Cycle-GAN Network Incorporated With Atmospheric Scattering Model for Dust Removal of Martian Optical Images

Dust particles in Martian atmosphere can significantly reduce visibility. This article proposes a physically guided neural network approach for dust removal of Martian images by incorporating atmospheric scattering model into the cycle-consistent generative adversarial network (Cycle-GAN) framework. The network consists of two primary modules: dust removal and dust addition, both of which combine neural networks and physical models. The dust-removal process estimates the scattering coefficient and transmission map and then physically implements dust removal with the atmospheric scattering model. The dust-addition process estimates the scene depth, which is then combined with the atmospheric scattering coefficients obtained by the dust-removal process to compute the transmission map. The transmission map is substituted into the atmospheric scattering model for dust addition of clean images. Moreover, the additional loss of transmission map and scattering coefficient furtherly enhances the consistency constraints. Martian clear and dust images collected by the Mars Curiosity rover are used to train and evaluate the new dust-removal approach. Extensive ablation experiments demonstrate the effectiveness of incorporating the physical model and the additional loss functions. Furthermore, the scattering coefficients learned by the network are validated with the Mie scattering theory, ensuring the physical plausibility of the estimated parameters. The key advantage of this approach is that it does not require paired images of the same scene with or without dust, which is often a limiting factor for supervised dust-removal techniques. By incorporating the physical scattering model into the Cycle-GAN framework, the network can learn to restore clear images from dust-affected ones in a more realistic and consistent manner.