Deep Learning-Based Segmentation of Connected Components in Arabic Handwritten Documents

Abstract

This work proposes a practical and powerful segmentation approach that allows touching or overlapping characters in adjacent text lines or words within Arabic manuscripts to be segmented correctly. It is the first deep learning-based method proposed to solve this problem. It is based on a modified U-Net named AR2U-net: an Attention-based Recurrent Residual U-net model trained to separate touching characters. It is trained on the LTP (Local Touching Patches) database to segment touching characters in a pixel-wise classification. The network labels pixels of the touching characters’ images in four classes: pixels of background, pixels of the first character, pixels of the second character, and those where characters touch. Once the segmentation is done, the separation of touching text lines or words can be done efficiently and speedily. We also propose a post-treatment to segment successive touching text lines in this work. Experimental results on the LTP database show that our proposed method is practical in copes with touching and overlapped characters separation. It achieves higher accuracy of 94.6% than those reported in the state-of-the-art.