Underwater image enhancement based on multi-branch residual attention network

Name	Quantity of convolutional kernel	Size of convolutional kernel	Step length	Filling	Quantity of output feature map	Size of output feature map
Convolutional layer	64	3 $\times$ 3	1	1	64	256 $\times$ 256
MCEM	64	1 $\times$ 1	1	0	64	256 $\times$ 256
Encoder 1/2	128/256	4 $\times$ 4	2	1	128/256	128 $\times$ 128/64 $\times$ 64
Residual attention module 1/2	256/256	3 $\times$ 3	1	1	256/256	64 $\times$ 64
Decoder 1/2	128/64	4 $\times$ 4	2	1	128/64	128 $\times$ 128/256 $\times$ 256
MCEM	64	1 $\times$ 1	1	0	64	256 $\times$ 256
Convolutional layer	3	7 $\times$ 7	1	3	3	256 $\times$ 256

Table 2. Average scores of evaluation metrics for LUSI400 test set
View table
View in Article
Table 2. Average scores of evaluation metrics for LUSI400 test set
Method PSNR/dB SSIM
ULAP 17.813 0.750
UDCP 12.833 0.548
U-GAN 24.934 0.852
Water-Net 26.389 0.878
FUnIE-GAN 23.308 0.831
U-shape Transformer 26.221 0.836
Ours 27.420 0.885

Table 3. Average scores of evaluation metrics for EUVP515 test set
View table
View in Article
Table 3. Average scores of evaluation metrics for EUVP515 test set
Method PSNR/dB SSIM
ULAP 19.626 0.735
UDCP 14.519 0.565
U-GAN 23.481 0.813
Water-Net 25.316 0.840
FUnIE-GAN 24.180 0.803
U-shape Transformer 25.236 0.822
Ours 26.159 0.851

Table 4. Ablation experiments results of network structure
View table
View in Article
Table 4. Ablation experiments results of network structure
Method PSNR/dB SSIM
w/o MCEM 26.301 0.701
w/o RAM 23.987 0.647
w/o CBAM 26.684 0.734
Ours 27.420 0.885

Table 5. Average evaluation index values of loss functions with different weight coefficients
View table
View in Article
Table 5. Average evaluation index values of loss functions with different weight coefficients
Distribution of
weight coefficients
PSNR/dB SSIM
$λ_{1} = 0$ ， $λ_{2} = 0.7$ ， $λ_{3} = 0.3$ 26.453 0.736
$λ_{1} = 1$ ， $λ_{2} = 0$ ， $λ_{3} = 0.3$ 27.165 0.728
$λ_{1} = 1$ ， $λ_{2} = 0.7$ ， $λ_{3} = 0$ 26.636 0.803
$λ_{1} = 1$ ， $λ_{2} = 0.7$ ， $λ_{3} = 0.3$ 27.420 0.885

Table 6. Operational performance analysis of different models
View table
View in Article
Table 6. Operational performance analysis of different models
Method Time/s Parameter/M FLOPs
ULAP 0.174 - -
UDCP 2.475 - -
U-GAN 0.053 57.17 38.97G
Water-Net 0.680 24.81 193.73G
FUnIE-GAN 0.023 7.01 10.23G
U-shape Transformer 0.071 65.64 66.21G
Ours 0.042 8.46 51.71G

Tools

Get Citation

Copy Citation Text

Zhuming CHENG, Jiaxuan LI, San'ao HUANG, Lichao HAN, Peizhen WANG. Underwater image enhancement based on multi-branch residual attention network[J]. Optics and Precision Engineering, 2025, 33(7): 1141

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites