CTIS spectral image reconstruction technology based on slit-scanning architecture

Area of Science:

• Optics and Photonics
• Computational Imaging
• Spectroscopy

Background:

• Computed tomography imaging spectrometers (CTIS) excel at dynamic detection but face spectral reconstruction challenges due to the missing-cone problem.
• Traditional CTIS algorithms and deep learning models trained on simulated data struggle with real-world spectral data accuracy.
• The discrepancy between simulated and real data limits the applicability of current CTIS reconstruction models.

Purpose of the Study:

• To propose a novel CTIS system using a slit-scanning architecture to enhance spectral reconstruction accuracy.
• To generate a real dataset for training CTIS reconstruction networks by acquiring accurate spectral cubes from real scenes.
• To develop and validate an advanced deep learning model for improved CTIS spectral reconstruction.

Main Methods:

• Developed a slit-scanning CTIS system with an adjustable aperture to limit the field of view and reduce diffraction overlap.
• Utilized the expectation-maximization (EM) algorithm for spectral reconstruction, enhanced by the slit-scanning architecture.
• Constructed a residual neural network incorporating multi-scale and attention mechanisms, trained on both simulated and real spectral imaging data.

Main Results:

• The slit-scanning CTIS architecture successfully acquired real spectral data cubes matching real-world diffractive images.
• The proposed deep learning model demonstrated superior spectral reconstruction accuracy compared to the EM algorithm and standard convolutional neural networks.
• The study validated the critical importance of using real spectral data for effective CTIS reconstruction.

Conclusions:

• The slit-scanning CTIS architecture offers a viable solution for accurate spectral data cube acquisition in real-world scenes.
• Training deep learning models with real spectral data significantly improves CTIS spectral reconstruction performance.
• The developed multi-scale, attention-based residual network represents a significant advancement in CTIS spectral reconstruction accuracy.