Path: utzoo!attcan!uunet!lll-winken!lll-tis!helios.ee.lbl.gov!pasteur!ames!pioneer!eugene
From: eugene@pioneer.arpa (Eugene N. Miya)
Newsgroups: comp.graphics
Subject: Re: Optical Character Recognition software?
Message-ID: <10160@ames.arc.nasa.gov>
Date: 11 Jun 88 02:44:31 GMT
References: <367@msn006.misemi> <11390012@hpldola.HP.COM>
Sender: usenet@ames.arc.nasa.gov
Reply-To: eugene@pioneer.UUCP (Eugene N. Miya)
Organization: NASA Ames Research Center, Moffett Field, Calif.
Lines: 78

This is a brief summary of my experiences in the OCR field up to today
when I had yet another Demo of OCR systems.  I've been asked to look at
one more and will report on it some time.

But first: My background: I've worked off and on for about 10 years
in digital image processing for remote sensing and computer graphics.
No, no, significant paper, but I understand the problems of IP, and I
know how hard OCR is.  So far I have tried the following
systems:
	DEST /DataCopy(first recommended by one of our Cray people, in fact, 
		one of those friend of a friend things)
	DataCopy
	1215 Terra Bella
	Mountain View, CA 94043
	
	Kurzweil (A Xerox Corporation, best know for all kinds of
		interesting equipment)
	185 Albany St.
	Cambridge, MA 02139
and today:
	Transimage
	910 Benicia Ave
	Sunnyvale, CA 94086
	(408)-733-411

You normally can't just contact these companies, they refer you to
places you can demo units.  (Like Western Office Supply).
They range in price from $3,000 for H/W and S/W to $20,000.
The first two are sheet readers and the Trans image is hand held.
The first two are 300 dot per inch resolution.  The DEST/Datacopy
uses fixed font information which keeps it cheap.  The recognition
miss rate on say variable space laser printer stuff is not helpful.
It's designed for reading fixed size typewriter faces.

The Kurzweil is a system wich requires training.  It takes several
passes.  It can handle variable spaced text.  Both I felt were
not worth the cost, I think the Kuzweil was something like $7,000
and getting cheaper.  Neither is super fast, but it was the error
rate which turned me off.  You have to deal with false positives
as well as true negatives (i.e., is that a blot or a period?)
I already mentioned the ells and 1s, zeroes and ohs.

Context is a very difficult thing to deal with.

Oh, I should say something about CPUs, all run using IBM PC clones
ATs or XTs, no PS2/OS2 support, and the DEST is the only Mac
connectable.  Well as of today I was most impressed with the
TransImage, but I suspect this is because it has 1,000 dots per inch
res.  Hand held is initially novel, but sucks in the long run.
(PC based, BTW).  It did my test sheet fairly well (including
special characters).

Now the reason why I posted this detailed explanation.

SO I can read text in with some rate of errors, yes it something of a
pain but it's only $3K.  The question is how to quantity the recognition
rate.  If I have to change 1 character (or edit) for every 100, is this
too high (able one per line?  Well the DEST/Kur* were 4-7 errors
per line on my sample text, much too high.  Why, my typo rate is
one per line and this is about what the TI1000 is.  Now a page is about
60 lines.  Should a the document just be handed to the secretary,
here type this? or should we go OCR?  I have to figure out error rate
per 1000 or so characters, and in the case of the TI1000 factor
time to scan as well.  Comments on error rate measurements?

Note: only a bureaucratic Agency or company really needs one of these
things.  I think WE can blow away $3K of taxplayers money to try it.
Like 90 page documents, etc.

Next Pallentair(sp).  $30K Tak-eye!  [sorry can't spell Japanese
phoentically well].

Another gross generalization from

--eugene miya, NASA Ames Research Center, eugene@aurora.arc.nasa.gov
  resident cynic at the Rock of Ages Home for Retired Hackers:
  "Mailers?! HA!", "If my mail does not reach you, please accept my apology."
  {uunet,hplabs,ncar,decwrl,allegra,tektronix}!ames!aurora!eugene
  "Send mail, avoid follow-ups.  If enough, I'll summarize."