The QCRI recognition system for handwritten Arabic

Abstract

This paper describes our recognition system for handwritten Arabic. We propose novel text line image normalization procedures and a new feature extraction method. Our recognition system is based on the Kaldi recognition toolkit which is widely used in automatic speech recognition (ASR) research. We show that the combination of sophisticated text image normalization and state-of-the art techniques originating from ASR results in a very robust and accurate recognizer. Our system outperforms the best systems in the literature by over 20% relative on the abcde-s configuration of the IFN/ENIT database and achieves comparable performance on other configurations. On the KHATT corpus, we report 11% relative improvement compared to the best system in the literature.