Multi-Scale Feature Fusion Dehazing Network Based on U-net

Fig. 2. Comparison before and after adding large kernel convolution. (a) Feature map before adding large kernel convolution; (b) feature map after adding large kernel convolution; (c) restored result before adding large kernel convolution; (d) restored result after adding large kernel convolution; (e) clear image

Download full size

Fig. 3. Large kernel convolution structure with dynamic weighting mechanism

Download full size

Fig. 4. Coordinate attention structure

Download full size

Fig. 5. Comparison of results of different algorithms on SOTS dataset

Download full size

Fig. 6. Comparison of results of different algorithms on HSTS dataset

Download full size

Fig. 7. Comparison of results of different algorithms on NH-HAZE dataset

Download full size

Fig. 8. Comparison of results of different algorithms on BeDDE dataset

Download full size

Fig. 9. Comparison of subjective results of ablation experiments

Download full size

Table 1. Experimental environment
View table
Table 1. Experimental environment
Hardware device Version environment
CPU i7-12700K
GPU NVIDIA GeForce RTX 3090
Memory 24 Gbit
Language Python 3.8
Framework PyTorch 1.10.0
CUDA 11.3
System Windows 10

Table 2. Objective metrics for different algorithms on SOTS dataset

View table

Table 2. Objective metrics for different algorithms on SOTS dataset

Algorithm	Indoor				Outdoor
Algorithm	PSNR /dB	SSIM	BRISQUE	DHQI	PSNR /dB	SSIM	BRISQUE	DHQI
DCP	16.52	0.6998	33.17	48.04	18.79	0.6912	35.79	46.92
AOD-Net	21.07	0.7950	27.17	50.13	21.81	0.7956	29.16	50.17
FFA-Net	34.89	0.8830	30.37	57.13	32.16	0.8798	25.41	55.33
RIDCP	21.29	0.8097	31.05	52.37	24.36	0.8145	27.37	52.06
Dehaze-UNet	29.31	0.8360	29.18	56.92	28.06	0.8894	30.17	56.31
MixDehazeNet	35.72	0.9120	25.71	53.19	35.18	0.8719	29.49	55.33
DEA-Net	37.81	0.9330	26.83	58.73	38.17	0.9236	26.33	59.01
Ours	40.63	0.9449	23.78	62.81	40.44	0.9443	24.17	60.17

Table 3. Objective metrics for different algorithms on HSTS and NH-HAZE datasets

View table

Table 3. Objective metrics for different algorithms on HSTS and NH-HAZE datasets

Algorithm	HSTS				NH-HAZE
Algorithm	PSNR /dB	SSIM	BRISQUE	DHQI	PSNR /dB	SSIM	BRISQUE	DHQI
DCP	15.17	0.6988	32.49	45.38	13.72	0.6178	39.28	43.72
AOD-Net	19.83	0.7316	28.71	48.77	16.98	0.6923	30.14	42.17
FFA-Net	31.33	0.7424	27.47	52.09	19.59	0.7137	27.38	52.39
RIDCP	25.89	0.8147	26.43	54.41	23.92	0.7814	29.18	50.25
Dehaze-UNet	35.71	0.8727	27.83	55.21	21.71	0.8326	33.42	53.22
MixDehazeNet	33.12	0.8951	23.22	50.19	24.16	0.8542	30.25	50.33
DEA-Net	36.79	0.9145	25.46	56.24	25.05	0.8933	29.17	54.16
Ours	38.01	0.9306	21.06	58.71	27.33	0.9122	24.73	57.43

Table 4. Objective metrics for different algorithms on BeDDE dataset
View table
Table 4. Objective metrics for different algorithms on BeDDE dataset
Algorithm BeDDE
VI RI
DCP 0.805 0.843
AOD-Net 0.785 0.792
FFA-Net 0.836 0.736
RIDCP 0.845 0.873
Dehaze-UNet 0.833 0.816
MixDehazeNet 0.816 0.839
DEA-Net 0.874 0.862
Ours 0.911 0.942

Table 5. Comparison of objective results of ablation experiments
View table
Table 5. Comparison of objective results of ablation experiments
Model DLKC PA₁ PA₂ PA₁+PA₂ PSNR /dB SSIM
1 39.36 0.9358
2 √ √ 39.93 0.9386
3 √ √ 39.77 0.9377
4 √ √ 40.63 0.9449

Table 6. Objective comparison of loss functions in ablation experiments
View table
Table 6. Objective comparison of loss functions in ablation experiments
L₁ L₂ L_SSIM PSNR /dB SSIM
√ 39.96 0.9398
√ 40.08 0.9411
√ √ 40.48 0.9425
√ √ 40.63 0.9449

Tools

Get Citation

Copy Citation Text

Qianyu Dong, Qiuxiang Yang, Yin Zhao. Multi-Scale Feature Fusion Dehazing Network Based on U-net[J]. Laser & Optoelectronics Progress, 2025, 62(6): 0637013

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites