Comparison of different preprocessing and feature extraction methods for offline recognition of handwritten arabic words

Preprocessing and feature extraction are very important steps in automatic cursive handwritten word recognition. Based on an offline recognition system for Arabic handwritten words which uses a semi-continuous 1-dimensional Hidden Markov Model recognizer different preprocessing combined with different feature sets are presented. The dependencies of the feature sets from preprocessing steps are discussed and their performances are compared using the IFN/ENIT-database of handwritten Arabic words. As the lower and upper baseline of each word are part of the ground truth of the database, the dependency of the feature set from the accuracy of the estimated baseline is evaluated.