PR2 model: An extended framework for Mongolian OCR pre-trained models with prompt tuning in Internet of Things

This paper presents the PR2 model, designed to address optical character recognition (OCR) for Mongolian, a low-resource language, in complex backgrounds. By integrating Mask-RCNN for text region detection, prompt tuning for language adaptability, and RoBERTa for post-processing text correction, the PR2 model significantly improves the recognition accuracy of Mongolian text. Experimental results show that the PR2 model achieves a character recognition rate of 91.2% on the MLWS2021 dataset, clearly outperforming traditional OCR methods. Despite challenges in inference speed and handling complex scenarios, the model demonstrates strong robustness in dynamic environments and complex backgrounds. This study provides an innovative solution for intelligent text recognition of Mongolian and other low-resource languages, with broad practical application potential, particularly in IoT environments.