Unsupervised single image dehazing with generative adversarial network

Abstract Most recent learning algorithms for single image dehazing are designed to train with paired hazy and corresponding ground truth images, typically synthesized images. Real datasets can help improve performance, but tough acquire. This paper proposes an unsupervised algorithm based on GAN alleviate this issue. An end-to-end network architecture is established fed unpaired clean signifying that the estimation of atmospheric light transmission not required. The proposed consists three parts: a generator, global test discriminator, local context discriminator. Moreover, dark channel prior attention mechanism applied handle inconsistency haze. We conduct experiments RESIDE datasets. Extensive demonstrated effectiveness approach which outperformed previous state-of-the-art methods by large margin.