Two Stage Dehazing Framework for Dense and Non-Homogeneous Dehazing

Wei Song, Yichang Gao, Jiahao Xiong, Hualiang Lin, Dong Li, Yun Zhang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 6722-6731

In real-world environments haze often causes a decrease in visibility leading to potentially severe consequences. Although current methods based on the assumption of homogeneous haze density have achieved commendable results dehazing techniques for non-homogeneous density haze still fall short in terms of visibility restoration and color accuracy. We observe that although single-stage methods have made significant strides a multi-stage enhancement can further improve dehazing in terms of both visibility and color restoration. In this paper we propose a two-stage dehaze framework named Two Stage Dehazing Framework. Our approach consists of a DehazeNet which does not require specifying a particular model for dehazing and can accept a hazy image as input producing an clear image of the same dimensions as the original. Two such DehazeNet are sequentially connected to form the final serial DehazeNet. Moreover to better approximate the output image to real-world scenes we propose the Multi-Scale Attention Head. Our method achieved third place in NTIRE 2024 Dense and NonHomogeneous Dehazing Challenge demonstrating outstanding performance metrics in the Peak Signal to Noise Ratio (PSNR) the Structral Similarity Index (SSIM) and the Mean Opinion Score (MOS). Related code will be acailable on \href https://github.com/sonwe1e/CVPRW2024Dehazing code .