Abstract
Underwater image enhancement is critical for vision-based underwater tasks, yet the field suffers from inconsistent benchmarking and a trade-off between processing speed and image quality. This paper presents a novel Fusion Generative Adversarial Network (FGAN) and a standardized public test dataset, U45. FGAN integrates traditional multi-scale fusion with deep learning, utilizing a dual-input generator and a relativistic discriminator. Our approach significantly reduces parameter count while achieving state-of-the-art performance and real-time processing speeds.
The Physical Challenge of Underwater Imaging
Capturing high-quality images underwater is hindered by the selective absorption and scattering of light. As depth increases, water absorbs longer wavelengths first: red light disappears at approximately 5 meters, followed by orange and yellow. This leads to the characteristic greenish or bluish color casts seen in raw underwater photography. Additionally, suspended particles cause light scattering, resulting in a "haze" effect that severely degrades contrast and visibility.
Proposed Method: Fusion Generative Adversarial Network (FGAN)
To address these issues, we propose FGAN, which leverages "hybrid intelligence" by combining traditional image processing with deep learning.
Dual-Input Generator Architecture
Unlike standard GANs, the FGAN generator accepts two inputs: the raw underwater image and a version pre-enhanced via traditional multi-scale fusion. This "head start" allows the network to focus on refining structural details and color accuracy rather than learning from scratch. The generator employs "Basic Blocks" with multiple kernel sizes (3x3 and 5x5) to capture multi-scale features effectively.
Relativistic Loss and Optimization
We utilize a Relativistic GAN (RaGAN) loss, which encourages the discriminator to estimate the relative realism of the generated image compared to the ground truth. The loss function is further refined with $L_{gt}$ (ground truth loss) and $L_{fe}$ (fusion-enhanced reference loss) to ensure high fidelity to the original scene.
The U45 Dataset
Standardized evaluation is a major hurdle in underwater vision research. We introduce U45, a curated dataset of 45 real-world images categorized by their primary degradation: Greenish, Bluish, and Haze-like. This provides a balanced benchmark for the community to compare algorithms objectively.
Key Results and Efficiency
FGAN demonstrates exceptional efficiency and performance:
- Speed: It processes a 256x256 image in just 0.0286 seconds, making it suitable for real-time applications.
- Compactness: The model uses only 1.11M parameters, a 97% reduction compared to previous models like UGAN (38.67M), without sacrificing quality.
- Practical Utility: Enhancement significantly improves downstream tasks. For instance, using a pre-trained YOLO-V3 detector on FGAN-enhanced images allowed for the detection of divers and objects that were previously invisible in raw footage.
Conclusion
The proposed FGAN framework successfully bridges the gap between traditional fusion methods and deep learning. By providing both a high-performance, lightweight model and the U45 benchmark dataset, we aim to advance the state of underwater vision for both research and practical marine applications.
Summary updated