Self-supervised learning for single-pixel imaging via dual-domain constraints.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Single-pixel imaging (SPI) is effective for compressive sensing but relies on supervised learning, which is data-intensive and limits generalization.
• Existing supervised methods for SPI reconstruction require extensive training data and struggle with complex scenarios.

Purpose of the Study:

• To develop a self-supervised learning method for SPI reconstruction that overcomes the limitations of conventional supervised approaches.
• To improve the efficiency and generalization capabilities of deep-learning-augmented SPI.

Main Methods:

• Introduced a self-supervised learning framework for SPI reconstruction by integrating the SPI physics model into a neural network.
• Employed dual-domain constraints, including a traditional measurement constraint and a novel transformation constraint for target plane consistency.
• Utilized the invariance of reversible transformations to provide an implicit prior, addressing the non-uniqueness issue of measurement constraints.

Main Results:

• Achieved self-supervised SPI reconstruction in diverse and complex scenes without requiring paired data, ground truth, or pre-trained models.
• Demonstrated the method's ability to handle underdetermined degradation and noise effectively.
• Reported approximately 3.7-dB improvement in Peak Signal-to-Noise Ratio (PSNR) compared to existing methods.

Conclusions:

• The proposed self-supervised learning method offers a robust and efficient solution for SPI reconstruction.
• This approach significantly enhances generalization and reduces the reliance on laborious data acquisition and training.
• The dual-domain constraint strategy provides a powerful tool for improving SPI performance in real-world applications.