Sat-DehazeGAN: an efficient dehazing model in water-sky background for river-sea transport

Gao, Xianglin

DOI: https://doi.org/10.1007/s00530-024-01599-z

IF: 3.9

2024-12-13

Multimedia Systems

Abstract:Traditional dehazing methods lost their dominance after the emergence of deep learning models, which has a significant effect on image dehazing. However, existing deep learning methods are mainly used for dehazing in land background, and they didn't perform well in water-sky environment. To solve this problem, we propose a Generative Adversarial Network-based method called Sat-DehazeGAN for dehazing images in water-sky background. Since paired real hazy and clean images with water-sky background are hard to capture in reality, datasets of unpaired real hazy and clean images are used in our experiments to get rid of the disadvantage brought by synthenic images. But in that case, the network may learn an under-constrained domain mapping from hazy images to haze-free images due to unpaired images of the training set. Therefore, we designed an innovative multi-scale attention module, and combine it with skip connections in the U-net structure to avoid the artifacts and loss of detail. In addition to ensure that the original haze-free and reconstructed haze-free images are more similar, a new loss called saturation loss was also proposed to improve the quality of reconstructed haze-free images. Experimental results finally demonstrate the effectiveness of the proposed method. It achieves a PSNR of 23.56, a SSIM of 0.923, and a CIEDE2000 score of 5.73 in dehazing images with a water-sky background, while maintaining a computational complexity of 70.57 G floating-point operations and 14.37 M parameters. Additionally, it delivers the best visual results compared to other state-of-the-art methods.

computer science, information systems, theory & methods