Focus-guided feature fusion network for lightweight image super-resolution

Single image super-resolution (SISR) has long faced the critical challenge of balancing model complexity and reconstruction accuracy. Existing research primarily focuses on lightweight network design, which reduces computational costs to some extent but often overlooks the latent information contained in low-resolution (LR) images. To address this issue, this paper proposes a Focus-Guided Feature Fusion Network (FocusSRNet), which introduces image focus information as a prior guidance signal into SISR tasks for the first time, thereby enhancing the sufficiency and discriminability of feature representations. The proposed network comprises three core modules: the Focus Measurement Unit (FMU), the Large Receptive Field Feature Extraction Module (LRFFEM), and the Asymmetric Channel Feature Enhancement Module (ACFEM), which collectively enable the network to capture more informative features and improve reconstruction quality. Experimental results demonstrate that FocusSRNet consistently outperforms existing state-of-the-art methods across multiple benchmark datasets, validating the effectiveness of the focus-guided mechanism in improving reconstruction quality. Overall, this study provides a new perspective and feasible approach for achieving lightweight and high-accuracy super-resolution reconstruction.