A novel approach for detecting and correcting segmentation and recognition errors in Arabic OCR systems

In this paper, we propose a new approach for detecting and correcting segmentation and recognition errors in Arabic OCR systems. The approach is suitable for both typewritten script recognition systems. Errors detection is based on rules of the Arabic language and a morphology analyzer. This type of analysis has the advantage of limiting the size of the dictionary to a practical size. Thus, a complete dictionary for roots, which does not exceed 5641 roots, the morphological rules and all valid patterns can be kept in a moderate size file. Recognition channel characteristics are modeled using a set of probabilistic finite state machines. Contextual information is utilized in the form of transitional probabilities between letters of previously defined vocabulary (finite lexicon) and transitional probabilities of garbled text. The developed detection and correction modules have been incorporated as a post-processing phase in an Arabic handwritten cursive script recognition system. Experimental results show a considerable enhancement in performance.