Low-sampling and noise-robust single-pixel imaging based on the untrained attention U-Net

Area of Science:

• Optics
• Image Processing
• Machine Learning

Background:

• Single-pixel imaging (SPI) uses structured light and a single-pixel detector (SPD).
• SPI is limited by white noise during detection, degrading image quality.
• Existing methods struggle with noise reduction and low sampling rates.

Purpose of the Study:

• To reduce noise in SPI using an untrained attention U-Net.
• To achieve high-quality imaging at low sampling rates.
• To improve the applicability of SPI in noisy environments.

Main Methods:

• Combining an untrained attention U-Net with the SPI model.
• Utilizing the attention mechanism to highlight image features and suppress noise.
• Employing numerical simulations and experimental validation.

Main Results:

• Effective reduction of various levels of Gaussian white noise.
• Superior imaging quality compared to existing methods at sampling rates below 10%.
• Demonstrated generalization without requiring pre-training.

Conclusions:

• The proposed method significantly enhances SPI performance in noisy conditions.
• Untrained attention U-Net offers a robust solution for low-sampling-rate SPI.
• This work broadens SPI's potential applications in complex noise environments.