Видео

Thanks for watching! :)

Yes, I agree with all that! Also, binning/pixel size affect the signal part of the SNR measurement, so you'll be able to get the histogram about 1/3 away from the left bar more quickly. It should decrease the integration time needed to get a similar result than if you have smaller pixels/don't bin.

I'm pretty sure you're correct that it doesn't matter when you do the binning for a CMOS camera (it was mentioned at the end). So, maybe it should be more clear that the video is about binning and pixel size (and how they're similar). To answer your question: most of the noise in an image should be shot noise (whether it comes from the DSO or light pollution). Increasing pixel size (however it is done- binning is one way) will take the signal of multiple smaller areas and combine them. That's where the benefit comes in. People often forget the numerator/signal in the SNR calculation. Combining pixels, or having larger pixels increases the "S" relative to the "R," which helps a lot with SNR. But the tradeoff is resolution.

I guess another way to think about it is that with a 2x2 bin and a CMOS camera, it's like stacking 4 pixels together (assuming equal illumination). On the other hand, if you had a CMOS camera with 2x the pixel size, it would be just as good (if not better b/c of less read noise) as the 2x2 bin in terms of SNR. Of course, with smaller pixels, you get to choose whether to bin or not, so there is more flexibility :)

@@deepskydetail thanks for taking the time to clarify. Sadly, I still don't understand why the SNR of single pixels is important enough to lose resolution by binning. We both have the same answer your question at 5:15: No, the signal of that DSO is the same, regardless of binning! So, why should I lose resolution on my DSO by binning the pixels if we both agree that this won't change the SNR of the DSO? Sure, by binning, I'll have big blocky and averaged pixels, but how can this possibly make the DSO look prettier? Sorry if I am daft, but what am I missing?

@@comeraczy2483 there is zero benefit from doing binning on CMOS sensor, the only reason you'd do it is in case you need faster file transfer rate from the camera or smaller files, but from a quality standpoint binned images on CMOS actually end up looking worse than normal ones.

This is a good question. There is a weird thing about Poisson shot noise. It doesn't split the same way the signal does. The standard deviation (i.e., noise measurement) only depends on the mean of the distribution. It is the square root of the signal. So, in this case if we have a pixel with a signal of 50, the noise becomes the square root of 50, or roughly 7.07.

@@deepskydetail if we're calculating the overall SNR however you need to sum the noise in quadrature, so that would be: SIGNAL*NrOfSources/sqrt(Noise^2*NrOfSources) -> 50*2/sqrt(7,07^2*2)=10, just like a single pixel with 100 signal.

​@@v0ldy54 We're not finding the SNR of the whole area though. We're finding the SNR of the individual pixels. That's what we're interested in, especially in very faint areas. If you take two images, one without binning, and the other with 2x2 binning, the 2x2 will be brighter. If you stretch the unbinned one to be as bright as the binned one, it will look noisier. You can actually see the difference around 13:00. The SX (and binned ZWO) is smoother (but less resolution) and the right one is noisier (with more resolution). I tried to equalize brightness and keep the stretching the same. The graph at 13:49 shows the actual calculated SNR for the 2x2 bin and unbinned ZWO, which confirms it.

@@deepskydetail ok, but the final image should be compared to the same output size , so it's the SNR over the same area of the final image that matters.

Glad to hear that! Thanks for the feedback!

Thanks for the sub! I appreciate it! I'm glad you liked the demo. When I saw that demo as I was preparing to make the video, it clicked with me too. There's so much I learn as I'm doing this that I think, "I've got to share this! It's really cool!" :)

Thanks for the shoutout! I appreciate it :)

Thanks for the comment. Could you please point in the video where I claim to be discovering something new 😉 Around 11:05-11:40 I talk specifically about sampling and the tradeoffs of resolution/sensitivity. The point of the video is to figure out why some of my viewers were having a hard time getting good SNR with the camera. It didn't start out as a video about sampling. Sampling just came about naturally as I was looking at the data from the two cameras' images. It's interesting about your comparison of the 294 and the 2600. If you have any data you'd like to share, I'd be more than happy to take a look!

Thanks for watching!

You can use it for OSC cameras too. I would just use the average brightness of the RGB channels (however you want to do it).

Yes, instructions for how to set it up are here: ruclips.net/video/IRe-d-Izo5o/видео.htmlsi=CPz-ws5aYLzGhCXs&t=691

@@deepskydetail thanks

Good questions! Based on what I understand from JASPs documentation, the underlying distribution is just a normal distribution. Other distributions (e.g., ex-gaussian or gamma) might be better, but looking at the overall histograms, a normal is a decent approximation (I'm also a bit lazy and didn't want to code things in jags or stan assuming different priors). There were the same number of samples/replicates for each combination.

