Black-Box Dehazing Network via Joint Loss Function

Single image dehazing is a notoriously challenging task in image processing due to the numerous unknown factors involved. Most existing dehazing methods are based on physical models of the haze formation process. However, real-world haze scenes cannot be accurately mathematically modeled due to the existence of various unquantifiable factors in the scene. Therefore, the dehazing methods based on physical models often perform poorly in complex haze scenes. In this paper, we propose a novel black-box dehazing equation. In this equation, the haze is modelled as an additional image interference layer, without explicitly reasoning about the physical model of haze formation. The dehazing process is modelled as removing this image interference layer. Based on this equation, we propose a novel network architecture called the Black-box Dehazing Network (BDN). Moreover, we propose a joint loss function for training this network. The joint loss function not only evaluates pixel-level differences between the dehazed image and the free image, but also measures differences in texture, color, and structure between the hazy image and its corresponding dehazed version as well as those between the hazy image and its haze-free version. In training, BDNet is only fed pairs of free images and their corresponding hazy images. The corresponding hazy patches are generated on-the-fly during network training. Experimental results demonstrate that the proposed method has the advantage of universality and outperforms existing state-of-the-art dehazing methods.