Improving and Measuring OCR Accuracy for Sinhala with Tesseract OCR Engine

Abstract

This research project proposes and implements a system to improve and measure the accuracy of the Sinhala OCR using the Tesseract OCR engine. The system implements modules to rectify the issues which are inherent to the Tesseract OCR engine when performing OCR for Sinhala language. During the course of the project, the world level accuracy was used to measure the accuracy of the output from the system. As a baseline to compare the results of the proposed system which implements tesseract OCR, the software the OCR Engine “පෙළ කැටෙත” was used. To improve the accuracy, a syntactical rule engine a module to detect and correct confusion character pairs and a rudimentary dictionary look up feature to detect and correct errors in word level has been implemented into the system. During the initial stage in the project which implemented only the Tesseract OCR library functionality, the output was less accurate when compared with the OCR Engine “පෙළ කැටෙත”. But as the features were built into the system, it yielded significantly improved results which improved the word level accuracy from the original 53.22% to 86.16%.