Simplified non-locally dense network for single-image dehazing
- PDF / 3,233,600 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 1 Downloads / 238 Views
ORIGINAL ARTICLE
Simplified non-locally dense network for single-image dehazing Zhihua Chen1 · Zhuoliang Hu1 · Bin Sheng2
· Ping Li3 · Jinman Kim4 · Enhua Wu5,6
© Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Single-image dehazing is an ill-posed problem. Most previous methods focused on estimating intermediate parameters for input hazy images. In this paper, we propose a novel end-to-end Simplified Non-locally Dense Network (SNDN) which does not rely on intermediate parameters. To capture long-range dependencies, we propose a Simplified Non-local Dense Block (SNDB) which is lightweight and outperforms traditional non-local method. Our SNDB will be embedded into a densely connected encoder–decoder network. To avoid gradients vanishing problem, we propose a simple branch network which only have five convolution layers. The effectiveness of our proposed network is proved through ablation experiment. In addition, we enhanced our training set by synthesizing colored hazy images, which helps restore the original color of the hazy image. The experimental results demonstrate that our network have better performance than most of the pervious state-of-the-art methods. Keywords Single-image dehazing · Dense · Non-local
1 Introduction
B B
Bin Sheng [email protected] Enhua Wu [email protected] Zhihua Chen [email protected] Zhuoliang Hu [email protected] Ping Li [email protected] Jinman Kim [email protected]
1
Department of Computer Science and Engineering, East China University of Science and Technology, Shanghai, China
2
Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China
3
Department of Computing, The Hong Kong Polytechnic University, Kowloon, Hong Kong
4
Biomedical and Multimedia Information Technology Research Group, School of Information Technologies, The University of Sydney, Sydney, Australia
5
State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China
6
Faculty of Science and Technology, University of Macau, Macau, China
Single-image dehazing, as a low-level vision task, is an active research topic and highly desired in computer vision. Because of the complexity of atmospheric composition, removing fog of a single image is an extremely difficult task. Before convolutional neural networks (CNN) attract researchers’ attention, a simple atmospheric scattering model [21,23,24,27] has been adopted to describe the relationship between hazy image and haze-free image: I(x) = J(x)t(x) + A(1 − t(x))
(1)
where x is the pixel on the image, I(x) is the observed intensity and J(x) is the scene radiance, A is the global atmospheric light, and t(x) is the medium transmission describing the portion of the light that is not scattered and reaches the camera. When the atmosphere is homogenous, the transmission map t can be expressed as t(x) = e−βd(x) , where β is the scattering coefficient of the atmosphere and d(x) is the distance between observed object and camera. As Eq. 1 shows that parameters a
Data Loading...