loren@tristan.llnl.gov (Loren Petrich) (07/14/90)
I'm planning a project to use a Neural Net or something similar to recognize overlapped star images and the like on an photographic or CCD image. I plan a number of things to simplify this task. The first is to have some way of "pointing out" images. This would be done by taking all the region of enhanced brightness around each peak. This will result in an image with much fewer pixels (one hopes!) and should be much more convenient for a pattern recognizer. I am planning further pre-processing of an image: one would be to rotate it until it has a specified orientation. This would be with the help of a moment expansion; the long axis would be rotated around to some specified orientation. And I am also planning some Principal Components Analysis. That would be to solve the eigenproblem imsq.mask = eigenvalue*mask where imsq = sum(parms) image(parms)*image(parms) One would select out the masks with the highest eigenvalues; these would select out the "components" of the image with the highest variation, and thus considerably lessen the load on the Neural Net. Has anyone else tried tricks like the ones I describe (the PCA trick is NOT original with me, I should say)? ^ Loren Petrich, the Master Blaster \ ^ / loren@sunlight.llnl.gov \ ^ / One may need to route through any of: \^/ <<<<<<<<+>>>>>>>> lll-lcc.llnl.gov /v\ lll-crg.llnl.gov / v \ star.stanford.edu / v \ v For example, use: loren%sunlight.llnl.gov@star.stanford.edu My sister is a Communist for Reagan