COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.require "memo.pub[let,jmc]" source
C00011 ENDMK
C⊗;
.require "memo.pub[let,jmc]" source;
.cb OPTICAL CHARACTER READING IS STILL AN UNSOLVED PROBLEM

	For many important applications optical character reading
is still an unsolved problem.  These include the reading of books,
magazines and newspapers in uncontrolled fonts.
With the advent of the newest generation of disk files, typified
by the IBM 3830 that stores 2.52 gigabytes per unit, it has become
economical to make a national computer-based
library that would store everything that has ever been printed and
make its text available from any suitable terminal in the country 
over the telephone lines.  Besides this, social science research
requires the ability to read large quantities of text into computers
for content analysis, and surely the CIA, whose function requires
monitoring thousands of foreign publications could use the capacity.
Another application is that military, government and business
organizations often receive correspondence that needs to be
redirected promptly.  If it could be read into a computer form,
it could be redirected electronically.  Even more important,
once it is in a file system that can be searched by computer, it
is much more likely to be found when it is relevant.
One of the advantages of ARPAnet messages is that one can search
one's files for a message menioning a certain subject or person.

	The history of the problem is as follows:

	In the 1950s there was much research in optical character
recognition supported by many government agencies.  Some promising
results were obtained, and commercial systems that could economically
read special "ocr fonts" were developed.  These machines have been
in widespread use since that time, but depend entirely on the fact
that the fonts and the manner of printing them are designed with
OCR in mind.

	Some time in the early 1960s, government research support
agencies decided to declare a victory in OCR and leave research
to the companies.  The victory celebration was premature.  %2Almost
twenty years later, there is
still no company that will take fixed price contracts to read
books into a computer%1.  There appears to be no current government
supported research, although there has recently been research in
holographic methods, supported in a search for applications of
holography.

	The main private research and development has been done
since the late 1960s by Information International Inc. of Culver
City, California.  This resulted in the early 1970s in a system
called Grafix I.  It is based on one of III's computerized
microfilm readers, a Digital Equipment Corporation PDP-10 computer
(KA-10 vintage), and a specially designed binary image processor.
The system includes terminals on which the system can display
gray-level images of rejected characters, so that "reject operators"
can key them in.  Two systems were sold some years ago, one to
the U.S. Navy Republication Facility, which uses it to republish
maintenance manuals for older aircraft and one to the British
Department of Health and Social Security, which uses it to read
hand-printed social security forms and is currently having the
system improved so as to read the card catalog of the British
Museum (the equivalent of our Libary of Congress).

	While Grafix I has read many books and magazines in several
languages with satisfactory error and reject rates.  A minimal system is
very expensive at α$2.5 million, although it is cost-effective for large
applications.  III is not currently motivated to attempt a lower cost
version without seeing more of a market.  Even Grafix I is cost-effective
compared to having the material retyped on keyboards, even if the work is
done overseas.  Cost-effectiveness does require keeping the system busy,
because a crew is needed to prepare material and process rejects.

	A few years ago the Kurzweil Company of Cambridge, Massachusetts
developed a widely publicized system selling for α$25,000
to enable the blind to read books in arbitrary fonts.  The system
depends heavily on the ability of the human to guess omitted or
mistaken characters and to recognize when the machine's output
doesn't make sense and try again.  At least in its present state,
the Kurzweil machine is inadequate to read books into a computer.

	We believe that the technical problems of making a
cost-effective character reading attachment for a computer are
solvable in less than five years with extremely high probability
of complete success.  However, it will require overcoming the
myth that the problem has already been solved and is unworthy
of scientific attention.

	A research and development program might include the
following:

	1. Requests for proposals to use present equipment - with III's in
house Grafix I as one obvious candidate, to read a substantial body of
material.  For example, keeping Grafix I busy for one shift for a year or
two with technical material in (say) computer science would shake down the
present algorithms.

	2. Announcement of a program that would be responsive
to unsolicited proposals from the universities and industry.
Perhaps there could be a Request for Proposals to develop a
vidicon based computer attachment that could read pages directly
from letters or books and which would not require microfilming
or a full-time staff.  Ideally, a secretary could take a letter
to a reading machine and read it into the office computer.

.<<paragraph on why now>>
.<<expensive as a whole though cheaper than people>>