Bioinspired Motion Perception Using the Wavelength-Sensitive Nonvolatile MoTe2 Homojunction

Abstract

Neuromorphic sensors with the capability to simultaneously sense and process dynamic images show great potential for data-centric in-sensor motion perception. However, the presence of persistent background noise and redundant computational burdens limits perception accuracy and efficiency. Here, we present wavelength-sensitive bioinspired neuromorphic devices that integrate the high-level preprocessing of the human retina and ultraviolet feature extraction of the fisheye, enabling efficient image compression and background filtering capabilities to enhance motion perception. By storing dynamic image information as photoresponsivity in our nonvolatile MoTe₂ homojunction, mass data can be compressed through multiplying trainable light intensity and photoresponsivity. This induces an n-fold reduction in backend computational requirements for n × n image motion perception while preserving a recognition accuracy of 100%. Furthermore, we successfully identify and track the motion of anemonefish obscured in the indistinguishable background, achieving enhanced recognition accuracy from 70% to 100%. This work showcases the substantial potential for developing intelligent vision systems with feature information extraction and high-level computational capabilities.