TY - JOUR
T1 - Stokes physical constraint method for improving polarization imaging-based vision task
AU - Song, Shaoting
AU - Mu, Tingkui
N1 - Publisher Copyright:
© 2025 Elsevier Ltd
PY - 2025/12
Y1 - 2025/12
N2 - Polarization vision captures polarization images to reveal scene properties inaccessible to conventional vision. However, the pixel-level error caused by sensor misalignment, ambient light leakage, and element imperfections often degrade the performance of polarization imaging-based vision tasks. To address this issue, we propose a novel Stokes Physical Constraints (SPC) method to model and mitigate the pixel-level error for improving vision tasks. Additionally, we introduce a new metric, the Constraint Level on Stokes Vector (CLSV) of each pixel, to quantify Stokes vector consistency and assess polarization image quality. Experiments on two publicly available datasets demonstrate the effectiveness of the SPC method and CLSV metric. The CLSV score increases from 38.24 % to 99.04 % in representative cases, indicating significant improvements in data quality. The SPC-processed data enhances the performance of polarization imaging-based vision tasks, including shape from polarization and polarization image fusion. In the 3D reconstruction task, neural networks trained on the SPC-processed data achieve better results in specular reflection and surface concavity-convexity. In the image fusion task, the SPC-processed data produces images with reduced artifacts, enhanced contrast, and improved visual quality, as confirmed by higher mutual information, spatial frequency, and visual fidelity scores. This work provides a robust framework for improving polarization imaging quality, with potential applications in biomedical imaging, remote sensing, and industrial inspection.
AB - Polarization vision captures polarization images to reveal scene properties inaccessible to conventional vision. However, the pixel-level error caused by sensor misalignment, ambient light leakage, and element imperfections often degrade the performance of polarization imaging-based vision tasks. To address this issue, we propose a novel Stokes Physical Constraints (SPC) method to model and mitigate the pixel-level error for improving vision tasks. Additionally, we introduce a new metric, the Constraint Level on Stokes Vector (CLSV) of each pixel, to quantify Stokes vector consistency and assess polarization image quality. Experiments on two publicly available datasets demonstrate the effectiveness of the SPC method and CLSV metric. The CLSV score increases from 38.24 % to 99.04 % in representative cases, indicating significant improvements in data quality. The SPC-processed data enhances the performance of polarization imaging-based vision tasks, including shape from polarization and polarization image fusion. In the 3D reconstruction task, neural networks trained on the SPC-processed data achieve better results in specular reflection and surface concavity-convexity. In the image fusion task, the SPC-processed data produces images with reduced artifacts, enhanced contrast, and improved visual quality, as confirmed by higher mutual information, spatial frequency, and visual fidelity scores. This work provides a robust framework for improving polarization imaging quality, with potential applications in biomedical imaging, remote sensing, and industrial inspection.
KW - Polarization image fusion
KW - Polarization imaging
KW - Polarization vision
KW - Shape from polarization
KW - Stokes vector
UR - https://www.scopus.com/pages/publications/105008436395
U2 - 10.1016/j.optlastec.2025.113408
DO - 10.1016/j.optlastec.2025.113408
M3 - 文章
AN - SCOPUS:105008436395
SN - 0030-3992
VL - 192
JO - Optics and Laser Technology
JF - Optics and Laser Technology
M1 - 113408
ER -