Broadband Single Pixel Video Imaging via Physics-driven Deep Reconstruction

Graphene/silicon Schottky photodetectors offer broadband responsivity and fast dynamics. However, the difficulty of fabricating large arrays limits their deployment in high-resolution broadband imaging. We address this challenge with a computational single-pixel imaging system that integrates structured optical encoding with a physics-driven deep unfolding reconstruction network. Our multilayer graphene/silicon detector provides broadband responsivity, high linearity, and a∼30 ns rise time, critical for synchronized multiplexed acquisition and thus essential for video-rate imaging. Reconstruction is performed through a deep-unfolding framework that combines 2D tensor-based gradient descent with attention-guided refinement for efficient learning. By embedding physical optical encoding within the forward model, our approach enables high-fidelity recovery, even from heavily undersampled data. Using this system, we demonstrate static imaging at a resolution of 256 ×256 pixels and dynamic video at a resolution of 64 ×64 pixels at >10 frames per second, spanning 375–1342 nm. Our work establishes the first broadband single-pixel video imaging with graphene/silicon detectors, underscoring the synergy between physics-driven reconstruction and versatile hardware in advancing compact and scalable broadband vision systems.