Deep Sky Workflows

A full explanation of wavelets involves math and more knowledge than I possess, but hopefully this gives you a higher level understanding. Let’s work on our lunar image.

Multiscale Linear Transformation

Multiscale linear transformation, or MLT for short, is often introduced as a way to reduce noise in your image. A lesser known function is for sharpening of image. At each level or layer, you can manipulate a bias setting from 0.001 to 15. For all practical purposes, the bias is the level of sharpening. Because higher layers involve more pixels per unit, small changes in bias have major effects on the image. The lower layers can handle higher bias because the impact is distributed across more pixels and affects a finer set of details.

One frustration I have with RegiStax is the lack of a repeatable approach to sharpening. Everyone seems to do it their own way and has their own opinions about what works. They are capable of producing amazing images, but I struggle to learn the technique and apply it to my own. With MLT, however, I have a very precise methodology to share with you.

Maybe I’m biased from my 8-bit computer programming days, but eight layers seems to give me the right amount of control over the process. The default of 4 doesn’t leave me enough room. Here’s my process:

1. Create a preview over an area with intersting details
2. Open MLT
3. Reset
4. Set the layers to 8
5. On the residual layer, set the bias to 0.001 and click preview to review the results
6. If I don’t see any result, I double the bias and let the preview refresh
7. I inspect the image to make sure it’s no oversaturated or clipping pixeles — when either happens, I back off on the bias level
8. As a general rule of thumb, lower layers can be sharpened more, so I always start with the next layer’s bias (for example, if layer 4 is at 0.42 bias, I’ll start working on layer 3 with 0.42)
9. I generally double until I hit saturation, then back off
10. After I have the preview in a satisfactory state, I apply MLT to the entire image

Tip When you find a good set of values, consider saving the process to a process icon so you can start with those settings the next time. Every image is different and may require tweaking but having some values that worked on other images can help you get to a result faster.

Here is the preview at the start of the session:

Starting MLT

Here the adjustment is too high and has oversaturated and clipped pixels:

Too much

And when it appears just about right:

Just right

After finding the sweet spot, I apply to the whole image and move on.

The finishing touches

Two more tools in my arsenal are LocalHistogramTransformation and HDRTransformation. As with other options, the key is to use preview and expereiment with different values until you find what works.

1. Save your work often as the new xisf format also embeds history
2. One trick I do is to apply a transformation, such as LHT, and the clone the image. I then undo the original image, and use PixelMath to blend the original with the clone. This gives me more control over how much is applied. For example, if I want to keep 70% of the original and 30% of new, I’ll use the expression 0.7*$T+0.3*hdr where $T is the target image I’m manipulating and hdr is the easy identifier I gave to the cloned image.
3. Another trick is to use multiple clones and blend them as described in this excellent article
4. Curves can also help manipulate the image by increasing contrast.

Here is my before and after.