TL;DR
This paper introduces CASF-Net, a novel deep learning approach for underwater image enhancement that tackles insufficient contrast and surface texture distortion. By proposing a Channel-Adaptive Correction Module (CACM) and a Spatial Multi-scale Fusion Module (SMFM), CASF-Net significantly improves image quality. The authors also introduce LHUI, a large-scale, high-resolution underwater image dataset to facilitate UIE research.
What Problem Does This Paper Address?
Underwater exploration and exploitation rely heavily on visual information, making underwater image enhancement (UIE) crucial. However, underwater images suffer from significant degradation due to light attenuation, scattering, and uneven illumination, resulting in low contrast, blurriness, and color distortion. While many UIE methods exist, they often overlook the specific challenges of insufficient contrast and surface texture distortion, leading to suboptimal results. Existing methods may also struggle with generalization across diverse underwater scenes and color domains.
Methods
The core of this work is the CASF-Net, a single-branch deep learning network designed for UIE. It comprises two novel modules:
- Channel-Adaptive Correction Module (CACM): This module addresses insufficient contrast and color distortion. It introduces a network-learned "channel-adaptive factor" that modulates features, allowing for adaptive correction of degraded image channels. This is inspired by prior work on dark channel priors and adaptive attenuation curves but integrated into a deep learning framework.
- Spatial Multi-scale Fusion Module (SMFM): This module tackles surface texture distortion and improves image saturation. It employs a multi-scale mechanism with channel separation and fusion techniques. Specifically, it uses strided convolutions and upsampling to capture features at different scales and then fuses them to reconstruct detailed textures and enhance color saturation.
The network also incorporates standard deep learning components like residual blocks for feature extraction. The training process utilizes a dual-domain L1 loss function (spatial and frequency domains) to ensure comprehensive feature learning.
Datasets
The authors introduce a new dataset called LHUI (Large-scale High-resolution Underwater Image Enhancement Dataset).
- Size: Contains 13,080 image pairs.
- Resolution: High-resolution images, with a significant portion in the 2K to 4K range.
- Diversity: Features a wide variety of underwater scenes (coastal areas, open seas, rivers, lakes, ports, aquariums, swimming pools) and six types of color degradation (blue, green, yellow, white, low-light, and other).
- Generation: Created by extracting frames from underwater videos captured with FIFISH V6 and V-EVO cameras and also by collecting images from underwater photographers. Reference images were generated by applying 20 existing UIE methods and selecting the best results based on observer assessments.
This dataset is positioned as larger and more diverse than existing UIE datasets like EUVP, UIEB, and LSUI, aiming to better support data-driven UIE research.
Key Insights and Ideas
- Channel-wise adaptation is key: The paper highlights that treating all image channels uniformly is a limitation of many existing UIE methods. Their CACM module demonstrates that learning channel-specific correction factors can significantly improve contrast and color fidelity.
- Multi-scale fusion for texture: Surface texture distortion is a subtle but important degradation. The SMFM module's approach of capturing and fusing features at different spatial scales effectively addresses this, leading to sharper details and improved saturation.
- Dataset diversity matters: The creation of LHUI underscores the need for large-scale, diverse datasets that reflect real-world underwater conditions to train robust UIE models that generalize well across various scenes and degradation types.
- Dual-domain loss for comprehensive learning: Incorporating both spatial and frequency domain losses during training helps the network attend to both pixel-level details and broader frequency characteristics, preventing the loss of important information.
Key Results
- Quantitative Performance: CASF-Net achieved state-of-the-art results on the LHUI dataset, outperforming 12 other UIE methods. It obtained the highest PSNR (26.41 dB), SSIM (0.9401), and lowest MSE (0.6881 × 10⁻²) among the tested methods.
- Qualitative Performance: Visual comparisons show that CASF-Net produces images with superior sharpness, contrast, and detail restoration, closely matching the ground truth. It also demonstrates better generalization across different color domains compared to other methods.
- Ablation Studies: Experiments confirmed the effectiveness of individual modules. The baseline model achieved a PSNR of 23.84. Adding CACM improved it to 26.17, and adding SMFM further boosted it to 26.41, indicating that both modules contribute significantly to the overall performance.
Limitations and Future Work
- Dataset Coverage: The authors acknowledge that the LHUI dataset, while large and diverse, may not comprehensively cover all possible underwater environments and conditions, particularly very low-light deep-sea scenarios.
- Future Directions: They plan to incorporate other general enhancement techniques, such as low-light enhancement, into their framework. They also aim to further improve their CACM and SMFM modules or introduce new modules to achieve even better enhancement effects.
Summary updated