Abstract
Simulating realistic underwater imagery is challenging due to complex light interactions like wavelength-dependent absorption and scattering. Traditional physical models are often oversimplified, while purely data-driven approaches struggle with the lack of paired datasets. This paper presents a hybrid framework that integrates a classical physics-based image formation model with a deep residual learning branch. By learning the "residuals"—the complex effects not captured by simplified physics—the proposed model achieves state-of-the-art simulation accuracy, outperforming standard image-to-image translation models like Pix2Pix and CycleGAN.
Introduction
Capturing high-quality underwater images is hindered by the unique optical properties of water. Light is absorbed and scattered based on turbidity and distance, resulting in low contrast and distinct color shifts (typically blue-green). While mathematical models exist to describe these phenomena, they often rely on parameters that are difficult to measure in situ. Furthermore, the absence of large-scale paired datasets (clear vs. underwater images of the same scene) limits the effectiveness of supervised deep learning. This research addresses these challenges by combining physical priors with the flexibility of neural networks.
Methodology
The proposed architecture employs a three-pronged approach to image simulation:
- Physics Branch: Estimates a depth map from an input RGB image and applies a classical dehazing equation to generate an initial, physically-grounded but incomplete degraded image.
- Residual Branch: A neural network specifically designed to learn the "residual"—the complex scattering and noise components that the simplified physics equation fails to model.
- Direct Branch: An encoder-decoder structure that refines the final output by predicting the image directly.
The final simulated image is the sum of the physics-based output and the learned residual. This "residual learning" strategy allows the network to focus on fine-grained deviations from the physical model rather than learning the entire transformation from scratch.
Datasets and Training
To overcome the lack of real-world paired data, the authors synthesized a ground-truth dataset using a complex physical model applied to existing indoor RGB-D datasets:
- NYU Depth v2: 50,000 indoor images with high-quality depth maps.
- Make3D: 534 RGB-depth pairs used to evaluate generalization across different resolutions.
Key Results
The hybrid model was compared against several baseline Image-to-Image (I2I) models, including Pix2Pix, CycleGAN, and UNIT.
Quantitative Performance
On the NYU Depth v2 dataset, the best variant of the proposed method achieved a Threshold Accuracy ($\delta_1$) of 0.792. This significantly outperformed purely data-driven models like CycleGAN (0.227) and even supervised models like Pix2Pix (0.743).
Qualitative Analysis
Visual evaluations demonstrate that the physics-informed approach produces more natural scattering effects and consistent color distortions compared to GAN-based methods, which often introduce artifacts or excessive sharpness.
Conclusion and Future Work
The integration of physical priors as a regularizer ensures that the generated images remain physically plausible while the residual branch captures complex real-world nuances. Future research aims to apply this "emulator" in reverse to solve the inverse problem of underwater image restoration. While the model currently relies on synthetic ground truth, it provides a robust foundation for generating high-fidelity training data for downstream underwater vision tasks.
Summary updated