Abstract
Training robust underwater image enhancement models is hindered by the lack of paired real-world data. We present Underwater Neural Rendering (UWNR), a data-driven framework that synthesizes realistic underwater imagery from clean terrestrial images. Our approach utilizes a Natural Light Field Retention (LFR) module to capture environmental styles and a Multi-branch Hybrid Unet (MHB-Unet) for high-fidelity rendering. We introduce the Large Neural Rendering Underwater Dataset (LNRUD), featuring 50,000 diverse images, to facilitate future research. Experimental results demonstrate that UWNR outperforms existing methods in both qualitative realism and quantitative metrics.
1. Introduction
Underwater image enhancement is critical for marine exploration and robotics. However, obtaining paired "degraded-clean" images in situ is nearly impossible. While synthetic data offers a solution, traditional methods based on simplified physical models or unstable Generative Adversarial Networks (GANs) often fail to capture the complex, non-uniform nature of underwater light scattering. This work proposes UWNR to bridge the gap between synthetic and real-world underwater imagery through implicit neural rendering.
2. Methodology
The UWNR framework avoids explicit physical modeling, instead learning degradation patterns from authentic underwater exemplars.
2.1 Natural Light Field Retention (LFR)
Inspired by Retinex theory, the LFR module extracts the ambient "style" of a real underwater scene. By applying multi-scale Gaussian low-pass filters, the module generates a light field map that preserves color casts and haze while removing high-frequency object details. This map serves as a spatial template for the rendering process.
2.2 Multi-branch Hybrid Unet (MHB-Unet)
The rendering engine, MHB-Unet, integrates three inputs: a clean terrestrial image, an estimated depth map (simulating distance-dependent attenuation), and the LFR map. The architecture employs spatial and channel attention mechanisms to prioritize regions with significant degradation and critical color channels. The model is optimized using a composite loss function including $L_1$ reconstruction, perceptual loss, Underwater Dark Channel (UDCP) loss, and a novel Light Field Consistency loss.
3. The LNRUD Dataset
To support large-scale training, we developed the Large Neural Rendering Underwater Dataset (LNRUD). By applying 5,000 authentic underwater light fields to clean images from the NYU and RESIDE datasets, we generated 50,000 high-quality synthetic pairs. LNRUD is significantly larger and more diverse than existing benchmarks, providing a superior training ground for enhancement algorithms.
4. Experimental Results
UWNR demonstrates significant improvements over state-of-the-art methods like UWGAN and UISA.
4.1 Quantitative Performance
On the NYU dataset, UWNR achieved a Fréchet Inception Distance (FID) of 221.93, outperforming UWGAN (236.23). This indicates that the distribution of UWNR-generated images is much closer to real underwater photography.
4.2 Qualitative Realism
As shown in the visual comparisons, UWNR avoids the artificial "over-greening" typical of GANs and the unrealistic brightness of physical models. The resulting images exhibit natural haze gradients and color attenuation consistent with real-world physics.
4.3 Computational Efficiency
The model is highly optimized, containing only 11.57M parameters. It achieves an inference time of 0.0023 seconds for a 1024x1024 image, enabling rapid generation of massive datasets.
5. Conclusion
UWNR provides a robust, efficient, and realistic solution for underwater image synthesis. By moving from explicit physical equations to implicit neural rendering with Light Field Retention, we provide a more accurate representation of underwater environments. The LNRUD dataset stands as a valuable resource for the community to advance underwater vision tasks.
Summary updated