Hidden source localization via learning-based reconstruction of thermal light-fields.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Thermal non-line-of-sight (NLOS) imaging faces challenges due to low surface reflectivity and diffuse scattering in the long-wave infrared (LWIR) spectrum.
• Conventional linear methods for denoising and deblurring thermal NLOS measurements have limitations in recovering weak signals.

Purpose of the Study:

• To develop a robust framework for thermal passive NLOS imaging using neural networks.
• To overcome the limitations of existing methods in reconstructing weak thermal signals from noisy scattered light fields.

Main Methods:

• A convolutional neural network (CNN) was developed and trained using synthetic data for signal reconstruction.
• The framework reconstructs noisy scattered light fields to enable thermal NLOS imaging.
• Experimental validation was performed in a miniaturized thermal NLOS imaging studio.

Main Results:

• The CNN successfully recovered extremely weak signals from real-life noisy thermal light fields.
• The proposed framework demonstrated improved standard deviation in source depth estimation by over 50% for multiple objects compared to comparable methods.
• The method outperformed approaches utilizing prior knowledge of scattering surfaces, enhancing depth estimation by approximately 25%.

Conclusions:

• The developed neural network framework offers a powerful solution for thermal passive NLOS imaging.
• The method enhances the accuracy of object localization in challenging LWIR environments without requiring prior knowledge of surface properties.
• This advancement has significant implications for various applications requiring non-line-of-sight thermal sensing.