Domain Randomization on Deep Learning Models for Image Dehazing

Abstract

Haze is a naturally occurring phenomenon that obstructs vision and affects the quality of images and videos. Recent literature has shown that deep learning-based image dehazing gives promising results both in terms of image quality and execution time. However, the difficulty of acquiring realworld hazy - clear paired images for training still remains a challenge. Widely available datasets use synthetically generated hazy images that suffer from flaws due to difficulty in acquiring accurate depth information to synthesize realistic-looking haze, causing a gap in the real and synthetic domain. In this paper, we propose the usage of domain randomization for image dehazing by generating a completely simulated training dataset for deep learning models. A standard UNET based dehazing model is trained on the simulated dataset without using any real-world data to obtain high quality dehazed images. The performance of the proposed approach is evaluated on the Sun-Dehaze dataset and RESIDE Standard (SOTS outdoor) dataset. We obtain favorable PSNR and SSIM scores on both sets and we also show how our approach yields better visual results compared to other learning-based approaches.