2D Time-Stretching Anisotropic Synapse Realizing In-Sensor Intensity-Spanning Visual Feature Fusion.

Abstract

High-dynamic-range (HDR) visual environments, where extremely bright and dark regions coexist, pose major challenges for conventional imaging systems that rely on multi-frame exposure fusion and cloud-based post-processing. These approaches often suffer from high latency, limited efficiency, and privacy concerns, making them unsuitable for real-time or edge-level intelligent vision. Here, a 2D Time-Stretching Anisotropic Synapse (2D TSAS) is reported that enables in-sensor intensity-spanning feature fusion from a single image frame. The 2D TSAS uniquely integrates two key features of NbOI₂ material: in-plane anisotropy, which gives rise to polarization-resolved optical responses, and a time-stretching photoresponse arising from multi-channel transition-relaxation. This dual-mode mechanism enables direct encoding and temporal integration of spatial-polarization and luminance features during photoexcitation. Leveraging this behavior, a neuromorphic preprocessing strategy is constructed for single-shot visual learning across extreme brightness domains. The system achieves accelerated model convergence with minimal training loss, reaching recognition accuracies of ≈95.41% on NWPU-RESISC45 and ≈95.39% on MNIST. This work offers a compact and efficient solution for contrast-adaptive intelligent vision in complex real-world environments.