Generative adversarial dehaze mapping nets

Single image haze removal is a challenging task with few effective constraints, which seriously affect performance of machine learning algorithms. In this paper, we propose a Generative Adversarial Dehaze Mapping Nets (GADMN) to estimate a medium transmission for an input hazy image. GADMN adopts Generative Adversarial Nets (GAN) based deep architecture, which maps haze-relevant features to medium transmission and uses the network to carry on the feedback restrain. We also propose a multiple-light scattering model, which adds artificial light source and diffuses reflection light emerged from reflected light in the mist. Since the interference light is estimated in this model, we name it Local Multi-scale Hierarchical Prediction Method (LMHPM), which is beneficial to recover the large luminance range image. Experimental result demonstrates that the proposed algorithm outperforms state-of-the-art methods, and exhibits better robustness and adaptability.