Experiments with unsharp masking on data from the 22nd

[michael.h.f.wilkinson]

I have just tried some variants of unsharp masking. In my more recent processing, I seem to have settled on a very narrow sigma, and fairly large strength in ImPPG, but others I have seen use much larger sigma, and lower strength. I tried such settings, with and without a second stage with my usual low sigma but smaller sigma than usual. The results are the following:

My standard process: sigma 1.15, strength=4.75

Sigma= 26.0, strength=2.5

Two step process: sigma= 26.0, strength=2.5, followed by sigma 1.15, strength=2.

The first has definitely got finer detail than the second. The third is also good, and has a bit more contrast on medium scales, but might look a bit overcooked. Not sure which is best. Opinions and comments very welcome.

Just spotted this: The latter two do have a ringing artefact along the limb. Not pleased with that, and wonder if it could be prevented easily

[JamesF]

I prefer the first to both of the others.  The AR looks much better to my eye, and just below the lower of the two large proms there's a tiny little prom feature that shows up much more clearly in the first than in the other two.

James

[michael.h.f.wilkinson]

I tend to agree, my older version is better. Now to get a decent GUI around my adaptive version

[Davey-T]

First one looks best to my un-expert eye, I occasionally have a fiddle with ImPPG settings but I've no patience really, still a newbie and pretty pleased with any images I get

Dave

[JamesF]

I've cropped down the images to the feature I was talking about which is in the centre of these three (in the same order as the originals)

     

I'd say that despite the slightly reduced contrast the definition is without doubt better in the first image.

James

[michael.h.f.wilkinson]

It is also a setting that allows processing of surface and prom detail without change. We are currently looking into various kinds of adaptive filter at work (including variants of Perona and Malik's anisotropic diffusion) which might improve matters further. The idea is to smooth/sharpen in such a way that the signal to noise ratio is roughly constant. At the moment we are focusing on radio and optical images of galaxies, but I can foresee extensions to solar and planetary

[JamesF]

Perona and Malik's anisotropic diffusion

This is the point at which my brain starts to dribble out of my ears...

https://en.wikipedia.org/wiki/Anisotropic_diffusion#Formal_definition

James

[michael.h.f.wilkinson]

This is the point at which my brain starts to dribble out of my ears...

https://en.wikipedia.org/wiki/Anisotropic_diffusion#Formal_definition

James

The principle is simple enough, but the Wikipedia article dives in at the deep end a bit too much. Put FAR more simply, we can interpret the grey levels in an image as the distribution of some substance (say a luminous ink). High grey level equates to a high concentration, low grey level to a low concentration. Gaussian blurring can be seen as the result you would get if you let this luminous ink diffuse in the image plane. The longer you wait, the higher the blur factor. Apart from convolution with a Gaussian (which is the quick way to compute the result), you could compute the same result by simulating the diffusion process. This is defined by a so-called partial differential equation which states that the rate of change of the brightness of each pixel depends linearly on the second spatial derivative (or Laplacian) of the image. The Laplacian has a high magnitude near peaks and valleys, and a low magnitude in smooth regions. The strength of the diffusion is determined by the diffusion constant, which is fixed. We start with the original image and then simulate the diffusion process, getting more and more blurry images.

What Perona and Malik proposed was to replace the fixed diffusion constant with a function that is low near strong edges in the image, and high where the image shows little detail. This means flat areas with noise are smoothed rapidly, but sharp edges are left more-or-less intact.

[JamesF]

Ah!  It's all so obvious now

Actually, that does make more sense to me.  I've never been scared of maths, but I think I've reached a point where not having done any formal mathematics for (ahem) years has resulted in me being unable to follow a lot of stuff when it's just written out in symbols.

James

[michael.h.f.wilkinson]

What we are doing is making the diffusion rate intensity driven. Low intensity means low S/N, high intensity means high S/H, so the fainter regions need more smoothing. We might want to balance that with edge information in some way too.

[Ewan]

I tend to agree, my older version is better. Now to get a decent GUI around my adaptive version

There is a De Ringing option in IMPPG Michael.

[JamesF]

I tend to agree, my older version is better. Now to get a decent GUI around my adaptive version

I seem to recall that you wrote it in some version of C.  If so then the Qt toolkit is worth a look, particularly because it is available on Windows, OSX and Linux.

James

[michael.h.f.wilkinson]

I seem to recall that you wrote it in some version of C.  If so then the Qt toolkit is worth a look, particularly because it is available on Windows, OSX and Linux.

James

Or FLTK, which I have. It is just a matter of time

[michael.h.f.wilkinson]

There is a De Ringing option in IMPPG Michael.

The de-ringing method is for the LR deconvolution, not unsharp masking, as I understand it, but I will check that later

[Pete Presland]

1st image for me as well.