TextDiff: Mask-Guided Residual Diffusion Models for Scene Text Image Super-Resolution

• DOI: 10.48550/arxiv.2308.06743
• Publication Date: 2023-01-01
• Authors: Baolin Liu; Zongyuan Yang; Pengfei Wang; Junjie Zhou; Ziqi Liu; Ziyi Song; Yan Liu; Yongping Xiong
• Keywords: Sharpening; Benchmark (surveying); Ground truth; Line (geometry)

The goal of scene text image super-resolution is to reconstruct high-resolution text-line images from unrecognizable low-resolution inputs. existing methods relying on the optimization pixel-level loss tend yield edges that exhibit a notable degree blurring, thereby exerting substantial impact both readability and recognizability text. To address these issues, we propose TextDiff, first diffusion-based framework tailored for super-resolution. It contains two modules: Text Enhancement Module (TEM) Mask-Guided Residual Diffusion (MRD). TEM generates an initial deblurred mask encodes spatial location MRD responsible effectively sharpening edge by modeling residuals between ground-truth images. Extensive experiments demonstrate our TextDiff achieves state-of-the-art (SOTA) performance public benchmark datasets can improve Moreover, proposed module plug-and-play sharpens produced SOTA methods. This enhancement not only improves results generated but also does require any additional joint training. Available Codes:https://github.com/Lenubolim/TextDiff.