A Generative Net for Haze Removal in UAV's Visual Fieldwork
Xiaolong Zhang, Zhihong Peng, Xiuxiang Jiang · 2018
Haze removal is a challenging problem for UAV's visual fieldwork where clear images are always needed. Estimating the medium transmission map of a given hazy image is the key to achieve dehazing. Inspired by the conditional generative adversarial nets, we consider the medium transmission map as a probability distribution based on the hazy image and propose a simple but effective deep network to generate it. After training this network on our own collected data set, we can directly estimate the medium transmission map from a single input image of any size and recover the corresponding haze-free one via atmospheric scattering model. Experiments on pair-wise hazy, haze-free images and real-world scenes show that our method is superior to existing models in performance. And analysis of effect of haze removal on object detection is carried out.