Multi-scale information distillation attention network for super-resolution reconstruction of remote sensing images
Bo Huang, Liaoni Wu, Yiqing Cao, Mingen Zhong · Journal of Measurements in Engineering · 2025
Super-resolution (SR) is an effective and reasonable way to improve the spatial resolution of remote sensing images, which serve as an important information carriers for Earth observations. Compared to natural images, the more complex spatial distributions and more detailed ground information contained within remote sensing data place higher demands on the feature-representation ability of the model. Moreover, considering the deployment of these systems on mobile hardware, the complexity of the model is also an urgent issue. To overcome these problems, this study proposes the multi-size information distillation attention network (MSIDAN) for super-resolution reconstruction of remote sensing images. In the designed residual block, a multi-size information-distillation module is designed to distill and fuse multi-level semantic features step-by-step while reducing the number of model parameters. After this, an enhanced contrast-aware channel attention mechanism is employed to perceive high-frequency information by automatically encoding the weight values of candidate features. A large number of comparative experiments on four typical remote sensing image datasets demonstrate that MSIDAN outperforms other state-of-the-art approaches in both quantitative metrics and visual qualities. Compared to the information multi-distillation network (IMDN), MSIDAN improves the Peak Signal-to-Noise Ratio (PSNR) by 0.03312 dB, 0.06031 dB, 0.05319 dB, and 0.03812 dB on the RSSCN7, WHU-RS19, NWPU VHR-10, and COWC datasets, respectively. Moreover, in comparison to other comparable CNNs-based approaches, MSIDAN achieves a more favorable balance by jointly considering SR performance and model size. This technology provides valuable support for small target measurement and opens new opportunities in the field.