Abstract

Automatic processing of bank cheques has been investigated for ten years at CENter for PAttern Recognition and Machine Intelligence (CENPARMI). Topics studied include the recognition of legal amount, courtesy amount and date. In the CENPARMI Cheque Processing System, sentence-to-word segmentation has to be done before the recognition of legal amount. Segmentation serves as the basis of recognition. However, the previous segmentation module is hindered by both under-segmentation and over-segmentation. To overcome this problem, a feedback-based approach is proposed. It is implemented in two steps. In the first step, segmentation is done according to the structural features found between the connected components in the legal amount. In the second step, a feedback process is introduced to re-segment the parts that could not be identified in the first step. Subsequently, the classifier is used to verify the re-segmentation result. The confidence value produced by the classifier is used to determine the best segmentation points. Two neural network classifiers have been introduced into the segmentation process. Their performances are compared. This approach is tested on a CENPARMI database with 680 images and the result indicates that the new approach has increased the correct segmentation rate by 15.0% from the previous approach. After segmentation, the legal amounts are recognized by a neural network classifier and a Hidden Markov Model (HMM) and a Multi-Layer Perceptron (MLP) hybrid classifier. With the new segmentation method, the performances of the two classifiers have increased by 9.7% and 8.9% respectively