Single Image Dehazing Based on Enhanced Generative Adversarial Network

Abstract

Single image dehazing is a classical and challenging problem in computer vision. However, existing GAN-based methods focus on designing more complex network to achieve better performance, which makes it difficult to converge stably during training in the discriminator. In this paper, we propose an enhanced generative adversarial network for single image dehazing. Specifically, we introduce ResNet in the generative network, which enhances the ability of feature extraction. Then, we use DenseNet in the discriminative network to improve the feature learning ability. Finally, we use perceptual loss to reduce the generated image and the real clear image spatial differences in the feature domain. Qualitative and quantitative comparison against several state-of-the-art methods on synthetic datasets demonstrate that our approach is effective and performs favorably for recovering a clear image from a hazy image.