)



Signal is what's in the pixels, Dunk ;) Take a look at the SubframeSelector script. It calculates a SNRWeight which gives you an idea of the quality of each sub wrt SNR. PI can only estimate noise, not measure it, so it's not perfectly accurate. High SNRWeight isn't always a good thing either. Light pollution reflected off a dirty big cloud will produce a high SNRWeight, and that's why we blink subs as well as measure them.

Don't forget that there's no such thing as a single SNR number for a sub. The SNR will be higher in bright areas and lower in dim areas due to the impact of shot noise.

Cheers,
Rick.

[Camelopardalis]

Cheers Rick, yeah I get that the signal is the precious pixel values :P

I'm assuming the noise estimation is based on the whole sub, and I roll with the premise that it's going to be more or less (within a margin of error) equally inaccurate for all the subs I collect.

Will take a look at SNRweight, I already use Subframe selector, for better or worse...

[Merlin66]

ImageJ will give you an SNR figure for the selected area of your image...
AA6 gives similar - average and standard deviation...I would think packages like Maxim would do likewise??

I'm not convinced that any/ many AP imagers actually finalise their image based on a rigorous SNR figure.....

[Camelopardalis]

Thanks Ken.

No rigor intended, more curiosity about when the noise becomes largely unobjectionable. Of course, that's subjective...

I have a few targets in the hit list that are fainter than the usual suspects I've been going for, so I've been looking for a yard stick really. The fainter targets will need more total exposure time to have acceptable noise levels when stretched, so it'd be useful to get a feel for what SNR/noise estimate I'm after.

[Astrofriend]

My order to use the tools are like this:

AstroImageJ (demosaic, calibrate, align, stack, my version only Canon EOS!):
http://astrofriend.eu/astronomy/tutorials/tutorial-astroimagej/tutorial-aij-01-introduction.html

Fitswork (rgb combining, adjust levels, background elimination):
http://astrofriend.eu/astronomy/tutorials/tutorial-fitswork/tutorial-fitswork.html

After that have done the above, a final touch with some photoprogram.


AIJ (AstroImageJ):
If you have an Canon DSLR you don't need the deBayering process because of the demosaicing in my macro for AIJ.

With that process in AIJ you lower the random noise and take away or reduce the static pattern without need of dark and bias frames, works very good.

Noise:
Where to measure the S/N in the image depends of what's important, the strong stars has very high signal and then high S/N. But if you measure the S/N in a nebula or the weak part of a galaxy, then the S/N is interesting to know. Most of the noise comes from background if not narrow band images.

/Lars

[RickS]

Dunk,

Why not make some measurements on a small area of an image you already have that's just around your level of minimum acceptability for noise. If you can figure out how many photons you need to capture in each pixel to get to that point then you can make inferences about future targets on the basis of a few sample subs.

Hint: previews, the Statistics process and PixelMath will do everything you need along with the understanding that shot noise is n^0.5 if you capture n photons (in a single sub or a whole stack) and SNR is n/n^0.5 = n^0.5. You can ignore other sources of noise so long as you're doing subs that are sky limited (and if you're not you should be!)

Ken might not believe me but I do take a fairly rigorous approach to SNR when I'm chasing very faint features. The jets of NGC 1097 would be a good example.

Cheers,
Rick.

[Merlin66]

Rick,
I believe you.....
So,what's the SNR in the faint stuff you see as acceptable....SNR>20?

[RickS]

Sorry for poking you, Ken, partly tongue in cheek :)

The incredibly dim R4 jet in NGC 1097 I got to a SNR of just over 9 and convinced myself and a few others that it was visible.

I have a half completed image of IC5148 that I really should finish up. It has a very faint Oiii halo that I have at a SNR of around 14 after 46 hours of 1800s subs. It's quite visible but not pretty.

SNR>20 sounds quite luxurious :lol:

[Camelopardalis]

Will give it a try, thanks Rick :thumbsup:

[RickS]

And report back, pls Dunk :)

[Slawomir]

Very interesting discussion.

