Lucy-Richardson Deconvolution: So what is it?

[sharkmelley]

What is motion analysis?

[michael.h.f.wilkinson]

Hi Michael,

I've been reading a bit about Moffat functions recently.  Apparently they are a better approximation to the point spread function of stars than Gaussians.  In other words, the shape of a star caused by seeing related atmospheric turbulence tends to be more like a Moffat rather than a Gaussian.  If this is true, it might be worth experimenting with a Moffat function kernel for your LR deconvolution.  If it's true for stars then it I guess it would also be true for the PSF of solar images.

Mark

Certainly an option. I would have to include a new algorithm, but that is probably no big deal. Things would become slower.

[NickK]

Well I've got my 2x1D curve fitting IIR working - simply add your own PSF and it will sort out the rest. http://stargazerslounge.com/topic/192900-atik-osx-drivers-r100-example-app/?p=2504036

I've also spent time researching the 2D non-symmetric version.. challenge is an understatement but I think there's slower form.. but it could be possible to 2D curve fit the coefficients however generating the IIR would take far longer than actually executing it! I'd look at using an external vector fit library.. but they're CPU based..

[michael.h.f.wilkinson]

Interesting progress. I am not sure if within the constraints of LR deconvolution the IIR approach is strictly necessary. In all plausible PSFs, distant pixels have very low weights, and should contribute very little. Truncating the PSF as is done in ImPP allows much more local processing, which is better for GPU implementations, on the whole. I was planning to make a Moffat-PSF along those lines.

[NickK]

From this exercise - the impulse is only normally about 16 pixels wide anyway before the difference drops to below 0.001.. It's that I'm coding on the GPU at the moment - make sense to me. It's possible to set the local memory or even a group of register variables as a sliding window - then you access a single input for each pass before writing a single result to make it even faster.

[NickK]

A non-optimised (code design or compiler) GPU implementation of the IIR - to apply... 0.00047 seconds for a 2D pass on the Lenna picture (AMD 6750M).