Cascade deep polarization network for precise image semantic segmentation.

Area of Science:

• Computer Vision
• Optical Imaging
• Machine Learning

Background:

• Optical polarization imaging captures rich light field data for scene analysis.
• Current methods use separate preprocessing, limiting semantic segmentation accuracy.
• Deep learning integration with polarization data is an active research area.

Purpose of the Study:

• To propose a novel method, the Cascade Deep Polarization Network (CDPN), for enhanced semantic segmentation.
• To integrate preprocessing modules directly into an end-to-end deep learning framework.
• To improve the accuracy and efficiency of target scene analysis using polarization imaging.

Main Methods:

• Developed a Cascade Deep Polarization Network (CDPN) integrating denoising, fusion, and enhancement modules.
• Input data includes angle of linear polarization, degree of linear polarization, and Stokes parameters.
• Employed self-supervised loss functions for collaborative training of preprocessing and backbone networks.

Main Results:

• The CDPN method significantly improved semantic segmentation accuracy.
• The proposed network maintained fast computation speeds.
• Experimental results validated the effectiveness of integrated preprocessing.

Conclusions:

• The CDPN offers a superior approach to semantic segmentation with optical polarization imaging.
• Integrating preprocessing into deep learning models is crucial for performance gains.
• This method advances target scene analysis through efficient and accurate polarization image segmentation.