Enhancing Image Restoration in Hazardous Weather with Blender-Rendered Datasets: A Data Augmentation Approach

The image restoration task in hazardous weather conditions lacks training data. The real degraded images lack paired reference images, and the degraded images synthesized from clean images are not realistic enough, which easily leads to domain gap problems. To address these problems, we propose a data enhancement method based on domain randomization for image restoration tasks in bad weather scenarios. This method digitally models the real degraded scenes and generates degraded images and corresponding reference images that conform to physical laws and imaging principles in a graphical rendering manner. Randomized visual attributes are introduced into the modeling system to reduce the domain gap between simulated data and real data, thereby enhancing the robustness and generalization ability of the image restoration model. In addition, this method also uses the simulation system to provide the image restoration algorithm with accurate prior data such as depth maps and degradation information to assist the learning of the image restoration model. Experiments show that by studying the characteristics of degraded images in the synthetic dust dataset Dusty-Cityscapes, this method can improve the generalization and convergence ability of the degraded image restoration model, so that degraded images from two different domains, rendered degraded images and synthetic degraded images, can obtain similar restoration effects.