Recognizing Handwritten Amounts from Brazilian Checks

Abstract

This paper discusses automated reading and processing of Brazilian bank checks. It highlights the main characteristics of Brazilian checks and the variety of writing styles that makes recognition more difficult than in countries where handwriting amounts are more uniform. Neural networks are used to recognize isolated digits obtained after segmentation. One set of neural networks was trained with digits from the NIST database and then evaluated using a database of numerical digits from Brazil. Other neural networks were trained with digits from both the NIST database and Brazilian digit samples extracted from real checks. These two approaches were utilized to determine the relative importance of the training set on the final accuracy levels. The same strategy is also applied to deal with other aspects of reading handwritten amounts in Brazilian checks such as delimiters and connected digits.