Prior Information Guidance for Image Dehazing
Fei Guo, Cong Han, Zhiyuan Wang, Jiaxu Wang · 2024
Restoration of images affected by severe weather conditions like heavy fog is indeed a current and significant topic within the field of computer vision. Despite the good performance of existing image dehazing methods, most rely on synthetic datasets due to the difficulty of obtaining real hazy images. These synthetic datasets lack the complexity found in real foggy conditions, leading to poor generalization of the dehazing algorithms. Moreover, existing end-to-end dehazing methods, which are often based on general frameworks for image inpainting or translation, do not adequately account for the unique characteristics of hazy images. This results in color distortion and incomplete dehazing with some fog remaining. To address these issues, we propose a novel image dehazing framework guided by prior information, comprising a physical prior module and an image prior module. Extensive experiments confirms that our approach achieves state-of-the-art performance across various dehazing datasets.