More and Less: Enhancing Abundance and Refining Redundancy for Text-Prior-Guided Scene Text Image Super-Resolution

Scene text image super-resolution (STISR) aims to enhance low-resolution text images, boosting downstream text recognition tasks. Recent STISR models leverage text recognizer for prior information, achieving superior performance via a novel strategy. However, we observe abundant erroneous prior information from the low-resolution (LR) text images processed by the text recognizer, which can mislead text reconstruction when fused with image features. Therefore, we propose a novel sequential residual blocks, termed sequence refinement blocks, to refine the merged features of text images and text priors during the reconstruction of LR images. Additionally, regarding the widespread problem of ignoring the contextual semantic information in the shallow features of text images in the STISR, We introduce a multi-scale feature module to supplement the fine-grained and coarse-grained information required in the reconstruction of LR text images, which can well resolve information loss and generate more accurate super-resolution text images. Our proposed method consistently outperforms baselines employing text recognizers ASTER, MORAN, and CRNN by 1–2 \(\%\) on TextZoom, and achieves impressive gains of 4–5 \(\%\) on the challenging hard subset when leveraging multi-modal recognizers like ABINet and MATRN. The generalization experiments on scene text recognition datasets demonstrate optimal 5–8 \(\%\) performance improvements over the baselines.