A rule-based post-processing approach to improve Persian OCR performance

Abstract

Optical Character Recognition (OCR) is a system to convert images including text into an editable text. Nowadays, the accuracy of these systems in images with simplestructure and high quality is high. However, the performance degrades for images with complex-structure, low quality, and in the presence of noise, scratches, pictures, stamps, or other non-textual symbols. This paper proposes a Persian OCR post-processing technique to increase the accuracy of the OCR systems dealing with real-world challenging samples. The proposed method extracts five features in each line of the text and uses seven proposed rules to investigate whether that line should be ignored or not. To evaluate the proposed method, Khana (structural based) and Bina (deep learning-based) Persian OCR systems have been utilized, a dataset containing 200 complex-structure images has been collected, and a dataset including 100 simple-structure images has been used. The accuracy of Khana and Bina in images with a complex-structure is 39% and 58%, respectively, while after applying the proposed post-processing method, the accuracy increases to 93% and 91%, respectively.