关键词: 图像去雾, 遥感图像, 无监督方法, 注意力机制, 特征增强

Abstract: To address the problem that existing dehazing methods are difficult to effectively suppress artifact generation and insufficient texture detail recovery caused by non-uniform haze distribution in real remote sensing images, unsupervised dehazing method for remote sensing images with non-uniform haze was proposed. First, based on the Cycle-Consistent Generative Adversarial Network (CycleGAN), Within the generator, a Residual Multi-scale Attention Mechanism (RMAM) was designed to expand the receptive field and enhance the extraction of multi-scale texture and structural information, enabling effective restoration of realistic texture details. Second, a Haze Distribution Enhancement Module (HDEM) was designed to explicitly enhance the expression of haze features. Combined with a dual-branch fusion strategy, it guided the network to accurately identify and process non-uniform haze, alleviating the artifact problem caused by insufficient haze distribution perception. Finally, a feature attention mechanism was embedded in the discriminator to strengthen its ability to distinguish the authenticity of local structures and textures, thus improving the restoration quality of dehazed images. Experimental results on the synthetic remote sensing datasets SateHaze 1k and RICE, the proposed method outperforms the best-performing baseline by 4.2% and 2.4% in Peak Signal-to-Noise Ratio (PSNR), and by 0.96% and 0.52% in Structural Similarity Index (SSIM), respectively. Experimental results on the real-world dataset RRSD300, the proposed method outperforms the best-performing baseline by 2.0% and 0.45% in Natural Image Quality Evaluator (NIQE) and Integrated Local NIQE (IL-NIQE), respectively. Extensive quantitative and qualitative results demonstrate that the proposed method effectively removes non-uniform haze, suppresses artifact generation, and restores texture details in remote sensing images.

Key words: image dehazing, remote sensing images, unsupervised method, attention mechanism, feature enhancement