Hi Nikolay,
One more idea: Integrate up-sample to alignment. So I want ask you to add nearest neighbor interpolation to StarAlignment.
Count with it implemented in StarAlignment (a new version will be released very soon).
However, instead of nearest neighbor I'll implement a binning upsample. You can try it with the IntegerResample process. A nearest neighbor interpolation is a disaster with even magnification factors, due to the fact that there is no central pixel in a neighborhood of even size. Along with that, with odd upsampling ratios the nearest neighbor algorithm will shift your images one pixel toward the top left corner. An integer binning doesn't have any of these problems. I have attached a comparison between both algorithms.
but StarAligment use Bicubic Spline ( it's non linear interpolation? ), so i got not so good result.
Sure, bicubic spline interpolation is linear. Pixel interpolations are implemented as convolutions in PixInsight, and convolution is a linear operation. The fact that the interpolation function is nonlinear (as a cubic function) has nothing to do
with the linearity of the interpolation with respect to pixel values (i.e., to the underlying relation between pixel values and incident light). In the interpolated image, each pixel is a
linear combination of a set of adjacent source pixels.
In your technique, only a discrete interpolation can work. This is because real-valued interpolations always cause some smoothing in the interpolated image. Smoothing happens because the interpolation function, which must be sampled in at least a 3x3 kernel, correlates adjacent pixels. This is just what you
don't wan't: your drizzling technique requires uncorrelated source pixels.
Do you need any files from me?
Indeed. As you have already created this technique, you have a data set that works well, as you've demonstrated. I think this technique is elegant and efficient. It is just the kind of things that I love to see implemented in PixInsight
