DSLR Dark Investigation - Puzzling Result

[Ignacio]

IanL: what you are effectively doing, by avoiding hot pixels, and now trying to deal with cold pixels, is calibrating your darks with some of their repetitive patterns. Try this: build a master dark with all of your frames. Then use it to calibrate the whole set of darks: dark frame - master dark + small pedestal if needed. Then stack it in increasing subsets as in your exercise, and look at the noise progression. It should decay exactly as expected from theory, at least at the beginning, since as you approach the total number of frames, noise approaches zero (and gets there at the end when the master dark is subtracted from itself).

Ignacio

[IanL]

That makes sense, I'll give it a try next weekend when I have sufficient time available.

[Juan Conejero]

Instead of measuring the means and standard deviations of the whole image, I thought I'd try to avoid any skewing of results due to hot pixels (which appear pretty consistent across all the test integrations as we'd expect).  I created eight previews in small regions with no hot pixels and use the preview aggregator to extract them to one image and then measured the mean and standard deviation of the result.  I repeated this for all of the test integrations plus the highest SNR single dark frame - taking care to duplicate the preview locations exactly for all images.

The mean and the standard deviation are non-robust statistics. I suggest you use robust estimators of location and scale such as the median and the median absolute deviation from the median (MAD), respectively, which will be immune to outliers such as hot and cold pixels. In this way you won't need to define previews and your analysis will be easier and more consistent.

For signal and noise analysis the relatively poor efficiency of MAD may become a problem. In this case you probably need more sufficient estimators (that is, estimators able to gather more information from the available data). For example, you can use the biweight midvariance or the Qn estimator of Rousseeuw and Croux. All of these estimators are available on the Statistics tool, and you'll find more detailed information in the reference documentation for the ImageIntegration tool. For noise estimates, you can also use our NoiseEvaluation script.

[IanL]

Thanks Juan,  I'll try those suggestions.  I did use the NoiseEvaluation script in some of my other testing, but I wasn't sure how valid the results would be for a low signal image like a master dark.