asi1600 or asi294?

[alacant]

OSCs: asi1600  asi294

DSO-wise; has anyone any hands-on with either? A side by side would be great.

Cheers and clear skies.

[Susaron]

I guess you want OSC not monochrome, do you?, If I had to change my qhy168C with one camera now, I would go for a 294. If you are thinking on a 1600 you will better try the Atik Horizon.

[alacant]

1 hour ago, Susaron said:

OSC not monochrome, do you?

Yes.

1 hour ago, Susaron said:

I would go for a 294

Have you tried one?

Cheers

[Adam J]

There is a big thread on cloudy nights about the ASI294c, essentially its excellent but you cant push it much past 3 to 4mins exposures without getting dark frame artifacts.....now personally I don't see any reason at all why you would want to go any longer than that as its very low read noise as per the ASI1600mc pro and even more sensitive. So in essence its a non issue so long as you use the camera approprietly with exposure of less than 4mins if your wanting to go longer your doing it wrong anyway. 

https://www.cloudynights.com/topic/608454-m81-m82-with-zwo-asi294/

https://www.cloudynights.com/topic/611039-help-me-with-this-artifact-zwo-asi294/

The ASI294mc pro is probably the only OSC camera I would consider purchasing baring the larger full frame CMOS sensors. 

Adam

[vlaiv]

22 minutes ago, Adam J said:

There is a big thread on cloudy nights about the ASI294c, essentially its excellent but you cant push it much past 3 to 4mins exposures without getting dark frame artifacts.....now personally I don't see any reason at all why you would want to go any longer than that as its very low read noise as per the ASI1600mc pro and even more sensitive. So in essence its a non issue so long as you use the camera approprietly with exposure of less than 4mins if your wanting to go longer your doing it wrong anyway. 

https://www.cloudynights.com/topic/608454-m81-m82-with-zwo-asi294/

https://www.cloudynights.com/topic/611039-help-me-with-this-artifact-zwo-asi294/

The ASI294mc pro is probably the only OSC camera I would consider purchasing baring the larger full frame CMOS sensors. 

Adam

Interesting read, I feel that it has something to do with two different regimes camera operates in. From what I've gathered in second thread - artifact arises when using unity gain that is still in high noise regime. I believe it is not artifact at all but the way read noise is presenting itself in image.

This post is specially indicative:

https://www.cloudynights.com/topic/611039-help-me-with-this-artifact-zwo-asi294/page-3#entry8467850

as it shows low noise vs high noise regime results

[alacant]

@vlaiv could you give me a plain English one liner on 'unity gain'? If I can understand that, I think I'll understand the 294 problems better. Cheers.

[Adam J]

3 hours ago, alacant said:

@vlaiv could you give me a plain English one liner on 'unity gain'? If I can understand that, I think I'll understand the 294 problems better. Cheers.

Its the gain value for which one photon hitting the chip produces a single electron of charge and that is then converted into a single bit of information. For lower gains it takes more electrons to produce one bit of data and at higher gains it takes less. 

Unity gain is often considered a best compromise between read noise, measurement resolution (quantization error) and dynamic range. 

 

[vlaiv]

3 hours ago, alacant said:

@vlaiv could you give me a plain English one liner on 'unity gain'? If I can understand that, I think I'll understand the 294 problems better. Cheers.

I'll try

It might not be quite one liner, but I'll do my best to make it simple. Hopefully you have good understanding of gain, and in this instance we will treat it as conversion factor between recorded electron count in each pixel and numeric value associated with that pixel (later turned in a shade of gray or color). Unity gain is simply one to one mapping. So if pixel captures 50e, actual pixel value of downloaded image will be 50 (units of some kind, usually called ADU). That is it, that is what unity gain is.

Why is unity gain important - there is simply no one liner to explain that. But if you want me to, I can give a go at that as well.

