A Robust Image Dehazing Model Using Cycle Generative Adversarial Network with an Improved Atmospheric Scatter Model

Learning-based dehazing methods have attained numerous achievements in dehazing images. However, during the dehazing, the image depth and the atmospheric scattering factor are completely ignored. Current models result in the poor quality in dehazing real-world haze images. In this study, we propose a robust image dehazing model (RIDehaze) based on cycle generative adversarial network (CycleGAN). In RIDehaze, a dehazing GAN leverages a transmission map estimator with an improved atmospheric scattering model to restore clear images, which contain more real-world physical characteristics. CycleGAN utilizes depth estimator to generate more realistic haze images to train the dehazing GAN, thereby improving its generalization. Meanwhile, considering that the discriminator in CycleGAN struggles to support network training in the later stages, we designed a multi-level discriminator. This design ensures that even in the later stages of training, the discriminator can still effectively guide the network in dehazing images. On the real-world haze images of the RESIDE and Haze4k datasets, RIDehaze obtains clearer and more natural haze-free images and can specially produce better visual effects of distant scenery. On the synthetic hazy images, RIDehaze obtains similar PSNR and SSIM to supervised learning methods and outperforms the other unsupervised learning methods. The code for the method in this article can be downloaded at "GitHub address", where the GitHub address is a hyperlink to: https://github.com/Paris0703/icann_RIDehaze/tree/main.