Two approaches for the recognition of printed Farsi documents based on sub-word shape recognition is proposed. First approach is based on recognition of sub-word shape as a whole and the second is based on the recognition of the body of sub-words. Sub-word body is const More
Two approaches for the recognition of printed Farsi documents based on sub-word shape recognition is proposed. First approach is based on recognition of sub-word shape as a whole and the second is based on the recognition of the body of sub-words. Sub-word body is constructed via removing dots and signs of the sub word. In second approach, information of dots and signs will be added after recognition of the body. Both approaches have two phases: training and test. In training phase, sub-words are clustered based on ISODATA algorithm. Initial centers of the clusters are computed through a hierarchical clustering algorithm. In first approach, sub-word recognition is performed in two stages: finding clusters close to the input sub-word and then finding the best match within the sub-words of these clusters. In the second approach another stage is required to find the final sub-word including dots and signs. Experimental results show that on clean images the first algorithm have better performance; 94% versus 93% in word level. But when dealing with low quality and noisy images, both algorithms are suffering from reduced accuracy. Sometimes this reduction is significant. The reasons of this behavior are inspected and some solutions are presented. Finally we compared both methods and inspected pros and cons of Farsi OCR based on sub-word shape.
Manuscript profile
