TSMPN-PSI: high-performance polarization scattering imaging based on three-stage multi-pipeline networks

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Polarization imaging offers advantages over traditional intensity imaging, especially in scattering environments.
• Deep learning (DL) has advanced polarization scattering imaging (PSI), but scattering media still degrade object information.
• Existing methods struggle with information loss caused by scattering, limiting PSI performance.

Purpose of the Study:

• To explore the relationship between multiple polarization feature learning strategies and PSI performance.
• To propose a novel multi-polarization driven multi-pipeline (MPDMP) framework for enhanced PSI.
• To develop a deep learning architecture for robust polarization scattering imaging.

Main Methods:

• A multi-polarization driven multi-pipeline (MPDMP) framework was developed to extract hierarchical representations from polarization feature maps.
• A three-stage multi-pipeline network (TSMPN) architecture, named TSMPN-PSI, was designed for PSI.
• A real-world polarization scattering imaging system was used to collect a dataset for model training and validation.

Main Results:

• The proposed TSMPN-PSI framework demonstrated superior generalization performance compared to existing methods.
• Experiments showed improved PSI results across varying imaging distances, target structures, and material compositions.
• The method effectively reconstructs target information even with significant degradation from scattering media.

Conclusions:

• The developed MPDMP framework offers a new approach for polarization scattering imaging.
• The TSMPN-PSI model shows significant potential for practical applications of PSI.
• This work advances the field of computational imaging in challenging scattering conditions.