Summary
An unsupervised conditional disentanglement network (UCDN) using unpaired datasets is proposed . Our method strengthens the constraints by introducing physics-based disentanglement . Unlike other unsupervised dehazing models, our method adapts to multiple concentrations of fog and performs well on datasets with different concentrations.
contribute:
1. The single image dehazing of CycleGAN is improved by adding a conditional disentangling network , which outperforms existing CycleGAN-based dehazing models on synthetic datasets.
2. A challenging large-scale natural image dehazing dataset is collected , which contains more than 3500 foggy images and 4000 fog-free outdoor scene images. Furthermore, according to different densities, we divide the hazy images into 3 subsets : light, medium and dense
Section 2.2:
In the dehazing problem, a haze-free image can be described as a superposition of hazy images of different concentrations, which explains why traditional CycleGAN [13] performs poorly on multi-concentration datasets.