Offline cursive script recognition of bank cheque amounts

Abstract

A method for offline recognition of the cursive legal amounts on bank cheques is proposed. This application presents a multi-scriptor environment with limited vocabulary, in view of which a global approach is employed. The process is treated from a communications perspective. After each word is isolated, logical filters extract salient features, emulating the reading process of a fast human reader. Using its positional data as magnitude, a vector component is created for each feature detected, and a word vector is hence synthesized. Statistical analysis of these vectors enables a multi-scriptor lexicon to be compiled from the training samples. An approximate Bayes' metric is then employed to map unknown word vectors to their most probable outcome value.