Image Dehazing Using Conditional Patch Generative Adversarial Network
摘要：*Restoring clear images from hazy images is an ill-posed problem and most recent researches based on atmospheric scattering model are not work well when facing real fog pictures. We propose an algorithm based on the Conditional Generative Adversarial Network（CGAN）, which is an end-to-end training network that outputs the dehazing results directly. The generator is designed as a skip connection U-Net structure, so that it can generate better results.We further modify the basic CGAN formulation by introducing the gradient loss, features loss, reconstruction loss. Instead of scalar loss, a patch-GAN loss is used to regularize the network model.Haze4 K and STOS datasets are used to train and evaluate the network. The experimental results show that the algorithm performs competitive result and better generality against the other works.