Unconstrained Handwritten Arabic Text-lines Segmentation based on AR2U-Net

Text-lines are hard to segment in the context of Arabic manuscripts, because of the narrowly spaced text-lines with touching or overlapping components, the varying spaces between words, the ascendant or descendant letters, special marks, and dots, calligraphy, etc. In this work, we proposed a system to automatically extract text-lines from images of unconstrained handwritten Arabic texts. Each text-line is detected by its baseline. The proposed system is based on text-line masks which are predicted by a deep neural network called AR2U-Net: a Recurrent Residual convolutional neural network based on the U-Net model with an Attention mechanism. We adjusted the AR2U-Net model to allow a pixel-wise classification and therefore to separate text-lines pixels from the background one. We tested it on BADAM: a Public Dataset for Baseline Detection in Arabic script Manuscripts that involves complex layouts as well as curved and arbitrarily oriented text-lines and overlaps between adjacent text-lines, words, or sub-words. Our model achieves the best performances with a Precision of 0.932% which competes with current state-of-the-art approaches.