Transforming Dust Storms into Clean on Mars Images Via Self Supervised Learning

The frequent occurrence of dust storms on the surface of Mars can seriously affect the imaging quality of Mars optical images, leading to incorrect detection and recognition of surface obstacles such as impact craters and rocks, which posing a serious threat to the normal landing and operation of Mars exploration vehicles. Therefore, research on removing dust storms from Mars images is of great significance for Mars exploration. However, there is seldom research in this field currently. The motion of the Mars vehicle results in a lack of paired clear and dust images under a unified landscape. Unpaired image dust storm removal is a challenging problem. We proposed a dust storm removal algorithm for Mars images. Based on the dehazing algorithm and atmospheric scattering model, we designed neural networks to predict the scattering coefficient, transmission map, and depth map. In addition, we designed a Cycle-constrain framework to attain efficient dust removal for unpaired clear and dust images aspired by CycleGAN. We compared our method with existing methods on the Mars Curiosity dataset and achieved superior performance.