How to align images with no "stars"

[Niall Saunders]

Ok folks,

Here is an interesting question, "How can I use the likes of Dynamic Alignment" to align two images that do not have any 'stars' ?"

Interestingly, these two images do not even happen to be astronomical images!! In actual fact they are images of my retinae, taken by my local optician over a period of several years - but, they could even be finger-print images, for example.

My 'thought-experiment' would involve analysing the two images, and accurately placing a 'synthetic star' at key points on each of the two images, typically using a 3 x 3 grid of such reference points on each image to cover the image, where each synthetic star is placed as accurately as possible over an easily identifiable point (such as the bifurcation point of two veins on my retina images, or a bifurcation point of two ridges on a fingerprint).

Obviously, I could now Dynamically Align the two 3x3 grids - but how could I apply the resultant 'alignment' to the images that did NOT contain the 3x3 overlay grids?

An immediately simple solution would be where the sources images had a 'Reference Grid' sumperimposed AS A LAYER. The DA process then gets applied to the layers, it obviously affects the image as a whole, so the background (i.e. the real image) is also then 'distorted into alignment' as well, and then the temporary layers are either made invisible, or are deleted.

So, that becomes a classic case FOR the inclusion of "Layers" in PixInsight - no longer an 'artistic' requirement, but now a 'scientific' requirement

However, in the meantime, given that Layers are NOT yet available in PI, is there an equivalent series of steps that can be implemented to achieve the same result?

Not only would this help me decide whether my opthalmologists are correct in advising me to invest heavily in a seeing-eye dog and a white cane, but I can also see how it could help align images that had more nebulosity than stars, but where key features in the nebulosity could be pin-pointed to aid the Dynamic Alignment process.

I can even see how planetary features could also be 'keyed' to aid alignment - perhaps this would be a step towards the Fourier Alignment process as used in the likes of the excellent Registax software?

Cheers,

[Nocturnal]

The FFTRegistration script doesn't do this already Niall? I have no data to try it with

[Niall Saunders]

Hi Sander,

I'm nowhere near any of my PI machines - hence the reason for postulating things as a 'thought experiment'.

Yes - assuming the FFT Registration script behaves in much the same way as the similar porcess in Registax, then it should work. And, probably more accurately than simply 'placing a dot at a bifurcation point', as I was considering doing.

However, if there is likely to be significant image distortion between the two images - as I expect there to be in the case of my retinal images, then I was imagining that the ability of DynamicAlignment to resolve these distortions might be the power that I was going to need.

I will need to look at the FFT script to see how it accomplishes things.

Even now though, I have visualised a similar scenario, where a user might be trying to re-align the likes of an AllSkyCam image sequence, where the 'normal' view of such a sequence, once animated, is to see the stars rotating overhead. However, perhaps the sequence becomes more interesting - for differing reasons - if the rotation is 'frozen' in the sky, and all other 'fleeting' objects (meteors, satellites, aircraft, UFOs, etc) appear over this 'frozen backdrop'. Perhaps FFT 'can' find points to assist with the image alignment, but perhaps it could be 'aided' by a user 'marking up' the reference points on each image. Yes it might be tedious, but the result might be worth the effort!

Really, the fundamental question is "How can I apply the RESULT of a Dynamic Alignment process (created by aligning image B to image A) to a non-associated image, C - accepting that images B and C must have the same physical dimensions ?"

I don't see that this can be done by 'saving' the DA Process Icon and re-applying an instance of that process - after all, this simply tells the DA process 'where' to look for each of the alignment stars in the two images. It does NOT (if I remember correctly) store the actual 'transformation matrix' to allow the transformation to be applied 'ad hoc'.

As I said, one obvious solution is to 'burn-in' some artificial alignment 'stars' onto the image, and then to align the data using those reference points. But that leaves the original data 'permanently scarred', which is not how we like to do things in PI, is it? Yes, the ability to use 'layers' would be a perfect workaround in this case - but not at the moment.

Can I use the 'alpha channel' to perform the alignment?

Sure, if I was working with a mono image (I am not) then I could add a 'layer channel' into - for example - the G channel, with my mono data in the R channel, and then I would just delete the G channel when finished.

Any thoughts anybody?

Cheers,

[Carlos Milovic]

Two answers:

- FFTRegistration: Try this first. It uses correlation on the fourier space to register the images, and the current implementation makes no star finding prefiltering, so you may use it directly. I used it to align the solar eclipse images (although it failed in some cases).

- DynamicAligment: If there is distortions, you may use bright and dark high frecuency features to align the images (I remember that there is a "invert sample" feature). There are some people that uses DA to register lunar images, so it is doable.

[Andres.Pozo]

- FFTRegistration: Try this first. It uses correlation on the fourier space to register the images, and the current implementation makes no star finding prefiltering, so you may use it directly. I used it to align the solar eclipse images (although it failed in some cases).

In my eclipse images, FFTRegistration failed in nearly all the images. I took five groups of five images with different exposure. FFTRegistration cannot align neither the images with the same exposure nor the images with different exposure. Registax also can not do it. I am at a loss about what to do.

[Carlos Milovic]

On the last eclipse (Rapa Nui) there was a bright star on the field (left on the images I processed... maybe a planet?). Since the frames had almost no rotation, only a shift, I used that single star with DynamicAligment. This worked pretty well.

FFTRegistration failed with some images, but in general produced fine results. The problem is that it registered the lunar disc, which moved across the different frames, so at the end the star registration was better (the solar corona is almost fixed in the background).