Dual-constraint reconstruction network with semantic-discriminative consistency for radon single pixel imaging

Area of Science:

• Optics and Imaging Technologies
• Computer Vision and Machine Learning
• Signal Processing

Background:

• Single-pixel imaging (SPI) offers advantages in terahertz, remote sensing, and hyperspectral applications.
• Radon SPI excels in classification via radon transform domain feature extraction.
• Low sampling rates severely degrade radon SPI performance, necessitating quality enhancement.

Purpose of the Study:

• To develop a deep learning-based reconstruction method for improving radon SPI quality at low sampling rates.
• To address semantic distortions and quality deterioration inherent in current reconstruction techniques.
• To enhance image detail restoration and suppress artifacts in low-sampling radon SPI.

Main Methods:

• A simulated dataset of moving targets (birds) was created using radon projection reconstruction principles.
• A dual-branch generative adversarial network (GAN) with semantic latent vector modulation was designed.
• Dual-constraint mechanisms ensured latent space semantic consistency and preserved image discriminative information.

Main Results:

• The proposed deep learning method effectively reconstructs high-quality images from low-sampling radon SPI data.
• Significant enhancement in image detail restoration capabilities was achieved.
• Hallucination artifacts were effectively suppressed, improving reconstruction fidelity.
• High-quality images were obtained at sampling rates of 2%, 5%, and 25%.

Conclusions:

• The developed deep learning approach successfully overcomes the limitations of radon SPI at low sampling rates.
• The method demonstrates robust performance in suppressing artifacts and restoring image details.
• The open-sourcing of the dataset and code will facilitate further research in this area.