Single-pixel imaging technology, as a novel computational imaging technique, features high sensitivity and interference resistance. By combining compressive sensing theory with single-pixel imaging technology, sampling time and storage resource consumption can be effectively reduced. However, in current research on FPGA-based single-pixel imaging reconstruction algorithms, researchers often struggle to achieve both algorithmic reconstruction quality and reconstruction speed. We introduce the alternating direction method of multipliers based on plug-and-play (PnP-ADMM) into FPGA-based single-pixel imaging systems to enhance both image reconstruction quality and speed. The numerical simulations and experiments demonstrate that the established PnP-ADMM system on chip (SOC) single-pixel imaging system can accurately reconstruct target scenes and preserve image details, with strong noise suppression capacity.

By incorporating the PnP-ADMM algorithm into a single-pixel imaging system, we break down the original large-scale optimization problem into multiple sub-problems to increase expedited computation speed. The algorithm also adopts a plug-and-play framework to introduce denoising operators to denoise reconstructed signals during algorithm iterations. Furthermore, we design a hardware structure for the PnP-ADMM algorithm, leveraging FPGA platforms to accelerate calculation processing. Additionally, based on block-based compressive sensing, we process target images in blocks, further speeding up calculation processing while effectively conserving hardware resource consumption.

The numerical simulation results demonstrate that the established PnP-ADMM algorithm achieves a PSNR of 24.82 dB and 29.64 dB for reconstructed images at sampling rates of 12.50% and 25.00%, respectively, while the SSIM reaches 0.75 and 0.88, respectively (Fig. 5). The test result of the designed hardware structure for the PnP-ADMM algorithm reveals that at a sampling rate of 25.00%, the duration of reconstructing a 256 pixel×256 pixel image by PnP-ADMM SOC using proximity operator, soft thresholding operator, and total variation operator is 0.369 s, 0.303 s, and 0.681 s, respectively (Table 2). This represents an improvement of 141.9 times, 172.1 times, and 80.3 times respectively compared to using only an ARM processor (Table 2). Furthermore, the imaging experiments for validation are completed by constructing a single-pixel experimental platform. The experimental results confirm that under the imaging conditions at a distance of 300 cm, the PnP-ADMM SOC single-pixel imaging system achieves a resolution of 0.445-0.500 lp/mm for images with 256 pixel×256 pixel resolution (Fig. 9).

We apply the PnP-ADMM algorithm to FPGA-based single-pixel imaging systems to enhance both image reconstruction quality and speed on the FPGA platform. Numerical simulation results demonstrate that the image quality reconstructed using the proposed PnP-ADMM inversion reconstruction algorithm surpasses that of the WQR-OMP algorithm and TVAL3 algorithm. The test results of the designed hardware acceleration structure for the PnP-ADMM algorithm show that this structure effectively accelerates the algorithm’s computation processing. Furthermore, the establishment of a single-pixel experimental platform confirms that the PnP-ADMM SOC single-pixel imaging system can accurately reconstruct target scenes and preserve target details, with strong noise suppression capability.