Polarization-guided diffusion model for physically inspired underwater image descattering

Area of Science:

• Computer Vision
• Optical Engineering
• Image Processing

Background:

• Underwater imaging faces challenges due to light scattering by water, blurring details and degrading vision task reliability.
• Existing descattering methods often fail in complex scenes due to limited polarization analysis or assumptions.
• Polarization differences between backscattered and object-reflected light offer a physical cue for scattering removal.

Purpose of the Study:

• To develop an advanced underwater image enhancement network.
• To improve the accuracy and effectiveness of scattering removal in complex underwater environments.
• To provide high-quality, reliable inputs for downstream underwater vision applications.

Main Methods:

• A two-stage approach: descattering prior generation and scattering removal.
• Utilizing a diffusion model for the first time to extract polarization information for high-quality priors.
• Incorporating an illumination estimation module to address water absorption and ensure global consistency.

Main Results:

• The proposed network achieves state-of-the-art quantitative and qualitative performance in underwater image enhancement.
• The diffusion model effectively mines polarization data for accurate scattering suppression.
• The illumination module enhances perceptual naturalness by compensating for absorption.

Conclusions:

• The diffusion model and polarization-driven approach significantly improves underwater image quality.
• This method offers a robust solution for scattering removal in diverse underwater scenes.
• The enhanced images are suitable for critical downstream computer vision tasks.