End-to-end image dehazing by joint atmospheric scattering and WGAN model
Hongguo Zhang, Qingbing Sang · 2023
The performance of most existing dehazing methods are limited by the independency of the transmission map estimation and atmospheric light. To ameliorate this, we present an algorithm based on both GANs models and atmospheric scattering. We first enhance the accuracy of transmission map estimation by constructing a multi-scale FC-DenseNet transmission map prediction network. To effectively avoids the issue of calculation errors caused by inaccurate prediction of the transmission map, a shallow residual dehazing network is also developed, which shares the learned weights of the transmission map prediction network. A multi-task loss function is constructed that dynamically optimizes the two networks simultaneously. To augment the generation of realistic images, a classical WGAN is used to fine-tune the generated transmission maps and dehazed images. The results demonstrate that our method has achieved significant improvements compared with the state-of-the-art (SOTA) methods.