Training neural networks for reading handwritten amounts on checks

Abstract

While reading handwritten text accurately is a difficult task for computers, the conversion of handwritten papers into digital format is necessary for automatic processing. Since most bank checks are handwritten, the number of checks is very high, and manual processing involves significant expenses, many banks are interested in systems that can read check automatically. This work presents several approaches to improve the accuracy of neural networks used to read unconstrained numerals in the courtesy amount field of bank checks.