[X] My Assistant

[Brian Burns]

Hello,

I'm a bit new here - I recently started on a project which I thought would be fun, to assemble movies of all the Voyager flybys. Of course, with 70k+ images, it would have to be automated as much as possible, and the results would be fairly crude. But there's enough information to make some rough black and white and color movies with one segment per target, and eventually combine them all into one movie.

The main task is getting the images properly centered, since the cameras don't point right at the target - there are still some jitters at the moment, so it needs some more stabilization. But in any case, here are some slightly bumpy rides along with Voyager - these are from versions 0.3-0.34 -

Ariel (Uranus)
https://www.youtube.com/watch?v=HeGgH34v8R4

Uranus
https://www.youtube.com/watch?v=0RLGadmvc40

Neptune
https://www.youtube.com/watch?v=HlU2_NOgoew

Assembling a color movie is a matter of combining the images that seem to go together (since it's not explicit in the data) - this is a first attempt at that. There are frames that lack certain channels, so colors are pretty off at points - later it could borrow a nearby image when a channel is missing.

Neptune, colorized
https://www.youtube.com/watch?v=c8O2BKqM0Qc

The goal is to generate the movies with as little manual annotation as possible - there are parts that need it though, like handling images with multiple targets, slowing down the movie at certain points, and turning off the centering at closest approach.

Anyway, just wanted to share what I've got so far - hopefully one day it will be able to make a nice, stable movie of all the flybys, with some appropriate music. Once the code is a little more stable I'll post the project on GitHub - it's written in Python with OpenCV, SciPy, and IMG2PNG.

-Brian

[Brian Burns]

I've got the centering and stabilization working fairly well now - here's a clip of the Voyager 1 Io approach (still working on the colors) -

http://imgur.com/a/8YWlS

The reseau removal marks from the CALIB images look like volcanoes going off all over the surface - those will need to be infilled from the prior frame where possible.

It uses a Hough circle detector (or a blob detector if that fails) then ECC maximization to a disc of the expected target size. It's fairly robust to noise and missing information - e.g. here are a couple of troublesome frames with the expected target size in yellow (based on the SPICE data) -

http://imgur.com/a/YTIlv
http://imgur.com/a/zy90J

One thing I'm not sure how to do is adjust the brightness of the images, as the CALIB images are so dark - at the moment I'm just doing histogram stretching on them, but that might be causing some of the brightness variations and odd color combinations.

---

The complete pipeline for the movies will look something like this -

Download - download archives from PDS
Unzip - decompress archive volumes
Convert - convert RAW imgs to pngs
Rotate - rotate 180
Flatfield - subtract good flatfields
Denoise - identify/eliminate noise where possible
Adjust - brighten
Center - center and stabilize images - first pass
Composite - align and combine channels using auto+manual info
Mosaic - align and combine images using auto+manual info
Dereseau - remove reseau marks cleanly - infill from prior frame where possible
Inpaint - fill in missing areas with pixels from prior frame
Undistort - geometric correction 800x800->1000x1000
Stabilize - find optimal solution - remove minor jitters
Annotate - add optional caption information
Clips - combine images into movies, one per target
Movies - combine clips into movies, add music

I was initially thinking it would need a crowdsourced type project to clean up the images, but I've come across 10 or so different types of noise that might be able to be removed or corrected automatically - then the inpainting step would try to fill in any gaps. Not sure how well it will work yet though.

The parts that would need to be crowdsourced are mainly the composites and mosaics, as there are so many of them, but some approximate solutions could be found using image alignment, then users could tweak the alignments.

There's no step in there yet for target spinning corrections, or position geometry corrections, or rotation problems (e.g. Voyager 1 Jupiter wide angle approach wobbles back and forth).

I get the feeling a complete movie might be an hour long or so by the end, unless some things are cut down, like the Jupiter cloud tops. But I guess you could produce a couple of movies - a short one with the highlights and a long one with everything, so I'll try doing that.