A terahertz time-domain super-resolution imaging method using a local-pixel graph neural network for biological...

Area of Science:

• Terahertz (THz) imaging
• Computational imaging
• Machine learning for scientific applications

Background:

• Low image acquisition speed of THz time-domain imaging systems hinders biological product analysis.
• Existing methods struggle with heterogeneous biological materials and require extensive training data.

Purpose of the Study:

• To develop a super-resolution method for THz time-domain imaging using a local pixel graph neural network.
• To improve image acquisition speed and analytical capabilities for biological products.
• To enable analysis of heterogeneous biological products with minimal training data.

Main Methods:

• A local pixel graph neural network was designed for THz time-domain imaging super-resolution.
• The Fourier transform was applied to graphs from low-resolution (LR) images for rotation and flip invariance.
• The network learned relationships between graph states and pixels for reconstruction, focusing on THz feature frequencies.

Main Results:

• The method achieved a root mean square error (RMSE) of 0.0957 for wood core images and 0.1061 for seed images.
• Reconstructed high-resolution (HR) images retained spatial details and useful signals from noisy high-frequency data.
• Super-resolution was achieved in both spatial and spectral domains, outperforming LR images.

Conclusions:

• The developed graph neural network effectively enhances THz imaging resolution and speed for biological product analysis.
• The method demonstrates robustness for heterogeneous materials and offers simultaneous spatial and spectral super-resolution.
• This approach significantly advances the potential of THz imaging in scientific and industrial applications.