Deep learning enabled reflective coded aperture snapshot spectral imaging.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Coded aperture snapshot spectral imaging (CASSI) acquires high-speed spatial-spectral data using compressive sensing.
• Existing CASSI designs, single-disperser (SD) and dual-disperser (DD), have limitations in spatial resolution or system size/cost.

Purpose of the Study:

• To develop a compact and efficient CASSI system with improved performance.
• To introduce a deep learning-based reconstruction algorithm for enhanced spatial-spectral data recovery.

Main Methods:

• Proposed a reflective CASSI (R-CASSI) system utilizing a mask and beam splitter for a compact optical path with a single prism acting as two dispersers.
• Developed a U-net-3D deep learning network with 3D convolution kernels for reconstructing spatial-spectral information.
• Released a real-world dataset for benchmarking.

Main Results:

• The R-CASSI system achieves a compact design by using a single prism as two dispersers.
• The U-net-3D network demonstrates state-of-the-art reconstruction performance, ensuring spatial and spectral consistency.
• The released dataset serves as a valuable benchmark for evaluating new reconstruction algorithms.

Conclusions:

• The R-CASSI system offers a promising solution for high-speed spatial-spectral imaging with reduced size and cost.
• Deep learning-based reconstruction significantly enhances the quality of recovered spatial-spectral data.
• The R-CASSI system and dataset pave the way for advancements in spectral imaging applications.