Development of Domain-Invariant Visual Enhancement and Restoration (DIVER) Approach for Underwater Images

TL;DR

DIVER addresses the challenge of robust underwater image enhancement across diverse water types and lighting by unifying empirical corrections with physics-guided restoration in an unsupervised, domain-invariant framework. The architecture partitions processing into illumination-aware pathways (IlluminateNet or SEF), a local adaptive correction stage (AOCM), and a physics-informed final refinement (Hydro-OpticNet with VeilNet and AttenNet) that jointly remove backscatter and compensate wavelength-dependent attenuation. Through dedicated losses and depth-informed modules, DIVER achieves consistent improvements in radiometric accuracy and perceptual quality across nine datasets, outperforming traditional priors and many deep-learning baselines on both reference and no-reference metrics, and enhancing downstream robotic perception. These results demonstrate the method’s generalization to varied underwater domains and its practical potential for improving AUV/ROV vision in real-world deployments.

Abstract

Underwater images suffer severe degradation due to wavelength-dependent attenuation, scattering, and illumination non-uniformity that vary across water types and depths. We propose an unsupervised Domain-Invariant Visual Enhancement and Restoration (DIVER) framework that integrates empirical correction with physics-guided modeling for robust underwater image enhancement. DIVER first applies either IlluminateNet for adaptive luminance enhancement or a Spectral Equalization Filter for spectral normalization. An Adaptive Optical Correction Module then refines hue and contrast using channel-adaptive filtering, while Hydro-OpticNet employs physics-constrained learning to compensate for backscatter and wavelength-dependent attenuation. The parameters of IlluminateNet and Hydro-OpticNet are optimized via unsupervised learning using a composite loss function. DIVER is evaluated on eight diverse datasets covering shallow, deep, and highly turbid environments, including both naturally low-light and artificially illuminated scenes, using reference and non-reference metrics. While state-of-the-art methods such as WaterNet, UDNet, and Phaseformer perform reasonably in shallow water, their performance degrades in deep, unevenly illuminated, or artificially lit conditions. In contrast, DIVER consistently achieves best or near-best performance across all datasets, demonstrating strong domain-invariant capability. DIVER yields at least a 9% improvement over SOTA methods in UCIQE. On the low-light SeaThru dataset, where color-palette references enable direct evaluation of color restoration, DIVER achieves at least a 4.9% reduction in GPMAE compared to existing methods. Beyond visual quality, DIVER also improves robotic perception by enhancing ORB-based keypoint repeatability and matching performance, confirming its robustness across diverse underwater environments.

Figures (8)

• Figure 1: Illustration of the underwater image formation process. As natural light penetrates the water column, shorter wavelengths (blue/green) dominate due to wavelength-dependent absorption, while scattering from suspended particles introduces haze and color distortion. The resulting captured image exhibits reduced contrast, spectral imbalance, and optical blur.
• Figure 2: The proposed DIVER architecture for domain-invariant underwater image enhancement. An illumination assessment first determines whether the input image is low-lit. Low-light inputs are routed to the CNN-based IlluminateNet and optimized using the losses $\mathcal{L}_G$ and $\mathcal{L}_L$, while adequately lit images are enhanced through the Spectral Equalization Filter (SEF). The resulting output is further refined by the Adaptive Optical Correction Module (AOCM), which enhances local contrast and suppresses chromatic speckle artifacts. Finally, the Hydro-OpticNet performs physics-guided restoration through VeilNet and AttenNet, which estimate backscatter and wavelength-dependent attenuation, respectively, producing radiometrically consistent results across diverse underwater domains. VeilNet is trained using the adaptive hubber loss $\mathcal{L}_H$, whereas AttenNet employs a composite loss comprising $\mathcal{L}_{L}$, $\mathcal{L}_{C}$, and $\mathcal{L}_{S}$.
• Figure 3: Visual comparison of underwater image enhancement results on three datasets: SeaThruakkaynak2019, OceanDarkmarques2020, and USOD10KHong2025. Each row corresponds to a SOTA methods: traditional approaches (IBLA peng2017, DCP he2010, UDCP drews2013, and ULAP song2018), deep learning-based methods (P2CNet rao2023, Phaseformer (PF) khan2025, WaterNet li2019, UDNet saleh2025adaptive, and the U-Shape Transformer (UST) peng2023), and our proposed method DIVER.
• Figure 4: Visual comparison of underwater image enhancement results on three datasets: U45drews2013, FISHTRACdawkins2024, and UIEBli2019. Each row corresponds to a SOTA methods: traditional approaches (IBLA peng2017, DCP he2010, UDCP drews2013, and ULAP song2018), deep learning-based methods (P2CNet rao2023, Phaseformer (PF) khan2025, WaterNet li2019, UDNet saleh2025adaptive, and the U-Shape Transformer (UST) peng2023), and our proposed method DIVER.
• Figure 5: Visual comparison of underwater image enhancement results on three datasets: UFO-120wang2019deep, EUVPislam2020, and LSUIpeng2023. Each row corresponds to a SOTA methods: traditional approaches (IBLA peng2017, DCP he2010, UDCP drews2013, and ULAP song2018), deep learning-based methods (P2CNet rao2023, Phaseformer (PF) khan2025, WaterNet li2019, UDNet saleh2025adaptive, and the U-Shape Transformer (UST) peng2023), and our proposed method DIVER.