FA-GAN: a feature attention GAN with fusion discriminator for non-homogeneous dehazing

Abstract

Recently, convolutional neural network has achieved a lot of attention for image dehazing tasks. Many deep learning-based methods can solve the homogeneous dehazing problems well. However, even if a well-designed convolutional neural network (CNN) can perform well on large-scaled dehazing benchmarks, it usually fails in the non-homogeneous dehazing tasks. There are two main reasons for this. First of all, due to its non-homogeneous nature, the haze distribution is too complex to remove the haze in dense areas. Second, real non-homogeneous haze datasets are limited; for example, the NH-Haze 2020 dataset contains only 45 training pairs. Therefore, learning mappings from limited data is extremely difficult. In order to solve these problems, we introduce a simple and effective ensemble-learning non-homogeneous dehazing method. Specifically, we propose a two-branch generative adversarial network to deal with the above problems separately and then map their different features through a learnable fusion tail which consists of several convolutional layers. And in order to make the generator produce a more natural and realistic dehazing image with less color distortion and fewer artifacts, we introduced a fusion discriminator which takes frequency information as additional priors. We present a large number of experimental results to demonstrate the effectiveness of our proposed method.