Lightweight Arbitrary-Scale Super-Resolution via Texture-Aware deformation

Abstract

Single-image super-resolution (SISR) has achieved remarkable progress through deep learning, yet mainstream SISR methods typically rely on fixed-scale up-sampling designs, struggling to balance reconstruction quality with computational efficiency across arbitrary scales, thereby limiting their practical flexibility. Although prior studies have attempted to incorporate positional and scale information for arbitrary-scale image super-resolution (ASISR), challenges remain in modeling cross-scale texture degradation characteristics. To address this, we propose two lightweight, structured plug-in modules that seamlessly integrate into existing SISR architectures, significantly enhancing their arbitrary-scale image modeling and reconstruction capabilities. Specifically, we design a Texture-Aware Deformation Up-sampling Module (TADUM), which captures scale-dependent texture deformation patterns by fusing position and scale-aware information to generate dynamic adaptive filters, enabling precise reconstruction at arbitrary scales. Furthermore, we introduce a Scale-Aware Image Refinement Module (SAIRM) that employs a multi-scale feature guidance mechanism and dynamic detail enhancement strategy to effectively maintain cross-scale visual consistency. Experimental results demonstrate that our approach significantly enhances reconstruction performance at non-integer scales while maintaining superior performance at standard integer scales, fully validating its efficiency, accuracy, and generalization in handling scale-sensitive tasks.