Thanks for the comment :) I think the benefit of going longer is most apparent in really faint areas. You help push down the noise in those areas, which really helps! Although stacking definitely helps with exposure times of all types!

Thanks for the comment! The bigger for is amazing, I think!

That would be interesting! If you do it, let me know!

I'm working on a video about the 294 right now as a follow up. Stay tuned!

Thanks! Glad you've come aboard :)

I think that makes sense, to me. Also, a lot of professional telescopes have quite long focal lengths. Bigger pixels also help to match the resolution. I think you're 100% right that pixel size is important to getting good SNR at long focal ratios and large aperture! It reminds me of an astro imaging channel live where they talked about why you'd want bigger aperture scopes.

Yeah, thinking of a planewave or something like it.

That's pretty interesting about your Canon camera. And of course, binning isn't always the best plan of action, depending on the pixel scale. Thanks for the comment!

Yes, that's a good point! :)

Be that guy! No shame :) This one? ruclips.net/video/T9IXodtjRgs/видео.html Heaven and Hell by Jeremy Blake (it's in the YT music library)

Thanks for the comment! I think there is some great information here. I agree that the angle and sampling of the sky are super important here, and it's a major problem in comparing the two images. I'm in the process of doing more rigorous tests, and all the conclusions in this video are very tentative! About the CMOS vs. CCD, this is really good info. Would you have some sources for me to look at and study this more? I've never thinking about the read noise to well depth ratio before, and it make sense on an intuitive level. Let's see if I'm understanding this: If you expose until the maximum without overexposing, you're read noise as a percentage of the actual signal is less than that of a CMOS camera? That's because in order to get the low read noise value in a CMOS camera, you're actually losing dynamic range, decreasing the full well depth? Anecdotally, one thing I noticed about my Starlight Xpress 825 was that it does seem to be less noisy than my 294MM, even though the read noise is greater in the CCD. I just felt that the images were smoother somehow, but I might be wrong, and it might be due to pixel size. Much to learn, and if you have some sites/books/articles to share, please do! You can always email me at deepskydetail at gmail.com!

@@deepskydetail Thanks alot! It was my only purpose to stimulate curiosity about that and I hope it useful... On the first I would like to point out that we are no more "taking pictures" but "making measurements" whit digital cameras. Our intent is to measure, as best as possible, the intensity of the energy (emitted in the form of photons) of a subject. Since our camera is an array of "micro-sensors" that we call pixels even if a pixel is something different in digital world. Anyway we can only make a measure of the total energy that falls in to a single pixel, per every pixel in a given time. We chose to represent these different energies in form of light reproduced from a monitor i.e. and this is quite obvious since we are capturing an emitted "light". Anyway this point out that, in a perfect world, we can only get that total energy and nothing more, no matter the system; in this case our best result is to get that exact energy. When we take such "measurements" we get some kind of errors in the process that alters the "real" data we want to capture. What we call "noise" in digital is exactly that, How much datas is certain of the source we intend to measure and what is not? That is the SNR which is more correct to call: certain Vs. uncertain. Readout noise is part of the electronic noise and is something related to the system that we can't measure or take out from the equation since is not fixed. When we say 5e of readnoise it can be +5 to -5 but can even be 0! The fact is that is different each time and we havo no way to measure it. Other form of shot noise we can deal with in some forms, i.e. dark current noise. We can take dark frames and then subtract these from the image to "correct" the values from the measurements but we can't remove the uncertanty of the readnoise from the dark frames too so another small error is contributing the final measurement (that's why the professionals use cryogenic cooling to cameras and they not take flats at all). Anyway the main difference from CCD's and CMOS is they're purpose, the first ones (CCD's where been invented to be memories for the computer at the beginning) where made for taking precise measurements and they involve complex electronics and clocking systems to work. For this reason they needs alot of energy to work. CMOS on the other end where born for industrial uses and they didn't need high degree of precision. At first they where used as counters or triggers for industrial machines. imagine a conveyor belt with a lot of cans of Campbell's soup, a CMOS underneth the belt could see the darkening of the lights for each can passing and count them. These kind of applications where the ones intended for CMOS use. Today we are too fiew users for the producing companies to think about making a CMOS sensor as precise as a CCD for measurements. We could be 1000000 people but they make 8 billions CMOS per year just for the mobile phone industry! Anyway we are getting smarter since we don't really take pictures anymore but we make statistics instead! That's why we are getting better and better even with sensors that are not made for our intent. It is important to underestand that the readnoise is an error we can't remove but we can bring to a certain value that we are confortable to accept and this is true for all the total noise that we can measure or deal with in some ways. As a fixed number in a range it is indeed a percentage and, higher full wells means that this is in a smaller and smaller percentage, the uncertanty on very low full wells becomes higher and higher as a percentage of the whole signal. We are adapting ourselves to smaller pixel sizes and this becomes even convenient for us 'cos we need shorter focal lenght to achieve the same spatial resolution per pixel respect to CCD's but at the cost of precision of the data and full well. We don't have the chance to know what any CMOS converter do behind the lines since this is an industrial secret, i.e. we nowdays take flats just for correcting the "vignetting" of the system but, whit CCD's we are also correcting variations of sensitivity or discrepancies from one pixel to another! This becouse the CMOS still have a map inside the electronics that correct this value but: how much is precise that correction? We don't know and we cannot change that. All these kind of trickery was not present at all in the CCD's. In the end that's mainly why Your CCD's frames seems more "smooth" to You, becouse they are! The data has an higher degree of precision. Sorry, I wrote too much... Anyway these are two great sources to read if You will: CMOS/CCD Sensors and Camera Systems, Second Edition. Author(s): Gerald C. Holst, Terrence S. Lomheim Photon Transfer Author(s): James R. Janesick