I tried to follow Rick’s instructions to measure SNR in a 2.5 hour test image of NGC 6744 taken from Paddington.
I measured mean background signal in a few featureless areas in the image and subtracted it from mean signal in a bright, medium and dim areas of the galaxy. Then I multiplied the resulting ADU values by gain so that in theory (I hope) should give me the number of photons in a 2.5 hour stack. Signal is still quite dim in 2.5hr exposure; it was only 27 photons in the bright area, 16 photons in medium-bright and only 1.4 photons in the dimmest parts that I could recognise, giving me SNR of 5, 4 and 1.2 respectively. I need more exposure!
Is the above workflow correct?
Suavi

A link to the image in question (although this stack had a touch of noise reduction applied): http://www.astrobin.com/full/305157/C/?nc=user (http://www.astrobin.com/full/305157/C/?nc=user)

It clearly looks very noisy…

[Merlin66]

Interesting methodology....
In spectroscopy all we do is to remove the background from the image and then measure the SNR in the result.
I have to say at a SNR <10 then the signal is very very low and difficult to differentiate from the actual data signal.

[Shiraz]

FWIW, it looks like many of us have a similar view of what constitutes an acceptable amount of noise in a finished "pretty pic" type astro image. I just did a quick check on 5 good quality images from IIS and a couple of my own better ones - in all of them, the noise level in the darker regions was consistently around 300-500DN RMS in a 16 bit representation. Also interestingly, the same sort of noise level seems to be acceptable in brighter regions - the extra signal gives better SNR, but the acceptable noise still seems to around 400DN(16bit). Would be nice if that could be checked by others on their true 16 bit data, rather than me using JPEGs off the web.

Of course, the signal level in any part of an image can be set at any desired value for aesthetic purposes, so measuring SNR is only of value for comparative testing or for validating theoretical approaches to system design - however, since there is some evidence that "good quality" images all have similar noise levels, measuring just the final noise may well be a useful thing to do as one part of "pretty pic" image quality assessment. Good thinking Dunk :thumbsup:. cheers Ray

[Camelopardalis]

Will do, as soon as I stop being a beach bum and knuckle down at the keyboard :D

[Camelopardalis]

Ray, yeah the end-result noise is what we perceive in the end, but obviously doesn't factor in the noise introduced and/or reduced in processing which can't readily be reversed to reveal the SNR of the original linear, single sub or stack.

I figure there's going to need to be a baseline for a pre-processed image to ensure decent results. If, say, I stack only 4 subs, it's only good for a sneak peek and not worth processing (no doubt obvious!), but for a given target I'm going to want to figure out a rough number of subs to aim for.

And this is without even considering proportions of L to RGB :eyepop:

[Slawomir]

Hi all,

While investigating plausibility of galaxy imaging from our heavily light polluted site I looked more closely at acceptable level of noise in featureless background in a master luminance.

It turns out that for sky-limited subs, measured level of noise follows very closely theoretical predictions, in spite of significant variations in sky glare from night to night in my location. The number of subs can be easily estimated with the following simple formula:

Number of subs = [ St dev (background) / St dev (desirable level of noise) ] ^ 2

I propose that an acceptable level of background noise in a master luminance is at or below 10 ADUs for galaxy imaging – or any imaging when we want to pull out faint bits that are just above the background noise.
In my location and with my 4” f/6 refractor at 1.18” pp, during the darkest nights I get St Dev for background signal of about 220 ADUs in a 5-minute Luminance sub. Therefore, to bring background noise down to about 10 ADUs, I need…

Number of subs = [ 220 / 10 ] ^ 2 = 500 5-minute Lum subs, which is about 40 hours of integration. To arrive at this result more quickly, I either need a faster setup or move to a darker location (or both).

[lazjen]

And if you had a 5 scope Dragonfly setup, all Lum filters, you might be able to squeeze that in one night. ;) Probably be better to go for a 10 scope Dragonfly instead to make it a more acceptable 4 hours, right? :)

[troypiggo]

Must say I'm loving this discussion. Excellent work Suavi. Might have to come back and ask some questions about how to actually measure all this stuff to apply it to my own gear.

