SPyRiT 3.0: an open source package for single-pixel imaging based on deep learning.

Area of Science:

• Computational optics
• Image reconstruction
• Deep learning applications

Background:

• Single-pixel imaging (SPI) relies on reconstruction algorithms to form images from limited measurements.
• Deep learning (DL) methods have advanced SPI reconstruction but pose reproducibility challenges.
• A need exists for tools to ensure reproducibility and benchmarking in DL-based SPI.

Purpose of the Study:

• Introduce SPyRiT, an open-source PyTorch toolbox for SPI.
• Facilitate reproducible benchmarking of DL reconstruction algorithms in SPI.
• Compare supervised and plug-and-play DL methods for SPI.

Main Methods:

• Developed SPyRiT, a versatile, open-source toolbox using PyTorch.
• Implemented various supervised and plug-and-play DL reconstruction methods.
• Evaluated method performance across different signal-to-noise ratios (SNRs) and datasets.

Main Results:

• Supervised DL methods trained on simulated data perform well on experimental data with sufficient SNR.
• Plug-and-play methods offer flexibility for lower SNRs through hyperparameter tuning.
• DC-Net demonstrates robustness to noise, comparable to plug-and-play methods without manual tuning.

Conclusions:

• SPyRiT enables rigorous benchmarking of DL reconstructions in SPI and related fields like ghost imaging.
• The toolbox supports further research and application of DL in computational optics modalities.
• SPyRiT promotes reproducibility and advancement in DL-based image reconstruction.