Organizations: State Key Laboratory of Physical Oceanography and the Faculty of Information Science and Engineering, Ocean University of China, Qingdao, 266100. · School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen university, Shenzhen, China.
Underwater images often suffer from severe degradation caused by light absorption and scattering, leading to color distortion, low contrast and reduced visibility. Existing Underwater Image Enhancement (UIE) methods can be divided into two categories, i.e., prior-based and learning-based methods. The former rely on rigid physical assumptions that limit the adaptability, while the latter often face data scarcity and weak generalization. To address these issues, we propose a Physics-Semantics-Guided Underwater Image Enhancement Network (PSG-UIENet), which couples the Retinex-grounded illumination correction with the language-informed guidance. This network comprises a Prior-Free Illumination Estimator and a Semantics-Guided Image Restorer. In particular, the restorer leverages the textual descriptions generated by the Contrastive Language-Image Pre-training (CLIP) model to inject high-level semantics for perceptually meaningful guidance. Since multimodal UIE data sets are not publicly available, we also construct a large-scale image-text UIE data set, namely, LUIQD-TD, which contains 6,418 image-reference-text triplets. To explicitly measure and optimize semantic consistency between textual descriptions and images, we further design an Image-Text Semantic Similarity (ITSS) loss function. To our knowledge, this study makes the first effort to introduce both textual guidance and the multimodal data set into UIE tasks. Extensive experiments on our data set and four publicly available data sets demonstrate that the proposed PSG-UIENet achieves superior or comparable performance against fifteen state-of-the-art methods.
Figures & tables
Fig. 1: Comparison of three Retinex-based UIE methods, including Retinexformer [ 5 ] , RetinexMamba [ 6 ] , and our PSG-UIENet. In terms of each image, the Peak Signal to Noise Ratio (PSNR), Structural Similarity Index (SSIM) [ 7 ] , and Learned Perceptual Image Patch Similarity (LPIPS) [ 8 ] values computed between the corresponding reference image and it are shown at the top-left corner. These results highlight the effectiveness of the Retinex theory [ 9 ] in alleviating underwater image degradation.
Fig. 2: Sixteen degraded-reference-text triplets contained in the LUIQD-TD. Each triplet consists of three components: a degraded image (top-left) and the associated reference image (top-right) and textual description (bottom). The data set spans diverse underwater scenarios, including coral reefs, marine life, divers, submerged wrecks, and underwater vehicles, thereby offering rich semantic and visual information for multi-modal UIE tasks.
Fig. 3: Statistical analysis of the textual annotations in the LUIQD-TD, including (a) the distribution of word frequencies, (b) the distribution of caption lengths, and (c) the distribution of VQA-based [ 39 ] image–text similarity scores.
Fig. 4: The architecture of our PSG-UIENet, which comprises two modules: (a) a Prior-Free Illumination Estimator that generates multi-scale light-enhanced representations and (b) a Semantics-Guided Image Restorer that performs multi-modal alignment, fusion and enhancement using a dual-branch structure.
Fig. 5: The architecture of the Semantics-Guided Encoder-Decoder Network. This network is built on top of a symmetric encoder-decoder network, which consists of a series of Transformer-Conv layers for joint global-local feature extraction. In addition, the cross-modal attention mechanism and a Cross-Attention FiLM Module (CFM) integrate textual semantics, enabling progressive image reconstruction with semantic and visual fusion.
Method
Test-T622
Test-U80
Test-S110
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
UDCP [ 11 ]
11.51
47.60
26.94
9.51
33.66
41.74
10.07
34.29
37.66
Retinex [ 14 ]
10.56
58.10
39.12
10.50
61.31
36.49
8.32
53.46
40.41
HE [ 43 ]
16.01
71.30
28.07
16.60
77.86
25.67
15.41
74.79
30.05
HLRP [ 15 ]
13.00
23.83
33.11
13.56
22.76
33.90
12.55
29.24
33.18
MMLE [ 16 ]
17.07
72.11
25.41
18.56
76.21
22.57
17.32
77.39
22.77
TABLE I: The results of the full-reference quantitative evaluation of the proposed PSG-UIENet and 15 baselines on three real-world test sets. The best and second best results are highlighted in the Red Bold and Blue Italic fonts, respectively.
