GRAPHICAL-DATA-PROCESSING RESEARCH STUDY AND EXPERIMENTAL INVESTIGATION

Peter E. Hart, John H. Munson · 1967

The report describes the continuing development of scanning, preprocessing, character-classification, and context-analysis techniques for hand-printed text, such as computer coding sheets in the FORTRAN language. The performance of topological feature extraction, combined with character classification by a learning machine, is described, and compared with the performance of other combinations. By performing intra-author testing (gathering the training data and test data from the same author), we have achieved a dramatic reduction in test error rate, to less than 10 percent on a limited sample. We describe an experiment in which a fragment of FORTRAN text is scanned, pre-processed, classified character by character, and subjected to context analysis to greatly reduce the recognition-error rate. Finally, we discuss advances in the method of analyzing arithmetic expressions, a key aspect of the context analysis.

Cited by 36 works. doi:10.21236/ad0661309

Authors