FAFMNet: A Lightweight Super-Resolution Network via Frequency-Aware and Multi-Scope Feature Fusion

Abstract

Single Image Super-Resolution (SISR), a fundamental task in computer vision, aims to reconstruct high-quality images from their low-resolution counterparts. To address the limitations of existing methods in recovering high-frequency details and achieving efficient feature modeling, we propose a lightweight and effective Frequency-Aware and Fusion-Modulated Network (FAFMNet). The proposed network jointly enhances local texture representation and global structure modeling by incorporating a Frequency-Aware Modulator (FAM) and a Multi-Scope Fusion Block (MSFB), thereby enabling efficient cross-scale and cross-frequency feature extraction. Specifically, the FAM module models global dependencies in the frequency domain, leveraging low-frequency preservation and high-frequency enhancement strategies to improve sensitivity to structural regions. Meanwhile, the MSFB module aggregates contextual information across multiple receptive fields and utilizes a sparse channel separation mechanism to achieve lightweight multi-scale feature representation. Extensive experiments on several benchmark datasets demonstrate that FAFMNet achieves superior reconstruction performance compared to existing methods while maintaining low parameter count and fast inference speed, validating the effectiveness and practicality of the proposed design.