An End-to-End Image Dehazing Method Based on Deep Learning

Image dehazing is a classic problem in computer vision. Most traditional methods use human-engineered features, such as dark channel prior, for dehazing. Recently, deep learning based approaches have been developed to solve this problem, but most of them are not end-to-end. In this paper, we propose an end-to-end learning method. This network consists of three parts to, respectively, estimate the transmission map, predict the global atmospheric light, and perform dehazing based on the estimated parameters. In order to train this network, we use Virtual KITTI dataset and NYU depth dataset to synthesize a training set composed of haze images and their corresponding transmission maps and global atmospheric light. Experiments demonstrate that our approach can obtain good performance on both synthetic and real haze images; moreover, the dehazed images have natural color and light contrast.