Learning deep transmission network for efficient image dehazing

Single image dehazing algorithms are recently attracting more and more attention from many researchers because of their flexibility and practicality. However, most existing algorithms have some challenges in dealing with images captured under complex weather conditions because the often used assumptions cannot always reflect true structural information of natural images in those situations. In this paper, we develop a deep transmission network to estimate the transmission map for efficient image dehazing, which automatically explores and exploits underlying haze-relevant features from RGB color channels and a local patch jointly for robust transmission estimation. Moreover, due to the fact that transmission values are affected by light wavelengths, a three-channel transmission map is considered in the proposed network so that this network can discover and utilize the chromatic characteristics for transmission estimation. We also investigate different network structures and parameter settings to achieve different trade-offs between performance and speed, and find that three color channels and local spatial information are the most informative haze-relevant features. This could explain why haze relevant priors or assumptions are often related to three color channels in most existing methods. Experiment results demonstrate that the proposed algorithm outperforms state-of-the-art methods on both synthetic and real-world datasets.