Persian Text Recognition using n-gram Language Models and Grammatical Refinement

Text recognition has been one of the growing research topics in recent years. Many of these researches have focused on recognition of letters and sub-words as a basis for identifying larger text structures such as words, phrases and sentences. This thesis presents a new method in which the recognized sub-words are combined in order to provide meaningful words and sentences in Farsi texts. Since there may be more than one meaningful combination, the potential meaningful sentences are filtered using Farsi grammatical rules. In the sub-word recognition stage, a double scan method is exploited while the words are extracted using a database of frequent Farsi words. In the last stage a 2 and 3-gram method as well as Farsi grammatical rules are employed to identify the most meaningful sentence from all potential candidates. Experiments have proved the accuracy of the exploited method to be more than 85 percent.