A Comparison Between Fourier and Hadamard Single-Pixel Imaging in Deep Learning-Enhanced Image Reconstruction

Numerous single-pixel imaging (SPI) schemes exist to reconstruct images, with the most notable schemes being Hadamard SPI (HSPI) and Fourier SPI (FSPI). To date, comparisons between both methods have been made only within the conventional optical image processing setting. However, with recent advancements in deep learning (DL), image restoration models have exhibited considerable performance improvements, which could potentially be reformulated to enhance existing SPI schemes. In this letter, we present the first-known comparison of conventional HSPI, FSPI, and their DL-enhanced variants, based on state-of-the-art nonlinear activation free network. The experiments were conducted by reconstructing the images of the STL-10 dataset, followed by evaluations on the Set11, Set14, BSD68, and Urban100 test sets. Our experimental results show that DL-enhanced FSPI and HSPI achieved substantial performance gains on peak-signal-to-noise ratio (PSNR) and structural similarity index measure. Interestingly, the improvement trend in PSNR for FSPI is inconsistent with HSPI due to the presence of reconstructed graphical artifacts at higher sampling rates.