OCR Post-Processing Text Correction using Simulated Annealing (OPTeCA)

OCR Post-Processing Text Correction using Simulated Annealing (OPTeCA) Gitansh Khirbat Computing and Information Systems The University of Melbourne Australia gitansh.khirbat@unimelb.edu.au Abstract This paper describes the system details and results of team “EOF” from the Uni- versity of Melbourne for the shared task of ALTA 2017, which addresses the problem of text correction for post-processed Op- tical Character Recognition (OCR) based systems. We developed a two stage sys- tem which first detects errors in the given OCR post-processed text with the help of a support vector machine trained using given training dataset, followed by rectify- ing the errors by employing a confidence- based mechanism using simulated anneal- ing to obtain an optimal correction from a pool of candidate corrections. Our system achieved a F1-score of 32.98% on the pri- vate leaderboard1, which is the best score among all the participating systems. 1 Introduction The dawn of digital age on mankind has laid the foundation of connectivity, fostering access and exchange of information practically anywhere in the world. Information can be present in any form, the most common being textual and graphical doc- uments. Capturing and curating documents such as magazines, newspapers, journals and scientific articles is the primary requirement for a digitized, inter-connected society. While most of the tex- tual documents can be stored with a decipherable textual component, the story is not the same for graphical documents which can be a collection of images containing scans of a textual document. Optical Character Recognition (OCR) is the process of identifying typed, handwritten or printed textual characters within a document con- taining scanned images or photographs with the 1https://www.kaggle.com/c/alta-2017- challenge/leaderboard help of various image processing and pattern recognition techniques (Tappert et al., 1990), (Gupta et al., 2007). The text obtained by OCR systems often suffers from low accuracy owing to irregularities in images, poor scans or simply the nature of arrangement of letters in a word. For example, reading “lwo” instead of “two”, “ia” in- stead of “is”, “m” instead of ”rn”, to name a few. These erroneous characters severely hamper the quality and readability of a converted document. Identifying and rectifying such erroneous charac- ters in every OCR-processed document manually is a tedious task due to the sheer volume of data. Consequently, a methodology is required to iden- tify such OCR errors and rectify them in order to enforce standards of purity and quality of the archived data. This need has motivated the shared task of ALTA 2017 (Molla and Cassidy, 2017). The task organizers have provided the original out- puts of an OCR system together with their cor- rected version for scanned Australian publications from Trove database2. Using this data, the par- ticipants are asked to automatically identify and rectify the OCR errors for documents in a separate test dataset. Considerable research has been conducted pre- viously to automatically correct text obtained by OCR systems using machine learning. (Lund and Ringger, 2009) (Lund et al., 2011) (William B. Lund, 2013) (Lund et al., 2014) introduced var- ious techniques to select …

Content truncated to 500 words