[comp.ai] pattern recognition software

cs313s19@uhccux.UUCP (Mike Morton) (11/23/87)

A friend does research work spotting humpbacks by recognizing their
dorsal fins.  The researchers finish each day by comparing the day's
photos with 300-400 photos of known whales to recognize individuals.
They're looking for a way to do this with a computer database.

They could code the data and enter them as numbers: size and shape of
fins, etc.  Then the database just needs to search for close matches. 
This could be done with a simple Basic program or spreadsheet macro; any
suggestions for a turnkey system which does this?

Better, but presumably harder to find or implement, would be a graphics
recognition system, scanning images or allowing them to be traced by
hand and entered.  I doubt there's anything like this available off-the-
shelf, but would be interested to hear about it if there is.

Solutions for the Mac are especially of interest, but any micro is OK.
Please reply by email.  Thanks in advance.

 -- Mike Morton // P.O. Box 11378, Honolulu, HI  96878, (808) 456-8455 HST
      INTERNET: cs313s19@uhccux.uhcc.hawaii.edu
      UUCP:     {ihnp4,uunet,dcdwest,ucbvax}!sdcsvax!nosc!uhccux!cs313s19
      BITNET:   cs313s19%uhccux.uhcc.hawaii.edu@rutgers.edu

UH2@PSUVM.BITNET (Lee Sailer) (11/25/87)

I can think of some pretty good ways to do this, but not with
database software, unless the matching problem is really simple.
     
The current masters of *sequence matching* are the molecular biologists,
who spend a lot of time matching LONG sequences of RNA, DNA, etc.
     
One approach
     
Can the fins be described with a simple sequence of tokens or symbols, like
<big gap> <small notch> <small gap> <big notch> <tip> ?  If so, then you've
got the DWIM (do what I mean) or spelling correction problem.  Given a
sequence of symbols, find the set of legal sequences that are close.
This turns out to be a graph search.
     
Another approach
     
Are accurate measurements needed to distinguish nearly identical fins?
If so, then a fin must be described something like this:
     
  gap of 15.2mm
  notch width 5mm depth 3mm
  gap of 45 mm
  notch width 3mm depth 5mm
  tip
  etc etc etc
     
If you think of a 'gap' as a notch with width 0, and the tip as a notch
of width and depth 0, then each feature characterized by a triple
of real numbers.  Using the <start> <stop> and <tip> as
landmarks, it ought to be possible to think up some way to convert
each fin to a point in N-space, and then to compute the distance
between a new fin and the 300-400 fins already in the database.
     

David_Bat_Masterson@cup.portal.com (11/26/87)

This request sounds vaguely familiar.  I thought I had seen a show about
a few students for a college doing a study of humpback whales.  They
also were having trouble keeping track of which whales were which (maybe
it was killer whales).  The way they went about handling it was to
classify the dorsal fin shape by things like size, shape, bites, extra
spots, barnacles, etc. (their fingerprint).  I forget if they used a
database system to keep track of this or just a file card approach.  If
you use a DB, this information could be entered into a relational database
for scanning purposes (Dbase perhaps).  This would not provide an automatic
mechanism for processing the photographs, but its a start.  Additional ideas
would be to implement an expert system as front end to this process.  The
expert system could be trained to ask the right questions about a photograph
to get a good classification.  On top of this could be added a laser scanner
(for about $3K) that would bring the photo into the database; there may be
database systems that would allow you to store the image of the whale right in
the database (I know the Amiga databases can).  Think about it, you can build
up from a basic capability, but don't try to do the whole thing at once.
				David_Bat_Masterson@cup.portal.com

Bob_Robert_Brody@cup.portal.com (11/27/87)

There is an organization I belong to re Moclips Cetological Society
which is non profit and centered around whales and whale sightings
and cataloging.  Maybe they could be of help re using databases to
maintain the catalogs.  You can call 206 378-4710.

The Whale Museum
P.O. Box 945
Friday Harbor, Washington 98250

Moclips Cetological Society is a non profit research and educational
corporation.

Bob Brody
Los Angeles