Image Dehazing Algorithm Based on Attention Mechanism and Markov Discriminant

Kezheng Lin; Jiahao Geng; Weiyue Cheng; Ao Li

doi:10.3788/LOP202259.1610003

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 16 >Page 1610003 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 16, 1610003 (2022)

Image Dehazing Algorithm Based on Attention Mechanism and Markov Discriminant

Kezheng Lin¹, Jiahao Geng^1、*, Weiyue Cheng², and Ao Li¹

Author Affiliations

¹School of Computer Science and Technology, Harbin University of Science and Technology, Harbin 150080, Heilongjiang , China

²Heilongjiang College of Business and Technology, Harbin 150025, Heilongjiang , China

show less

DOI: 10.3788/LOP202259.1610003 Cite this Article Set citation alerts

Kezheng Lin, Jiahao Geng, Weiyue Cheng, Ao Li. Image Dehazing Algorithm Based on Attention Mechanism and Markov Discriminant[J]. Laser & Optoelectronics Progress, 2022, 59(16): 1610003 Copy Citation Text

show less

Fig. 1. Attention module

Download full size

Fig. 2. Diagram of Inception mechanism

Download full size

Fig. 3. Network framework

Download full size

Fig. 4. Predictive transmission network

Download full size

Fig. 5. Image processing process

Download full size

Fig. 6. Comparison of indoor PSNR

Download full size

Fig. 7. Comparison of indoor SSIM

Download full size

Fig. 8. Comparison of outdoor PSNR

Download full size

Fig. 9. Comparison of outdoor SSIM

Download full size

Fig. 10. Comparison of algorithms in outdoor. (a) Foggy images; (b) CAP results; (c) DCP results; (d) Dehaze results; (e) MSCNN results; (f) results of proposed algorithm

Download full size

Fig. 11. Comparison of PSNR

Download full size

Fig. 12. Comparison of SSIM

Download full size

algorithm 1 A&P-dehaze algorithm

input：Foggy image

output：Defogging image

1）Input foggy imageI；

2）Use formula（7） to extract shallow features and get feature map $F_{s}$ ；

3）First， $F_{s}$ is downsampled by residual network，and attention mechanism is used to allocate weight $F_{d} = H_{A M} [F_{s} + F (F_{s}, W_{l})]$ ， $H_{A M}$ is the whole function of attention mechanism， $F (F_{s}, W_{l})$ is the residual；

4）Then the deconvolution residual network is used for upsampling to get $F_{b}$ ， $F_{b} = F_{d} + F (F_{d}, W_{l})$ ；

5）The transmittance map is obtained by using the mapping function， $t = H_{M A P} (F_{b})$ ；

（6）Using Inception module，the atmospheric light value of foggy image is predicted $A = F_{I n c, n} [. . . F_{I n c, 1} (I)]$ ；

（7）The defogging image $J$ can be obtained by using the atmospheric scattering model，

$J (x) = \frac{I (x) - A [1 - t (x)]}{t (x)}$ ；

8）Using PantchGAN to judge whether it is true or false；

9）Further training the network，repeat formula（8） until the loss function of the network is optimal，and the training is completed；

10）Save the optimal model.

Table 0. [in Chinese]

Algorithm	SSIM	PSNR
CAP	0.8524 0.8705 0.8756 0.8069 0.8764	18.96 18.97 21.34 17.12 20.86
DCP
Dehaze
MCSNN
Proposed algorithm

Table 1. Comparison results on SOTS dataset

Algorithm	SSIM	PSNR
CAP	0.7859 0.8095 0.8886 0.8632 0.8938	18.24 15.99 22.94 19.61 22.36
DCP
Dehaze
MCSNN
Proposed algorithm

Table 2. Comparison results on HSTS dataset

Algorithm	CAP	DCP	Dehaze	MSCNN	Proposed algorithm
Time /s	1.42	9.86	1.78	1.70	0.93

Table 3. Comparison of average running time of different algorithms

Kezheng Lin, Jiahao Geng, Weiyue Cheng, Ao Li. Image Dehazing Algorithm Based on Attention Mechanism and Markov Discriminant[J]. Laser & Optoelectronics Progress, 2022, 59(16): 1610003

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information