Urdu Digital Text Optical Character Recognition

Abstract

This thesis introduces an innovative word-level Optical Character Recognition (OCR) model designed specifically for digital Urdu text recognition. Leveraging the power of transformer-based architectures and attention mechanisms, the proposed model was trained on a comprehensive dataset comprising approximately 160,000 Urdu text images. Remarkably, the model achieved a commendable character error rate (CER) of 0.242, indicating its superior accuracy in recognizing Urdu characters. The key strength of the model lies in its unique architecture, incorporating the permuted autoregressive sequence (PARSeq) model. This advanced approach enables context-aware inference and iterative refinement, leveraging bidirectional context information to enhance recognition accuracy. Additionally, the model's ability to handle a diverse range of Urdu text styles, fonts, and variations further enhances its applicability in real-world scenarios. While the model demonstrates promising results, it does have some limitations. Blurred images, non-horizontal orientations, and the overlay of patterns, lines, or other text can occasionally lead to suboptimal results. Additionally, trailing or following punctuation marks may cause noise in the recognition process. Addressing these challenges will be a focal point of future research. The proposed model's exceptional performance and its ability to adapt to various text styles make it a valuable tool for applications that require accurate and efficient Urdu text recognition. Future work will focus on refining the model, exploring data augmentation techniques, optimizing hyperparameters, and integrating context-aware language models to further improve its overall performance and robustness.