A dozen companies, including AT&T Co, IBM Corp, Litton Industries Inc, TRW Inc and Westinghouse Electric Corp are in competition for a contract with the US Postal Service to develop a computerised handwritten address recognition system. Also bidding is the AEG AG of Germany, which already supplies about 75% of the scanning and sorting equipment currently in Post Office use, the New York Times reports. Sorting handwritten mail is one of the service’s most expensive and tedious tasks, and it accounts for some 20% of the daily volume (555m items) received. It is hoped that by 1995 a system will be in place capable of deciphering at least half of this load. This would be a major advance on the equipment presently in use in many big post offices which can read only typewritten or machine-printed envelopes. It would have applications too for any business involved in sifting large volumes of mail. The winning formula will employ a combination of advanced optical scanning and artificial intelligence technology. A pilot system has been developed by the Post Office’s own researchers at the Center of Excellence For Document Analysis and Recognition in New York. This can read enough of a handwritten address to assist with automated sorting, but can process letters at the rate of one per second which falls well below the Post Office’s standard rate of 13 per second. The researchers’ goal will be to develop a system capable of recognizing enough of an address to place a bar code on the envelope. This will represent the nine digits of the ZIP code plus the last two numbers of the street address.
Neural network
They must also bear cost in mind as the Post Office is keen to minimise hardware expenses. Teaching computers to recognise patterns in different styles is a challenge as they can only recognise exact replicas of images stored in memory. Thousands of written and printed forms have to be analysed in order to identify the eccentricities and commonalities necessary for the computer to interpret effectively; a further device is then needed to help the computer to set characters in context. Even the latest pen-based computers, which enable users to input information by writing with an electronic stylus instead a keyboard, have problems with interpretation, although they have an advantage in being able to follow the writer’s sequence of strokes. They are also at a disadvantage in terms of speed since a human hand produces only a few characters per second, and an effective postal recognition system would need to work through around 1,000. AT&T’s solution to this problem involves a neural network of thousands of tiny algorithms that extracts features of known characters, and then further classifies them against algorithms of known combinations of features. The system eventually chosen may well be combined with another innovation recently tested in Tampa, Florida. At present, the mail rejected by automated sorting systems is channelled off for coding by postal workers at keyboards. The new method involves the scanning of mail illegible to the machine and the transmission of its image by video to a keyboard operator either at the same post office or at an office hundreds of miles away. The keyboard operator then codes the video image to which the network itself applies the appropriate bar code. Given that the cost of automated sorting of 1,000 letters is around $3, as opposed to the $40 per 1,000 for hand-sorting, there are distinct economic advantages in making maximum use of technology. Even so, the Post Office does not expect to eliminate humans entirely from the process and is demanding a 50% success rate for its first scrawl-decipherers. The first round of contracts will be issued over the next few weeks. This will reduce the number of bidders to four, a figure which will probably be halved again next year.
