Arabic Handwriting Recognition

This thesis explores a number of different techniques for use in the field of Arabic Handwriting Recognition. A review of previous work in the field is conducted, and then various techniques are explored in the context of classifying town names from the IFN/ENIT database. A baseline-finding algorithm using Principal Components Analysis is implemented, and the change in performance from reducing the influence of certain word features is also demonstrated. Several simple methods of town name classification are investigated, including a scheme using Tangent Features. These model the variations in the training examples in order to improve generalisation, and perform with 94% accuracy on a small 10-class lexicon. Moment invariants are considered as useful features for classification, but fail to surpass the performance of simpler methods. An approach where town names are split into parts and traced to recover temporal information is conceived, and found to have encouraging performance and several useful properties.