Technology developed at IBM laboratories on three continents is helping
the company to gain a foothold in the emerging worldwide market of postal
sorting
In Brief:
Responding to a business opportunity offered by the global growth in
mail-order items, IBM laboratories are developing technologies for automatic
sorting of mail. Optical character recognition technology, video imaging
methods, fuzzy search engines, parallel computing techniques and specialized
hardware have combined to build production machines for sorting flats thin
packages larger than conventional letters and prototype technology for sorting
parcels. Continuing research will ensure IBM a niche in one of the worlds
largest retail markets.
Sorting mail is a task that machines do more effectively than humans,
because automatic readers can scan destination addresses far faster than even
the most skilled individuals. Post offices around the world have used such
technology for more than 20 years to sort letters, the simplest type of mail to
scan.
Recently, spectacular growth in the mail-order business has sparked interest
in automatic sorting for two other types of mail: thin packages, called flats,
that are larger than the standard envelopes used for letters; and parcels.
Collaborative projects at IBM involving the Almaden and Haifa research
laboratories and the Boeblingen development laboratory have set the scene for
the company's entry into those potentially profitable segments in one of the
worlds largest retail markets. Overseeing the young business is the World Wide
Mail Automation (WWMA) group, based in Bethesda, Maryland, and managed by Dave
Bausch, which IBM's government industry solutions unit set up in 1995.
A handful of companies, including AEG and Alcatel, dominate letter-sorting
technology. A few years ago, the same companies introduced sorters for flats.
But until recently, no company had developed means of automatically sorting
parcels. The reasons: prohibitive volumes of data and the 3-D nature of parcels.
The surfaces of flats typically contain 10 times as much information as letters,
and parcels 100 times as much. "The system has to scan the entire surface
and make the tough decision about which data is the destination address,"
explains Ilan Efrat, manager of parcel and mail sorting systems at the Haifa
lab.
The key to automatic reading is optical character recognition (OCR). A team
at Almaden headed by Moidin Mohiuddin has developed OCR technology that
progressively breaks up addresses into lines, individual words, characters, and
finally a set of features representing each character. The latter are fed to a
neural network program that outputs the most probable ASCII code for each
character.
At that point, a software system called CheckMate comes into play. It first
recognizes the zip code on the address. ãCheckMate is a highly flexible
system that can be adapted to different address formats and national address
styles," explains Mohiuddin.
Once CheckMate has identified the basic zip code, it searches the address
block for information such as the street address that will enable it to
consign the mail piece to a postal route. It does so by conducting a ãfuzzy
search" through a database of addresses to identify the one that most
closely resembles that found by the OCR. That type of search even permits the
system to handle incorrect zip codes.
The initial version of Almaden's technology applied to printed addresses and
those handwritten in block letters. Recently, under a First-of-a-Kind solution
project the team has extended its OCR technology to cursive script. In February,
Mohiuddin's group demonstrated the ability to recognize cursive 5-digit zip
codes, along with city and state names. Now, the team is incorporating the new
technology into CheckMate, to provide a system that can route a satisfactory
percentage of cursive addresses to the correct postal routes.
The technology has found application in sorting flats. Late in 1994, Mueller
Martini, a Swiss company that specializes in feeders for newspapers, decided to
try to gain a foothold in the market for sorting and handling flats. The company
approached IBM Germanys development laboratory in Boeblingen, near Stuttgart,
seeking the necessary technology.
Boeblingen developers incorporated Almaden's OCR and CheckMate into a sorting
system that uses parallel processing and specialized hardware designed at
Boeblingen. A line camera placed beside the conveyor belt carrying the flats
films an image of each flat, which it passes to a network of Motorola VME cards.
Each image is sent to one card, where Almaden technology is used to identify the
extended zip code and ensure the proper routing of the flat as it leaves the
conveyor belt.
The Boeblingen group had a feasibility model of its flats sorter running by
May 1995. Impressed, Swiss Post ordered a prototype machine for its Zurich
office, which Mueller Martini installed last September. Then, in December, says
Manfred Vodegel, who heads the Boeblingen team, ãIBM and Mueller Martini
closed a contract for another four machines after a competitive runoff against
the incumbent supplier."
A tougher problem
Parcels present a tougher problem for address recognition devices for two
reasons. Many contain several blocks of information for example, the senders
address, company logos, bar codes and itemized lists of contents as well as the
destination address. "You cant assume the address is always in the middle,"
says Efrat. In addition, parcels have depth, as well as horizontal dimensions.
That severely complicates the process of imaging their data blocks.
To overcome those difficulties, the Haifa team has devised a prototype
system that incorporates significant additions to the technology developed for
flats. To image data on the surface of boxes of different heights, the scanner
has a depth of field adjustable from 0 to 800 millimeters (0 to 31.5 inches).
Two precision motors adjust the lens focus and the angle of the slanted
illumination according to the measured parcel height.
To pick out the destination address, the system uses software that includes
several related searching procedures. The simplest focuses on keywords. "TO"
plainly reveals a destination address, while "FROM" indicates
something other than a destination. Another technique, batch handling, aims to
identify consignments of several items of merchandise passing through the system
in quick succession a shipment of televisions, for example. The technique first
identifies common logos or originating addresses. Then it seeks the location of
text blocks that are different for each of the items presumably the destination
addresses.
Once the address has been identified, it is stored in a database, using the
identification code as a key. Any other system in the postal environment that
can read the identification code can retrieve the relevant sorting information
from the database. The Haifa technology can also read bar codes, a necessary
ability because some senders now affix their own bar codes to parcels and flats
before dropping them off.
One of the key advantages of IBM's solution is that the core technology can
be applied to letters, flats and parcels. Combined with IBM's unique
image-capturing system and address block location routines, the address
recognition engine offers customers a powerful and versatile solution. WWMA is
actively engaged worldwide in bidding this solution in response to tenders from
international postal agencies, as well as commercial shippers.
