HLFFNet: A Hierarchical Lightweight Feature Fusion Network for Remote Sensing Image Super-Resolution

Recent advances have been made in remote sensing image super-resolution (RSISR). However, many techniques require extensive computational resources, which poses challenges for resource-constrained edge devices computing capabilities. To tackle this, we propose a hierarchical lightweight feature fusion network (HLFFNet) to improve remote sensing image resolution. Specifically, we initially replace redundant convolutional operations with depthwise separable convolutions. This not only reduces computational and parameter overhead while improving efficiency but also maintains or even enhances model performance. This design is highly suitable for deep learning applications requiring efficient and lightweight computations. Subsequently, we design a feature distillation residual module (FDRM), which leverages distilled features, residual connections, and attention mechanisms to improve the network's ability to represent high-quality features. Furthermore, we introduce hierarchical feature fusion blocks (HFFBs) for superior feature fusion. This structure extracts features layer by layer through cascaded FDRMs and lightweight attention fusion module (LAFM), followed by feature fusion at the final stage, enabling the subsequent network to better capture critical detail information. Finally, we adopt an efficient reconstruction structure that combines convolution, PixelShuffle, and SE modules to address detail recovery and computational efficiency in upsampling. This approach boosts image resolution and refines output feature quality using attention mechanisms. Comprehensive experiments were conducted on two remote sensing datasets. The results demonstrate that our approach successfully achieves a balance between complexity and performance.