Navigating Style Variations in Scene Text Image Super-Resolution through Multi-Scale Perception

Scene Text Image Super-Resolution (STISR) aims to simultaneously improve the resolution and legibility of text images, generating images that will significantly impact the performance of downstream tasks. In spite of the many advances that have been made, most of the existing methods have difficulty in balancing the attention to the text region in the image and to the rest in a reasonable way. This results in a poor restoration performance on images with complex texture structures or strong stylization due to its inability to accurately localize the text on images. To address the problem, we propose a novel Multi-scale Converged Super-Resolution Network MCSRN. In this framework, a Spatial Text Adjustment Module (STA) is designed to help eliminate the influence of stylized images on text recognition, obtaining advanced semantic guidance. To avoid issues such as distortion and pixel drift, we put forward a Multi-scale Alignment Fusion Module (MAF), which aligns and fuses images with the high-level semantics generated by the STA at different scales, facilitating more comprehensive interaction between text and image features. Finally, we combine the Upscale Module with the graph attention to get the super-resolution output. Through extensive experiments on TextZoom benchmark dataset and four scene text recognition datasets IC15, CUTE80, SVT, SVTP, our proposed MCSRN achieves competitive performance, highlighting the superior overall quality and text detail of our generated images, as well as its outstanding accuracy when applied to downstream recognition tasks.