Tomographic reconstruction from planar thermal imaging using convolutional neural network.

Area of Science:

• Thermography
• Non-destructive testing
• Artificial Intelligence

Background:

• Volumetric temperature reconstruction is challenging due to infrared's non-penetrating nature.
• Traditional methods struggle with internal temperature mapping compared to ionizing radiation techniques.

Purpose of the Study:

• To develop and evaluate a deep learning approach for thermal tomography.
• To reconstruct internal temperature distributions using planar infrared thermal images.

Main Methods:

• Utilized an autoencoder deep neural network trained on synthetic 3D models.
• Generated synthetic data of a cylindrical phantom with varying heat source locations and times.
• Applied initial backprojection reconstruction followed by the deep learning model.
• Validated with experiments on a real object (pork) and complex scenarios with multiple heat sources.

Main Results:

• The deep model demonstrated capability in reconstructing internal temperature distributions.
• Achieved accurate heat source localization and spatial similarity in volumetric models.
• Evaluated performance using metrics like signal-to-noise ratio.

Conclusions:

• Deep learning, specifically autoencoders, offers a viable solution for thermal tomography.
• This method advances non-invasive internal temperature mapping capabilities.