PGHE-Net: physics-guided dehazing network with high-frequency enhancement for remote sensing images

As atmospheric haze degrades the visibility of optical remote sensing images (RSIs), image dehazing has become a crucial preprocessing step in geographic monitoring and advanced computer vision tasks. Despite the fact that current convolutional-neural-network-based dehazing technologies have achieved remarkable results on natural images, detail blurring and color distortions still occur when applied to RSIs. We proposed a physics-guided dehazing network with high-frequency enhancement (PGHE-Net), which consists of a channel-attention-based high-frequency enhancement (CAHE) and a physics-guided refiner (PGR). CAHE is designed to compensate for the loss of high-frequency information in features. It utilizes multiscale feature fusion and gated attention mechanisms to explore haze features with different distributions and concentrations. To further enhance the restoration of features in hazy regions, PGR leverages a physical model to guide the interaction between spatial attention and channel attention. This synergy enables the optimization of the reconstructed images in a hierarchical and stepwise manner at different scales. Experiments conducted on four public benchmark RSIs demonstrate the superiority of the proposed PGHE-Net in RSIs dehazing.