Fully Hardware‐Implemented Neuromorphic Visual System Based on a Single‐Layer Memristive Spiking Neural Network

ABSTRACT The integration of sensing, memory, and computation into a unified hardware platform is crucial for energy‐efficient neuromorphic vision at the edge. However, it remains challenging due to the lack of complete hardware‐level implementations. Inspired by the retina‐to‐cortex hierarchy of the biological visual system, we propose a fully hardware‐implemented neuromorphic visual system based on single‐layer memristive spiking neural network that seamlessly integrates these three functions. In this architecture, Pt/NbO x N y /TiN‐based memristive visual neurons directly serve as input neurons, encoding light intensity and colour into spikes, interconnected Pt/TaO y /TaO x /Pt‐based memristive synaptic array performs computing in memory following precise weight mapping via a closed‐loop training module, while Pt/NbO x N y /TiN‐based memristive output neurons implement leaky integrate‐and‐fire dynamics to enable tunable computation and recognition. This highly integrated, fully analog architecture eliminates the need for buffer units and analog‐to‐digital∖ digital‐to‐analog converters, thereby reducing system latency by an order of magnitude and lowering energy consumption by 26.7%. As a proof of concept, the system demonstrates real‐time obstacle avoidance in an autonomous driving task without requiring frame‐based image storage or digital post‐processing. This work establishes a scalable, low‐latency, low‐energy, and fully hardware‐integrated paradigm for real‐time neuromorphic vision, advancing practical edge‐deployed intelligent sensing systems.