Enhancing Underwater Light Field Images via Global Geometry-Aware Diffusion Process.

Area of Science:

• Computer Vision
• Image Processing
• Optical Imaging

Background:

• Underwater imaging is crucial for marine research and exploration.
• Acquiring high-quality 4-D light field (LF) images underwater presents significant challenges due to light scattering and absorption.
• Existing methods struggle with color distortion and loss of spatial-angular information in underwater LF images.

Purpose of the Study:

• To develop a novel framework for enhancing underwater 4-D LF imaging.
• To leverage the spatial-angular structure of LF data for improved image quality.
• To mitigate color distortion and enhance visual fidelity in underwater scenes.

Main Methods:

• Proposed GeoDiff-LF, a diffusion-based framework built upon SD-Turbo.
• Introduced a modified U-Net architecture with adapters for geometric cue modeling.
• Implemented a geometry-guided loss function using tensor decomposition and progressive weighting.
• Developed an optimized sampling strategy with noise prediction for efficiency.

Main Results:

• GeoDiff-LF effectively mitigates color distortion in underwater images.
• The framework leverages diffusion priors and LF geometry for superior performance.
• Extensive experiments show outperformance over existing methods in visual fidelity and quantitative metrics.

Conclusions:

• GeoDiff-LF represents a significant advancement in underwater 4-D LF image enhancement.
• The integration of diffusion models and geometric priors offers a powerful approach.
• The proposed methods push the state-of-the-art in underwater imaging applications.