Using Prominent Regions in Search Space Reduction for Recognition of Printed Farsi Subwords

In the most common Lexicon reduction methods، lexicon words are clustered based on their holistic shape features and then each query word image is classified into the closest cluster. As the errors at this stage propagate to the subsequent stages، relevant clusters should be selected with a high degree of accuracy. In this paper we present a novel verification method which decides on the validity of the recognized clusters based on a proposed confidence measure. The level of confidence to the selected clusters is measured using local shape features in the verification phase، where it is determined that the selected cluster is acceptable or not. For this purpose، some local shape features of the input subword image are compared to the “prominent regions” of the corresponding cluster. The prominent regions of a cluster are some local regions that discriminate the members of that cluster compared to the other clusters. The proposed verification method along with some predefined rules is used to reduce the lexicon size of Farsi subwords. The experiments conducted on a set of 6895 common Farsi subwords show that our proposed method significantly reduces the search space while preserving the accuracy in an acceptable rate.