Abstract
Underwater image enhancement (UIE) is vital for high-level vision-related underwater tasks. Although learning-based UIE methods have made remarkable achievements in recent years, it's still challenging for them to consistently deal with various underwater conditions, which could be caused by: 1) the use of the simplified atmospheric image formation model in UIE may result in severe errors; 2) the network trained solely with synthetic images might have difficulty in generalizing well to real underwater images. In this work, we, for the first time, propose a framework SyreaNet for UIE that integrates both synthetic and real data under the guidance of the revised underwater image formation model and novel domain adaptation (DA) strategies. First, an underwater image synthesis module based on the revised model is proposed. Then, a physically guided disentangled network is designed to predict the clear images by combining both synthetic and real underwater images. The intra- and inter-domain gaps are abridged by fully exchanging the domain knowledge. Extensive experiments demonstrate the superiority of our framework over other state-of-the-art (SOTA) learning-based UIE methods qualitatively and quantitatively.
Introduction
Underwater photography is notoriously difficult due to the physical properties of water. Light is absorbed and scattered by water molecules and suspended particles, leading to two primary degradations: severe color casts (typically deep blue or green) and haze-like turbidity. While deep learning has significantly advanced underwater image enhancement (UIE), existing methods often struggle with consistency across diverse conditions.
Two major hurdles persist in the field. First, many models rely on simplified "atmospheric" image formation models that fail to accurately represent underwater light behavior. Second, there is a significant "domain gap"; because it is nearly impossible to obtain "ground truth" clear images for real-world underwater scenes, models are often trained on synthetic data that lacks the complexity of the actual ocean environment. SyreaNet addresses these issues by integrating physical guidance with a novel domain adaptation strategy.
Methodology
SyreaNet introduces a comprehensive framework that bridges the gap between synthetic training data and real-world conditions through three key innovations:
1. Revised Physical Model
Unlike standard atmospheric models, SyreaNet utilizes a revised underwater image formation model. This model accounts for the fact that the attenuation of scene radiance and backscattering (the "haze" effect) occur at different rates underwater, providing a more accurate mathematical foundation for image restoration.
2. Physically Guided Disentangled Network (PGDNet)
The core architecture, PGDNet, is designed to "disentangle" an image into its constituent physical components: the clear image ($J$), backscattering ($B$), transmission ($T$), and the ambient white point ($W$). A critical feature of PGDNet is its use of Wavelet Transforms to separate low-frequency information (color and luminance) from high-frequency information (edges and textures). This allows the model to correct heavy color casts in low-frequency channels without compromising fine structural details.
3. Domain Adaptation (DA) Strategies
To overcome the synthetic-to-real domain gap, SyreaNet employs a complex DA strategy that facilitates knowledge exchange between domains. A Physically Guided Synthesis Module (PGSM) creates realistic synthetic images, while an adversarial training setup ensures that the features learned from synthetic data are effectively transferred to real-world underwater images.
Experimental Results
The framework was trained using a combination of 20,000 synthetic images (derived from datasets like NYUDepth and KITTI) and 5,000 real underwater images from the UIEB and EUVP datasets.
Quantitative Performance
SyreaNet demonstrates significant quantitative superiority over state-of-the-art (SOTA) methods. In color restoration accuracy tests on the SQUID dataset, SyreaNet achieved an RGB Error of 5.158—less than half the error rate of major competitors like UColor and GL-Net, which typically scored between 10 and 12.
Qualitative Comparison
Visually, SyreaNet excels at removing color casts from distant objects, providing a consistent and natural look across the entire depth of a scene. While other models often fail to correct the background or introduce artifacts, SyreaNet maintains clarity and color fidelity.
Downstream Task Utility
The practical utility of the enhancement was proven through SIFT feature matching. Images processed by SyreaNet yielded 208 feature matches compared to only 66 for raw images, indicating a substantial improvement for downstream tasks such as underwater robotic navigation and mapping.
Conclusion
SyreaNet represents a significant advancement in underwater image enhancement by successfully merging physical modeling with deep learning and domain adaptation. While the framework's complexity—involving multiple modules like PGSM, PGDNet, and UID—presents a challenge for real-time implementation on low-power autonomous underwater vehicles (AUVs), it establishes a robust foundation for future research. The integration of physically guided disentanglement and cross-domain knowledge exchange sets a new benchmark for restoring clarity to the world's oceans.
Summary updated