Polarization Guided HDR Reconstruction via Pixel-Wise Depolarization.

Area of Science:

• Computer Vision
• Image Processing
• Optics

Background:

• Digital cameras struggle with saturated pixels due to limited dynamic range, making restoration ill-posed.
• Existing high dynamic range (HDR) reconstruction methods using bracketed exposures are prone to ghosting artifacts from image misalignment.
• Polarization cameras offer spatially-aligned, temporally-synchronized images for ghost-free HDR, but current methods underutilize polarization information (DoP, AoP).

Purpose of the Study:

• To develop a robust and generalized HDR reconstruction method that fully exploits polarization information.
• To address the ill-posed nature of HDR reconstruction by classifying pixels based on their reconstruction difficulty.
• To propose a novel pixel-wise depolarization strategy for improved HDR image quality.

Main Methods:

• A pixel-wise depolarization strategy is proposed, classifying pixels by their ill-posedness level for tailored reconstruction.
• A network-physics-hybrid pipeline is introduced, integrating a neural network with physical principles for HDR reconstruction.
• The method fully exploits the Degree of Polarization (DoP) and Angle of Polarization (AoP) encoded in polarized images.

Main Results:

• The proposed approach achieves state-of-the-art performance in HDR reconstruction.
• Experimental results demonstrate superior performance on both synthetic and real-world image datasets.
• The pixel-wise strategy effectively mitigates ill-posed reconstruction issues and ghosting artifacts.

Conclusions:

• The developed polarization-guided HDR reconstruction method offers significant improvements over existing techniques.
• The pixel-wise depolarization strategy and hybrid pipeline enhance robustness and generalization.
• This work advances ghost-free HDR imaging by effectively leveraging polarization data.