Polarization-driven camouflaged object detection: a multimodal fusion network with iterative polarimetric feature...

Area of Science:

• Computer Vision
• Optical Imaging
• Artificial Intelligence

Background:

• Camouflaged object detection (COD) performance degrades under complex backgrounds and dynamic illumination.
• Traditional visible-light imaging struggles to differentiate objects based on material and surface optical properties.

Purpose of the Study:

• To propose a novel polarization-driven multimodal fusion network (PMFNet) for high-precision COD.
• To address the limitations of existing methods in challenging imaging scenarios.

Main Methods:

• Developed a feature rectification module using polarization differences based on surface scattering properties.
• Implemented a polarization-guided iterative refinement mechanism to correct RGB texture degradation with polarization features.
• Introduced a polarization adaptive fusion module for context-aware complementary enhancement of RGB and polarization features.

Main Results:

• The PMFNet demonstrated robust camouflaged object detection performance under adverse illumination and complex backgrounds.
• Experimental results on public datasets showed superior performance compared to state-of-the-art COD methods.

Conclusions:

• The proposed PMFNet effectively fuses complementary features from polarization and RGB modalities.
• This approach significantly enhances COD accuracy and robustness in challenging environments.