Abstract

This thesis presents a study to process Arabic handwritten cheques. It includes the development of a unique set of databases that constitute a solid base for research in this domain. The databases are unique in terms of their source, domain and tags validation process. First, they are originated from real-world bank cheques, which, to the best of our knowledge, has never been reached in a university setting. Second, it constitutes the only databases in the domain of Arabic handwritten cheques so far. To the best of our knowledge, there is no database that provides training and testing samples for Arabic cheques. Third, it involved a unique tagging validation process that takes advantage of the embedded redundancy in the format of the cheque to verify the tagging process. A grammar to validate Arabic legal amounts and translate them to numerical values is also included. This work includes an efficient method to derive one-dimensional feature sequence that preserves the dynamics of the original two-dimensional images. This is very important to accommodate small variations while modeling two-dimensional signals. The thesis provides a detailed description of an improved graph representation of sub-word images, a more efficient method to extract dynamic information from two-dimensional images and a clear positioning of the applicability of other curve-ordering criteria, e.g. vision rules. In addition, this thesis includes a significant improvement in the discrimination power of HMM which allows the differentiation between short sub-words and longer ones that share significant initial observation sequences. It also allows the HMM to properly classify incomplete observation sequences. The improvement is achieved by introducing a new parameter to the HMM called the termination probability. Included in this work are tests that prove the applicability and efficiency of the above contributions. At the time of this dissertation, our survey indicates that this work is the only research in the literature which handles images of handwritten sub-words extracted from Arabic cheques. The results of this study show a 94.36% sub-word recognition rate on the top 10 choices. Error analysis indicates some errors caused by the pre-processing (48%), feature extraction (28%) and classification (24%) modules