Juan, what would happen to your optimization procedures if image noise in Canon DSLR images was not proportional to exposure time? Would your optimization routines fail?
Not at all. The optimization algorithm is absolute; it does not depend on any physical property of the sensor, neither on any acquisition conditions. It will always find the optimum scaling factor that minimizes small-scale noise induced by dark subtraction in the calibrated image
With our current dark scaling routine, nonlinearity of the sensor for very bright pixels causes undercorrection of hot pixels for some cameras. This happens because the optimization algorithm finds the best dark scaling factor for noise minimization, while optimal hot pixel removal involves subtracting darks with higher scaling factors close to one. We have designed a variant of the optimization algorithm (multipoint dark scaling) that should fix this problem very well. Hopefully It will be released with the next version of PixInsight.
Also we already have a couple of reports that the wavelet based noise estimation sometimes fails for DSLRs.
Those reports refer to a few odd issues with the MRS noise evaluation algorithm implemented in the ImageIntegration tool (multiresolution support noise evaluation). MRS noise evaluation is extremely accurate but has the drawback that it can fail in some cases. The evaluation algorithm used in the ImageCalibration tool (k-sigma iterative noise thresholding on the first wavelet layer) is much simpler and 100% robust. It cannot fail unless the image has absolutely no noise, which obviously cannot happen when we are subtracting a dark frame. Even if you try to fool the algorithm providing noise-free images, the optimization will converge "graciously" toward a zero dark scaling factor
Just a gut feeling that this may be the cause for the non-improvement compared to DSS despite of PIs sophistated procedures. I have not yet had the time to investigate...
Well, one possibility can be that DSS is carrying out a very nice job
On one hand, we are speaking of "exquisite" algorithms and features, whose actual repercussions on final images are difficult to evaluate. On the other hand, with the exception of automatic dark scaling there is not too much room for improvement in the image calibration stage, since the procedures are rather simple and calibrating is just subtracting and multiplying things after all.
Where PixInsight really shines --in my opinion-- is in image registration and integration, where we provide superior algorithms and implementations. Both tools will receive important improvements in the short term, which will make them even more powerful and versatile. Whether these routines can have a clearly measurable result depends basically on the characteristics and quality of the raw data, in my opinion.
