Push the limit of scene text recognition using character and text length guided text super-resolution

Although scene text recognition has achieved remarkable progress in recent years, its performance remains limited when dealing with low-resolution (LR) scene text. To mitigate this issue, some recent approaches have adopted super-resolution (SR) as a preprocessing step. Nevertheless, these approaches tend to treat SR and recognition as independent tasks, often overlooking their inherent complementarity, which may restrict the full potential of the system and hinder further performance improvements. To overcome this limitation, we propose a unified framework that performs SR and scene text recognition simultaneously, enabling the two tasks to mutually reinforce each other. In addition, we seek to more effectively exploit the prior information inherently present in scene text images to guide the SR network toward better reconstruction performance. Specifically, we introduce an end-to-end architecture that integrates SR and recognition modules within a unified framework and adopts an iterative strategy to facilitate mutual enhancement between the two tasks. Inspired by human visual perception, we further incorporate two perceptual priors, Character Features and Text Length, collectively referred to as the CF-TL priors. These priors leverage semantic and structural cues to enhance the reconstruction of text images and improve recognition accuracy. Extensive experiments conducted on benchmark datasets demonstrate that our method significantly outperforms existing approaches in terms of recognition accuracy. These results highlight the effectiveness of our framework and its potential to push the limit of more robust and accurate scene text recognition systems for low-resolution inputs.