PixInsight Forum

PixInsight => Wish List => Topic started by: dfrost on 2019 February 15 22:14:37

Title: Noise reduction like this would be nice!!!
Post by: dfrost on 2019 February 15 22:14:37
A noise reduction tool in PI that would do this would be amazing!!! 
Title: Re: Noise reduction like this would be nice!!!
Post by: bulrichl on 2019 February 16 02:00:46
Jon Rista has made an excellent guide about Noise Reduction with PixInsight processes:

https://jonrista.com/the-astrophotographers-guide/pixinsights/effective-noise-reduction-part-1/ (https://jonrista.com/the-astrophotographers-guide/pixinsights/effective-noise-reduction-part-1/)
(followed by part 2 and 3)

His introduction should be taken to heart by each astrophotographer:

"The first rule of noise reduction is, it’s noise REDUCTION. All too often, astrophotographers will try to eliminate noise, however this distinction between noise elimination and noise reduction is foundational to good image processing. The goal for you as an image processor should be to reduce the noise in the image. If you eliminate noise, you lose something, one way or another. If you eliminate high frequency per-pixel noise, the data takes on a plastic look, what I call plasticization. If you eliminate lower frequency noise, you are usually going to eliminate some detail as well. Noise reduction is a balancing act, walking the fine line between obliteration and preservation."

PixInsight provides a wide range of powerful noise reduction tools. Jon's guide is meant to show how to apply them in the right way.

Title: Re: Noise reduction like this would be nice!!!
Post by: pfile on 2019 February 16 10:12:22
i mean??? ML/neural net algorithms are a hot topic and i don't really see why something new should not be considered for implementation in PI.

having said that these algorithms can be incredibly compute-intensive and most of them want to run on a graphics card, so that would be a new direction for PI.

Title: Re: Noise reduction like this would be nice!!!
Post by: Juan Conejero on 2019 February 16 13:02:41
These algorithms, and the development shown on this paper (https://sites.skoltech.ru/app/data/uploads/sites/25/2018/04/deep_image_prior.pdf) in particular, are extremely interesting.  Despite the fact that these algorithms are very intensive and require GPU-based implementations to be feasible, an implementation on PixInsight would be certainly possible. Our C++ development platform is mature and more than sufficient to host a project like this. GPU support can be implemented for specific tasks by linking a PixInsight module with the required external libraries. The advantage of a PixInsight implementation is the availability of a rich, stable and powerful multiplatform infrastructure. PixInsight modules can be released as open source or closed source, free or commercial products. Our development frameworks are released as open source (https://gitlab.com/pixinsight/PCL) under a liberal BSD-like license (https://gitlab.com/pixinsight/PCL/blob/master/COPYING.md) that allows practically everything.

However, unfortunately, don't expect this to happen. The probability of a group of academic people being interested in PixInsight is basically zero. I am unable to implement something like this, mainly because it is probably beyond my technical capabilities, and even if I were able to manage the difficulties of such a complex project, this is something that I cannot afford. There are much more important things that must be done, including urgent tasks required to keep PixInsight alive and able to compete, and I am completely alone to do everything. So this is just impossible from any practical point of view.
Title: Re: Noise reduction like this would be nice!!!
Post by: Ignacio on 2019 February 21 09:10:24
Very interesting. I have some AI background and have been thinking about this lately. Will look into the technical details. Thanks.

Title: Re: Noise reduction like this would be nice!!!
Post by: dld on 2019 February 21 09:34:49
Here are some projects I stumbled upon while procrastinating :P

Noise2Noise: Learning Image Restoration without Clean Data (https://github.com/NVlabs/noise2noise) from NVidia labs (GitHub repo, python/TensorFlow)

MPI for Intelligent Systems - Pixel Group (http://webdav.is.mpg.de/pixel/): Blind/Non-blind deconvolution. The "Non-stationary Correction of Optical Aberrations" article seems pretty relevant to our community needs for restoring optical aberrations.

Of course I don't expect any of these to be implemented! But you never know, this posting may trigger a gathering of skilled PI volunteers to implement something new for the PI community!

Happy reading!