A Bioinspired Optic-Neural Device with Neuromorphic Spatiotemporal Visual Perception for Dynamic Object Recognition.

Neuromorphic vision with integrated sensing, memorizing, and computing is considered a potential way to break through the high latency and storage bottlenecks of existing machine vision. However, simulating the human vision system with a single device to achieve spatiotemporal visual information perception remains a challenge. Hence, we demonstrated a bioinspired optic-neural device that achieves dynamic object recognition by simultaneously implementing encoding and sensing functions on molybdenum disulfide (MoS2). This device exhibits excellent optical synaptic plasticity originating from charge trapping and recombination in the device. The electrical output is weighted by the number of optical inputs, which reflects the history information on optical inputs. This mechanism is similar to the spatiotemporal information encoding ability of the lateral geniculate nucleus in the human vision system, enabling the device to remember historical visual information as well as achieve spatial resolution. Benefiting from the optic-neural device nonlinearly encoding and mapping a series of light inputs when arranged in an array, we achieved spatiotemporal visual information perception for dynamic object recognition similar to that of the human vision system. High classification accuracy up to 96.3% in five dynamic object data sets with high energy efficiency and low computational load are achieved based on such neuromorphic vision system. Our research findings have profound significance for promoting the advancement of existing machine vision systems.