Pixel-Based Long-Wave Infrared Spectral Image Reconstruction Using a Hierarchical Spectral Transformer.

Area of Science:

• Spectroscopy and Imaging Technologies
• Artificial Intelligence in Scientific Applications
• Optics and Photonics

Background:

• Long-wave infrared (LWIR) spectral imaging is vital for applications like gas monitoring and fire detection.
• Current systems like the Uncooled Snapshot Infrared Spectrometer (USIRS) offer real-time imaging but suffer from low spectral resolution and high noise.
• Deep learning shows promise for improving LWIR imaging, but data scarcity and limitations in existing network architectures hinder progress.

Purpose of the Study:

• To develop a novel deep learning architecture for enhancing the spectral resolution of LWIR images.
• To address the challenges of noise and limited training data in LWIR spectral imaging.
• To improve the capture of both local and global spectral correlations in LWIR data.

Main Methods:

• Proposed a pixel-based Hierarchical Spectral Transformer (HST) deep learning architecture.
• Trained the HST model using publicly available single-pixel LWIR spectral databases.
• Evaluated the HST's performance on simulated and real-world LWIR datasets.

Main Results:

• The HST architecture effectively enhances spectral resolution in LWIR spectral images.
• The model demonstrates robustness in mitigating noise and improving image quality.
• Successful performance was achieved even with limited training data, showcasing the method's effectiveness.

Conclusions:

• The Hierarchical Spectral Transformer (HST) offers a powerful solution for improving LWIR spectral imaging quality.
• The proposed method effectively addresses limitations in spectral resolution and noise in LWIR data.
• HST provides a robust framework for advancing LWIR spectral imaging applications through deep learning.