Prior Segmentation of Old Arabic Manuscripts by Separator Word Spotting

Because of the low quality of old manuscripts, the complexity of Arabic script and the different writing styles, segmenting them is a challenging problem. This work aims to preprocess these manuscripts to be correctly segmented into independent words for text recognition. The idea is to spot separator words, detach them from neighboring words if necessary and use them to segment text-lines into words. To locate separator word in these document images, we proposed a word spotting method based on Generalized Hough Transform. This method is performed using convex theory points. Around a window centered on the group of votes of the separator word, it detects all connections below text-line baseline, analyses terminal letter morphology and tries to separate between touching or overlapping components. We tested the proposed system on Arabic historical manuscripts from the 19th century onwards conserved in the Tunisian National Archives. Experiments show very encouraging results.