Semantic Reference Matching and Diffusion Learning for Intelligent Super-Resolution in Visual Information Systems

To enhance spatial detail in remote sensing images, super-resolution (SR) has become essential. Conventional single-image SR suffers from limited information, often producing overly smoothed results. Reference-based SR leverages high-resolution reference images to mitigate this issue, but real-world scenarios face cross-sensor discrepancies and temporal land-cover changes, causing concept omission and mismatch that hinder effective reference utilization. To address these challenges, the authors propose a diffusion-based SR framework with semantic reference matching, namely Semantic Reference Matching and Diffusion Learning for Intelligent Super-Resolution in Visual Information Systems (SRM-DL). It comprises two key modules: (a) Concept Activation, which uses diffusion priors to recover missing structures, and (b) Attribute Concentration, which used a local–global dual-branch alignment to robustly incorporate semantically consistent reference information while suppressing mismatches. Multiscale consistency constraints further align reference and target features across spatial and semantic domains. Extensive experiments validated its effectiveness and practical potential in multimedia visual information systems.