[hughgilhespie]

54 minutes ago, vlaiv said:

I'll try

It might not be quite one liner, but I'll do my best to make it simple. Hopefully you have good understanding of gain, and in this instance we will treat it as conversion factor between recorded electron count in each pixel and numeric value associated with that pixel (later turned in a shade of gray or color). Unity gain is simply one to one mapping. So if pixel captures 50e, actual pixel value of downloaded image will be 50 (units of some kind, usually called ADU). That is it, that is what unity gain is.

Why is unity gain important - there is simply no one liner to explain that. But if you want me to, I can give a go at that as well.

Yes please!

[vlaiv]

1 minute ago, hughgilhespie said:

Yes please!

Ok, you asked for it

There are 3 different settings for gain, we will examine each one of those settings and their impact in mapping / conversion process. One is e-/ADU value greater than one (lower gain then unity), Unity gain, and e-/ADU value less then one (higher gain than unity).

Let's see first what happens with lower gain then unity, or e-/ADU value greater than one.

With modern CMOS cameras it is often case that full well capacity is higher than there are levels in ADC bit count. For example with ASI1600 FW is around 20K, but ADC is 12 bits. 12bits allows for 4096 levels of signal to be recorded (0-4095, 2 to the power of 12). Values from ADC are always integer numbers (no fractions). This is where e-/ADU conversion factor steps in. It determines "usable" full well capacity, and also dictates how much "read noise" there will be. I've put "read noise" in quotes because not all of it is true read noise, some of it is quantization noise, and that type of noise depends on chosen conversion factor.

Here is an example for this case, so we can analyze what will happen. Let's say that we choose e-/ADU to be equal to 2. Pixel gathers 100e of signal and what happens next is that this number is divided by e-/ADU factor and resulting number is recorded as integer. So pixel value at this gain setting will be 50ADU. From this, and fact that ADC is 12 bit, you can clearly see that saturation signal is at 8190, so in effect your full well capacity is not maximum 20K (20,000) but rather 8190, as any higher number of electrons will simply be recorded as maximum value allowed by ADC range, and that is 4095, and that represents 8190e (4095 times e-/ADU).

Let's see what else happens. If we for example have odd number of electrons captured in pixel - say 75e, what would be ADU count for that pixel? We do the same, divide electron count with e-/ADU factor and record value. Result is 37.5 ADU. But we said at the beginning that ADU count can only be integer value, no fractions allowed as result of ADC. So instead of recording 37.5 actual recorded value will be 37 (rounded down). That is problematic since if we try to get true electron count from our recorded value (multiply by e-/ADU), we will end up with 74 instead of 75. So simply by doing conversion we introduced 1e of noise.

What is worse, in case of e-/ADU being equal to 2, even numbers have no quantization noise, odd numbers have 1e of quantization noise, so such kind of noise is not random! Non random noise is bad kind of noise.

In reality when quantization noise is smaller than true read noise, two "mix up" and you get kind of noise that behaves almost like random noise, but if you take for example e-/ADU value of 4, you will get 0, 1, 2, 3 noise values (all possible reminders when dividing with 4) and if read noise (intrinsic one) is less than that you will start getting more and more "badly" behaving noise. So we can conclude that lower gain does bring in more usable full well capacity but increases "read noise", and the more it increases, worse it becomes (in terms of being non random).

This can clearly be seen on gain vs noise graph:

Now let's see what happens when we choose e-/ADU to be less than one, or higher gain than unity. First of all, we still continue to limit our FW capacity. For e-/ADU value of let's say 0.4 in our example max signal that we will be able to record is 4096 * 0.4 or 1638e. But here we also have quantization noise in some cases. Let's examine e-/ADU value of 0.4, and take that we recorded 7e in pixel, what ADU value would that give us? 7e / 0.4 = 17.5 - again we come up with fractional number that can't be recorded and it needs to be rounded down - thus introducing quantization noise. For this regime when we select gain higher than unity it turns out that there are e-/ADU factors that never cause problems with rounding. These are 1/2, 1/3, 1/4, 1/5 ..... It is worth noting that any kind of e-/ADU you choose that is less than 1, resulting quantization noise is going to be less than 1 and is almost always masked by proper read noise, so while quantization noise is there it's contribution is not so bad that you should worry much about it.

