Reproduce “TextDiff: Mask-Guided Residual Diffusion Models for Scene Text Image Super-Resolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Baolin Liu, Zongyuan Yang, Pengfei Wang, Junjie Zhou, Ziqi Liu, Ziyi Song, Yan Liu, Yongping Xiong
TextDiff introduces a mask-guided residual diffusion model for scene text image super-resolution, effectively addressing text edge blurring and distortion by precisely reconstructing high-resolution text. The model achieved state-of-the-art text recognition accuracy, improving average performance on TextZoom by 2.0% for ASTER, 0.3% for MORAN, and 0.6% for CRNN, and notably reduced perceptual similarity metrics like LPIPS by up to 12%.
Key insights
- Modeling residuals with a conditional diffusion process, guided by an initial deblurred image and an explicit text mask, effectively mitigates generative diversity while enhancing text structure and sharpness.
- Training the diffusion model to predict the clean residual rather than noise, combined with deterministic sampling, proves more effective for text super-resolution, ensuring structural consistency with fewer sampling steps.
- Integrating a CRNN-based perceptual loss into the overall objective significantly aligns the super-resolved output with downstream text recognition requirements, improving both visual quality and machine readability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.