Method
Test-L622
Test-U80
Test-S110
Test-C60
Test-R53
PAUQA ↑
UIF ↑
PAUQA ↑
UIF ↑
PAUQA ↑
UIF ↑
PAUQA ↑
UIF ↑
PAUQA ↑
UIF ↑
UDCP [ 11 ]
36.71
55.91
29.73
36.40
29.92
32.01
28.19
25.41
33.07
3.24
Retinex [ 14 ]
36.83
46.41
34.38
49.83
31.79
39.43
29.20
49.85
37.52
2.61
HE [ 43 ]
41.32
45.10
37.82
44.04
34.26
45.01
34.69
41.50
34.95
1.67
HLRP [ 15 ]
36.83
0.57
38.12
0.59
37.33
0.91
30.16
3.57
28.53
1.26
MMLE [ 16 ]
45.26
37.88
43.49
30.99
43.15
35.05
37.06
24.95
33.15
24.67
TABLE II: The results of the non-reference quantitative evaluation of the proposed PSG-UIENet and 15 baselines on five real-world test sets. The best and second best results are highlighted in the Red Bold and Blue Italic fonts, respectively.
Fig. 6: The results produced by 15 baselines and our method in terms of a degraded image in the Test-L622 test set. Here, the PSNR, SSIM and LPIPS values, computed between a degraded or enhanced image and the reference image, are shown at the top-left corner of the image.
Fig. 7: The results produced by 15 baselines and our method in terms of a degraded image in the Test-U80 test set. Here, the PSNR, SSIM and LPIPS values, computed between a degraded or enhanced image and the reference image, are shown at the top-left corner of the image.
Fig. 8: The results produced by 15 baselines and our method in terms of a degraded image in the Test-S110 test set. Here, the PSNR, SSIM and LPIPS values, computed between a degraded or enhanced image and the reference image, are shown at the top-left corner of the image.
Fig. 9: The results produced by 15 baselines and our method in terms of a degraded image in the Test-C60 test set. Here, both the PAUQA and UIF values are shown at the top-left corner of the degraded image and enhanced images.
Fig. 10: The results produced by 15 baselines and our method in terms of a degraded image in the Test-R53 test set. Here, both the PAUQA and UIF values are shown at the top-left corner of the degraded image and enhanced images.
Method
Test-L622
PSNR ↑
SSIM ↑
LPIPS ↓
PAUQA ↑
UIF ↑
w/o IE
23.49
89.70
9.56
51.27
67.43
w/o IR
22.44
86.47
11.60
48.73
77.70
w/o TA
23.69
90.11
8.90
51.10
67.14
w/o CFM
23.62
89.73
9.21
50.78
67.67
Ours
24.07
90.19
9.11
51.34
67.70
TABLE III: Comparison between our method and its four variants obtained by removing a key component in PSG-UIENet.
Method
Test-L622
PSNR ↑
SSIM ↑
LPIPS ↓
PAUQA ↑
UIF ↑
w/o Text
23.66
89.93
9.21
51.48
67.41
w/ MHSA
23.25
89.52
9.61
51.03
70.03
Ours
24.07
90.19
9.11
51.34
67.70
TABLE IV: Comparison between our method and its two variants obtained by removing the text path or replacing the cross-attention mechanism with the MHSA mechanism.
θ
Test-L622
PSNR ↑
SSIM ↑
LPIPS ↓
PAUQA ↑
UIF ↑
0
23.82
89.93
9.19
51.33
67.23
0.25
22.08
88.78
9.73
50.34
67.55
0.5 (Default)
24.07
90.19
9.11
51.34
67.70
0.75
23.23
89.58
9.55
51.05
68.52
1.0
22.87
88.99
10.29
50.32
67.52
TABLE V: Comparison of five masking ratio ( θ ) values.
Shandong Technology and Business University, Yantai, China · Shandong University of Finance and Economics, Jinan, China · Dalian Maritime University, Dalian, China
College of Future Information Technology, Fudan University, Shanghai 200433, China · Institute of Artificial Intelligence (TeleAI), China Telecom, China · Faculty of Robot Science and Engineering, Northeastern University, Shenyang 110819, China +1