Organizations: School of Biomedical Engineering & Imaging Sciences, King’s College London, London, UK · Institute of Psychiatry, Psychology & Neuroscience, King’s College London, London, UK · Department of Mathematics, Imperial College London, London, UK · Department of Computer Science, Durham University, Durham, UK
Light scattering, absorption and turbidity degrade underwater imagery through complementary frequency tendencies: large-scale, channel-dependent appearance distortions and local high-frequency detail loss. This band asymmetry motivates different forms of restoration for global colour and local structure. Our primary contribution is SandBAR, a novel unified band-asymmetric architecture for underwater image restoration. Rather than treating frequency processing as an isolated module, SandBAR organises the restoration pipeline around band asymmetry at the block, encoder and objective levels. Within this architecture, we introduce a novel band-asymmetric sandwich block with a reserved post-attention repair slot and the Split-band Parallel Asymmetric Repair Kernel (SPARK), a novel module that applies channel-aware large-support correction to the low band, which predominantly carries large-scale appearance distortions, and lightweight local refinement to the detail-bearing high band, thereby restoring global colour while preserving fine structure. To deliver the global cast to the depth at which the network represents global structure, a learnable low-frequency guidance path injects a low-pass view of the input into the deep encoder, and a three-scale spatial-frequency objective supervises the same axis throughout the decoder. Building on the reserved post-attention slot, we further introduce Post-attention Scaffold Training (PaST), a novel removable nonlinear scaffolding scheme that trains the permanent SandBAR graph under band-specific capacity and runs it at inference without that capacity. SandBAR reaches 29.89 / 28.92 / 22.77 dB on EUVP / LSUI / UIEB and establishes a new three-domain mean state of the art on the benchmark family, exceeding the strongest published result by approximately 1.0 dB. PaST adds no parameters or FLOPs at inference and preserves matched latency.
Figures & tables
Figure 1: Visual comparison of SandBAR and a competing method. Top: input and SandBAR. SandBAR restores global colour while preserving texture detail.
Figure 2: SandBAR architecture and its constituent designs. (a) Encoder–decoder backbone with a learnable low-frequency guidance path injected at encoder levels 2 and 3, and three supervised output heads. (b) The sandwich block: two feed-forward slots around attention, with SPARK occupying only the post-attention slot. (c) SPARK: content-adaptive band split, asymmetric per-band repair, gated recombination and residual closing. (d) PaST: after joint training the scaffold is removed and its bias folded, and what remains at inference is the permanent scaffold-free block.
Figure 3: PaST preserves performance while reducing parameters and inference cost. Removing SPARK before inference restores the permanent graph’s parameter count, FLOPs and latency.
Figure 4: Samples for visual comparison showcase how SandBAR achieves more faithful colour restoration and clearer structural detail by repairing the low- and high-frequency bands with separate operators, while the compared baselines either retain the input’s cast or soften surviving detail.
EUVP
LSUI
UIEB
3-domain
Method
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
mean PSNR ↑
ROP [ 19 ]
15.34
0.714
17.38
0.806
18.48
17.07
ADPCC [ 32 ]
15.20
0.692
16.20
0.763
–
–
WWPF [ 29 ]
15.95
0.648
17.90
0.739
–
–
UIE-WD [ 21 ]
17.80
0.760
19.23
0.803
–
–
TACL [ 20 ]
20.99
0.782
22.97
0.828
–
–
Table 1: Comparison on the 4,800-pair underwater restoration benchmark (published results from the source papers [ 3 , 27 ] ).
Params ↓
GFLOPs ↓
Latency (ms) ↓
3-domain
Graph
@256
@256
@512
mean PSNR ↑
SPARK retained
28,586,687
146.86
27.62
81.52
27.20
Post-fold
24,088,479
132.15
21.53
61.89
27.19
SandBAR w/o PaST
24,088,479
132.15
21.44
61.91
26.79
Table 2: PaST inference accounting. The SPARK-retained row evaluates the trained SandBAR model before strip-and-fold, with the frequency scaffold kept at inference. After strip-and-fold, SandBAR recovers the permanent graph’s parameter and FLOP counts and matched latency, while preserving its three-domain mean PSNR.
Figure 5: Removing SPARK at inference is visually inert; training with it is not. The first two columns evaluate the trained SandBAR model before and after strip-and-fold. The third column shows their absolute difference amplified 10× ; the fourth applies the same amplification to the difference between the post-fold graph and the plain baseline. Depending on the domain, the mean absolute difference is 20 – 139× larger in the fourth column, so the near-black third column is a property of strip-and-fold rather than insufficient amplification.
Training scaffold
PSNR ↑
SSIM ↑
Without band decomposition
27.069
0.91248
SandBAR
27.193
0.91493
Table 3: Effect of band decomposition. The unsplit variant retains the two repair branches. Scores are equal-domain means over EUVP, LSUI and UIEB, evaluated from floating-point predictions after scaffold removal.
Figure 6: What the band scaffold learns. Channel-mean magnitudes of the SPARK band split ( zl and zh columns; each map normalised independently) on one test image from each domain, captured from the training graph at the second encoder stage. The split is exact (Eq. ( 4 )): the low branch carries smooth illumination and colour-cast structure, while the high branch isolates contours, bubble trails and rock texture. After strip-and-fold this machinery is absent at inference (Sec. 3.3 ); the plain graph that remains improves 472 of the 490 test images in both bands (Gaussian split, σ=10 ).
Configuration
Mean PSNR ↑
Module
Backbone
pre-attn.
post-attn.
SPARK
SandBAR
26.43
26.72
DFFN [ 14 ]
SandBAR
26.25
26.52
SPARK
Restormer [ 28 ]
26.15
26.44
Table 4: Where band repair belongs. Three-domain mean PSNR under the controlled 100 -epoch budget. Each row compares the same module in the pre-attention and post-attention feed-forward slots of one backbone, with all other settings held fixed. SPARK and SandBAR are ours, while DFFN and Restormer are existing designs.
Variant
Splitter
Band ops
LF guid.
PSNR ↑
Δ↑
(a)
SandBAR full
adaptive
asym.
✓
26.72
—
(b)
fixed Gaussian
fixed
asym.
✓
25.73
−3.69%
(c)
symmetric ops
adaptive
sym.
✓
26.29
−1.63%
(d)
no LF guidance
adaptive
asym.
—
26.28
−1.65%
Table 5: Architectural ablation. Model (a) is the full SandBAR scaffold. All variants are trained under the same controlled 100 -epoch budget as Table 4 , with each variant changing exactly one design decision and all other settings held fixed. Δ is the three-domain mean PSNR relative to model (a).
Key Laboratory of Measurement and Control of Complex Systems of Engineering, Ministry of Education, School of Automation, Southeast University, Nanjing 210096, China · Advanced Ocean Institute of Southeast University, Nantong 226010, China
King Fahd University of Petroleum & Minerals, Dhahran 31261, Saudi Arabia · SDAIA-KFUPM Joint Research Center for Artificial Intelligence, Dhahran 31261, Saudi Arabia
National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University · School of Computer and Artificial Intelligence, Zhengzhou University · Aerospace Information Research Institute, Henan Academy of Sciences +2