@@deepskydetail Sorry but I would like to make another, mabye useful, consideration... There is another difference between CCDs and nowdays CMOS and is the quantization noise. The CCD has a true 16bit converter while all actual CMOS are double reading 14bit that means they read the data twice ( and I don't know if double readings is also doubling the readout noise...) with a different gain converter and interpolate the results to "obtain" a 16bit approximation. Try to shot a good dynamic subject with large light differences with the same scope, sampling angle and same lenght single exposure that is quite long i.e. 300secs (And You will probably going to saturate the CMOS not becouse is much sensitive but 'cos it has a smalle full well) with CCD and CMOS and than look at the histogram... The CMOS one will be much "smaller" and "spiky" than the CCD one. Try then to increase the non linearity of the representation on fixed steps on both images and You will see that, in the CCD, You will retain more and more "tones" and capacity of distinguish more light differences while the CMOS will "appear" to be in a very small area with less tones to show. This is another effect of the precision of lecture of the datas...

Thanks for the info and sources! I now really want to test those ideas, like comparing the histograms. Sounds like some ideas for new videos ;) A also checked your astrobin images, and was blown away! Great images!!

What cameras do/did you have if you don't mind me asking? I think that generally the newer CMOS cameras have less read noise. But the CCDs tended to have bigger pixels.

@deepskydetail I'm sorry, I should have been clear. First digital cameras from fujifilm. They switched to cmos for costs. Hopefully my new move shoot move will help.

Yeah, the qe for these chips (e.g., 533, 2600 etc) do drop off quite a bit. SII is around 50% qe. You can still have two pixel sizes with your 533 though by binning! Bin 2 mode on the 294MM I think is just 2x2 binning from what I understand (with some gain setting manipulation going on!). Also, seeing and aperture size also influence resolution! Let me know if you get the full 47M (I've actually never tried it myself; I generally stick to bin2 mode)!

Thank you!!

Thanks for the comment! Some responses: 1) Pixel size does matter as well. Let's consider just the light gathering ability, as you mention. -Let's say a given area on the imaging circle gets 4 photons of light on average each sub-exposure. If there is only one pixel (and of course only considering shot noise and a perfect imaging setup), the SNR is going to be 4/sqrt(4) or 2. The average SNR of the pixel is 2 (2 divided by 1 pixel). If there are 4 pixels, then the each pixel will have an SNR of 1 (signal = 1, noise = sqrt(1), SNR = 1). The average SNR of the four pixels is 1 (1+1+1+1 divided by 4 = 4/4 = 1). The 4 pixels are going to be noisier than the 1 big one. Instead of 1 big pixel with 1 value, you have 4 small pixels with slightly different values. That will look noisier. 2) I think there are too many variables between the two systems to draw conclusions, the test should be done with the same setup. -I agree! I hope in the future to do follow up tests. I tried to make it clear that the conclusions and test is flawed and I need more data. 3) binning has zero benefit to a CMOS image SNR This is, as far as I know, not true. You still get the benefit of adding together the signal. However, as you mention you add in the read noise of all the pixels instead of just one pixel. But read noise in newer cameras are pretty small, so you still benefit like you do with CCD binning. Altair has a good explanation here about CMOS binning: www.altairastro.help/info-instructions/cmos/how-does-binning-work-with-cmos-cameras/

