Unpaired Image Dehazing for Real Hazy Images

Recently, image dehazing has achieved good performance benefiting from deep learning, but most of them are aimed at synthetic hazy images that have corresponding haze-free images. However, in some practical applications, the acquired images always contain haze in hazy weather, which makes it difficult to obtain pairs of haze-free images. Therefore, image dehazing for real hazy images is very important and challenging because there are no corresponding haze-free images. Therefore, we propose an unpaired image dehazing method to solve the dehazing problem of unpaired real hazy images, which contains the unpaired training module and the supervised training module. For the unpaired training module, we first use a cycle generative adversarial network to generate dehazed images. Then, we use perceptual loss to make the features of dehazed images and unpaired haze-free images more consistent to improve image dehazing performance. Finally, we use instinct properties of clear images to further improve image dehazing performance. For the supervised training module, we design a pseudo labelling scheme to train the model using dehazed images generated from the unpaired training module as pseudo haze-free images, so that we can conduct supervised training to further increase the performance of image dehazing. In addition, we collect a real hazy image dataset from real high-speed railway scenarios, named the HRHI (High-speed Railway Hazy Image) dataset, which contains real hazy images and unpaired haze-free images. To verify the generality of our method, we also collect a real sandstorm image dataset from real high-speed railway scenarios, named the HRSI (High-speed Railway Sandstorm Image) dataset. Expensive experimental results on the HRHI and generic hazy image datasets show the effectiveness of our method, and the experimental results on the HRSI dataset prove the generality of our method.