TherNet: Thermal Segmentation Network Harnessing Physical Properties.

Area of Science:

• Computer Vision
• Image Processing
• Thermal Imaging

Background:

• Precise segmentation of thermal infrared images is vital for surveillance, medical diagnostics, and intelligent transportation.
• Existing methods often oversimplify by treating thermal images as grayscale, ignoring crucial physical factors.

Purpose of the Study:

• To develop a novel thermal infrared segmentation framework, TherNet, that integrates thermal imaging effects and material physical information.
• To enhance segmentation precision by addressing limitations of current grayscale-based approaches.

Main Methods:

• Elucidating the impacts of object radiation, inter-object thermal exchange, atmospheric scattering, and camera thermal inertia.
• Developing four modules to model or rectify these physical processes within the TherNet framework.
• Creating two large-scale infrared datasets: TI-Cityscapes and TBRSD for validation.

Main Results:

• Achieved State-of-the-Art (SoTA) performance across multiple infrared semantic segmentation and blind road segmentation datasets.
• Demonstrated the critical importance of leveraging physical properties for improved segmentation.
• TherNet framework showed superior results compared to existing methods.

Conclusions:

• TherNet offers a significant advancement in thermal infrared image segmentation by incorporating physical principles.
• The framework provides robust benchmarks and innovative perspectives for future research in the domain.
• Accurate segmentation of thermal images is enhanced by considering physical imaging effects and material properties.