An image dehazing method based on multi-scale fusion
Dehazing is an important preprocessing technique in computer vision. It can make vision system adapt to different weather conditions. Traditional dark channel prior (DCP) based methods ignore the edge between the objects, which result in halo artifacts. To overcome the problem, this paper proposes a...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Science Press (China Science Publishing & Media Ltd.)
2024-09-01
|
| Series: | Shenzhen Daxue xuebao. Ligong ban |
| Subjects: | |
| Online Access: | https://journal.szu.edu.cn/en/#/digest?ArticleID=2683 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Dehazing is an important preprocessing technique in computer vision. It can make vision system adapt to different weather conditions. Traditional dark channel prior (DCP) based methods ignore the edge between the objects, which result in halo artifacts. To overcome the problem, this paper proposes a multi-scale fusion method to estimate the transmission map. Firstly, we obtain multi-scale transmission layers in a way of different sizes of minimum filter, and then apply multi-scale fusion technique to obtain the transmission map based on the local smooth assumption. Secondly, we obtain the global air-light by selecting the brightest pixel in the smallest transmission map area. Finally, we obtain the final dehazed result in a way of atmospheric scattering model. In the experiment, we compare the results of proposed method with those of different methods with respect to the visual effects and quantitative metrics to illustrate the effectiveness of the proposed method. The results show that in typical four haze images, by applying our proposed method, the visual effects of dehazed results are better than those of compared methods. In quantitative evaluation, we further compare PSNR and SSIM metrics of our method with that of traditional methods and deep learning methods in D-Hazy dataset. The results show that proposed methods can achieve higher scores than those of prior based and deep learning-based methods. |
|---|---|
| ISSN: | 1000-2618 |