Novel Heterostructure-based Artificial Neuron Enables High-precision Multi-Color Near-Infrared Object Recognition

Near-infrared (NIR) photon detection and object recognition excel in low-light and adverse conditions, enabling robust applications in night vision, autonomous driving, surveillance, and medical imaging. However, conventional approaches face energy inefficiency and data bottlenecks due to separated photodetectors and von Neumann computing.

Artificial sensory neurons based on infrared-sensitive volatile memristors offer a promising approach. Yet most existing photoresponsive memristors are limited to visible or ultraviolet light, hindering multi-color near-infrared sensing and recognition.

Recently, a research team led by Dr. WANG Jiahong from the Shenzhen Institutes of Advanced Technology (SIAT) of the Chinese Academy of Sciences, developed an artificial sensory neuron based on a vanadium carbide/oxide (V₂C/V₂O_5-x) heterostructure via topochemical conversion, enabling multi-color near-infrared response and high-precision object recognition in complex scenarios.

The study was published in Advanced Materials on September 12.

In this study, the researchers engineer a 2D V₂C/V₂O_5-x heterostructure with a natural fusion interface through a precisely controlled mild-oxidation topochemical conversion of V₂CT_x. This unique integration of metallic V₂C and dielectric vacancy-enriched V₂O_5-x grants the heterostructure NIR responsivity and threshold-type volatile resistance switching (RS) ability.

The V₂C/V₂O_5-x memristor demonstrates robust volatile capability, showing low coefficients of variation (C_v) of merely 1.62% and 1.7% for the set and reset voltages, respectively, and its threshold voltage can be effectively modulated by the power density and wavelength (785-1550 nm) of NIR light.

Moreover, the correlation between wavelength and threshold firing voltage is consistent with the photoelectric response, demonstrating tunable photoelectric control of the V₂C/V₂O_5-x memristor via photonic parameter modulation.

"Our photoelectric programmability enables multi-color infrared discrimination through characteristic threshold voltage signatures, and the distinct wavelength responses can be encoded in the artificial sensory neuron for near-infrared object recognition." said Dr. WANG.

By integrating the V₂C/V₂O_5-x memristor-based sensory neuron with the YOLOv7 model, the researchers achieved 87.7% and 89.6% average accuracy for vehicle and pedestrian recognition on the FLIR driving dataset, demonstrating robust perception in complex driving environments.

The study presents a highly promising memristor-based neuromorphic system that significantly enhances efficiency and accuracy in object detection and recognition, paving the way for advancements in autonomous systems, robotics, and intelligent environments.

Schematic illustration of a 2D vanadium carbide/oxide heterostructure-based artificial sensory neuron for multi-color near-infrared object recognition. (Image by SIAT)