[Slawomir]

I also have a formula for such multi-scope set-ups Chris ;)

actual imaging time = ( amount of money spent on the gear * time spent tweaking ) / (number of telescopes^2 * real amount of money needed to be invested to make it work)




Glad you find it useful Troy :thumbsup:

[Camelopardalis]

You're way ahead of me Suavi :thumbsup: I've got a couple of subs open in PI for analysis...

[Atmos]

Those 40 hours in your light polluted back yard would likely be swamped by 4 hours of great dark sky time :P

[Slawomir]

Hi all,

Encouraged by my initial tests, I decided to decrease the length of Lum subs from 5 minutes to 3 minutes only. On average, Lum subs are sky-limited in about 2 minutes in my location with my rig, but when it occasionally gets darker longer subs might be beneficial, so I decided to stick with 3 minutes.

Last night I was able to collect nearly 3 hours of 3-minute Lum subs and the data seem to be better than with 5-minute subs.

I selected the best (darkest) sub and used it as a reference for making stacks with various number of subs. Removing 10 brightest subs did not make things better and it looks like unless a sub is really bad (read: with bright background), it is better to include all in a stack with significant number of subs.

Perhaps bringing background noise down to st.dev. = 10 ADUs can be done in just 10 hours with 3-minute sky-limited subs, as opposed to 40 hours with 5-minute subs with my rig in my location.

A link to a stack of 55 3-minute Lum subs: http://cdn.astrobin.com/images/thumbs/1d2c69646833e92c36fa13d6110ff45b.16 536x16536_q100_watermark.jpg


Suavi

[Astrofriend]

Hi Suavi,
I did this simple S/N calculator, it's far from perfect but still get ideas how exposure length, readoutnoise, background light and vignetting sum up in the S/N relation at the end.

http://astrofriend.eu/astronomy/tutorials/tutorial-dithering/tutorial-dithering.html#part06

With a very low noise camera it doesn't influence so much on the noise if divided on lot of short sub images, but you get better dynamics. But heavy to process all those sub images.

I try to correct it when I found something wrong in the calulations.

At my place my exposure length limit are from 30 to 60 seconds. F/5.3, ISO800, Canond 6D. When our summer house observatory is ready for use I hope I can extend it to 120 to 300 seconds.

/Lars

[Slawomir]

Very thorough investigation Lars :thumbsup:

I haven't thought about not using bias but it makes sense what you wrote - since I dither my subs I will see if not using bias for calibration makes things a bit better. I'm sky limited also in about 60s on average nights, and 120s after 11pm on a dark night and when Suncorp switches lights off, and so does a large portion of fellow Brisbane dwellers :lol:

[Astrofriend]

Hi Suavi,
Earlier I used Fiswork a lot: http://www.fitswork.de/software/softw_en.php , that software subtract a constant that compare to the dslr camera's bias by automatic. To my Canon DSLR it had been working perfect. You don't have a DSLR so I don't know how well it will perform about bias compensation.

If you find it interesting you can take a look how I do the calibration today:

http://www.astrofriend.eu/astronomy/tutorials/tutorial-astroimagej/tutorial-aij-01-introduction.html

I took a look at your photos at Astrobin, very great looking photos. Your telescope looks to have the same design as my TS130 which is bought from Germany. TS are very common in Scandinavia.

I have an older version of this:
http://www.teleskop-express.de/shop/product_info.php/language/en/info/p3374_TS-Photoline-130mm-f-6-62-Triplet-FPL53-Super-Apo---3--Rack-and-Pinion-Focuser.html

Very satisfaid with it.

/Lars

[Slawomir]

Thank you Lars for the links and for your very kind feedback - I am glad you like my attempts at imaging DSOs.

Most of the images on Astrobin were taken with an f/7 102mm TS doublet, but the most recent few were taken with CFF's splendid 105mm at f/6 and one at f/4.5.

TS makes very good telescopes for sure, and the only reason I upgraded mine was that it was a doublet and I wanted a small quality triplet.

These days I exclusively use PixInsight for all data processing, perhaps occasionally I may experiment with FitsLiberator though.