PnP-ADMM based efficient hardware implementation of a flexible denoising single-pixel imaging system

Single-pixel imaging, as a novel computational imaging method, boasts high sensitivity and interference resistance, offering significant application potential. This paper addresses the noise robustness requirements for real-time imaging environments in current single-pixel imaging systems. A hardware-amenable PnP-ADMM algorithm and its corresponding IP core were developed for implementation in an SPI hardware system. A dedicated computing system, which leverages FPGA technology alongside the PnP-ADMM IP, was designed specifically for enhancing the image recovery using different denoising operators. Thorough evaluation results demonstrate that the proposed PnP-ADMM inversion reconstruction algorithm, under Gaussian noise intensity of 0.05, achieves an average PSNR of 16.7312, 19.1865, and 20.3925 dB for sampling rates of 6.25%, 12.5%, and 25%, respectively. The average SSIM values correspondingly reach 0.40 and 0.44. Experimental results of the hardware system indicate that the reconstruction time for a resolution of 256 x 256 images using the proximal operator, soft-thresholding operator, and TV operator as different denoising operators are approximately 0.350, 0.284, and 0.659 s, respectively. Compared with using only an ARM processor, this represents improvements of 150.1, 184.1, and 82.5 times, respectively. At a 25% sampling rate and imaging distance of 300 cm, the established PnPADMM single-pixel imaging system can achieve a resolution of 0.445-0.5 lp/mm. Our dedicated SPI system outperforms existing works in terms of image quality at the same sampling rate and achieves a larger image size when compared with previous work. Owing to its compact form as compared with conventional computers, the FPGA-based system is poised to expand the application scope of an SPI system into the realms of the IoT and outdoor settings. (c) 2025 Optica Publishing Group. All rights, including for text and data mining (TDM), Artificial Intelligence (AI) training, and similar technologies, are reserved.