Robust Pixel-wise Dehazing Algorithm based on Advanced Haze-Relevant Features

Guisik Kim and Junseok Kwon

Abstract

The dehazing algorithm aims to remove haze from an image and has been widely used as a pre-processing step in several computer vision applications. The performance of the computer vision algorithm, however, is significantly affected by weather conditions. Hence, conventional algorithms do not work well when weather conditions vary severely over time. This paper proposes an effective haze removal algorithm based on a single image, which is robust to the varying weather conditions. Unlike conventional methods, which estimate "global" atmospheric light, we find the "local" atmospheric light by assuming that it is absorbed and emitted differently at each atmospheric particle. We divide the image into two components, namely, illumination and reflection, according to the retinex theory, and obtain the pixel-wise atmospheric light from the illumination. To estimate an accurate transmission, we initialize it by using a combination of haze-relevant features that have been proven experimentally to be highly correlated with haze. The aforementioned processes interact with each other in producing accurate dehazing results. We solve the dehazing problem via convex optimization and obtain the optimal solution. The experimental results demonstrate that the proposed algorithm outperforms state-of-the-art methods, using the benchmark dataset with regard to contrast, detail, and visibility.

Session

Posters

Files

PDF iconPaper (PDF)
PDF iconSupplementary (PDF)

DOI

10.5244/C.31.79
https://dx.doi.org/10.5244/C.31.79

Citation

Guisik Kim and Junseok Kwon. Robust Pixel-wise Dehazing Algorithm based on Advanced Haze-Relevant Features. In T.K. Kim, S. Zafeiriou, G. Brostow and K. Mikolajczyk, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 79.1-79.12. BMVA Press, September 2017.

Bibtex

            @inproceedings{BMVC2017_79,
                title={Robust Pixel-wise Dehazing Algorithm based on Advanced Haze-Relevant Features},
                author={Guisik Kim and Junseok Kwon},
                year={2017},
                month={September},
                pages={79.1-79.12},
                articleno={79},
                numpages={12},
                booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
                publisher={BMVA Press},
                editor={Tae-Kyun Kim, Stefanos Zafeiriou, Gabriel Brostow and Krystian Mikolajczyk},
                doi={10.5244/C.31.79},
                isbn={1-901725-60-X},
                url={https://dx.doi.org/10.5244/C.31.79}
            }