A deep bilateral learning based GAN for non-homogeneous dehazing
Peizhe Li, Weiwei Kong, Jiamei Dai · 2023
Affected by haze, images often face color distortion, resolution reduction and other image quality degradation problems. The existing dehazing methods based on convolutional neural network(CNN) often perform well on large-scale synthetic datasets, but lack robustness in the processing of real haze images. This is because haze images in reality are often non-homogeneous. Due to the fact that the haze texture faced by the real haze processing is more complex, it is easier to destroy the texture details. Meanwhile, the paired training images are difficult to collect, and the small-scale data set is easy to lead to overfitting. To address these challenges, we propose a dehazing approach based on ensemble learning, DB-GAN, which uses Res2Net pre-trained by ImageNet as the encoder in the knowledge adaptation branch to improve the generalization ability of the network and avoid overfitting. In the data fitting branch, deep bilateral learning is used to learn the structure of the features from the full-resolution and low-resolution inputs, respectively, to better learn the color features and boundary features. We then map the different features by a fusion tail. Finally, we demonstrate the effectiveness of our approach through extensive experimental results.