Histogram-based lines and words decomposition for arabic omni font-written OCR systems; enhancements and evaluation

Abstract

Given font/type-written text rectangle bitmaps extracted from digitally scanned pages, inferring the boundaries of lines hence complete words is a preprocessing vital to whatever OCR system while the recognition process itself as well as the post processing necessary for producing the recognized text. Histogram-based methods are commonly used due mainly to their relative implementation simplicity and computational efficiency, however, some authors report about their vulnerability to some idiosyncratic textual structure complexities, and noise. This paper elaborates on this approach to produce a more robust algorithm for lines/words decomposition esp. in Arabic, or Arabic dominated, text rectangles from real-life multifont/multisize documents. Trying to evaluate this algorithm, this paper also presents the results of extensive experimentation made on about 1800 documents fairly distributed over different kinds of sources with different noise levels at different scanning resolutions and color depths. © Springer-Verlag Berlin Heidelberg 2007.