Spatially variable resolution single-pixel imaging reconstruction based on diffusion transformers

Area of Science:

• Computational imaging
• Signal processing
• Machine learning for imaging

Background:

• Single-pixel imaging (SPI) relies on compressive sensing but suffers quality degradation at low measurement rates.
• Existing SPI reconstruction methods struggle to preserve details in critical image regions with limited data.

Purpose of the Study:

• To develop an advanced SPI reconstruction framework that improves image fidelity, especially in visually important areas, under low measurement conditions.
• To introduce diffusion transformers (DiTs) into SPI reconstruction for the first time, leveraging their generative capabilities.

Main Methods:

• A two-stage cascaded architecture, SVR-DiT-Net, is proposed.
• A spatial variant resolution module (SVR) dynamically allocates measurement resources using progressive pixel-sharing and a prioritized differential loss function.
• A novel image-conditioned guidance mechanism based on diffusion transformers (DiTs) is employed, using SVR's output for iterative denoising and reconstruction.

Main Results:

• The SVR-DiT-Net achieves significant improvements in both global and regional image quality assessments.
• Reconstruction quality is notably enhanced in visually critical areas, even with a limited number of measurements.
• The framework demonstrates adjustable reconstruction quality across different image regions.

Conclusions:

• The proposed SVR-DiT-Net effectively addresses the challenge of SPI reconstruction quality degradation at low measurement rates.
• The integration of diffusion transformers offers a new generative approach for high-fidelity SPI.
• This work provides a novel perspective for generative model-based single-pixel imaging research.