Abstract

Interest in online handwriting recognition is increasing due to market demand for both improved performance and for extended supporting scripts for digital devices. Robust handwriting recognition of complex patterns of arbitrary scale, orientation and location is elusive to date because reaching a target recognition rate is not trivial for most of the applications in this field. Cursive scripts such as Arabic and Persian with complex character shapes make the recognition task even more difficult. Challenges in the discrimination capability of handwriting recognition systems depend heavily on the effectiveness of the features used to represent the data, the types of classifiers deployed and inclusive databases used for learning and recognition which cover variations in writing styles that introduce natural deformations in character shapes. This thesis aims to improve the efficiency of online recognition systems for Persian and Arabic characters by presenting new formal feature representations, algorithms, and a comprehensive database for online Arabic characters. The thesis contains the development of the first public collection of online handwritten data for the Arabic complete-shape character set. New ideas for incorporating relational information in a feature representation for this type of data are presented. The proposed techniques are computationally efficient and provide compact, yet representative, feature vectors. For the first time, a hybrid classifier is used for recognition of online Arabic complete-shape characters based on the idea of decomposing the input data into variables representing factors of the complete-shape characters and the combined use of the Bayesian network inference and support vector machines. We advocate the usefulness and practicality of the features and recognition methods with respect to the recognition of conventional metrics, such as accuracy and timeliness, as well as unconventional metrics. In particular, we evaluate a feature representation for different character class instances by its level of separation in the feature space. Our evaluation results for the available databases and for our own database of the characters' main shapes confirm a higher efficiency than previously reported techniques with respect to all metrics analyzed. For the complete-shape characters, our techniques resulted in a unique recognition efficiency comparable with the state-of-the-art results for main shape characters.