Weakly Supervised Image Dehazing Using Generative Adversarial Networks
Kavinder Singh, Vishruth Khare, Vishwas Agarwal, Sourabh · 2022
With the increase in air pollution, the quality of captured images has degraded to a large extent. This makes removing haze a prerequisite for a lot of scientific applications like monitoring traffic in real-time, celestial observation, and different sorts of tracking systems. Few methods like prior based have already been proposed and have demonstrated some success. However, they are not truly applicable in real-life applications because of the artifacts they produce in the outputs. Along with this, there is a dearth of paired images of hazy and cleared images; due to this, their testing capabilities are not completely tested. We have attempted to merge the advantages of the learning-based and prior-based models. Our model comprises three steps; first, it uses gamma correction to restore visibility; second, it takes advantage of adversarial learning to improve image realness; and lastly, a fusion block is used to blend different dehazing outputs. Extensive experimentations prove that our model outperforms most state-of-the-art dehazing models and produces visually pleasing dehazed images.