Computational spectral imaging: a contemporary overview.

Area of Science:

• Optics
• Computer Vision
• Data Science

Background:

• Spectral imaging (SI) captures 3D spectral data cubes for object and material identification.
• Acquiring 3D SI with conventional 1D/2D sensors is challenging due to hardware limitations.
• Computational spectral imaging (CSI) offers an alternative using 2D encoded projections to retrieve 3D SI.

Purpose of the Study:

• To review advancements in computational spectral imaging (CSI).
• To explore the integration of deep learning (DL) with CSI for improved performance.
• To highlight the combination of optical design and DL for advanced spectral data analysis.

Main Methods:

• Utilizing 2D encoded projections for data acquisition in CSI.
• Employing computational recovery processes to reconstruct 3D spectral images.
• Integrating deep learning algorithms for enhanced SI reconstruction and high-level task execution.

Main Results:

• CSI enables snapshot optical systems, reducing acquisition time and storage costs.
• Deep learning significantly improves SI reconstruction from 2D projections.
• DL-powered CSI can perform classification, unmixing, and anomaly detection directly from encoded data.

Conclusions:

• CSI, particularly with deep learning, offers a powerful approach to spectral imaging.
• The synergy between optical system design and DL algorithms advances spectral data analysis.
• Future research directions involve combining physical optics with computational DL for complex tasks.