Abstract

This thesis presents the first automatic date processing system developed on a Canadian real-life standard cheque database. This system can process unconstrained handwritten dates written in English or in French, and it can also be applied to the recognition of any handwritten dates with similar format on many other kinds of documents. A knowledge-based module has been proposed for the date segmentation and a new cursive month word recognition system has also been implemented based on a combination of classifiers. The interaction between the segmentation and recognition stages has been properly established by using a multi-hypotheses generation and evaluation module. In addition, a verification module with two levels is designed in the postprocessing stage to correct some errors and reject invalid results, which further improves the reliability of the system. The segmentation of the date zone can be implemented in the knowledge-based segmentation module, the multi-hypotheses generation and evaluation module, or the verification module. An effective neural network ensemble system is proposed in this knowledge extraction stage to differentiate handwritten alphabetic words from numeric strings (A/N). We investigate the use of effective features extensively, and propose several new methods in the design of neural networks, creation of neural network ensembles, and combination methods for the ensembles created. For date recognition, the new cursive month word recognizer is implemented by combining a Hidden Markov Model classifier (HMM) with two Multi-Layer Perceptron (MLP) classifiers