High-resolution iterative reconstruction at extremely low sampling rate for Fourier single-pixel imaging via...

Area of Science:

• Optics and Imaging Technologies
• Computational Imaging
• Artificial Intelligence in Imaging

Background:

• Fourier single-pixel imaging (FSPI) faces a trade-off between imaging efficiency and quality.
• High-resolution imaging in FSPI typically requires more measurements, reducing efficiency.
• Existing reconstruction methods struggle to balance speed and resolution.

Purpose of the Study:

• To develop a novel, high-quality reconstruction method for FSPI using diffusion models.
• To achieve high-resolution FSPI reconstruction at extremely low sampling rates.
• To overcome the inherent efficiency-quality trade-off in traditional FSPI.

Main Methods:

• A score-based diffusion model was designed to learn data distribution priors.
• The model was iteratively constrained using real-sampled low-frequency Fourier spectrum data.
• The method integrates learned priors with spectral consistency for reconstruction.

Main Results:

• The proposed diffusion model method significantly improved image quality compared to traditional FSPI and U-Net.
• At a 1% sampling rate, the method showed a 241% improvement in edge intensity score for a coin experiment.
• Simulations and experiments validated the superior performance of the diffusion model approach.

Conclusions:

• The diffusion model offers a high-quality reconstruction solution for FSPI, especially at low sampling rates.
• This method has the potential for high-resolution imaging without sacrificing imaging speed.
• The approach broadens the applicability of FSPI in practical scenarios requiring both speed and detail.