TL;DR
SandBAR is a new architecture for underwater image restoration that treats global color casts and local detail loss as distinct frequency-based problems. Its core innovation, Post-attention Scaffold Training (PaST), allows the model to learn from complex frequency-aware modules during training and then "strip" them away for inference, achieving state-of-the-art results with zero additional computational cost.
What Problem Does This Paper Address?
Capturing clear images underwater is notoriously difficult. Light doesn't travel through water the same way it does through air; different wavelengths are absorbed at different rates (causing that familiar blue-green tint), and suspended particles scatter light (causing blur and low contrast).
The authors identify a phenomenon they call band asymmetry:
- Low-frequency bands carry large-scale color and illumination distortions.
- High-frequency bands contain the fine textures and edges that are easily lost to turbidity.
Existing deep learning models often treat these issues uniformly, using the same operations for both color correction and detail recovery. This often leads to a "tug-of-war" where fixing the color might smudge the details, or vice versa. Furthermore, models that do try to handle frequencies separately often become computationally heavy, making them impractical for real-time use on underwater drones or cameras.
Methods
The authors propose SandBAR, a unified architecture designed around band asymmetry at every level—from individual blocks to the overall encoder-decoder structure.
1. The Sandwich Block and SPARK
Instead of a standard transformer block, SandBAR uses a "sandwich" design with two feed-forward slots. The second slot, placed after global attention has been applied, is reserved for a novel module called SPARK (Split-band Parallel Asymmetric Repair Kernel).
SPARK performs an exact two-band split of the image features:
- Low-band branch: Uses channel attention and large-support convolutions to fix global color casts.
- High-band branch: Uses lightweight local refinement to preserve delicate structures without over-processing them.
2. PaST: Removable Scaffolding
This is the paper's most clever trick. SPARK is a powerful but complex nonlinear module. To keep the final model efficient, the authors introduce Post-attention Scaffold Training (PaST).
During training, SPARK acts as a "scaffold," helping the network learn how to handle different frequencies. Because SPARK is designed as a residual module, it can be mathematically "folded" or removed after training. The complex frequency machinery is stripped away, leaving behind a simpler, permanent graph that retains the performance gains of the complex version but runs much faster.
3. Architecture-Wide Coordination
Beyond the blocks, SandBAR uses a learnable low-frequency guidance path to inject color information directly into the deep encoder and a three-scale spatial-frequency objective to supervise the restoration process across both image space and frequency space.
Datasets
The model was trained and evaluated on a massive 4,800-pair underwater restoration benchmark consisting of three major datasets:
- UIEB: (Underwater Image Enhancement Benchmark)
- EUVP: (Enhancing Underwater Visual Perception)
- LSUI: (Large-Scale Underwater Image)
The evaluation used a 490-image paired test set across these three domains.
Key Insights and Ideas
- Functional Asymmetry: The authors found that frequency repair works best after global context has been established by the attention mechanism. Placing SPARK in the post-attention slot consistently outperformed placing it before attention.
- Nonlinear Scaffolding: While "re-parameterization" (like RepVGG) is common for linear layers, this paper proves that nonlinear modules can also serve as removable training scaffolds if they are structured as residuals.
- Matching Operators to Content: Giving the low-frequency band "stronger" operators (like channel attention) is more effective than treating both bands with the same high-power tools.
Key Results
SandBAR establishes a new state-of-the-art (SOTA) across all three major benchmarks.
- Performance: It achieved a three-domain mean PSNR of 27.19 dB, exceeding the previous strongest result by approximately 1.0 dB.
- Efficiency: Thanks to PaST, the model reduces its parameter count by 15.74% (from 28.59M to 24.09M) and its FLOPs by 10% at inference time compared to the version that keeps the SPARK module, with virtually no loss in restoration quality.
- Qualitative: Visual results show that SandBAR is uniquely capable of removing deep-sea color casts while keeping the "fuzz" on a sea fan or the texture on a turtle's shell sharp.
Limitations and Future Work
While SandBAR is highly effective, the authors note that the current scaffolding scheme is specifically designed for the post-attention slot. They suggest that extending removable nonlinear scaffolding to other parts of the network or other restoration tasks (like deblurring or low-light enhancement) is a promising direction for future research. Additionally, while the model is efficient, the initial training phase still requires the full "scaffolded" memory overhead.
Summary updated