| In this thesis, we have conducted research on off-line automatic recognition of Arabic handwritten text. Automatic recognition of Arabic handwritten text has many applications, like forms processing, postal address and zip code recognition, etc. Recognition performance of Arabic handwritten text lags far behind compared to Latin and Chinese scripts. Cursive nature of Arabic script, overlapping characters and ligatures, presence and variability of dots/diacritics, lack of benchmarking databases are among the reasons that make Arabic handwriting recognition a challenging task.;Research on Arabic handwriting recognition has mainly focused on the recognition of isolated characters, numerals and words. Few works have been reported on the recognition of Arabic handwritten text. Moreover, statistical methods (like Hidden Markov Models (HMM), Neural Networks, etc.) have been more common in the reported results on Arabic handwriting recognition. Structural methods have remained largely unexplored in this regard. In addition, some recent multi--classifier based systems have tried to integrate both statistical and structural techniques.;We present several novel techniques for different phases of Arabic handwriting recognition using a structural approach. We present three segmentation algorithms: segmentation of a page into lines, a line into words/sub--words and a word into characters/graphemes. In addition, we introduce polygonal approximation based representation and modeling of Arabic text. The recognition system is based on a proposed dissimilarity measure. Polygon--based modeling of text, prototype selection using set--medians, lexicon reduction using dot--descriptors etc. are some of the techniques that are applied to Arabic handwriting recognition for the first time. These techniques are applied to Arabic handwritten isolated--characters, digits, words and text recognition. Hence, this thesis addresses the lack of research in structural methods for Arabic handwriting recognition. In addition, the presented techniques may provide alternate sources of information to be utilized in multi--classifiers systems.;The proposed techniques are applied to different databases. We have obtained competitive results compared to the state of the art performance in statistical approaches and have surpassed the results based on structural techniques. In addition, the proposed techniques can be further extended. Moreover, we believe that several avenues of research in Arabic handwriting recognition have been initiated in this thesis. |
