Single-pixel imaging reconstruction based on a complementary frequency-domain filter mask with classifier-free...

Area of Science:

• Optics and photonics
• Computational imaging
• Machine learning for image processing

Background:

• Single-pixel imaging aims for high-quality image reconstruction with minimal measurements.
• Current deep learning methods are limited by optimizing image-domain loss, hindering low-measurement performance.

Purpose of the Study:

• To develop an advanced single-pixel reconstruction algorithm overcoming limitations of existing deep learning approaches.
• To enhance image reconstruction quality at low measurement rates using frequency-domain analysis.

Main Methods:

• Proposed a single-pixel reconstruction algorithm utilizing a complementary frequency-domain filter mask classifier model.
• Designed a regulation mask of complementary filters and integrated classifier-free guidance.
• Leveraged frequency-domain multi-dimensional information for restoring image details.

Main Results:

• Achieved an average peak signal-to-noise ratio of 28.82 dB on the MNIST dataset at a 10% measurement rate.
• Demonstrated superior performance in high- and low-frequency detail restoration.
• Validated excellent performance across diverse dataset scenarios.

Conclusions:

• The proposed complementary frequency-domain filter mask classifier model with classifier-free guidance offers a significant advancement in single-pixel imaging reconstruction.
• The algorithm effectively restores image details at low measurement rates, outperforming existing methods.
• Further research into parameter adjustment schemes provides valuable insights for practical applications.