Chongfan Technology

News

All Categories >

Home page

2026

Progress has been made in research on polarized vision.

Author:

Existing polarization-aware low-light enhancement methods suffer from several limitations: they rely on paired low-light–normal-light datasets, exhibit limited cross-scene generalization, and struggle to jointly recover radiance and polarization information under complex low-light conditions.

Recently, a team from the Institute of Automation, Chinese Academy of Sciences, in Shenyang has made progress in the field of polarization vision by proposing ZeroPIE, a zero-reference polarization‑based framework for low‑light enhancement, thereby offering a novel technological approach to reconstructing low‑illumination polarization images.

Starting from the physical mechanisms of polarization imaging, the research team derives four categories of illumination-invariant priors—degree of linear polarization, angle of linear polarization, geometric invariants, and color invariants—and uses them to bridge low-light and normal-light polarization images. Simultaneously, they construct a physics‑guided residual denoising diffusion model that is trained solely on normal-light polarization data, enabling polarization image enhancement without requiring ground‑truth low-light training data while improving the robustness of radiance and polarization recovery under complex lighting conditions.

The relevant research findings were published in IEEE Transactions on Pattern Analysis and Machine Intelligence.