In the United States alone, there are literally billions of bank checks written every year. The processing of these documents in a timely fashion is both expensive and labor intensive. Magnetic Ink Character Recognition (MICR) has dramatically improved the efficiency of the check clearing operation. However, this technology still requires a great deal of manual effort. The recognition and encoding of check amounts into the MICR format is still largely performed by humans.; Much research has gone into the development of systems to automate the recognition of amounts on bank checks. While considerable progress has been made, these efforts have not yet yielded accuracy levels comparable to those of humans. Consequently, they have been of little practical value, as the cost of making mistakes in this domain is very high.; This research proposes an adaptive writer dependent solution to the amount recognition problem. From the beginning, the primary design goal was the minimization of the error rate. To this end, several verification techniques were employed. This system attempts to read both the legal amount and the courtesy amount, thus allowing cross verification from independent sources.; This solution employs multiple parallel classifiers for the discrete recognition of both cursive script words and digits. Various decision fusion methods are then employed to arrive at decision consensus. Syntactic validation provides additional confidence in the parsing stage of the legal amounts.; This work demonstrates the feasibility of a writer dependent automated amount recognition system with a rejection rate as low as 25.97%, and an error rate that is comparable to that of humans.
展开▼
