Deep Multi-Scale Transformer for Remote Sensing Image Restoration
Yanting Li · 2024
Under adverse weather conditions, remote sensing imagery is susceptible to distortions caused by severe weather, thereby compromising subsequent observation efforts. The existing transformer-based image restoration methods rely on a single-input-single-output structure, neglecting information from other scales. In this paper, we propose an effective deep multi-scale Transformer network for remote sensing image restoration. Specially, by incorporating a Shallow Convolutional Module, we enrich information across multiple scales. Additionally, we leverage a Fast Fourier Transformation-based module to process frequency features, and the Restormer Transform Block to extract deep-level features. To seamlessly integrate these multi-scale features, we employ the Asymmetric Feature Fusion approach. Experimental results demonstrate the effectiveness of our method in both image deraining and dehazing.