Arabic Optical Character Recognition (OCR) Evaluation in Order to Develop a Post-OCR Module

摘要

Optical character recognition (OCR) is the process of converting an image of a document into text. While progress in OCR research has enabled low error rates for English text in low-noise images, performance is still poor for noisy images and documents in other languages. We intend to create a post-OCR processing module for noisy Arabic documents which can correct OCR errors before passing the resulting Arabic text to a translation system. To this end, we are evaluating an Arabic-script OCR engine on documents with the same content but varying levels of image quality. We have found that OCR text accuracy can be improved with different stages of pre-OCR image processing: (1) filtering out low-contrast images to avoid hallucination of characters, (2) removing marks from images with cleanup software to prevent their misrecognition, and (3) zoning multi-column images with segmentation software to enable recognition of all zones. The specific errors observed in OCR will form the basis of training data for our post-OCR correction module.