Deconvolution with ImageJ + DeconvolutionLab

[michael.h.f.wilkinson]

Just a few experiments:

Wavelet:

Lucy Richardson using 1.5 sigma Gaussian 20 iterations:

Lucy Richardson using 1.5 sigma Gaussian 100 iterations:

The results are much the same as the previous LR Deconvolutions, and the process takes up about the same time. There is a batch-processing option, but I have not quite got that working. The other advantage of ImageJ with DeconvolutionLab is of course that it is free. It is a bit fiddly, and occasionally throws odd error messages, and forgets which image is the input. I should be able to sort that out (along with batch processing). You have to add a PSF image, rather than allowing you to select PSFs. I tackle this problem by creating an image with a single dot, and convolving that with a Gaussian of the right size. Other PSFs can also be created that way. A bit cumbersome at first, but the BIG advantage is batch processing, which will safe lots of time once I sort it out. I will keep you posted.

[sharkmelley]

The results look pretty similar.  The wavelets may have slightly more contrast than the other two.

Mark

[Pete Presland]

interesting results, i think the 3rd image has a much better balance to it.

[michael.h.f.wilkinson]

Thanks guys. The third is the most subtle, and shows most detail. I have been experimenting a bit more, and have frankly been driven to despair by ImageJ/DeconvolutionLab's fickle behaviour. The first time you open the plugin, it generally works, the second time, it complains the PSF image is not the same size as the image you want to deconvolve, even though it just worked seconds before. There are also problems with image types being incompatible (for some reason, AS!2 stores grey scale 16 bit images as 3x16bit RGB, which ImageJ interprets more or less as a stack of three (or a composite), whereupon DeconvolutionLab refuses to deconvolve it with a grey scale PSF. It also sometimes produces strangely darkened results. I am almost tempted to implement LR deconvolution in C myself, to make a standalone program which would have a commandline interface only (LR deconvolution is not that hard, especially if you have various math libraries at your disposal). I have also just found an OpenCV implementation here:

http://gigadom.wordpress.com/2012/05/14/re-working-the-lucy-richardson-algorithm-in-opencv/

Which I will try to get working. Really want to get a batch mode program to speek up stuff

[michael.h.f.wilkinson]

After removing a few coding horrors from the above code example I have got something working in OpenCV. I will experiment more later

[michael.h.f.wilkinson]

Just realised that the code I found is incorrect. It just performs iterated unsharp masking rather than the real deal posted here:

http://en.wikipedia.org/wiki/Richardson%E2%80%93Lucy_deconvolution

That description is quite straightforward to implement, however.

[Montana]

Great results :)

Alexandra

[michael.h.f.wilkinson]

Thanks, Alexandra, but not so great user interface and stability. I think I will write my own code from scratch and if all goes well post a batch processor version of RL deconvolution; code, makefiles for Linux (and perhaps MasOSX), and Windows executable at least).

[sharkmelley]

Good luck!

Mark

[tvganesh]

@Supernova -    This is Ganesh, the author of the post mentioned above - https://gigadom.wordpress.com/2012/05/14/re-working-the-lucy-richardson-algorithm-in-opencv/

I could not get the algorithm to work effectively. If anybody managed to get this to work I would appreciate if they could share it with me.

Thanks

Ganesh

[Space Cowboy]

Are we talking registax wavelets for the first image? If so dyadic or Gaussian?

[GreatAttractor]

Hi Ganesh,

Inspired by Michael's posts I got interested and managed to implement the L-R deconvolution, which eventually led to development of ImPPG (see LucyRichardsonGaussian() in lrdeconv.cpp). I used http://en.wikipedia.org/wiki/Richardson%E2%80%93Lucy_deconvolution for reference.

[tvganesh]

Very impressive results!! Thx.  - Ganesh