Haifeng Hu

Transformer-based image super-resolution has become a critical technique for enhancing visual quality in digital photography, remote sensing and video surveillance. However, the quadratic computational complexity of self-attention has led to a surge of token compression strategies, where an inherent conflict arises between coarse-grained token representations and high-fidelity image reconstruction. Existing methods typically overlook the guiding role of semantic and structural priors in token dependency modeling, resulting in suboptimal attention allocation and limited token compression ratios. To address these issues, we propose a novel Semantic-Aware and Structural-Focused lightweight Super-resolution Network. Specifically, a semantic-aware token interaction module is designed to efficiently associate semantically similar yet spatially distant regions, facilitating complementary information aggregation from low-resolution inputs. Subsequently, pre-modeled structural priors queried from semantic cues are leveraged to guide the reconstruction of fine-grained details. Moreover, we design a structural-focused token aggregation module that adaptively focuses on similar structures, significantly reducing redundant computations among irrelevant tokens. In the feedforward process, group-wise global-local interaction mechanisms are employed to enable efficient information propagation while maintaining lightweight computation. Extensive experiments demonstrate their effectiveness and efficiency. Compared with the strongest baseline, our method achieves up to 0.44[ ]dB improvement in PSNR while delivering [ ] faster inference speed.