Self-Supervised Denoising Framework for Single-Pixel Imaging Based on Image Fusion and Physical Prior

Abstract

Single-pixel imaging (SPI) is strongly affected by a hybrid Poisson–Gaussian noise. Clean labels are difficult to obtain in many SPI experiments, while conventional label-free denoisers generally operate after reconstruction and do not explicitly use the bucket-measurement physics. To address this issue, we propose a self-supervised SPI denoising framework that combines dual-frame image fusion with a physics-guided loss. The method forms two noisy reconstructions from paired measurements of the same static scene, constructs a fused estimate, and constrains residual learning through measurement-domain cross-consistency and image-domain fusion-decomposition consistency. Experiments on Set11, Kodak24, and two real SPI systems show that the proposed method achieves a favorable overall tradeoff among full-reference metrics, no-reference indicators, and local gray-value fidelity compared with representative conventional and learning-based baselines. The results indicate good robustness under severe mixed-noise conditions in both simulated and experimental settings.