Multi-scale feature fusion distilled attention network for efficient image super-resolution

Efficient single image super-resolution (SISR) seeks to enhance reconstruction quality while minimizing computational complexity and model parameters for practical deployment on resource-constrained edge devices. A critical limitation of existing approaches lies in the restricted receptive fields of conventional convolutions, which hinder attention mechanisms and backbone networks from effectively capturing non-local features, ultimately degrading super-resolution performance. To solve this challenge, we propose a novel Multi-scale Feature Fusion Distilled Attention Network (MFFDAN) that synergizes multi-scale feature extraction with lightweight attention mechanisms. Our framework introduces three key innovations. First, we proposed a Multi-scale Convolution (MConv) module that dynamically extracts hierarchical features with multi-scale receptive fields while maintaining parameter efficiency, subsequently extended into a Multi-scale Feature Fusion Distillation Block (MFFDB). Second, we proposed a Distilled Spatial Attention (DSA) mechanism that enhances spatial feature interactions through multi-scale downsampling and distillation-based information refinement. Third, a lightweight Depth-wise Channel Attention (DCA) module optimized via depth-wise convolutions for efficient channel feature selection is proposed. These components are systematically integrated through residual connections to form the Multi-scale Feature Fusion Distilled Attention Block (MFFDAB), the core building block of our architecture. Extensive experiments demonstrate that MFFDAN achieves state-of-theart performance while maintaining superior computational efficiency, achieving an optimal balance between reconstruction quality (PSNR/SSIM), model complexity (parameters/Multi-Adds), and inference speed across benchmark datasets. The proposed innovations provide a practical solution for deploying high-quality SISR models on edge computing platforms.