An OCR System for Arabic Calligraphy Documents

Abstract

This paper introduces to get good accuracy for Arabic OCRresultsforolddocumentsandcalligraphydocuments.While our developed system has provided accurate results for modern Arabic documents, when we used that system for old Arabic documents, we got a steep degradation in performance, (around 25% accuracy compared with 85% for modern Arabic documents).MarketofArabicOCRforolddocumentsislargeand it deservers higher attention even more than the modern documents, which in many cases are already distributed in digitized format. Therefore, in this paper, we addresses the challengesofArabicOCRforolddocuments.Wemadethreemain modifications for our system. Firstly, we eliminated the word segmentation step and run the OCR process on the complete line. With this modification, we managed to avoid large number of segmentation errors on the word level but had to change our recognition approach to be a dictionary based. The second modification, we changed the used features to histogram gravity basedone.Thistypeoffeatureprovidedmuchbetterperformance especially in the challenging cases of old documents such as low quality printing effects, wavy baselines and heavy noisy documents. Third, we used a hybrid model that integrate Neural NetworkswithHMMtoprovidebetterdiscriminationbetweenthe shapes of Arabic ligatures. The paper starts with fast review for the baseline system and the added enhancements of such OCR. Then introduce to the new developed OCR efforts for Arabic old documents and calligraphy documents.