Remote Sensing Single Image Dehazing With Histogram Degradation Representation Learning

Haze severely degrades the visual quality of remote sensing images, resulting in the loss of fine details and reduced reliability in subsequent tasks. Although Transformer-based approaches have recently achieved notable progress in image dehazing, most existing methods mitigate computational complexity by constraining self-attention to channel interactions or fixed spatial regions. This design inevitably restricts their capacity to capture long-range spatial dependencies, which are crucial for effectively modeling haze distributions. To address this limitation, we propose a Histogram-guided Remote Sensing Image Dehazing Transformer (HRSformer). The framework is built upon two key components: a Histogram Degradation Representation Block (HDRB), which dynamically selects spatial feature ranges and collaboratively processes distant pixels with similar degradation patterns to suppress global haze interference, and a Multi-Expert Modulation Block (MEMB), which employs a mixture-of-experts strategy to adaptively capture multi-scale and multi-granularity contextual features in a sample-specific manner, thereby improving the reconstruction of fine local textures. Extensive experiments on multiple public remote sensing dehazing benchmarks demonstrate that HRSformer consistently outperforms state-of-the-art methods in both quantitative metrics and visual quality, validating its effectiveness and superiority.