MADN: Multi-Attention With Diffusion Network for Scene Text Image Super-Resolution

Scene Text Image Super-Resolution (STISR) enhances the recognition accuracy of degraded textual imagery that requires fine-grained character reconstruction. Despite incorporating text-specific prior knowledge, existing methods suffer from suboptimal feature discrimination and inadequate sequential dependency modeling, resulting in information dilution and spatial discontinuity that compromise textual legibility. In this paper, we propose a Multi-Attention with Diffusion Network (MADN), a hierarchical attention framework that addresses these limitations through dual-domain feature refinement. Our Attention-Guided Feature Extraction (AGFE) module implements Channel Attention (CA) for adaptive feature recalibration and Spatial Attention (SA) for region-specific localization, enabling selective amplification of text-critical features while suppressing background interference. We further introduce Multi-Attention Residual Blocks (MARB), which integrate Bidirectional Long Short-Term Memory (BLSTM) networks for modeling sequential dependencies and refining multi-dimensional features. This multi-attention-based model acquires inter-character contextual relations while preserving the fine-grained textual structure. MADN adopts a hybrid dual-branch architecture, comprising a Super-Resolution (SR) branch equipped with AGFE and cascaded MARB modules to process low-resolution inputs and a Guidance Branch, along with a Text Prior Enhancement Module (TPEM) for forward diffusion and text feature generation. The Feature Fusion Module (FFM) fuses multi-scale features of two branches, enabling information exchange across the branches and refinement of features through diffusion mechanisms. Extensive experiments on the TextZoom dataset demonstrate state-of-the-art performance with significant improvements in Recognition Accuracy, PSNR, and SSIM. The code for MADN is available at: https://github.com/cuee-mdap/madn-net