For years I have used a fixed-gain CCD camera, first an Sbig ST-10, then QSI 532, both with the wonderful Kodak KAF-3200M chip. The 532 has died, and finding a QSI 632 on the market, which also has the same chip, I bought it only to discover that it operates with either a high (.9) or Low (2) gain setting instead of the fixed 1.29 of the prior cameras, bringing an unwanted complexity to my program.
I use 2X binning. The nominal well-depth of the chip is 55K, but binned, the readout limit is 110k. Doing linear testing, the high gain shows a linear response out very nearly to 65k ADU R.995, as does the 532, while the low gain on the 632 breaks down dramatically at about 48k ADU R.998. The latter is presumably explained by the readout limit of 110k when binned.
In circumstances where I was willing to extend exposure, which is another topic, I could deepen my e count using the low gain and hence reduce the shot noise by a bit over one part in a thousand near the top of the linear range. However, and this is my question, in plotting the incremental increase in well-depth with each incremental increase in exposure, there is a significantly larger variation observed with the low gain setting than with the high gain, and if it is real, it can largely offset, even overwhelm the gain from the deeper well depth provided by the low gain setting. In high gain mode, the average deviation in the incremental gain in ADU reading across the whole chip in the linear range is ~.0014, while in the low gain, it is .0056.
It has been suggested that this may be due to some clocking element inherent in the camera where the exposures are not exactly the same incremental duration due to some rounding off of a granular time cycle – the minimum exposure being .001”. However, that does not seem to fit for several reasons. The time increments at high gain are half a second, while the low gain increments are one second. Any variation from such rounding error should be larger in proportion with the shorter exposures, not smaller, as any rounding would be diluted. Also, rounding exposures to the nearest .001” could not cause variations as large as one percent, which are evident in the low gain data.
I have read that readout noise should be lower in a high gain mode, but I have no idea if it might be large enough an element to account for this. In fact, I have yet to find a good explanation for exactly what causes readout error. If it is readout noise, then it is real, and I should use the high gain setting to reduce it.
I am working at my outer technical intelligence limit and perhaps beyond it. I suspect also I am straining at nates, as the difference between the available well depths is not significant enough to be material in the kind of photometry I am doing, which is focused on untransformed volume rather than ultra-precision.
Thoughts anyone?
Hi, James:
Before trying to offer any advice, you've used some phrases and terminology that I don't really understand in this context:
- Mark
Hi, James:
Okay, that info really helps.
I applaud you – with enthusiasm – for your work characterizing your new camera. Not only are you getting a better handle on how the new camera behaves, but you (and everyone you share this with, including me) are learning new things in the process.
Let's talk about flicker. If your panel is flickering at 60Hz, then the 0.5 second exposure time increment nominally adds 29 to 31 flicker cycles. The difference between 29 and the nominal 30 flicker cycles is potentially a variation of up to 1/30 of the incremental flux that you add for the exposure, meaning that your flux "steps" have an inherent 3% variation in their size, depending on depth of the flicker and synchronization between your exposure start time and local power frequency. I compare 3% to your quoted number (0.0015 or so), and think that you're actually doing quite well.
I do suggest that you gather data separately on linearity and noise. It's difficult to measure noise using a methodology intended to gather linearity info.
That said, a good linearity methodology fits a curve to your plot of ADU vs exposure time. You'll get one set of fitting statistics using a linear fit and a different set of fitting statistics using a quadratic fit. Calculate the residuals (errors) between the fitting line(s) and your raw data. Then we should have a conversation about those residuals instead of your point-to-point increments. The comparison of those residuals between the linear fit and the quadratic fit will tell you a lot about the overall linearity of the camera's response.
I encourage you to make your own measurements of gain and read noise. It isn't unusual for individual cameras to have gain and read noise numbers somewhat different from what's on the vendor spec sheets. Measuring gain and read noise can be done together, and gives you a better understanding of camera behavior than measuring point-to-point variation in the linearity curve.
Mark
Thanks Mark, I will look into pursuing those ideas.