Efficient Kronecker-based deep unfolding network for single-pixel imaging with hardware implementation.

Area of Science:

• Computational Imaging
• Deep Learning
• Image Reconstruction

Background:

• Single-pixel imaging (SPI) is an emerging computational imaging technique.
• Deep learning has enhanced SPI reconstruction quality and speed.
• High computational costs for large images hinder practical SPI applications.

Purpose of the Study:

• To develop an efficient deep learning network for single-pixel imaging (SPI) reconstruction.
• To address the challenge of increasing computational costs in large-scale SPI.
• To enable practical, high-quality SPI reconstruction.

Main Methods:

• Proposed K-CPP Net, a gradient-descent-based deep unfolding network.
• Coupled a Kronecker-based SPI model with a Chambolle-Pock-inspired primal-dual optimization algorithm.
• Utilized Kronecker product for efficient sensing matrix representation and a lightweight denoising module with residual convolutional blocks and channel-wise self-attention.

Main Results:

• Substantially reduced computational costs for SPI reconstruction.
• Enabled full-image SPI training and inference.
• Achieved superior reconstruction quality, suppressing distortion and blur while preserving fine details.
• Demonstrated improved computational efficiency compared to existing methods.

Conclusions:

• K-CPP Net offers an efficient and effective solution for single-pixel imaging reconstruction.
• The method significantly lowers computational demands, making large-scale SPI more feasible.
• Implementation on FPGA provides a practical reference for real-world SPI systems.