Efficient Reference-Based Super-Resolution via Adaptive Texture-Smooth Skipping and Transformer Architecture

Reference-based super-resolution (RefSR), which enhances low-resolution (LR) images by leveraging texture information from high-resolution reference (Ref) images, offers a promising solution for rapid, high-fidelity reconstruction from degraded measurements in computational imaging. However, existing RefSR methods predominantly focus on improving reconstruction performance while neglecting computational efficiency, limiting practical deployment. In this work, we observe that texture-smooth regions in LR images can be well reconstructed without Ref assistance, motivating an efficient RefSR method named skip low-frequency super-resolution (SLSR). We design a Low-Frequency Skip Module (LSM) that adaptively bypasses texture-smooth patches based on texture complexity measured by angular second moment (ASM). Furthermore, we propose a Multi-Information Enhanced Swin Transformer (MEST) that jointly leverages target features, gradient features, and transferred texture features to extract high-frequency details faithfully. A gradient loss is introduced to preserve gradient fidelity. Extensive experiments on CUFED5, Sun80, and Urban100 datasets demonstrate that SLSR achieves highly competitive performance while significantly reducing computational costs. Notably, on CUFED5, SLSR outperforms the second-best method by 0.16 dB in terms of PSNR with the lowest FLOPs. The efficient and high-fidelity reconstruction capability of SLSR suggests potential for future applications in computational imaging and nanostructured material characterization, where rapid texture restoration is critical.