HDR Wavelet Transform Question

[LD]

I just followed the HDRWavelet tutorial on an older M33 image of mine. The tutorial made it easy to follow. I first used A Trous Wavelet to see what layer I wanted to stop at, then moved over to ADR Wavelets. But when finished I wound up with holes in the brighter stars and the galaxy core. You can see the original and reworked image here:

http://www.freewebs.com/drivewayastronomy/workspace.htm

I am quite sure I checked luminance mask. Is there a way to prevent those holes?

I am also amazed at the detail it brings out, not sure if it's an esthetically more pleasing image though. Of course, this was a first and rough attempt. But I also could see how using this to bring up the detail in a luminance for an LRGB image would make a big difference with the RGB adding the "softness" back with color.

Thanks for any help,
Larry

[Carlos Milovic]

Hi Larry

I'm not sure f the cause of these "holes", but is seems to be something related to overflow. Were you working over a 16bit integer image? I may be wrong with this, becouse ithe algorithm performs a rescalation, and the temporally working images are all 32bits floating point ones...
Another way to go, is that maybe this is an artifact due to the small range of scales you are using. Try increasing to 5 layers, and see if this behavior disappears.

Finally, the best thing I can think of, is to build a star mask. Extract the luminance of your image, and with ATrous Wavelets disable all scales larger than 16 or 32 pixels. Now, set the deringing to the maximum to every surviving layer, and apply it to the image. This is a quick way to build it (something similar to a high-pass filter). After the wavelets, you may adjust the mask with curves or the morphological filters, and finally invert it.

By the way, the luminance mask that is used in HDRWT protects the shadows, not the highlights. It means, it is not internally inverted.

[LD]

Carlos,
Thanks, as always, for that very fast and detailed reply. I'll try your suggestions. Once again, one of the things I really appreciate about Pix is the ability to play with different processing strategies.
BTW, is that true of the automated mask in noise reduction? Or is that automatically inverted to protect highlights and only affect the background?
Thanks again,
Larry

[Jack Harvey]

If the automated mask that is in ACDNR does work like this it would be a great stand alone tool to create a custom "Quick Mask"

[Carlos Milovic]

Hi Larry

Yes, there is a inverted luminance mask inside ACDNR. You can modify some parameters of it, and preview the effects. For example, you may disable the first 2 wavelet layers of the mask, to obtain a smoother image. (BTW, ACDNR's algorithm has a powerfull edge detection algorithm that in most cases makes innecesary to use a [inverted] luminance mask... but, of course, since SNR depends on luminosity, is always a good idea to help the process and use it).

There are more processes that includes such kind of masks internally... ExponentialTransforms, UnSharpMask, etc. So far, HDRWT is the only process that uses a non-inverted mask, i.e. it protects shadows instead of highlights.



Hi Jack

In fact, creating a mask like the one used by ACDNR is quite simple. Just extract the luminance, then adjust it with HistogramTransform, and finally use ATrousWavelets to smooth it (disabling some layers).

A more refined mask generation process will be included very soon, following Juan's design for a deringing support in the Deconvolutions process. It will be a quite easy to use, but very powerfull way to design star masks.

[LD]

Carlos and Jack,
That makes sense--what I would have figured about inverted masks. But I never even thought about using wavelets to refine the mask. Actually, just using a simple inverted mask with the noise reduction in the LE version was brought me into Pix in the first place. Well, looks like something new to play with! (Always is with this program.)
Best,
Larry

[LD]

I have tried redoing the image from the above post to see where the holes in the stars come from. They appear after several iterations. I do not see them after two iterations in the 5x5 spline, but they appear after three iterations. I could do a bit more with 5x5 peak, but the stars also seem to bloat a bit more. Don't know if this makes sense.

The original image was a FITS image, median combined in AstroArt3, then opened in Pix Core 1. I resaved the image as a FITS 32 floating point image before doing any other work--histogram transform, curves transform then ADRWavelets.

Again, don't know if this information is helpful.

Larry