Neural network based text recognition for engineering drawing conversion

Telephone companies and other utilities have a large number of engineering records stored on paper as tabular drawings. The author presents a neural network based system for the recognition of drafted or handprinted text in scanned images of the tabular engineering drawings. The system consists of a table structure interpretation module and a text string recognition module. The structure interpretation module uses a-priori knowledge about the structure of tabular drawings. The text recognition module uses a single character feedforward classifier that is trained using high-order discriminant analysis and error backpropagation, and a text string interpreter that uses pre-defined grammars for table entries