Two-Branch Feature Fusion Image Dehazing Algorithm Under Brightness Constraint

Fig. 3. Data comparison before and after Gammar correction. (a) Comparison of brightness distribution of hazy images in datasets before and after Gamma correction; (b) comparison of brightness distribution of standard fog free images in datasets before and after Gamma correction; (c) comparison example of hazy images before and after Gamma correction; (d) comparison of standard fog free images before and after Gamma correction

Download full size

Fig. 4. Partial experimental results of SOTS dataset. (a) Hazy images; (b) DCP; (c) CAP; (d) DehazeNet; (e) AOD-Net; (f) GCANet; (g) FFA; (h) TBD; (i) proposed algorithm; (j) standard fog free images

Download full size

Fig. 5. Partial experimental results of DENSE-HAZE dataset. (a) Hazy images; (b) DCP; (c) CAP; (d) DehazeNet; (e) AOD-Net; (f) GCANet;(g) FFA; (h) TBD; (i) proposed algorithm; (j) standard fog free images

Download full size

Fig. 6. Partial experimental results of NH-HAZE 2020 dataset. (a) Hazy images; (b) DCP; (c) CAP; (d) DehazeNet; (e) AOD-Net; (f) GCANet;(g) FFA; (h) TBD; (i) proposed algorithm; (j) standard fog free images

Download full size

Fig. 7. Partial experimental results of NH-HAZE 2023 dataset. (a) Hazy images; (b) DCP; (c) CAP; (d) DehazeNet; (e) AOD-Net; (f) GCANet; (g) FFA; (h) TBD; (i) proposed algorithm; (j) standard fog free images

Download full size

Fig. 8. Partial results of image rocovery on real no-reference hazy images of NH-HAZE 2021 dataset. (a) Hazy images; (b) DCP; (c) CAP; (d) DehazeNet; (e) AOD-Net; (f) GCANet; (g) FFA; (h) TBD; (i) proposed algorithm

Download full size

Table 1. Details of real haze datasets used in experiment

View table

Table 1. Details of real haze datasets used in experiment

Datasets	Training set	Validation set	Test set	Image size /（pixel×pixel）
DENSE-HAZE	45	5	5	1600×1200
NH-HAZE 2020	45	5	5	1600×1200
NH-HAZE 2021	25	5（w/o GT）	5（w/o GT）	1600×1200
NH-HAZE 2023	40	5（w/o GT）	5（w/o GT）	6000×4000

Table 2. Quantitative comparison between proposed method and five ablation methods

View table

Table 2. Quantitative comparison between proposed method and five ablation methods

Algorithm	ImageNet pre-training	Brightness correct	PSNR /dB（↑）	SSIM（↑）
1）Res2Net+DFB	√	√	21.38	0.6955
2）Swin Transformer+DFB	√	√	22.25	0.6841
3）ConvNeXt+DFB	√	√	22.84	0.7220
4）ConvNeXt SK DFB	×	√	18.52	0.6631
5）ConvNeXt SK DFB	√	×	22.51	0.7250
6）Proposed algorithm	√	√	22.93	0.7293

Table 3. Results comparison of loss function ablation experiment
View table
Table 3. Results comparison of loss function ablation experiment
Group $L_{l 1}$ $L_{l 1, a}^{}$ $L_{p e r c}$ $L_{M S - S S I M}$ $L_{a d v}$ PSNR /dB（↑） SSIM↑
1 √ √ √ √ √ 22.93 0.7293
2 √ × √ √ √ 22.87 0.7293
3 √ √ √ √ × 22.79 0.7289
4 √ √ √ × × 22.60 0.7259
5 √ √ × × × 22.44 0.7130

Table 4. Quantitative comparison over synthetic and real haze datasets for different methods

View table

Table 4. Quantitative comparison over synthetic and real haze datasets for different methods

Algorithm	SOTS		DENSE-HAZE		NH-HAZE 2020		NH-HAZE 2021		NH-HAZE 2023
Algorithm	PSNR /dB	SSIM	PSNR /dB	SSIM	PSNR /dB	SSIM	PSNR /dB	SSIM	PSNR /dB	SSIM
DCP	17.05	0.8626	10.01	0.4127	12.26	0.4521	11.49	0.6377	10.89	0.4984
CAP	18.27	0.7970	11.38	0.4302	13.01	0.4289	12.63	0.6437	10.77	0.5319
DehazeNet	22.38	0.8770	11.73	0.4382	12.82	0.4520	12.84	0.6493	10.96	0.5047
AOD-Net	22.14	0.8942	12.69	0.4526	14.81	0.4619	16.29	0.6840	14.57	0.5505
GCANet	30.23	0.9652	12.46	0.4323	14.01	0.4803	12.66	0.6231	11.73	0.5194
FFA	35.32	0.9834	16.27	0.5308	17.60	0.6218	19.87	0.7920	18.23	0.5840
TBD	34.86	0.9848	16.36	0.5820	21.53	0.6955	21.66	0.8421	20.10	0.7323
Proposed	35.01	0.9901	16.78	0.5886	22.93	0.7293	21.01	0.8798	25.11	0.7790

Table 5. Quantitative comparison of dehazing results on real no-reference hazy images for each algorithm

View table

Table 5. Quantitative comparison of dehazing results on real no-reference hazy images for each algorithm

Metrics	DCP	CAP	DehazeNet	AOD-Net	GCANet	FFA	TBD	Proposed algorithm
$e ↑$	2.450	1.448	1.936	2.295	3.274	0.911	3.427	3.552
$\bar{r} ↑$	1.999	1.511	1.758	1.815	2.036	1.361	4.258	4.515
$σ ↓$	0.098	0.005	2.622	1.666	0.067	0.007	0.016	0.285
$F A D E ↓$	0.301	0.596	0.196	0.616	0.288	0.880	0.271	0.196

Tools

Get Citation

Copy Citation Text

Jinqing He, Xiucheng Dong, Xianming Xiang, Hongda Guo, Yaling Ju. Two-Branch Feature Fusion Image Dehazing Algorithm Under Brightness Constraint[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1837007

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: Jan. 2, 2024

Accepted: Feb. 5, 2024

Published Online: Sep. 14, 2024

The Author Email: Xiucheng Dong (dxc136@163.com)

DOI:10.3788/LOP240436

CSTR:32186.14.LOP240436

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Details of real haze datasets used in experiment

Table 1. Details of real haze datasets used in experiment

Table 2. Quantitative comparison between proposed method and five ablation methods

Table 2. Quantitative comparison between proposed method and five ablation methods

Table 3. Results comparison of loss function ablation experiment

Table 3. Results comparison of loss function ablation experiment

Table 4. Quantitative comparison over synthetic and real haze datasets for different methods

Table 4. Quantitative comparison over synthetic and real haze datasets for different methods

Table 5. Quantitative comparison of dehazing results on real no-reference hazy images for each algorithm

Table 5. Quantitative comparison of dehazing results on real no-reference hazy images for each algorithm