@@deepskydetail afaik since they're 4 independent and unrelated sources of noise, where you're averaging shot noise from adjacent pixels you should sum thoise noise sources in quadrature. That means that a single pixel gets 4 signal/√4 noise =2. 4 pixels with 1 noise each will get 1*4 / √(1²*4) = 2. It's the same, which makes sense since shot noise is only dependent on the amount of light and nothing else. Still works for bigger numbers (just pointing out since 1 squared might look weird): with 100 signal on 1 pixel you get √100= 10 noise, with 25 signal on 4 pixels (each with 5 shot noise) you get 25*4/√( 5² *4) ) = 10. Again, you can check for example on the DPReview comparison that you have very small difference between an A7RIV with 60mpx and an A7III which only has 24, if shot noise changed the result would be vastly different (I'll link in another comment in case stupid RUclips blocks the link). The problem with CMOS binning is that you're not gaining a real benefit in SNR, mathematically it's exactly the same as simply reducing the image size, which can be done in post after you stacked the full resolution files, so there is no reason to bin a CMOS unless you need faster transfer speed during acquisition (which is definitely not the case for deep sky astro

@@deepskydetail www.dpreview.com/reviews/image-comparison?attr18=lowlight&attr13_0=sony_a7iv&attr13_1=sony_a7iii&attr13_2=sony_a7riii&attr13_3=sony_a7riv&attr15_0=raw&attr15_1=raw&attr15_2=raw&attr15_3=raw&attr16_0=6400&attr16_1=6400&attr16_2=6400&attr16_3=3200&attr126_2=1&attr199_3=1&normalization=compare&widget=1&x=0.086367625548513&y=-0.14081228556828976 Here is the comparison, of course the A7III wins because read noise is lower, but that's not even a single sotp difference even with pixels that are almost 3 times smaller in area.

lol!😂 I would definitely be the one getting the short end of the stick in that scenario!

Thank you, so much!! :)

Thanks, Doug! :)

I do enjoy my 294mm too. But I agree that flats with the 294MM's flats are difficult. It's hard to get longer flat exposure times, especially in L. A lot of times, I end up having to redo them!

@Ben_Stewart @deepskydetail Don't stress too much over the flats, find a flat panel with extremely dim setting, set NINA to the recommended ADU value with a maximum deviation of 2% and let the app shoot your flats with dynamic exposure. Mine are around 10s of exposure. Guess what, they correct the lights so well that the SPCC in pixinsight sometimes doesn't even do much since the channels are already properly corrected by the flats. This camera indeed goes nuts at short flats or bias frames, so f that, give her what she needs: long flats and dark flats instead of bias frames. Been using it for over an year and I love the results. The small pixels start becoming a little problematic at my 1200mm focal length, but so far deconvolution does take care of the oversampling pretty well, boosting the sharpness of the image. I would not recommend this camera to an even deeper focal length tho.

Great comment! About the different setups/nights, I completely agree with you, as I stated a few times in the video, the comparison is quite flawed and I really want more data do test things with. That being said, most of the variables that I do have data for related to the equipment, moon phases etc. are in the viewer's favor. About the image scale, I also agree with you, which is why I mentioned binning and the tradeoffs with resolution at the end of the video ;) Thanks!

@@deepskydetailThanks for the response. Perhaps you can go back and normalize for pixel scale and see how that impacts your results and update your video to keep it as up to date with your latest thinking as possible. Bummer about your injury, I hope you heal quickly. I think this is critical because I see this being missed by many people and even web sites. I think it would be a huge service to the AP community for them to know that in order to compare systems you need to compare at equivalent pixel scale.

That's a good idea!! If I were to guess, I'd think that the results will show that the SNR per unit area is very similar for both cameras (one has slightly higher qe, the other has lower read noise). I also think that with digital images, the overall SNR of the image will change how good it looks, and pixel size is one (out of many) important factors to consider in the overall SNR. Even if the SNR per unit area is the same, an image with bigger pixels might get better overall image SNR faster than one with smaller pixels, which of course is why binning might be considered. Sorry for the rambling, but I guess what I'm saying is an image using bigger pixels, all things equal, will get a smoother looking image faster than one with smaller pixels. It's the human perception at the end of the day that will make the judgment. The tradeoff, of course, is resolution (i.e. start zooming in on the bigger pixel image, and things might start getting blocky).

@@deepskydetail actually this has been debated on cloudynights extensively, and the general consensus is that you can always downscale the higher resolution image to the equivalent pixel scale and improve the SNR, even after stacking, and so there is really no advantage of the larger pixel scale camera unless you are suffering from read noise. The only reason a larger pixel scale camera is better is to swamp read noise more quickly but as you know, CMOS cameras have very low noise so this is unlikely to be true in this case even with smaller pixels. I think if you wrk the math you’ll find this to be true. So really the main difference is down to Ha sensitivity.

@stevenmiller5452 I see what you're saying, and I agree! That does make sense the way you've explained it. The thing is, generally, the people who have contacted me using these CMOS cameras aren't binning, downsampling etc., and they are wondering why the SNR is so bad (and consequently why it takes them so long to get an image they want). They could downsample. They could bin. And that's ok! It's in the video as a solution! But their expectation (based on what I think is marketing) is "the cameras should be faster" by default (i.e., without binning/downsampling), and they're worried something is wrong (when it really isn't).