A semi-supervised video dehazing method based on CNNs

Recently, image dehazing methods based on deep learning have achieved good results, 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, but it is difficult to get the corresponding haze-free images. So the image dehazing for real hazy images is a very important and challenging problem due to the absence of the corresponding haze-free images. In this paper, we propose an unpaired image dehazing method to solve the problem of unpaired real hazy images, which includes two parts: one is the unpaired training module and the other one is the supervised training module. For the unpaired training module, we first use cycle generative adversarial network to generate the dehazed images; then we use perceptual loss to make the features of the dehazed images and the unpaired haze-free images more consistency to improve the image dehazing performance; Finally, we propose to introduce instinct properties of clear images to further improve the image dehazing performance. For the supervised training module, we design a pseudo labeling scheme to train the model by using the dehazed images generated from the unpaired training module as pseudo haze-free images, so that we can conduct supervised training to further improve the performance of haze removal. Besides, we dedicate to collect a real hazy images dataset from real high-speed railway scenarios, named HRHI (High-speed Railway Hazy Image) dataset, which contains real hazy images and unpaired haze-free images. In order to verify the generality of our method, we also collect a real sandstorm images dataset from real high-speed railway scenarios, named HRSI (High-speed Railway Sandstorm Image) dataset. Expensive experimental results on HRHI dataset and the generic hazy image dataset show the effectiveness of our method compared with other state-of-the-art image dehazing methods, and the experimental results on HRSI dataset prove the generality of our method.