Abstract:
This thesis investigates the effect of the structure of the Hidden Markov Models (namely number of states) on the performance of an offline recognition system for handwritten Arabic names.
According to the proposed hypothesis which state that we can improve the performance for the obtained best fixed number of states using variable states model near to that fixed state model in search space. The estimation of these variable state models is based on image length and Chain Codes length. The idea is to use some factors to limit the search space. Using the proposed hypothesis two classes of models has been designed; image length based variable number of states model called AHOP, and Chain Codes based variable number of states models called AHOC. The thesis has studied finding the best number of states through searching in a group of functions. Each function in this group calculates the number of states from the length of Chain codes and the best fixed number of states.
Extensive experiments have been conducted using SUST dataset for Arabic names. The proposed method outperformed the fixed number of states models by 8% in these experiments.
