Research on Single Image Dehazing Enhancement Method Based on CycleGAN

Abstract

Although researchers have made great progress in image dehazing recently, there are still great challenges in balancing the suitable universality and dehazing accuracy. In this paper, we propose the dark channel prior cycle dehaze network (DCP-Cycle-Dehaze) to single image dehazing. This network is based on CycleGAN, which adds DCP loss based on dark channel prior knowledge and improved cycle perceptual loss to achieve image dehazing function. DCP-Cycle-Dehaze mainly enhance the dehazing capacity of model by enhancing the sensitivity of network for haze features during training. It further improves the performance of the CycleGAN network framework in image dehazing tasks, and makes the network still reach the accuracy of supervised training without unsupervised training. We conduct simulation experiments on four representative data sets: O-HAZE, I-HAZE, RESIDE and D-Hazy. The experimental results show that DCP-Cycle-Dehaze network we proposed has achieved very good results in outdoor environment, the results on the O-HAZE dataset exceed the best results of NTIRE2018; moreover, it also has better results on the indoor environment. The experimental results prove the effectiveness of our method from a quantitative and qualitative perspective.