The following is a press background fact sheet on handwriting recognition distributed by IBM's Advanced Systems Development Division on September 15, 1964.


Experimental optical scanners capable of recognizing handwritten numbers have been developed and demonstrated by International Business Machines Corporation. These machines make it possible to enter handwritten information directly into computer systems.

For more than half a century, most source information entering data processing systems has required manual transcription. This is the slowest and most error-prone step in a modern data processing operation.

In applications where some control can be exerted over the preparation of source documents, and where an immediate response is not required from the computer, optical scanning of data recorded on documents at the source promises to improve the speed and accuracy of data entry. One example of this type of an application is a sales check prepared by a department store clerk.

Perhaps the most economical method of recording data at low-volume sources would be to write it with ordinary pen or pencil and with minimum constraint so as not to inhibit performance of the writer's primary job. Because of the many variations in human handwriting, optical scanners have only been able to recognize simple pencil marks, and highly constrained handwriting.

Several years ago, Evon C. Greanias, an IBM engineer, began development of a scanning technique that would provide wide tolerance for common character shapes and for character positioning on documents. A "curve-following" approach, demonstrated in several experimental numeric handwriting readers, departs significantly from the scanning methods used in other character recognition machines for printed characters.

While other methods involve scanning with one relatively fixed routine, in the curve follower technique, the contours of each number are traced with a moving beam of light. As the beam moves across the area to be scanned, analog voltages corresponding to the shape of the handwritten number are generated and analyzed to detect the presence of important shape features.

This technique is flexible enough to allow wide variations in size, registration, orientation, and shape of the numerals, as long as they are written within the area on the document that corresponds to the scanner's optical field.

Scanning starts with a search pattern that looks for a number on the document. When a number is intercepted, the search pattern is stopped, and beam control logic causes the flying spot scanner to whirl around the edge of the number with a rapid, spiral motion.

To accomplish recognition, the contours of each number are traced two or more times by the small beam of light. The beam deflection signals are analyzed in terms of two properties of the line edges traced by the light: relative location within the character and approximate direction.

One special subroutine causes the beam to re-explore poor lines with a modified black-white threshold to accommodate weak writing. Another subroutine causes the beam to seek specific shape features in selected regions to resolve any ambiguities that exist after normal scanning.

When identity cannot be established by measurements of the character's outer contour, the scanner is directed to obtain more information from measurements inside the character.

The only constraints imposed by the scanner are that writers must avoid large embellishments such as curlicues and "tails," excessive slanting, large breaks in the lines that compose the number, writing too small or too large beyond a 4-to-1 range in height and running numerals together.

If erasures are made with reasonable care, they do not effect reading performance.

Several handwriting recognition projects have been completed or are underway in IBM's Advanced Systems Development Division. The reading capabilities of the first of these experimental handwriting scanners were tested at Tufts University, Medford, Mass., in 1962. Classroom experiments established the fact that people write numbers in a wide variety of ways. The performance of the scanner used at Tufts in recognizing this broad sampling of numbers was very encouraging and led to two field tests with later experimental models of the reader.

In an experiment which began at the Higbee Company department store in Cleveland, Ohio, in March 1964, another model of the scanner has been reading sales checks of customers' purchases in selected departments.

As customers make purchases at Higbee's, sales clerks write pertinent information on IBM card sales checks. This handwritten information indicates the quantity, department number, merchandise number, amount of the transaction, etc.

At the end of the day, the sales checks are collected from participating departments and scanned by the experimental reader, and the information is processed by an IBM 1401 data processing system. An important objective of this test is to study the feasibility of providing reports to store management without having to keypunch sales data into cards before processing.

Still another scanner is being demonstrated at the IBM Pavilion at the New York World's Fair.

The scanner system at the World's Fair reads dates written on cards by visitors and retrieves featured headlines published by The New York Times as far back as its first issue, September 18, 1851. The card bearing the handwritten date is placed in the scanner, and the recognized number is transmitted to an IBM 1460 data processing system in which the nearly 40,000 headlines from the Times are stored. After identification, the headline is flashed on an overhead display and printed on a souvenir card for the visitor.

Through field tests such as these, numeric handwriting scanners are providing data needed to establish the kind of performance and flexibility required for a practical data entry device.

Return to exhibit