Lightweight Image Super-Resolution Using Fine-Grained Feature Distillation in a Dense Residual U-Net

In recent years, convolutional neural networks (CNNs) have achieved remarkable success in single image super-resolution (SISR). However, existing methods often suffer from high model complexity, limited suitability for edge devices, and insufficient fine-grained feature extraction. To address these challenges, we propose a lightweight architecture called Fine-Grained Feature Distillation Dense Residual U-Net (FDDRU), which enhances fine-grained representation while significantly reducing parameter overhead. The model incorporates a Fine-Grained Feature Distillation Block (FFDB) and a Fine-Grained Shallow Residual Block (FSRB), enabling efficient collaboration between depth-wise and point-wise convolutions to improve reconstruction quality. Built upon a U-Net backbone, FDDRU further integrates a Dense Residual Connection Mechanism (DRCM), a Multi-Level Information Supplementation Mechanism (MISM), and a Bottom Module (BM) to strengthen feature propagation and information retention. For training, we adopt a hybrid loss function that combines L1 loss with a structural rigidity loss, jointly optimizing pixel-level accuracy and local structural consistency. Extensive experiments on standard benchmarks demonstrate that FDDRU outperforms state-of-the-art methods, achieving superior reconstruction performance with minimal model complexity.