Dehazing-DiffGAN: Sequential Fusion of Diffusion Models and GANs for High-Fidelity Remote Sensing Image Dehazing

Abstract

Atmospheric haze significantly reduces image quality by obscuring important visual details, which in turn impairs the performance of computer vision tasks. Many existing dehazing methods struggle in heavy haze conditions, often resulting in images with poor visual quality, unrealistic features, or visible artifacts. To overcome these limitations, we propose Dehazing-DiffGAN, a novel deep learning framework that combines diffusion models with generative adversarial networks (GANs) in a dual-phase pipeline. In the Dehazing Phase, the diffusion model progressively removes haze by simulating atmospheric scattering, effectively capturing both local structures and global context. In the Enhancement Phase, a GAN-based enhancement module improves image quality by emphasizing fine details and textures, enabling precise restoration of edges and high-frequency components. This synergistic fusion approach allows Dehazing-DiffGAN to generate images of superior visual quality, often indistinguishable from real, haze-free images. Extensive experiments show that Dehazing-DiffGAN achieves state-of-the-art performance on several challenging benchmarks. On the high-resolution SateHaze1k dataset for remote sensing, our model achieves a peak signal-to-noise ratio (PSNR) of 27.21 dB, outperforming the next best method at 26.84 dB, and obtains a superior Learned Perceptual Image Patch Similarity (LPIPS) score of 0.0472 compared to 0.08. On the ultra-high-resolution DNH-HAZE dataset (6000×4000 pixels), Dehazing-DiffGAN achieves the best LPIPS score of 0.267, significantly improving over the second-best score of 0.326. These results highlight the effectiveness of Dehazing-DiffGAN in improving image clarity and confirm its robustness and adaptability across a variety of dehazing tasks. The code is available at https://github.com/riotulab/Dehazing-DiffGAN.

Topics & Concepts