Color-Constrained Dehazing Model

Shengdong Zhang, Yue Wu, Yuanjie Zhao, Zuomin Cheng, Wenqi Ren; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 870-871

In this paper, we address the insufficiency of the popular atmospheric scattering model (ASM) used in the image dehazing problem. Unlike ASM assumes the global uniform atmospheric light and attenuation coefficients and thus often introduce unrealistic color after dehazing, we propose a novel dehazing model by relaxing the global uniform atmospheric assumption to local with additional color constraints to ensure more appealing and realistic dehazed results. More precisely, we make the modeling process as an optimization problem, whose cost function is composed of color constraint, local smooth of transmission map and atmospheric light. Consequently, we are able to generate more realistic dehazed images comparing to ASM, implying that deep neural networks trained with these samples could effectively learn how to dehaze images of complicated cases, especially when the global atmospheric assumption fails. Our extensive experimental studies also confirm that the proposed dehazing model outperforms the state-of-the-art methods by a noticeable margin on all three public benchmarks including HazeRD, RESIDE, and O-HAZE in terms of SSIM and PSNR.