Odd behaviour of "Additive with Scaling"

[sharkmelley]

I was using ImageIntegration to stack some TIFF light frames with the default "Additive with Scaling" option for output, which is the recommended setting for integrating light frames.

There was quite a dramatic change in the colour balance - especially in the background light pollution.

Here's a single exposure:


Here's the stack:


I would expect the stack to look exactly like the single exposures but less noisy.

The documentation ( https://pixinsight.com/doc/tools/ImageIntegration/ImageIntegration.html#description_004 ) is not very clear on Output Normalization.  It indicates it is trying to match backgrounds somehow.  But the backgrounds of the exposures were already very well matched.

The solution is easy - set it to "No normalization".  But I would like to understand what is going on and whether or not it is intentional behaviour.

By the way, this was non-linear data, as you can see.  I don't know if that has anything to do with it.

Mark

[johnpane]

One problem I see with doing the normalization at the time of integration is that the images have been registered, leaving black bands around some edges that can distort the statistics on which the scaling is based. If the backgrounds were matched prior to registration they won't be any more afterwards unless they were already zero.

The more movement between subframes (e.g. aligning on a comet), the worse this problem may be.

It seems better to do the normalization prior to registration. 

Edit: ...and ideally the normalization would be done in a color-preserving way.

[sharkmelley]

If we don't understand what is going on (and I still don't), I think the safest approach for those who are concerned about consistent colour over the whole dynamic range in their images is use the safest option i.e. "No Normalization".

It still makes sense to use normalisation during pixel rejection but not for the output of the integrated image.

Mark

[Juan Conejero]

Hi Mark,

Some critical algorithms implemented in the ImageIntegration tool have been conceived, designed and adjusted to work with linear data. For example, noise evaluation does not make sense, in general, for nonlinear data. Moreover, default parameter settings have been defined after extensive testing on linear data sets.

The transformations applied for output normalization are described in section 1.4.3 of the ImageIntegration reference document. The main goal of the ImageIntegration task is to maximize signal to noise ratio in the output image for a given data set, not to preserve color balance or color consistency. In our tests, additive with scaling output normalization improves SNR consistently in integrated light frames for most data sets, that's why it is enabled by default and recommended in the documentation.

Color adjustments such as background neutralization and white balancing can be performed very easily and accurately with a variety of tools, including HistogramTransformation, BackgroundNeutralization, ColorCalibration, PhotometricColorCalibration, etc. However, these concepts are basically of no concern to ImageIntegration.

[sharkmelley]

Hi Juan,

Thanks for the explanation!

Mark