A Novel Scene Text Image Super-Resolution Approach

doi:10.1016/j.dsp.2025.105651

A Novel Scene Text Image Super-Resolution Approach

2025-10-14

0

PRE

AI

李洪均 (Hongjun Li) *

S

Shangfeng Li

DOI：10.1016/j.dsp.2025.105651

Original

Original request for help

Save

Abstract

En 中文

Scene text image super-resolution (STISR) seeks to reconstruct high-quality, super-resolved (SR) images from low-resolution (LR) inputs. While existing state-of-the-art approaches have achieved notable gains in image quality, they frequently fail to account for the unique geometric and structural characteristics inherent to text images. Consequently, these methods often struggle in scenarios involving complex structures, stroke adhesion, or character deformation. As a result, text recognition on low-resolution images remains challenging. To address these limitations, we propose a novel STISR method based on Character Skeleton Structure Priors (CSSP). Specifically, we employ a skeleton extraction algorithm to distill the principal structural features from LR text images and deeply integrate these skeleton cues with the LR inputs. This integration guides the network to attend to the distribution and relationships of character strokes. In addition, we design a new sequence reconstruction block that dynamically captures and enhances key structural features. This facilitates more effective use of skeleton information and promotes fine-detail restoration. Furthermore, we introduce a skeleton-aware loss term into the training objective to enforce structural consistency, which further improves the model’s ability to recover character integrity and readability. Extensive experiments on the TextZoom dataset, using the ASTER recognizer, show that our CSSP model surpasses current state-of-the-art deep learning approaches by 1.1% in recognition accuracy. These results clearly demonstrate the superiority and effectiveness of our proposed method.

A Novel Scene Text Image Super-Resolution Approach

Abstract

Journal

Organization