Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Restore Text First, Enhance Image Later: Two-Stage Scene Text Image Super-Resolution with Glyph Structure Guidance

Minxing Luo, Linlong Fan, Wang Qiushi, Ge Wu, Yiyan Luo, Yuhang Yu, Jinwei Chen, Yaxing Wang, Qingnan Fan, Jian Yang

TL;DR

TIGER addresses the persistent trade-off between readability and image quality in scene-text super-resolution by decoupling glyph restoration from full-image enhancement. It introduces a two-stage, text-first paradigm where a region-level text restoration module reconstructs accurate glyph structures that guide a ControlNet-like diffusion-based image restoration stage, all trained with a two-phase strategy on synthetic and real data. The authors also contribute UZ-ST, an extreme-zoom Chinese scene-text dataset with precise LR–HR alignment, enabling robust evaluation under challenging degradations. Across Real-CE and UZ-ST benchmarks, TIGER achieves state-of-the-art performance in both image quality metrics and text readability (OCR-A), demonstrating the effectiveness of explicit glyph-structure guidance for coherent, artifact-free text in high-resolution scenes. The work advances practical scene-text SR with broad implications for document restoration, accessibility, and navigation tasks.

Abstract

Current image super-resolution methods show strong performance on natural images but distort text, creating a fundamental trade-off between image quality and textual readability. To address this, we introduce TIGER (Text-Image Guided supEr-Resolution), a novel two-stage framework that breaks this trade-off through a "text-first, image-later" paradigm. TIGER explicitly decouples glyph restoration from image enhancement: it first reconstructs precise text structures and uses them to guide full-image super-resolution. This ensures high fidelity and readability. To support comprehensive training and evaluation, we present the UZ-ST (UltraZoom-Scene Text) dataset, the first Chinese scene text dataset with extreme zoom. Extensive experiments show TIGER achieves state-of-the-art performance, enhancing readability and image quality.

Restore Text First, Enhance Image Later: Two-Stage Scene Text Image Super-Resolution with Glyph Structure Guidance

TL;DR

TIGER addresses the persistent trade-off between readability and image quality in scene-text super-resolution by decoupling glyph restoration from full-image enhancement. It introduces a two-stage, text-first paradigm where a region-level text restoration module reconstructs accurate glyph structures that guide a ControlNet-like diffusion-based image restoration stage, all trained with a two-phase strategy on synthetic and real data. The authors also contribute UZ-ST, an extreme-zoom Chinese scene-text dataset with precise LR–HR alignment, enabling robust evaluation under challenging degradations. Across Real-CE and UZ-ST benchmarks, TIGER achieves state-of-the-art performance in both image quality metrics and text readability (OCR-A), demonstrating the effectiveness of explicit glyph-structure guidance for coherent, artifact-free text in high-resolution scenes. The work advances practical scene-text SR with broad implications for document restoration, accessibility, and navigation tasks.

Abstract

Current image super-resolution methods show strong performance on natural images but distort text, creating a fundamental trade-off between image quality and textual readability. To address this, we introduce TIGER (Text-Image Guided supEr-Resolution), a novel two-stage framework that breaks this trade-off through a "text-first, image-later" paradigm. TIGER explicitly decouples glyph restoration from image enhancement: it first reconstructs precise text structures and uses them to guide full-image super-resolution. This ensures high fidelity and readability. To support comprehensive training and evaluation, we present the UZ-ST (UltraZoom-Scene Text) dataset, the first Chinese scene text dataset with extreme zoom. Extensive experiments show TIGER achieves state-of-the-art performance, enhancing readability and image quality.
Paper Structure (23 sections, 8 equations, 10 figures, 11 tables)

This paper contains 23 sections, 8 equations, 10 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methodology
  4. Architecture Overview
  5. Text Restoration stage
  6. Image Enhancement stage
  7. Dataset and Benchmark
  8. Experiment
  9. Implementation Details
  10. Comparison Results
  11. Quantitative Results
  12. Qualitative Results
  13. Ablation Study
  14. Conclusion
  15. Reproducibility Statement
  16. ...and 8 more sections

Figures (10)

  • Figure 1: We present TIGER (Text–Image Guided supEr-Resolution), a novel framework for scene text super-resolution. Its ‘text-first, image-later’ paradigm ensures accurate glyph restoration and consistently high overall image fidelity and visual quality.
  • Figure 2: The framework of TIGER, which includes the Text Restoration stage (stage 1) and the Image Enhancement stage (stage 2). Stage 1 recover accurate glyph structures from text regions. Stage 2 uses them to guide full-image restoration for coherent text and background.
  • Figure 3: Overview of UZ-ST (UltraZoom-Scene Text). (a) Real-CE LRs show only mild degradation (red box), while UZ-ST LRs exhibit stronger degradation (red box), enabling a more comprehensive evaluation. (b) Coarse-to-fine alignment: images are sorted by focal length, each warped to the next higher-focal neighbor using an estimated homography matrix, then refined to the 200 mm GT.
  • Figure 4: Qualitative Evaluation on Real-CE and UZ-ST.
  • Figure 5: Qualitative Results of Ablation Study with stage 2 fixed as the baseline.
  • ...and 5 more figures