Adaptive Dual-domain Learning for Underwater Image Enhancement

Lintao Peng, Liheng BianAAAI 2025 · DOI

TL;DR

This paper introduces SS-UIE, a novel deep learning approach for enhancing underwater images. It tackles the challenge of inconsistent degradation across different image regions and spectral bands by employing a dual-domain adaptive learning strategy with linear complexity, outperforming existing methods in both quality and efficiency.

What Problem Does This Paper Address?

Underwater images are notoriously difficult to work with due to the way light interacts with water. Dissolved impurities and suspended matter cause light to be absorbed and scattered unevenly. This leads to images that suffer from:

  • Color Casts: A dominant hue (often blue or green) that distorts the true colors.
  • Artifacts: Unwanted visual distortions.
  • Blurred Details: Loss of sharpness, especially in distant objects.

Existing methods for enhancing these images often struggle with the fact that the degradation isn't uniform. Some areas of the image might be more affected than others, and different color channels (spectral bands) can be degraded differently. Furthermore, regions with fine details are harder to reconstruct accurately. Many current deep learning approaches, while powerful, can be computationally expensive and don't effectively address these spatial and spectral inconsistencies simultaneously.

Methods

The core of SS-UIE lies in its novel Spatial-Spectral block (SS-block), designed for adaptive learning with linear complexity. This block is built upon two key modules:

  • Spectral-Wise Self-Attention (SWSA): This module operates in the frequency domain using Fast Fourier Transforms (FFT). It applies a learnable filter to mix spectral information across different frequency components, allowing it to model inter-spectral dependencies and capture degradation levels across spectral bands.
  • Multi-scale Cycle Selective Scan (MCSS): Inspired by State Space Models (SSMs) like Mamba, this module is adapted for vision tasks. It uses a selective scan mechanism to achieve a global receptive field with linear complexity, enabling it to model degradation levels in different spatial regions.

These two modules are combined in parallel within the SS-block. This parallel design is crucial as it facilitates complementary interactions between spatial and spectral features, allowing the block to simultaneously capture degradation levels in both domains. By stacking these SS-blocks, the SS-UIE network can adaptively enhance images based on their specific degradation characteristics.

Additionally, the paper introduces a Frequency-Wise Loss (FWL). This loss function specifically targets the reconstruction of high-frequency details, which are often lost in degraded underwater images. By narrowing the frequency-wise discrepancy between the enhanced and ground truth images, FWL encourages the network to pay more attention to these crucial details, leading to sharper and more visually appealing results.

Datasets

The paper uses the following datasets for training and evaluation:

  • UIEB: A dataset of 890 underwater images with corresponding labels, used for training and testing.
  • LSUI: A dataset of 3879 underwater images, randomly partitioned for training and testing.
  • U45: A dataset of 45 underwater images used for non-reference evaluation, testing the model's generalization capabilities.

Key Insights and Ideas

  • Dual-Domain Adaptive Learning is Key: The most significant insight is that effectively enhancing underwater images requires simultaneously considering both spatial (where in the image) and spectral (what color information) degradation. Treating these independently or uniformly leads to suboptimal results.
  • Parallelism for Efficiency and Effectiveness: Combining spatial and spectral processing modules in parallel within the SS-block is more efficient and effective than a serial approach. This parallel design allows for better feature interaction and reduces computational complexity.
  • Frequency Domain for Detail Recovery: Focusing on the frequency domain with techniques like SWSA and the FWL is crucial for recovering fine-grained details and textures that are often lost in degraded images.
  • Linear Complexity is Achievable: The paper demonstrates that it's possible to achieve powerful image enhancement with linear complexity, making the approach more practical for real-time applications and resource-constrained environments.

Key Results

  • State-of-the-Art Performance: SS-UIE achieves state-of-the-art (SOTA) performance on both full-reference (UIEB, LSUI) and non-reference (U45) evaluation metrics, including PSNR, SSIM, LPIPS, FID, UIQM, and UCIQE.
  • Efficiency: The proposed SS-UIE method requires significantly fewer parameters and FLOPs compared to other SOTA methods, making it computationally efficient. It achieves an inference time of approximately 0.05 seconds per image on a GPU.
  • Visual Quality: Qualitative results show that SS-UIE produces images with fewer color casts and artifacts, and recovers high-frequency local details better than existing methods. The enhanced images are closer to the ground truth.
  • Ablation Studies: Experiments confirm the effectiveness of each component of SS-UIE, with the parallel SS-block design and the FWL contributing significantly to the overall performance.

Limitations and Future Work

  • Computational Cost of DFT: While the paper emphasizes linear complexity for the SS-block, the use of DFT/FFT within SWSA still has a computational cost, though it's generally efficient for image processing.
  • Generalization to Extreme Conditions: While the U45 dataset shows good generalization, further testing on a wider variety of extreme underwater conditions could be beneficial.
  • Real-time Applications: The paper mentions potential for real-time applications, and future work could focus on further optimizing the model for even faster inference on embedded systems or mobile devices.
  • Exploring Other State Space Models: The MCSS module is inspired by Mamba. Further exploration of other advanced SSM architectures could potentially lead to even more powerful enhancement capabilities.

Summary updated

BibTeX
@inproceedings{peng2025adaptive,
  author = {Lintao Peng and Liheng Bian},
  title = {{Adaptive Dual-domain Learning for Underwater Image Enhancement}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i6.32692},
}
Figure