Single pixel imaging via unsupervised deep compressive sensing with collaborative sparsity in discretized feature...

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning

Background:

• Single-pixel imaging (SPI) offers low-cost sensing beyond the visible spectrum.
• Sub-Nyquist sampling in SPI is hindered by spectrum truncation and discretization, limiting traditional reconstruction.
• Existing deep learning methods often require retraining for different SPI parameters.

Purpose of the Study:

• To develop a robust deep compressive sensing (CS) framework for single-pixel imaging reconstruction.
• To enable collaborative sparsity in feature space for improved image recovery.
• To create an unsupervised learning approach that avoids retraining for varying sensing parameters.

Main Methods:

• A novel deep compressive sensing (CS) framework with a compression network was designed.
• The network was trained in an unsupervised manner, independent of specific sensing physics.
• Validation was performed using numerical simulations, physical experiments, and fluorescence imaging.

Main Results:

• The proposed method achieved comparable image fidelity to sCMOS cameras at sampling ratios as low as 4%.
• It maintained the inherent advantages of SPI, such as low cost and broad spectral range.
• The unsupervised and self-contained nature facilitates downstream applications in compressive imaging.

Conclusions:

• The unsupervised deep CS framework effectively addresses sub-Nyquist sampling limitations in SPI.
• This approach enhances image reconstruction quality and broadens the applicability of SPI.
• The technique shows significant promise for advanced sensing and imaging applications.