Learning to remove sandstorm for image enhancement

In the dusty weather, the suspended sand particles in the air lead to various degrees of degradation for the captured images. Commonly, severe color deviations, low contrast, and blurring details arise in the images. In this paper, a method equipped with color correction and an unsupervised learning network is proposed to enhance sandstorm images. First, the color correction is carried out by color compensation followed by the radical histogram equalization which enables the recovery of the natural color of an image and relieves the pressure of subsequent network that focuses on the enhancement of details. Second, the three sub-networks including atmospheric light with accurate initialization, transmission map, and clear image estimation layers are introduced to remove haze-like effects and reconstruct the clear image by unsupervised learning, which solves the lacking of paired sandstorm images and clear counterparts. To perform a comprehensive study of the state-of-the-art sandstorm image enhancement methods, we propose a Sandstorm Image Enhancement (SIE) dataset which can benchmark the performance of different methods and make it possible to train sandstorm image enhancement networks. The experimental results demonstrate that our method outperforms several state-of-the-arts both qualitatively and quantitatively.