Neural networks based on in-sensor computing of optoelectronic memristor

Abstract

The separation band of perception, storage, and computation modules in vision systems based on traditional von Neumann architectures leads to latency and power consumption problems in data transmission, which severely limits the computational power. In recent years, in-sensor computing has gained significance in enhancing the computational performance of machine vision systems. It integrates sensing, storage and computation and is an important way to break out of the Von Neumann architecture. This study introduces an optoelectronic memristor-based image recognition algorithm to improve recognition efficiency by performing image feature extraction in a hardware array. The experimental results show that the network achieves the best accuracy of 93.26% after 30 epochs, and the loss of accuracy after weight quantization is about 1%.