Robust metrics for evaluating arabic OCR systems

Evaluating OCR software is an important task to test the accuracy and to compare the strengths and weaknesses of distinct OCR systems. Research in the past has been done to compare the performance of OCR systems, all metrics used before dependent only on the character accuracy. This paper introduces word-based accuracy as well as character accuracy metrics to evaluate and to compare the performance of Arabic OCR systems. Different types of real scanned images such as newspapers, books, and journals image samples are used. The experimental results proved that such metrics are useful for evaluation of Arabic OCR systems and for Arabic text retrieval applications.