A Large-Scale Font-Diverse Sindhi Ligature Recognition System

Sindhi is an ancient language rooted in the lower Indus River Valley. Many literary works have been produced in it; however, it is considered a low-resource language for Deep Learning. The applications of Deep Learning in the English, and Arabic Text Detection and Recognition are well known. However, little to no work has been conducted on the extended Perso-Arabic Sindhi script. This is mainly due to the absence of benchmark datasets. Consequently, no models have been published for Sindhi OCR. This study aims to bridge the research gaps by making the following contributions: (i) It produces a collection of 22,597 unique Sindhi ligatures. (ii) It provides a complete pipeline for data synthesis. (iii) It produces a large-scale and font-diverse Sindhi ligature-recognition dataset. (iv) It builds a baseline model for Sindhi ligature recognition that achieves state-of-the-art accuracy of 91.85%. The model can be used as a benchmark to assess the performance of future models, or as a baseline to build a complete pipeline for the Sindhi OCR.