High-quality single-pixel imaging via denoiser-driven regularization

We report denoiser-driven single-pixel imaging (SPI) for robust object reconstruction under low sampling ratios and noisy conditions. The method replaces the proximal step in fast iterative shrinkage thresholding algorithm (FISTA) with a convolutional neural network-based denoiser (e.g., U-net). This replacement enables the proposed method to achieve high-quality reconstruction at low sampling ratios in noisy scenarios, outperforming conventional FISTA in terms of fidelity and robustness. To validate effectiveness of the proposed method, an optical experiment is conducted at a low sampling ratio in noisy conditions. Experimental results clearly demonstrate that the proposed method successfully reconstructs target objects with distinct details and conventional FISTA fails to recover recognizable structures. It is believed that the denoiser-driven SPI provides a practical and effective way for high-quality SPI reconstruction in realworld environments.