So what is so special about unity gain? It does not introduce quantization noise (multiply with 1, divide with 1 - number stays the same), and it is setting with highest FW capacity that does not introduce quantization noise in the mix. This means that it is good place in terms of exposure length - that will not saturate as quickly as higher gain settings, and lowest gain that is not starting to boost "read noise" in a bad way.

Hope all of this makes sense.

[drjolo]

36 minutes ago, vlaiv said:

In reality when quantization noise is smaller than true read noise, two "mix up" and you get kind of noise that behaves almost like random noise, but if you take for example e-/ADU value of 4, you will get 0, 1, 2, 3 noise values (all possible reminders when dividing with 4) and if read noise (intrinsic one) is less than that you will start getting more and more "badly" behaving noise. So we can conclude that lower gain does bring in more usable full well capacity but increases "read noise", and the more it increases, worse it becomes (in terms of being non random).

Are you sure this read noise increase at low gain is due to quantization error? I always thought that slope is getting larger when camera electronics noise limits the total read noise. Read noise is sum of two components: sensor noise and electronics noise. Difference is that electronics noise is constant, and sensor noise is amplified when gain/iso increases.  http://www.clarkvision.com/articles/iso/

[vlaiv]

9 minutes ago, drjolo said:

Are you sure this read noise increase at low gain is due to quantization error? I always thought that slope is getting larger when camera electronics noise limits the total read noise. Read noise is sum of two components: sensor noise and electronics noise. Difference is that electronics noise is constant, and sensor noise is amplified when gain/iso increases.  http://www.clarkvision.com/articles/iso/

I think it is as you say combination of multiple factors. I don't think that gain amplification has such impact on read noise (increasing / decreasing) level because read noise is expressed in electrons and not ADU units, therefore it is converted back into signal level before measurement of noise (at least it should be). It can happen that there is "modulatory" phase between noise sources, and not just additive - then gain setting would have "amplification" effect.

There is quantization noise impacting read noise, but also remember that quantization noise impacts any and all signal measured, so when you shoot in real conditions, light, dark and read signal combine into signal and this is what quantization noise acts on. So while on read noise graph we can see consistency with quantization noise, in reality to measure it, one would need to do it on flats with different gain settings. Take let's say 256 flats, stack 255 and then subtract from one that is left over, measure noise and multiply with 17/16 (or something like that, need to check my math) and you will get noise level. Repeat on different gain settings and create graph - subtract read noise graph, and you should have quantization vs gain graph.

[alacant]

Wow...

Just staying above the clouds for a moment to get this in perspective, this is a 4 hour shot with the 294. It's sort of snap the mono brigade cite; you only get this quality in mono. 

It doesn't seem to be as bad as the cloudynights guys say. I'd certainly not get anywhere near with my dslr no matter how long or dark I had.

And yes, I'm gonna do my homework. 

[vlaiv]

43 minutes ago, drjolo said:

Are you sure this read noise increase at low gain is due to quantization error? I always thought that slope is getting larger when camera electronics noise limits the total read noise. Read noise is sum of two components: sensor noise and electronics noise. Difference is that electronics noise is constant, and sensor noise is amplified when gain/iso increases.  http://www.clarkvision.com/articles/iso/

Actually, I had a think about this, and you are right, it does not have to be modulatory, simple additive noise from electronics, can cause 1/x kind of graph combined with gain setting.

(sensor_noise * amp + electronic_noise) / amp = sensor_amp + electronic_noise / amp

Above simplified (and inaccurate formula, noises don't add like that) shows that there is 1/ (amp or gain) component which explains curve. There is however still quantization noise involved (it has to be, because results are always integer values) - but this shows that working at lower gain is not as bad as it might have been. With increase in "masking" noise, quantization noise with large e-/ADU factor becomes less of bad noise.

[Adam J]

18 minutes ago, alacant said:

Wow...

Just staying above the clouds for a moment to get this in perspective, this is a 4 hour shot with the 294. It's sort of you can only get that detail in mono. 

