Abstract

A segmentation based courtesy amount recognition (CAR) system is presented in this thesis. This system consists of four modules, which are preprocessing, segmentation, recognition and syntactical analysis. The noise is first removed in the preprocessing module. Then, the segmentation module seeks to extract characters (digits, punctuation marks and so on) from the numeral strings. The recognition module aims at classifying each pattern into one of the ten numerals ('0'-'9') and double zeros ('00'), rejecting ambiguous patterns. Finally, the syntactical analysis module parses the recognition results to provide an acceptable courtesy amount. Two sets of features, shape features and spatial features, are developed to locate the punctuation marks. In addition, a hypotheses-then-verification approach for segmentation of the numeral strings is introduced. The emphasis is on finding all possible segmentation hypotheses and then verify them. A two-level segmentation module is proposed, namely the global segmentation level and the local segmentation level. At the global segmentation level, the intent is to find the potential broken numerals and group them together. On the contrary, the local segmentation level seeks to split the touching digits. Two classifiers are combined into the recognition module. The isolated digit classifier divides the input patterns into ten numerals ('0'-'9'), while the holistic double zeros classifier intends to recognize the cursive and touching double zeros as an atomic symbol. The proposed courtesy amount recognition system has been trained on the database of CENPARMI checks and tested on the database of real checks. The system reads 66.5% real checks correctly at 0% misreading rate