K-FISTA: a Kronecker-domain FISTA unrolling network for flexible single-pixel imaging across sampling ratios and...

Area of Science:

• Computational Imaging
• Signal Processing
• Machine Learning for Imaging

Background:

• Single-pixel imaging (SPI) offers hardware simplicity and spectral flexibility but struggles with high-quality reconstruction across diverse sampling ratios and resolutions.
• Existing SPI reconstruction methods face challenges in adapting to varying acquisition parameters and maintaining image fidelity.

Purpose of the Study:

• To develop a flexible and robust single-pixel imaging reconstruction framework.
• To address the limitations of current SPI reconstruction techniques in handling varying sampling ratios and image resolutions.

Main Methods:

• Proposed K-FISTA, a Kronecker-domain fast iterative shrinkage-thresholding algorithm (FISTA) unrolling network for SPI.
• Integrated learnable modules for momentum extrapolation, gradient descent, and prior modeling within a FISTA-inspired unrolling paradigm.
• Employed tensor-form gradient descent for measurement fidelity, cross-stage state momentum refinement (CSMR), and a deep restoration prior with attention mechanisms.

Main Results:

• K-FISTA demonstrated accurate, efficient, and flexible SPI reconstruction across sampling ratios from 1% to 50% and multiple resolutions (128x128 to 512x512).
• Achieved significant PSNR/SSIM improvements on the Set11 dataset, e.g., 22.93 dB/0.6426 at 1% and 28.33 dB/0.8475 at 4% sampling.
• Showcased robustness in severely undersampled scenarios and practical adaptability to varying acquisition settings.

Conclusions:

• K-FISTA provides a unified and adaptable solution for high-quality single-pixel imaging reconstruction.
• The proposed deep learning framework effectively handles diverse sampling ratios and resolutions, outperforming existing methods.
• K-FISTA represents a significant advancement in enabling practical and versatile single-pixel imaging applications.