Visibility restoration for real-world hazy images via improved physical model and gaussian total variation

Under real-world haze conditions, the captured images not only suffer from the haze but also are affected by the noise, which significantly deteriorates the visibility of images. However, most of existing haze removal methods mainly focus on the haze degradation and fail to consider the noise interference.
To address the above issue, a research team led by Hailing XIONG and Yun LIU published their new research on 15 Feb 2024 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
The team proposed a novel unified variational model consisting of multiple effective constraints that simultaneously obtains the haze free image, the transmission map and the noise map. The proposed model can achieve both haze removal and noise suppression. Compared to existing research result, the proposed algorithm guarantees the visibility while suppressing the hidden noise.
In the research, they carefully devised a novel variational model which consists of total variation regularization term, Gaussian total variation regularization term and L2 norm regularization term to respectively constrain the scene radiance, the transmission map and the overall noise map. By adopting the re-weighted optimization strategy, the proposed variational model is solved to obtain the haze-free image. Compared to previous dehazing methods, our proposed unified variational model can achieve haze removal while suppressing the noise amplification.
The experiments are performed on real-world hazy images. Extensive experimental data demonstrate our proposed unified variational model can achieve superior dehazing effects, significantly improving the quality and visibility of real-world hazy images. In addition, our proposed model also shows the ability of noise suppression. In future work, we plan to explore the parameter-adaptive unified models or networks that can adaptively adjust the parameters or automatically learn the model parameters, thereby enhancing the robustness of the model.
DOI: 10.1007/s11704-023-3394-0