🚀 go-pugleaf

In article <ALIEN.92Aug13121817@vulcan.essex.ac.uk> alien@essex.ac.uk ("Adrian F Clark") writes:
>While MEM -- and particularly the MEDC package -- seems to be pretty
>effective for astronomical imagery, I'd like to raise a small word of
>warning regarding MEM for general restoration applications.  MEM works
>by taking a uniform image and `pulling' non-uniformities out of it
>only where there is statistical evidence for doing so.  Unfortunately,
>the staistical models used by all the implementations of MEM that I've
>seen only work well for pulling bright regions from a dark
>backgrouond.  Given a image with the opposite contract (e.g., a
>bright-field electron micrograph) I've yet to see a good job.

The results from MaxEnt are only as good as your prior assumptions. In
astronomy, it is reasonable to say 'my prior expectation of what my ccd
measures is typically 10 photons/s, but may be from 1e-4 to 1e5 photons/s'.
This is obviously an assymetric distribution, because of the knowledge of
a minimum signal level. (i.e. there are no negative photons out there). It's
not invariant under inverting the contrast, and rightly so.

If you want such invariance, then it's easy to attain; either throw away
the positivity constraint if you don't know what the zero level is, or
introduce a maximum level if you have upper and lower bounds for you image.

The MEDC code supports both of those.


David Robinson. (drtr1@cus.cam.ac.uk)