A generative adversarial network incorporating dark channel prior loss used for single image defogging

Deqiang Cheng; Yangyang You; Qiqi Kou; Jinyang Xu

doi:10.12086/oee.2022.210448

Journals >Opto-Electronic Engineering >Volume 49 >Issue 7 >Page 210448 > Article

Opto-Electronic Engineering
Vol. 49, Issue 7, 210448 (2022)

A generative adversarial network incorporating dark channel prior loss used for single image defogging

Deqiang Cheng^1、2, Yangyang You², Qiqi Kou^3、*, and Jinyang Xu²

Author Affiliations

¹Engineering Research Center of Underground Space Intelligent Control, Ministry of Education, Xuzhou, Jiangsu 221000, China

²School of Information and Control Engineering, China University of Mining and Technology, Xuzhou, Jiangsu 221000, China

³School of Computer Science and Technology, China University of Mining and Technology, Xuzhou, Jiangsu 221000, China

show less

DOI: 10.12086/oee.2022.210448 Cite this Article

Deqiang Cheng, Yangyang You, Qiqi Kou, Jinyang Xu. A generative adversarial network incorporating dark channel prior loss used for single image defogging[J]. Opto-Electronic Engineering, 2022, 49(7): 210448 Copy Citation Text

show less

Fig. 1. Framework of adversarial generation network

Download full size | View in the Article

Fig. 2. Dark channel feature comparison.(a) Original images; (b) Dark channel feature

Download full size | View in the Article

Fig. 3. Dark channel feature intensity distribution. (a) Intensity distribution; (b) Average intensity distribution of 5000 images

Download full size | View in the Article

Fig. 4. Framework of the proposed algorithm

Download full size | View in the Article

Fig. 5. Qualitative comparison on synthetic images

Download full size | View in the Article

Fig. 6. Qualitative comparison on real hazy images

Download full size | View in the Article

Fig. 7. Quantitative comparison with control group on SOTS test-set & synthetic images of HSTS test-set

Download full size | View in the Article

Fig. 8. Qualitative comparison with control groups on real images of HSTS test-set

Download full size | View in the Article

Fig. 9. Quantitative comparison with control group on real hazy images of HSTS test-set

Download full size | View in the Article

卷积层	Conv	Res	Conv	Res	Conv	Res	Upconv	Res	Upconv	Res	Conv	Tanh
输入通道数	3	64	64	128	128	256	256	128	128	64	64	3
输出通道数	64	64	128	128	256	256	128	128	64	64	3	3
卷积核尺寸	7		5		3		4		4		7
步长	1		2		2		2		2		1
边界填充	3		1		1		1		1		3

Table 1. Parameters of generator network

View in the Article

卷积层	Conv1	Conv2	Conv3	Conv4	Conv5
输出通道数	64	128	256	512	1
卷积核大小	4	4	4	4	4
步长	2	2	2	1	1

Table 2. Parameters of the discriminator network

View in the Article

数据集	HSTS		SOTS-outdoor		SOTS-indoor
评价指标	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
DCP^[7]	17.22	0.80	17.56	0.82	20.15	0.87
BCCR^[9]	15.09	0.74	15.49	0.78	16.88	0.79
CAP^[10]	21.54	0.87	22.30	0.91	19.05	0.84
MSCNN^[15]	18.29	0.84	19.56	0.86	17.11	0.81
D-Net^[16]	24.49	0.92	22.72	0.86	21.14	0.85
AOD-Net^[17]	21.58	0.92	21.34	0.92	19.38	0.85
GFN^[18]	22.94	0.87	21.49	0.84	22.32	0.88
DEnergy^[26]	24.44	0.93	24.08	0.93	19.25	0.83
本文算法	25.35	0.96	25.17	0.96	23.70	0.82

Table 3. Quantitative results of each algorithm on SOTS test-set & synthetic images of HSTS test-set

View in the Article

数据集	D-HAZY		HazeRD		BeDDE
评价指标	PSNR	SSIM	PSNR	SSIM	VSI	VI	RI
DCP^[7]	15.09	0.83	14.01	0.39	0.946	0.911	0.965
MSCNN^[15]	13.57	0.80	15.58	0.42	0.947	0.892	0.972
D-Net^[16]	13.76	0.81	15.53	0.41	0.952	0.890	0.972
AOD-Net^[17]	13.13	0.79	15.63	0.45	0.954	0.896	0.970
CycleGAN^[22]	13.55	0.77	15.64	0.44	0.942	0.866	0.961
RefineDNet^[39]	15.44	0.83	15.61	0.43	0.960	0.907	0.971
SM-Net^[24]	15.32	0.81	15.55	0.40	0.961	0.899	0.969
本文算法	15.39	0.82	15.59	0.44	0.967	0.899	0.967

Table 4. Quantitative results of each algorithm on D-HAZY & HazeRD & BeDDE test-set

View in the Article

		Pic1	Pic2	Pic3	Pic4	Pic5	Pic6	Pic7	Pic8	Pic9	Pic10
对照组1	e	0.0162	-0.0506	0.0868	0.1629	0.3869	0.3391	0.6262	1.0014	0.9579	0.0160
	r	1.0100	1.0435	1.1291	1.0450	1.6743	1.4286	1.4641	1.4606	1.8278	1.0928
	p	0.0000	0.0000	0	0.0019	0	0.0002	0	0.0002	0.0003	0.0000
对照组2	e	0.3360	0.6421	0.2709	0.0774	0.3833	0.4338	0.7845	1.8348	0.6207	0.4650
	r	1.4605	1.4222	1.2640	1.2047	1.6926	1.4382	1.3379	1.5126	1.6021	1.1351
	p	0.0102	0.1155	0.0001	0.0081	0	0.0072	0.0002	0.0081	0.0016	0.0286
本文算法	e	0.3534	0.6061	0.8870	0.1232	0.4504	0.6265	1.2378	2.0714	1.1443	0.5197
	r	1.6284	1.5576	1.6631	1.3572	2.1599	1.6863	1.5793	1.7814	1.9561	1.3026
	p	0.0082	0.1015	0.0001	0.0020	0	0.0079	0.0011	0.0108	0.0013	0.0153

Table 5. Quantitative results of the control groups & proposed algorithm on real images of HSTS test-set

View in the Article

	DCP^[7]	CAP^[10]	MSCNN^[15]	D-Net^[16]	AOD-Net^[17]	GFN^[18]	CycleGAN^[22]	本文算法
Device	CPU	CPU	GPU	CPU	GPU	GPU	GPU	GPU
Run time	1.74	0.72	3.41	3.33	0.56	6.10	2.96	1.37

Table 6. Run time of each algorithm on SOTS test-set

Download Citation

Save the article for my favorites

Paper Information