Dense Residual Network for Text Image Super-Resolution

Scene Text Image Super-Resolution (STISR) aims to enhance the resolution and visual quality of low-resolution scene text images, thereby improving the accuracy of downstream text recognition. Given the unique characteristics of text images, single-image super-resolution (SISR) methods based on natural images often fail to achieve satisfactory results on text images. In this paper, we propose a Dense Residual Network for Text Image Super-Resolution (TRRDB). The network extracts sequential information from text images through a sequence of residual blocks and incorporates this sequential information into the super-resolution network using densely stacked residual modules, guiding the super-resolution of text images. Results demonstrate that our proposed method significantly improves image recognition accuracy and enhances visual quality on the TextZoom benchmark dataset, particularly for images that were originally challenging to restore.