Camera Labs

Thought I make some random noises here while it's still fresh in my mind...

If you think comparing one DSLR against another is complicated, try figuring out astro imaging CCDs too.

I thought as a first step, I'll look at two sensors, the Atik 320E and 383L Plus compared to a Canon 18MP crop sensor body. The two sensors were picked purely as they happened to be two I randomly came across as representative of more affordable models. Affordable is used here as a relative term! The Canon 18MP crop sensor body could be the new 60Da, or one of the 550D, 600D, 60D, 7D preferably with filter modification for astro use.

The 320E is small, around 5x crop factor and weighs in at 2MP. The 383L+ is about microFourThids size at a tad over 8MP. My gut feel was that something like the 320E was just too small to be worth bothering with as an upgrade to replace a DSLR. The 383L+ is starting to get interesting, apart from the price which is above a new 5D mk2.

As the sensors are different sized, at different pixel densities, how do I compare them? I made an assumption of a 100mm objective. Further assume the focal length could be freely adjusted such that it gave the same pixel field of view. From there you can calculate the focal ratio and differences between them in exposure time. As it turned out, this wasn't too significant since their pixel sizes were in a similar ball park. The Canon and 320E were near enough the same, with the 383L+ pixels around 20% bigger per side.

Making the very big assumption that all three have the same colour filter for now, quite simply bigger area is better. Factoring in the speed differences for a normalised 1 arcsecond per pixel, the 383L+ is about 80% of the output of the Canon, and the 320E is a pitiful 12%.

But that isn't the whole story and assumes the sensors are all equally sensitive. But are they? I'll have to drop the 320E from comparison here as I can't find data on it. The 383L+ is based on a Kodak KAF-8300 sensor which has a handy datasheet. Canon don't publish their data, but you can find measurements for various models here. Looking at the charts for both they are pretty similar. The KAF-8300 has roughly the same blue response, with slightly higher peak green and red response. Overall the colour filter characteristics aren't that different, not at least a stop different to be significant.

My verdict: for one shot colour filter imaging, a DSLR is ball park competitive with similar size CCD imaging devices, with the DSLR being at a significantly lower cost.

What if you don't want to do one shot colour imaging? If you want to do narrowband or multiple pass colour filter imaging, then the advantage swings rapidly in the direction of mono CCDs. Take what is probably the worst case scenario for a DSLR: H-alpha lines at around 650-something nm. Assuming the DSLR has been modified, we get a QE around 25%, but due to the colour filter, only 1 in 4 detectors are effective. The KAF-8300 has a mono sensitivity around 650nm of about 45% depending on which variant of the sensor is used. It's advantage is it isn't hobbled by the presence of a colour filter at this point, so it gets all the benefit. The smaller sensor of the 320E only gives a relative mono response chart, where 650nm is about 0.55x the peak response in green. In the absence of this info, I'll assume it is comparable to the KAF-8300.

So factoring in the H-alpha sensitivity and the colour filter loss of the DSLR, how does that shake things up? The 383L+ is now a whopping 6x more effective than the DSLR, and the 320E is an estimated 85% of the DSLR.

Verdict: for narrowband, a mono CCD will give a significant advantage providing it is competitive in sensor size. A small mono sensor will negate that potential advantage.

What does this mean to me? I'll stick with DSLR imaging for now, since a small CCD does not offer significant value over a DSLR for their cost. A big CCD does, but I'll be paying for it!

Big warning: I didn't take into consideration noise effects here. The CCDs are cooled so reduce thermal noise effects, so they may provide further benefit over a DSLR. However that is far too complicated for me to work out.

Bonus points if anyone actually bothers to read all the above.

_________________
Canon DSLRs: 7D, 5D2, 1D, 600D, 450D full spectrum, 300D IR mod
Lenses: EF 35/2, 50/1.8, 85/1.8, 135/2+SF, 28-80 V, 70-300L, 100-400L, TS-E 24/3.5L, MP-E 65, EF-S 15-85 IS
3rd party: Zeiss 2/50 makro, Samyang 8mm fisheye, Sigma 150 macro, 120-300 f/2.8 OS, Celestron 1325/13
Tinies: Sony HX9V.

I find your reasoning rather soothing, and I agree with your conclusion.

