A Novel Arabic Baseline Estimation Algorithm Based on Sub-Words Treatment

Abstract

Baseline detection is an essential preprocessing step for many OCR systems, it has a direct effect on the efficiency and reliability of characters segmentation and features extraction stages, which contribute strongly to yielding higher recognition accuracy. For Arabic handwritten, the conventional methods which extract baseline as straight line are ill-suited because some Arabic words may be contracted from two or more sub-words (PAWs), and the distribution of these sub-words can produce different slant angles within the same word. Focused on the source of the problem, we propose a novel Arabic baseline estimation algorithm in which the PAW level is the real basic block to be processed rather than word level. Experimental results using IFN/ENIT [1] database demonstrate the efficiency of the proposed algorithm.