ABSTRACT Owl's hierarchical and context‐sensitive scotopic vision has promised major advances in optoelectronic synapse devices. However, the limited performance of visual artificial intelligence under ultra‐weak illumination remains a critical bottleneck for nighttime recognition in autonomous driving, where conventional sensors exhibit poor classification accuracy. Here, we developed a scene‐aware, three‐level adaptive vision system that integrates van der Waals photodetectors and optical synapses within a unified architecture. Diverse interfacial dynamics in vdW heterostructures enable reconfigurable optoelectronic characteristics, giving rise to abundant synaptic plasticity and good detection performance, including a responsivity of 10 4 A W −1 , and a specific detectivity of 10 11 Jones. On top of that, we construct a three‐level visual‐adaptive hardware system with reconfigurable learning efficiency by integrating a scene‐aware phototransistor circuit to monitor contextual environmental information, 6 × 6 optical synapse arrays for adaptive feature preprocessing, and artificial neural networks for image classification. The owl‐inspired platform achieves a classification accuracy of 84.8% in darkness and dynamically reconfigures its learning efficiency across illumination levels from 0 to 100 mW cm −2 . This work establishes a scene‐aware paradigm for bioinspired night‐vision hardware, bridging neural‐computation principles with 2D optoelectronics to advance intelligent perception in autonomous systems.
Gao et al. (Sun,) studied this question.