As far as cooling goes though, you can if you don't already, remove the battery and run it directly from a power source. I see some do that to neglect heat during long time exposures and of course "have to do it" in places where it is very cold, where batteries don't last.


Some thoughts on CMOS noise vs temperature:

In some of the google hangouts I have used my 7D in live view mode with a screen capture program instead of a webcam. (Damn logitech drivers...! :d)
And the camera does get VERY hot after around 30 min, you can properly feel it when touching the camera, the noise performance during these "heat-strokes" are also very apparent. However it very quickly disappears when the heat load is removed, ie turning live view off. So I would think that the temperature vs noise is best shown in a log curve?

Some kind of simple heat spreader that is mounted on the camera screw on the bottom should suffice for outside use. You could take two old (Square type) Intel or AMD stock heat spreaders, get someone who can TIG weld, and then drill a hole in the middle and insert a screw. The messy part would be connecting the heat spreader to the camera, you need something in between to get proper contact and "goop/thermal-compound" would go all over the place, so a soft thermal film of some kind would be easier.

_________________
Canon 5D MkII -|- Sigma 24mm f/1.8 | Carl Zeiss 35mm f/2.0 | Canon 50mm f/1.4 | Canon 100mm f/2.8 Macro | Lensbaby Composer
Canon A-1 -|- Canon 28mm f/2 | Canon 35mm f/2.8 | Canon 50mm f/3.5 Macro

Hi popo,

I've got no way to quantitatively justify what I'm about to say (maybe your research will be more productive) but my gut feeling is that repeatability is key and this is where DSLRs fare badly. There's really no way to ensure that the temperature of a DSLR sensor is constant and that seriously affects one's ability to use Dark calibration frames. I guess one could interleave Lights and Darks during imaging but that cuts down camera time by a factor of two and even then doesn't really work well as statistically you need several Darks for each temperature of the sensor to stop the Darks themselves introducing a significant noise signal. Christian Buil refers to thermal noise on the page you linked in Figure 1 but has no way to measure the temperature dependence. However from the KAF-8300 datasheet we can see that dark current for that sensor doubles for every 5.8°C increase in temperature so it is very easy to anticipate dark current, and by implication noise, varying by a factor of 8 times for a typical range of DSLR operating temperatures. Even over a single session variation can be more than noticeable as I showed many moons ago in this post using a 40D.

The Atik 383L+ has a Peltier effect cooler so not only does one operate the sensor at a fixed temperature, making production and use of a standard set of accurate calibration frames a snap (if you'll excuse the pun), but that fixed temperature can be set at a level which can reduce the thermal current by a similar factor (or more) offering maybe a factor of 8 x 8 = 64 times reduction over that of a DSLR. The numbers I've used are "back of a fag packet" in character and I'm sure you could improve on them but I think they convey a very good sense of why an actively cooled and temperature stabilised astronomical CCD camera is worth the outlay for all but maybe the shorter exposure times.

Bob.

_________________
OM-D E-M5 + ED 12-50mm, H-PS14042E, H-F007014E, L-RS014150E, M.ZUIKO DIGITAL 45mm 1:1.8.
Gitzo GT1541T, Arca-Swiss Z1 DP ball-head.
Astrophotography: TEC 140 'scope, FLI ML16803 camera, ASA DDM60 Pro mount.

Hi popo,

I've been scratching my head trying to figure out how to show how well my MK16803 camera's "Lights" clean up after calibration with the idea of backing up my assertion that a temperature stabilised cooled CCD is the way to go for long exposure work. What I've done is compare an 8 second exposure with a 1500 second exposure after both have been calibrated. Not having a pair of "lights" shot in quick succession of the same subject I've chosen to use a couple of darks. I chose an 8 second Dark at random (number 10 as it happened) from a sequence of 25 I took to produce a master Dark. Similarly with a 1500 second Dark. I calibrated each Dark using a master Bias frame and the master Dark made from the full set of 25. That sounds a bit incestuous as the master Dark is being used to calibrate a Dark which was one of the 25 used to produce said master but I'm working on the assumption that the random thermal noise (what we are interested in here) of the other 24 will mean that the procedure is valid provided one isn't looking for accuracy to the last 10%. Here are the statistics from PixInsight for a random block within each image:

