Zero-order diffraction-guided spectral-wise transformer: a lightweight and efficient framework for computed...

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Computed tomography imaging spectrometry (CTIS) captures 3D spectral data but suffers from artifacts due to limited projections.
• Existing deep learning methods, often CNN-based, require large datasets and significant computational resources, risking overfitting.

Purpose of the Study:

• To develop a lightweight and efficient deep learning framework for CTIS datacube reconstruction.
• To improve reconstruction quality while minimizing computational cost and data requirements.

Main Methods:

• Introduced the Zero-order Diffraction-guided Spectral-wise Transformer (ZDST) framework.
• Integrated Full Projection Connect Reshaping (FPCR), Spectral-wise Multi-head Self-Attention (S-MSA), and Zero-order Diffraction-guided Mechanism (ZDM) modules.
• Utilized a custom-built filter wheel system to collect a real-world dataset.

Main Results:

• ZDST achieved high reconstruction quality with low computational resource consumption.
• The model demonstrated effectiveness with limited datasets, reducing reliance on extensive parameters.
• Experimental results validated the model's performance under real-world conditions.

Conclusions:

• The ZDST framework offers a superior solution for CTIS datacube reconstruction compared to existing methods.
• The integration of specific modules enhances model fitting capability and reduces data dependency.
• ZDST presents a promising approach for practical hyperspectral imaging applications.