Exploiting Light Polarization for Deep HDR Imaging from a Single Exposure.

Area of Science:

• Computational photography
• Image processing
• Computer vision

Background:

• High dynamic range (HDR) imaging aims to capture a wider range of light intensities than standard sensors.
• Traditional HDR methods use multiple exposures and tone mapping, while recent research explores single-exposure techniques using data-driven models or polarimetric cameras.
• Estimating HDR from a single exposure remains challenging due to limitations in sensor dynamic range.

Purpose of the Study:

• To present a novel HDR reconstruction method using a single polarimetric filter array (PFA) camera with an external polarizer.
• To combine classical HDR algorithms with data-driven solutions for polarimetric images.
• To enhance dynamic range and accurately reconstruct HDR scenes from a single capture.

Main Methods:

• Utilized a PFA camera with an external polarizer to increase dynamic range and simulate varied exposures.
• Developed a pipeline integrating standard HDR bracketing algorithms with data-driven approaches for polarimetric images.
• Introduced two convolutional neural network (CNN) models: one for estimating scene properties from PFA data and another for improving tone mapping.

Main Results:

• The proposed method demonstrated effective HDR reconstruction on synthetic and real-world datasets.
• Achieved a peak signal-to-noise ratio (PSNR) of 23 dB on the entire test set.
• Outperformed state-of-the-art methods by 18% in PSNR.

Conclusions:

• The novel approach successfully leverages light attenuation from polarimetric filters for accurate HDR reconstruction.
• The combination of PFA imaging, external polarization, and CNN models offers a significant advancement in single-exposure HDR imaging.
• The method provides a robust and effective solution for capturing scenes with extreme lighting variations.