Night Sky SuperResolution (instead of star trails) - possible?

Question

I don't know if this is possible, or runs into fundamental limits of sensors and photo sharpness.

I want to "pretend" that I have a fancy night sky motorized rig - through software.

1. Take a few thousand 5-second exposure nigh sky pictures with a wide angle consumer camera (laying perfectly still)
2. Align the images AND correct for lens distortion.
3. Apply Superresolution to stack the aligned images.
4. Profit!

I've had great success making star trails using StarStackX - no problem - but that was a straight stack-and-blend with no alignment.

I can also extract the location of the various star trails using some very brain-dead math and lots of laptop CPU time. Which feels like it should be enough to do a pretty good model of the lens distortion + atmosphere distortion... or maybe no "model", and instead a straight transform.

Hugin and PTGui looked promising, but barfs on the number of control points I'm feeding to it.

ALTERNATIVELY - it may be impossible. If you imagine inverting the system, it would be like putting a cheap camera on a slow motorized mount, letting it take 3,000 nearly completely overlapping photos of a landscape, and then telling it "make one huge high-resolution panorama"... which AFAIK would be nearly impossible to do. I could align the photos, but I'm unaware of any stitching software that outputs significantly improved superresolution results. I'm hoping the nearly point-like light sources of stars isn't quite as hard.

Answer 1

Superresolution techniques require pretty good source data to start with, and that source data usually needs slight offsets between each frame (dithering.) Without a tracking mount, you will see field rotation in the corners of the frames, and that will greatly diminish your ability to align and stack, let alone apply superresolution. Distortion mapping can be done to correct distortion among frames, however that is algorithmically complex, and I only know of one software package that really does that well.

Your looking to surmount a number of very high hurdles to produce "superresolution wide field" astrophotography without a mount. You could get significantly better results, and actually make the concept viable, if you just add a tracking mount. That would eliminate most problems (assuming no landscape was visible in the field). For superresolution ("drizzling", as it's usually called in the astro world) to be effective, you need to dither between each sub frame, and you need LOTS of sub frames. To gain an increase in resolution, at the cost of noise, you can drizzle 20-30 sub frames of the same region of sky, and that would improve resolution. Drizzling works in such a way that it trades off the noise reduction you get with straight stacking to improve resolution. To get both resolution enhancement and noise reduction via stacking, you need more on the order of 200 sub frames...of the same region of sky.

Now, if you are looking to do a high resolution, low noise panoramic view of the night sky (milky way), then your probably looking at getting multiple panels of overlapping regions of the sky, with about 200 sub frames each. If you want a four panel panorama, you're looking at 800 sub frames. Assuming you're going with 30 second subs, that would be nearly 7 hours of total exposure time. Ultra wide panoramas like this usually require a few extra panels to plug in the holes that occur as the panels are distortion corrected (warped) to fit properly with each other...so instead of four panels, you might actually need seven, which extends the exposure time into the 12 hour range, so multiple nights.

Drizzling can certainly improve the detail of a final image, but it is really not actually all that necessary in the end. With a tracking mount, you can get away with only using one frame per panel in a mosaic if you have dark skies (which are usually necessary for milky way imaging.) Drizzling is usually more effective for wide field DSO (which is much narrower than your average milky way image). With a medium-wide focal length of 50mm to 250mm, one frame per panel, and an effective tracking mount, you can get some very nice results:

Answer 2

The problem with wide angle lenses is that many of them have a mustache subfrequency in their bulb distortion that makes it impossible to stack the images. The image alignment that works in some parts of the frame does not work in others. In fact for this type of distortion, you will notice a smeared look on the image corners after only stacking a few photos. I think you can escape this with the Zeiss lenses but they are a bit more expensive. From my experiences if you are trying to get a wide angle shot in one image, you can get very good results by using a full frame camera at iso 6,400 with a single shot and little noticeable noise. I have done very will with 24mm lenses at f/2.8 with 15 second exposures and unless you zoom in a long ways, you cannot notice the star drift. Dithering is also another technique that can be tried, but again you run into the stacking problems which can be tricky with wide angle lenses.

Answer 3

Sounds like a great idea to me, but the proof is in the software. You'd put the camera on a tripod, take your whole lotta pictures, and then hand it to the software. The software would:

• Pick some bright stars and track them across images to do loose mutual alignment
• Determine the tracking parameters (e.g. where's the North Pole, what the field of view was, etc)
• Now, assume the appearance is constant, you're just moving the camera: do the superresolution calculations.

Given that almost all celestial objects would be point sources, tracking the points across the images would give you precise motion of the camera versus the point source. Once you had this for all of the objects across all the frames, you could feed that back into more precise camera positioning, and then get even better celestial object location.

I'm guessing you'd need to use astrophotography-specific software, because a solution for a field of point sources is likely to be different from a solution for areas of shading.

Kickstarter, anyone?