It doesn't seem to be as bad as the cloudynights guys say. I'd certainly not get anywhere near with my dslr no matter how long or dark I had.

And yes, I'm gonna do my homework. 

Actually the guy complaining about the camera in one thread is the same chap who took that picture....I actually replied to it and suggested he was being a tad picky. 

Overall I think you would be better off going mono...but if you have some real reason to go OSC then I would get the 294.

[alacant]

On 27/04/2018 at 14:52, Adam J said:

I actually replied to it and suggested he was being a tad picky

Hi. OK, thanks. Try as i may, I really can't find out what it is they are talking about. Red patch, cold finger margin... Could you point out what it is that is supposed to be wrong? e.g. the 4 hour shot  in question or WHY? TIA.

[ollypenrice]

I don't like to post negative criticism of other imagers' work but if the M81/82 image were mine I'd be thinking that it was very colour cold and that there was limited Ha from the cigar galaxy. I'd also be thinking about how to find more contrast in the upper end of the brightness range. I don't know how much the camera has to do with this - well depth etc etc. I'm just trying to answer your question.

What I do find myself wondering, though, is why you'd want to saddle yourself with the difficulties of using an RGGB matrix throughout the month, irrespective of the state of the moon.

Olly

[kens]

On 4/27/2018 at 21:06, vlaiv said:

So what is so special about unity gain? It does not introduce quantization noise (multiply with 1, divide with 1 - number stays the same), and it is setting with highest FW capacity that does not introduce quantization noise in the mix. This means that it is good place in terms of exposure length - that will not saturate as quickly as higher gain settings, and lowest gain that is not starting to boost "read noise" in a bad way.

Hope all of this makes sense.

There is nothing special about unity gain in terms of quantization noise. You get quantization noise because you are converting the analogue output of the amplifier into a digital value. That analogue signal is an imperfect rendition of the original discrete charge on the photoreceptor, unless you have zero read noise. With read noise in the order of 1e-  the amplifier output cannot represent the discrete charge in any way that looks discrete. Even getting an analogue amplification that precisely represents unity gain is a fluke.

Quantization noise will be LSB/sqrt(12) always.

[vlaiv]

54 minutes ago, kens said:

There is nothing special about unity gain in terms of quantization noise. You get quantization noise because you are converting the analogue output of the amplifier into a digital value. That analogue signal is an imperfect rendition of the original discrete charge on the photoreceptor, unless you have zero read noise. With read noise in the order of 1e-  the amplifier output cannot represent the discrete charge in any way that looks discrete. Even getting an analogue amplification that precisely represents unity gain is a fluke.

Quantization noise will be LSB/sqrt(12) always.

Analog value is made up out of individual electric charges, always. When we are talking about electricity and millions of electrons flowing, then we can talk about "analog" signal. Here we are talking about discrete number of electrons so it perfectly makes sense to talk about "stepped" analog signal, or electron count.

Where did you come up with LSB/sqrt(12) expression? I just run a test and created 2 images with pure Poisson noise - equivalent to what you would get from light signal in single exposure. First was with average value of 1 per pixel, second was with average value of 2 per pixel. I then quantized it with e-/ADU factor of 2, then restored to "original" and subtracted two. What is left is of course noise that was introduced.

First one gave standard deviation of 0.49992 and second one 0.4955, these two figures are not the same, and I can't fanthom how any of those numbers can be related to LSB/sqrt(12) expression.

These noise values are expected with Poisson noise and expected value of 1, there will be fairly equal number of 0 and 1 values with occasional 2, 3 and very rare above, so about half of numbers will have noise of 1 and other half noise 0 - about 0.5 in average. With Poisson and expectation value of 2, more larger values are to be expected (like 2) so it shifts a bit down from 0.5 because of this 0 and 2 have 0 noise, while 1 and 3 have 1 noise.

