Online Handwritten Persian Isolated Letter Recognition by Using Discrete Markov Models and Language-Based Features

In this paper, an online Persian isolated letter recognition by using discrete Markov models and language-based features is presented. In the proposed method, the preprocessing includes the extraction of dots, the removal of ending-bends, the normalization of the main body dimension, and the main body smoothing. The direction of the pen for the main body, and also the location and the number of secondary movements and dots of letters are used as the observable features in order to train and test models. No alphabet grouping has been used in the proposed method. In order to produce appropriate discrete observation sequences for the hidden Markov models, the proposed codebook is based on the characteristics of the Persian language. To improve the recognition accuracy, the proposed post-processing checks the recognized character in terms of the number and location of the secondary movements and dots. The final recognition rate is obtained more than 98% for the test dataset and outperforms other methods applied to this dataset