The 8 second exposure, after calibration, is showing a mean of about 2.70 compared with 3.16 for the 1500 second exposures with a 16 bit (65535) range. In my book that's impressive and might have been even more so had I figured a way to sample from exactly the same block in both images. To show visually what the noise looks like I produced two 100% crops of exactly the same region, the 8 second on the left and the 1500 second on the right:

I produced it by applying an extreme histogram stretch to the 8 second exposure to show some noise (bear in mind the mean was only 2.7 so the stretch really was extreme) and then applied exactly the same stretch to the 1500 second exposure. Exported and cropped and composited in PhotoShop with no further processing. You can see the extra noise in the 1500 second exposure if you look carefully but bear mind that if these exposures were actually "Lights" then one would typically take several (5 or more hopefully) in order to reduce the background noise still further (Winsorised Sigma Clipping in PixInsight ).

So what's the point of all this? Hopefully I've shown some of what I see when processing my images from the ML16803 camera, namely that they clean up incredibly well after a really good set of calibration frames is applied. Because I always shoot at -25°C I can use calibration frames made months earlier and, of course, those calibration frames can be made regardless of weather which means making a stack of 25 x 1500 second Darks is easy peasy as MaximDL can run the exposures unattended overnight. My assertion is that there is no way one could obtain a comparable clean up from a DSLR if for no other reason than one can't use an exactly matching set of calibration frames because the sensor temperature can vary so much. Only an assertion I'm afraid and I don't have a DSLR to play with now but if you want to try and replicate the methodology using your own DSLR I'd be interested to see if you could get anywhere close to the minimal noise I see in a 1500 second exposure after calibration!

Hope this helps your decision even if it doesn't exactly help the bank balance.

Bob.

_________________
OM-D E-M5 + ED 12-50mm, H-PS14042E, H-F007014E, L-RS014150E, M.ZUIKO DIGITAL 45mm 1:1.8.
Gitzo GT1541T, Arca-Swiss Z1 DP ball-head.
Astrophotography: TEC 140 'scope, FLI ML16803 camera, ASA DDM60 Pro mount.

Hi Bob, that's an interesting way to show an example of the noise, although I'm finding it really hard to put that into context. I think I'm going to have to have a go doing the same thing and see what I get...

_________________
Canon DSLRs: 7D, 5D2, 1D, 600D, 450D full spectrum, 300D IR mod
Lenses: EF 35/2, 50/1.8, 85/1.8, 135/2+SF, 28-80 V, 70-300L, 100-400L, TS-E 24/3.5L, MP-E 65, EF-S 15-85 IS
3rd party: Zeiss 2/50 makro, Samyang 8mm fisheye, Sigma 150 macro, 120-300 f/2.8 OS, Celestron 1325/13
Tinies: Sony HX9V.

Left: RAW image without bias/dark correction.
Right: Same image with correction.

Here we go, and this is also the first time I ever tried using bias and darks in PixInsight. ISO800 30 seconds on a 450D. 10 frames of each taken over a year ago, used to correct one of the darks. I wont bother with the crop since all you see when you stretch the display enough is an abstract art work, and it is the scale that is of importance.

I am wondering how to interpret the results though. I'm not sure the Mean is so meaningful, except if you need maximal detail in the dynamic range available. Given I'm heavily light pollution limited that isn't the worst for me. I think the standard deviation is more meaningful as to the level of visible noise in the image. Not sure what AvgDev means in this context though, since the value seems to be the same as Mean. Regardless, my values are quite a bit higher. Maybe it would help if I had also taken 25 of each too, which I think is 1.6x better than 10? Although I may have more proportionate benefit in self referencing here.

_________________
Canon DSLRs: 7D, 5D2, 1D, 600D, 450D full spectrum, 300D IR mod
Lenses: EF 35/2, 50/1.8, 85/1.8, 135/2+SF, 28-80 V, 70-300L, 100-400L, TS-E 24/3.5L, MP-E 65, EF-S 15-85 IS
3rd party: Zeiss 2/50 makro, Samyang 8mm fisheye, Sigma 150 macro, 120-300 f/2.8 OS, Celestron 1325/13
Tinies: Sony HX9V.

