A new approach for training a physics-based dehazing network using synthetic images

In this study, we propose a new approach for training a physics-based dehazing network, using RGB images and depth maps gathered from a 3D urban virtual environment, with simulated global illumination and physically-based shaded materials. Since 3D scenes are rendered with depth buffers, full image depth can be extracted based on this information, using a custom shader, unlike the extraction of real-world depth maps, which tend to be sparse. Our proposed physics-based dehazing network uses generated transmission and atmospheric maps from RGB images and depth maps from the virtual environment. To make our network compatible with real-world images, we incorporate a novel strategy of using unlit image priors during training, which can also be extracted from the virtual environment. We formulate the training as a supervised image-to-image translation task, using our own DLSU-SYNSIDE (SYNthetic Single Image Dehazing Dataset), which consists of clear images, unlit image priors, transmission, and atmospheric maps. Our approach makes training stable and easier as compared to unsupervised approaches. Experimental results demonstrate the competitiveness of our approach against state-of-theart dehazing works, using known benchmarking datasets such as I-Haze, O-Haze, and RESIDE, without our network seeing any real-world images during training. The DLSU-SYNSIDE dataset and source code can be accessed through this link: https://neildg.github.io/SynthDehazing/. (c) 2022 Elsevier B.V. All rights reserved.