[kens]

The output of the amplifier stage is an analogue signal just as you describe. It is a continuous waveform with rises, decays, harmonics, distortions etc... and that is what is input to the ADC for digitization. It doesn't matter that it is stepped as it is an imperfect count of electrons. The error in converting the charge on the sensor to a voltage and amplifying it is sufficiently large to render the idea of a perfectly stepped waveform unrealistic. Consider a read noise of 2e- applied to any charge of  discrete electrons. The analogue is lucky to represent the correct number of electrons let alone a precise value to within a small fraction of one. There are also errors in the comparison process in the ADC as it is all in the analogue domain.

LSB/sqrt(12) is the standard model for quantization noise. See https://en.wikipedia.org/wiki/Quantization_(signal_processing)#Quantization_noise_model and scroll down to quantization noise model.

[vlaiv]

7 hours ago, kens said:

The output of the amplifier stage is an analogue signal just as you describe. It is a continuous waveform with rises, decays, harmonics, distortions etc... and that is what is input to the ADC for digitization. It doesn't matter that it is stepped as it is an imperfect count of electrons. The error in converting the charge on the sensor to a voltage and amplifying it is sufficiently large to render the idea of a perfectly stepped waveform unrealistic. Consider a read noise of 2e- applied to any charge of  discrete electrons. The analogue is lucky to represent the correct number of electrons let alone a precise value to within a small fraction of one. There are also errors in the comparison process in the ADC as it is all in the analogue domain.

LSB/sqrt(12) is the standard model for quantization noise. See https://en.wikipedia.org/wiki/Quantization_(signal_processing)#Quantization_noise_model and scroll down to quantization noise model.

Also quote from Wiki (on that particular model):

"For complex signals in high-resolution ADCs this is an accurate model. For low-resolution ADCs, low-level signals in high-resolution ADCs, and for simple waveforms the quantization noise is not uniformly distributed, making this model inaccurate.^[17] In these cases the quantization noise distribution is strongly affected by the exact amplitude of the signal."

I completely agree with you when we are talking for example of 16bit recording of audio from microphone. But when it comes to imaging and capturing very faint nebulosity, we are talking here order of magnitude 0.01 - 0.001e from target per second, or in short exposure of 60-120s, total signal from target in one exposition ~1e, This is by no means high ratio of signal to resolution of ADC and LSB affected, even if you account for sky brightness that does dithering (think narrowband - not much sky signal), and dark noise, and gaussian model for read noise.

Also bear in mind that read noise can be significantly less than 2e-, just take a look at read noise figures at high gain, or maybe sCmos sensors, all having below 1e read noise.

Problem with this type of noise is not that it is noise - but that it is badly behaving noise. It behaves badly at signal levels that are low compared to resolution of ADC. Stacking will not improve SNR in presence of such noise, to the level one might expect. If you want to see that sort of problem "in action" - look at "onion rings" when imaging Jupiter in 8bit format.

[alacant]

Is there any way I can put a filter in the train of a 294 between a 16mm length OAG and the sensor?

[alacant]

14 hours ago, ollypenrice said:

using an RGGB matrix

?

[vlaiv]

I do it like this:

ASI1600 + Filter drawer + male / female inverted + spacer + 16mm OAG + 2" nosepiece.

ASI1600 has the same housing (at least it looks the same) as 294, so I suspect it will work similarly. I use filter drawer because I use 1.25" filters, and those need to be fairly close to sensor to avoid vignetting.

If you are going to use any filters, they should probably be larger format, like 31/36 unmounted or 2", because 294 is a larger chip.

What kind of filters do you plan to use with OSC camera? If LPS type, then put it before OAG - it will not cause trouble with guiding.

If you want to use narrow band or RGB filters, why bother with OSC in the first place?

[vlaiv]

4 minutes ago, alacant said:

?

I think Olly implied that with Mono camera with filters you can utilize more imaging time in the course of the month with relation to Moon being present, as not all colors are impacted equally by the Moon. You shoot luminance when there is no moonlight, and you can shoot red when there is young moon out. With RGGB matrix (or in another words OSC camera) - you will be shooting all colors each time you image, and depending on moon, they will be impacted differently.