Text Feature-Aware Network for Scene Text Image Super-Resolution

Scene Text Image Super-Resolution (STISR) aims to restore high-frequency details from low-resolution text images to improve scene text recognition accuracy. However, existing methods predominantly rely on character sequences predicted by recognizers for guidance, leading to imbalanced feature extraction across different characters. To address this issue, we propose an efficient framework named TFA-Net, which consists of two branches: a Semantic Guidance Branch and a Super-Resolution Reconstruction Branch. In the Semantic Guidance Branch, we design a Text Feature Enhancement (TFE) module to enhance and fuse the intermediate features extracted by the recognizer with low-resolution image features, thereby generating high-quality semantic guidance. In the Super-Resolution Reconstruction Branch, we propose a novel Efficient Attention Recurrent Residual Block (EARRB) to improve feature extraction efficiency and enhance super-resolution reconstruction capability. Extensive experiments on benchmark datasets demonstrate that TFA-Net achieves superior performance compared to state-of-the-art methods.