Arabic handwritten text line segmentation using a multi-agent system and a directed CNN
Mohsine El Khayati, Youssfi Elkettani · 2021
Segmentation of text lines is a critical phase in Arabic handwriting recognition systems. Incorrect line segmentation generates errors in later phases, leading to total changes in the meaning of the text. Line segmentation in printed Arabic documents is an easier task compared to handwritten ones. Segmentation of lines from Arabic handwriting is difficult due to many challenging issues including overlaps, touches, skew/tilt in lines, diacritics, etc. To implement this phase, this paper proposes a robust approach (A2) that is based on a multi-agent system and Convolutional Neural Networks (CNN). The proposed approach is an enhanced version of the approach proposed in [1] (A1). In A1, the agents rely mainly on a morphological analysis algorithm to segment lines. Instead of that, the new approach adapts the External Features-based CNN (EFNet) architecture [2] inside the multi-agent system in order to enhance the performance of the segmentation algorithm. In terms of segmentation score, A2 brought better results compared to A1 on two benchmark databases (KHATT and HAPD). It also outperformed other recent methods in the literature.