Artificial sensory neurons enable high-precision, multicolor, near-infrared object recognition

Near-infrared (NIR) photon detection and object recognition are key technologies for all-weather target identification. Traditional NIR detection systems that rely on photodetectors and Neumann computational algorithms are energy inefficient. Artificial sensory neurons based on infrared-sensitive volatile memristors offer a promising solution.

In a study published in Advanced Materials, a team led by Dr. Wang Jiahong from the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences developed an artificial sensory neuron based on vanadium carbide/oxide (V2C/V2O5-x) heterostructures with topochemical transformation, enabling multicolor near-infrared response and high-precision object recognition in complex scenarios.

The researchers designed a two-dimensional V2C/V2O5-x heterostructure with a natural fusion interface through precisely controlled mild oxidative topochemical transformation of V2CTx. This unique integration of metallic V2C and dielectric vacancy-rich V2O5-x resulted in NIR responsiveness and threshold-type volatile resistive switching (RS) capability of the heterostructure.

The V2C/V2O5-x memristors demonstrated robust volatility capabilities with low coefficients of variation of only 1.62% and 1.7% at set and reset voltages, respectively. Its threshold voltage can be effectively modulated by the power density and wavelength of the NIR light. The correlation between wavelength and threshold firing voltage is consistent with the photoelectric response, indicating tunable photoelectric control of V2C/V2O5-x memristors by photonic parameter modulation.

“Our optoelectronic programmability enables multicolor infrared discrimination through distinctive threshold voltage signatures, allowing us to encode distinct wavelength responses in artificial sensory neurons for near-infrared object recognition,” said Dr. Wang.

An artificial neural network architecture based on the modular RS characteristics of multicolor NIR and the YOLOv7 algorithm model achieved an average recognition accuracy of 89.6% for cars and 85.9% for people on the FLIR dataset.

This study presents a promising memristor-based neuromorphic system that significantly increases the efficiency and accuracy of object detection and recognition, paving the way for advances in autonomous systems, robotics, and intelligent environments.