Typefaces and Ligatures in Printed Arabic Text: A Deep Learning-Based OCR Perspective

Arabic script is complex, with multiple shapes for the same characters in different positions. Another challenge of the script, in the context of recognition, is ligatures. A combination of a specific two or more character sequence takes a different shape than what those characters normally look like when they appear in a similar position. Deep learning-based systems are widely used for text recognition these days. In this work, we investigate the performance of deep learning systems for two alternative modeling choices: using characters as modeling units and using character shapes as modeling units. Moreover, we also investigate the impact on text recognition with mixed typefaces, where the training and test sets have samples from multiple typefaces, and discuss the effect of font families on recognition performance. We extend this by studying the effectiveness of the text recognition system in recognizing text from unseen typefaces, i.e., text in the test set is from a typeface not available in training. Finally, we present a methodology to automatically detect ligatures in printed Arabic text. We conducted experiments on the publicly available APTI dataset of printed Arabic text and report the findings and discuss the results.