Reconstructing obscured information under extreme noise via multichannel single-pixel imaging

In experimental mechanics, imaging is a vital method for information acquisition. A persistent challenge in this field is photographing scenes with an extreme dynamic range, a problem rooted in the intrinsic limitations of conventional sensors. To address the critical challenge of reconstructing information obscured by localized overexposure in high-dynamic-range (HDR) imaging, a novel computational framework has been proposed that synergistically integrates multi-detector single-pixel imaging (SPI) with blind source separation (BSS) and a self-supervised adaptive fusion algorithm. By processing pixel-level aligned one-dimensional (1D) measurements from multiple photodetectors (PDs), the method separates target information and noise directly in the signal domain prior to image reconstruction, overcoming the inherent entropy limits of single-image processing. The fusion process is guided by no-reference quality metrics optimized via a genetic algorithm without requiring ground-truth data. Validated in challenging scenarios, including strong backlighting, QR codes under acrylic protection, and specular reflections, our approach significantly suppresses artifacts and enhances contrast and detail visibility. Quantitative results demonstrate a maximum SNR enhancement of 624% under extreme noise conditions compared with SPI utilizing a single PD. This work establishes a robust solution for high-fidelity imaging in extreme dynamic range environments.