A semi-supervised video dehazing method based on CNNs

Abstract To solve the problems that no corresponding clear videos are available for real haze videos of high-speed comprehensive detection train and a supervised method cannot be used for training, a semi-supervised video dehazing method based on convolutional neural networks (CNNs) is proposed. First of all, this paper introduces the limitations of the existing dehazing algorithms and the ideas to solve such limitations. Secondly, the video dehazing model proposed is introduced, which consists of a dynamic haze generator and a dehazing device. In addition to the experimental design, the process of acquiring and constructing the railway video dataset, experimental loss constraint items, evaluation methods, and experimental configurations are presented. Specifically, the process of acquiring and constructing the railway video dataset is to process the video data acquired during the running process of high-speed comprehensive detection train into multiple scene materials according to the haze concentration, and to synthesize haze videos using a clear video haze adding tool. Finally, we compare the proposed method with the classical dehazing methods based on the synthetic and real haze video datasets quantitatively and qualitatively. Experimental results show that the video dehazing method proposed herein is very effective in dehazing