Device–System Co‐Design of a Memristor‐Based 3D Convolutional Spiking Neural Network for Dynamic Vision Processing

ABSTRACT

Dynamic image processing is a vital component of artificial intelligence and Internet of Things (IoT) applications. However, conventional image processing systems often suffer from high computational complexity and substantial power consumption. Memristors, as emerging nanoscale devices, offer promising advantages in terms of high‐density integration and energy‐efficient computing. In this work, we present a three‐dimensional convolutional spiking neural network (3D CSNN) system capable of efficiently processing dynamic visual data. A bio‐inspired spike encoder is implemented using volatile NbO _x memristors, and its encoding characteristics are dynamically adjustable by introducing a time window strategy. In parallel, we introduce a high‐yield, stable anodic oxidation method for fabricating non‐volatile TaO _x memristors. These devices are integrated into an array to perform 3D convolution operations, enabling effective extraction of both spatial and temporal features from dynamic input. By combining the low energy consumption of memristor‐based analog circuits with the high precision of digital logic, the digital‐analog hybrid system achieves a classification accuracy of 97% in gesture recognition tasks. This work establishes a scalable hardware paradigm for low‐power neuromorphic computing, providing new opportunities for next‐generation artificial vision and edge intelligence.