Enhancing Underwater Light Field Images via Global Geometry-aware Diffusion Process

TL;DR

GeoDiff-LF addresses the challenge of underwater 4-D light field image enhancement by tailoring diffusion priors to the LF geometry. It introduces geometry-aware adapters (Conv-Adapter and EPIT-Adapter), a global geometry regularization via tensor Tucker decomposition, and an efficient noise-prediction mechanism to enable few-step sampling on a frozen SD-Turbo backbone. The approach achieves state-of-the-art performance on paired and real underwater LF datasets, improving color fidelity and depth-consistency while maintaining computational practicality. The work demonstrates the value of embedding explicit geometric structure into diffusion-based restoration, with practical impact for high-quality underwater imaging and downstream analysis.

Abstract

This work studies the challenging problem of acquiring high-quality underwater images via 4-D light field (LF) imaging. To this end, we propose GeoDiff-LF, a novel diffusion-based framework built upon SD-Turbo to enhance underwater 4-D LF imaging by leveraging its spatial-angular structure. GeoDiff-LF consists of three key adaptations: (1) a modified U-Net architecture with convolutional and attention adapters to model geometric cues, (2) a geometry-guided loss function using tensor decomposition and progressive weighting to regularize global structure, and (3) an optimized sampling strategy with noise prediction to improve efficiency. By integrating diffusion priors and LF geometry, GeoDiff-LF effectively mitigates color distortion in underwater scenes. Extensive experiments demonstrate that our framework outperforms existing methods across both visual fidelity and quantitative performance, advancing the state-of-the-art in enhancing underwater imaging. The code will be publicly available at https://github.com/linlos1234/GeoDiff-LF.

Figures (9)

• Figure 1: The whole pipeline of the proposed GeoDiff-LF. For the training phase, we first employ a noise map predictor to generate a noisy sample $\mathbf{X}_{\tau}$ from the underwater LF image $\mathbf{Y}_0$ and timestep $\tau$, where $\tau(\tau<T)$ is the start timestep. We then feed this noisy sample, together with the corresponding underwater image $\mathbf{Y}_0$ and noise level $\bar{\alpha}_{\tau}$, into our adapter-augmented U-Net $\epsilon(\cdot)$ to predict the added noise $\epsilon$. To thoroughly capture global geometric structure, we construct an intermediate reconstruction $\mathbf{X}_{\tau-1}$ from the predicted noise $\epsilon_\theta (\mathbf{X}_{\tau},\mathbf{Y}_0,\tau)$ and undergo additional regularization throughout the restoration process. While testing LF underwater images, we start at timestep $\tau$ to generate the noisy sample $\mathbf{X}_{\tau}$, then predict the corresponding noise to obtain the enhanced result.
• Figure 2: Architecture of our adapted SD-Turbo. Two types of adapters are introduced: a Conv-Adapter for capturing multi-dimensional features and an EPIT-Adapter for multi-view correlation.
• Figure 3: The entire process of global geometry regularization. We explore the intrinsic geometric structure of 4-D LFs from tensor-dimension similarity, which is measured by the core tensor of similar 3-D blocks from 4-D LFs.
• Figure 4: Visual comparisons of the enhanced center SAI by different methods on the scene Boat in the LFUB dataset. It is recommended to view this figure by zooming in.
• Figure 5: Visual comparisons of the enhanced center SAI by different methods on the scene Airplane in the LFUB dataset. It is recommended to view this figure by zooming in.