Distribution, Directional, structural and concavity features for historical Arabic handwritten recognition: a comparative study

In the process of automatic handwritten recognition especially in Arabic historical documents, the feature extraction is an important step, which find a set of measured values that accurately discriminate the input handwritten words or characters. In this paper, we try to determine how features designed for Arabic handwritten recognition can be efficient in Arabic historical documents by conducting a comparative study of four types of features (distribution, directional, structural and concavity features). The recognition process is based on Hidden Markov models with HTK toolkit and sliding window features. Words HMMs are learned using embedded training based on character HMM models. Experiments are performed on the benchmark Iben Sina database of Arabic historical documents.