2007/09/24

ADASS XVII: Automatic Image Registration for the Future (W. Hack)

Warren Hack is a Sr. Software Developer at STScI.

Image registration is on the primary problems facing new archives of astronomical data
Image registration is assumed to be "relative astrometry". Astrometry.net is the next step after this registration process.

Scope of the Problem. Need to combine multiple exposures into one and allow astronomers (HLA is doing this, isn't it?)

Algorithm Research: what's available?
  • 2 papers since 2004 deal with algorithm development in the areas of image registration
  • 5 papers on planetary and geo-science observations
  • hundreds of papers from the medical field.
  • a reference paper is 2003, Image and Vision Computing, 21, 977
    • Area based
    • Feature based: DOAFind, SExtractor
  • in astronomy a few techniques:
    • cross correlation (area based)
      • does not work well in non aligned images, or on variability
    • catalog matching (feature based)
      • not work on extended sources,
      • accuracy relies on position of sources
      • manual tuning is required to work well
    • multi-resolution (wavelets)
      • cosmic-rays are difficult to remove
      • accuracy when working with large sources is limited

We'd like something new. We need to combine both approaches and seeking consensus.
Over the last 2 years I have taken an integrated approach: MIRA, the multiresolution Image Registration Algorithm.
  • MIRA relies on a combination of factor
  1. It determines the relative astrometry by reading WCS
  2. Generate multi-resolution views
  3. Detect and identify source in the image
  4. Compute distance matrix
  5. Select matching pairs
  6. Solve ---> iterate
How is image registration used? We create a media filter from all images. Smoothing images can be effective at determining the object centers for further processing.
This multi-resolution method propagates the source identification thru all resolutions.

Once can also use Contour and Edge detection to compare image. Reference: On Geoscience and Remote Sensing 37,2351

Chain encoding is then used to get a representation of an object which is rotation and scale independent.

We also use invariant moments to characterize the source: scale invariance, rotation, etc.. (the usual)

Implementation of MIRA:
  • Once sources are identified, then we solve for the relative registration.
  • Tests were run on extended source and on point sources in crowded fields
    • Orion: offset a few pixels
    • 47 Tuc: offset few tens of pixels
    • M87: issue with the jet, but alignment was actual found
  • In all these test MIRA was run with no input parameters from the user!

MIRA has been tested with super-associations from CADC.
MIRA seems to be able to automatically do image registration, regarless of the type of image.
MIRA show that "cross-pollination" from other fields.