Recognition of Urdu ligatures - a holistic approach

This paper presents an effective segmentation-free and scale-invariant technique for recognition of Urdu ligatures in Nastaliq font. The proposed technique relies on separating the main body of ligatures from the secondary components and training a separate hidden Markov model for each. Features capturing projection, concavity and curvature information of ligatures are extracted using right-to-left sliding windows and are fed to the models for training. The system trained and evaluated on a total of more than 2,000 frequently occurring Urdu ligatures from a standard database realized a recognition rate of 97.93%.