Hi popo,

Looks like you are getting value from those calibration frames but the big question is how much? I didn't include the uncalibrated statistics for the 8 and 1500 second exposures in my original post because my purpose was to show how little thermal noise was being added by the very long exposure time to the final result. However it might make interpretation of your own result clearer if I do include them so here's the full set of four, uncalibrated in the top row and calibrated, as seen before, in the bottom row; 8 seconds first and 1500 seconds second. You can see a very significant bias signal in the uncalibrated images which makes comparing the Mean signal between calibrated and uncalibrated images pretty meaningless, if you'll excuse yet another bad pun!

A spot of Googling elicited that "image noise definition ean be axioniatically shown to be proportional to the standard deviation of an image of a homogeneous region" (reference). That fits quite well with the above statistics with the 8 second exposure already exhibiting low noise before calibration. It also illustrates quite powerfully how calibration has improved the 1500 second image with the StdDev dropping from about 121 to 25. Just numbers, of course, and as I remarked before it makes a lot of sense to combine several subs to further reduce the effects of thermal noise. Comparing the StdDev figures for the calibrated images moving from 8 to 1500 seconds increases the StdDev from about 5 to 25. Frankly I don't know whether perceived noise is proportional to the StdDev or to some power of StdDev so I'll show the histogram stretches again. For folk who didn't read the previous post the 8 second exposure on the left was given a really extreme stretch (much more than would be applied during regular processing) just so that the noise pattern can be seen. Exactly the same stretch that was applied to the 8 second exposure was applied to the 1500 second exposure on the right so the amount of noise in the two images can be visually compared.

Looking carefully at those 100% crops I suspect that perceived noise really is directly proportional to the StdDev. If so an increase in noise (thermal plus read) by a factor of 5 for an exposure change of a factor of nearly 200 is a good result. Maybe a mathematician can come to the rescue on this one as my graduate maths studies are a long way behind me and statistics was always a dark art so far as I was concerned.

I'm hesitant to try and interpret your own numbers as I'm already straying well above my pay grade in trying to interpret my own but I do note that not only are your StdDev figures higher in each channel but the improvement after calibration is proportionally much less though still worthwhile I think.

Bob.

_________________
OM-D E-M5 + ED 12-50mm, H-PS14042E, H-F007014E, L-RS014150E, M.ZUIKO DIGITAL 45mm 1:1.8.
Gitzo GT1541T, Arca-Swiss Z1 DP ball-head.
Astrophotography: TEC 140 'scope, FLI ML16803 camera, ASA DDM60 Pro mount.

I think I made one minor mistake in my previous results. The bias/darks apply to the frame before debayer, so the stats for each colour channel have not been re-interpolated to fill in the gaps. I'm not familiar with the debayer function in PixInsight and it's rather late for me to RTFM, so in a blind play I picked something at random and the result from that gave more significant improvement. I wont post it yet as I'll figure out how to make sure I do it right before I do that.

While I'm passing, here's a paper on the 5D2's performance, with the claim it is in many respects better than dedicated astro CCDs apart from the operating temperature hindrance.

_________________
Canon DSLRs: 7D, 5D2, 1D, 600D, 450D full spectrum, 300D IR mod
Lenses: EF 35/2, 50/1.8, 85/1.8, 135/2+SF, 28-80 V, 70-300L, 100-400L, TS-E 24/3.5L, MP-E 65, EF-S 15-85 IS
3rd party: Zeiss 2/50 makro, Samyang 8mm fisheye, Sigma 150 macro, 120-300 f/2.8 OS, Celestron 1325/13
Tinies: Sony HX9V.

Hi popo,

That was interesting, especially about the read out noise vs. ISO and the dark noise.

In some ways it would be surprising if a sensor such as that in the 5D2 wasn't near the top of the game given the R&D spend that Canon can bring to bear. I guess the difficulty in calibrating a DSLR could be overcome with work. The article refers to a similar pattern I saw with the 40D so step one would be to take several shots to let the sensor get up to a constant temperature relative to ambient. Then, provided one had calibration frames for, say, every 5°C change in ambient temperature one could just select the appropriate set during development provided the ambient temperature remained reasonably constant for each set of lights.

Not for me, I'm afraid. I've become averse to Bayer arrays and their associated AA filters for astronomical work. Just as well given my current investment!

Bob.

