Unpaired Image Dehazing Network using smoothed Dilated Convolution Network and Self-Supervised CycleGAN

The purpose of this study is to investigate methods restoring hazy images to haze-free images. Most dehazing studies have used datasets that consist of pairs of images, one hazy and one haze-free of the same scene, for training purposes. However, in the real world, it is almost impossible to acquire this kind of data where the hazy image and the haze-free image are perfectly matched except for the haze. Therefore, this paper aims to develop a network that removes haze using hazy images and images without haze that are not paired. The proposed model uses the CycleGAN architecture with this unpaired data. In order to improve the haze removal performance, we propose a dehazing model consisting of a smoothed dilated convolution, a perceptual loss function and a rotational loss function under self-supervised learning. For objective performance evaluation of the proposed techniques, we conducted experiments on the D-HAZY dataset and with real hazy images. The performance of the proposed method was demonstrated through qualitative and quantitative analysis.