_________________
OM-D E-M5 + ED 12-50mm, H-PS14042E, H-F007014E, L-RS014150E, M.ZUIKO DIGITAL 45mm 1:1.8.
Gitzo GT1541T, Arca-Swiss Z1 DP ball-head.
Astrophotography: TEC 140 'scope, FLI ML16803 camera, ASA DDM60 Pro mount.

If I won the lottery, a fat CCD would be one of the first items on my shopping list. But given the models that start to look interesting often have a 4 figure price tag, and the first figure isn't often a "1", it's not likely to be on my shopping list any time soon for something I'd use relatively little.

Actually I've been finding it quite easy NOT to use any darks/bias correction so far. But how much noise is it buying me? Putting all that aside, my biggest gain will be in improving the mount so we'll see where to go from there.

_________________
Canon DSLRs: 7D, 5D2, 1D, 600D, 450D full spectrum, 300D IR mod
Lenses: EF 35/2, 50/1.8, 85/1.8, 135/2+SF, 28-80 V, 70-300L, 100-400L, TS-E 24/3.5L, MP-E 65, EF-S 15-85 IS
3rd party: Zeiss 2/50 makro, Samyang 8mm fisheye, Sigma 150 macro, 120-300 f/2.8 OS, Celestron 1325/13
Tinies: Sony HX9V.

A very interesting read, which is bringing back all sorts of memories when I researched into this.

I struggled for ages to reconcile the lack of MP in dedicated CCD cameras compared to DSLRs and eventually released that its more about the pixel size, density and QE of the sensor as well as the FOV generated.

Despite that I have still settled on the Kodak KAF8300 chip albeit in the form of the QHY9, as I can get that, a filter wheel and filters for the price of the Atik version.

I think I have decided that I want mono as I want to be able to do narrowband and take advantage of the clear nights that coincide with the fill moon (which seems to happen quite a lot!!)

I think repeatability is key and for OSC DSLRs are probably the best thing to start with in terms of cost.

Right, Im off for a lie down in a darkened room....

_________________
Lee Diggle
My Astronomy Blog | My Photography Gallery

.
Sweet deal, Lee. Mono camera plus filter wheel for only about £400 more than a 60Da at today's prices.

Bob.

_________________
OM-D E-M5 + ED 12-50mm, H-PS14042E, H-F007014E, L-RS014150E, M.ZUIKO DIGITAL 45mm 1:1.8.
Gitzo GT1541T, Arca-Swiss Z1 DP ball-head.
Astrophotography: TEC 140 'scope, FLI ML16803 camera, ASA DDM60 Pro mount.

How could I miss the QHY9 especially as I've settled on the QHY5 as my guide cam! The pricing is certainly more attractive... ARGH!

_________________
Canon DSLRs: 7D, 5D2, 1D, 600D, 450D full spectrum, 300D IR mod
Lenses: EF 35/2, 50/1.8, 85/1.8, 135/2+SF, 28-80 V, 70-300L, 100-400L, TS-E 24/3.5L, MP-E 65, EF-S 15-85 IS
3rd party: Zeiss 2/50 makro, Samyang 8mm fisheye, Sigma 150 macro, 120-300 f/2.8 OS, Celestron 1325/13
Tinies: Sony HX9V.

It's certainly a contender, if you can get past all the issues QHY have had in the past. I know a couple of fellow astrophotographers who have the QHY9 and it certainly is a cracking deal MA are offering.

I just need to bite the bullet and order it, the only thing stopping me at the moment is the light nights and knowing it won't get much use until autumn......

_________________
Lee Diggle
My Astronomy Blog | My Photography Gallery

This complicates matters since I had given up on worthwhile affordable CCDs and dropped 2k on a lens! If I got the CCD I ideally need to start looking at scopes again I think, unless anyone knows of an of the shelf Canon EF lens to CCD adapter?

_________________
Canon DSLRs: 7D, 5D2, 1D, 600D, 450D full spectrum, 300D IR mod
Lenses: EF 35/2, 50/1.8, 85/1.8, 135/2+SF, 28-80 V, 70-300L, 100-400L, TS-E 24/3.5L, MP-E 65, EF-S 15-85 IS
3rd party: Zeiss 2/50 makro, Samyang 8mm fisheye, Sigma 150 macro, 120-300 f/2.8 OS, Celestron 1325/13
Tinies: Sony HX9V.