In-Sensor Vision Convolution Based on Self-Powered Optoelectronic Memristors for Energy-Efficient Binarized Spiking Neural Networks

Neuromorphic vision systems seek compact, energy-efficient hardware that integrates sensing, preprocessing, and inference for robust perception in dynamic environments. Inspired by the biological retina, we present self-powered optoelectronic memristors based on an Ag/Sb2Se3/FTO architecture that integrates optical detection, analog preprocessing, and synaptic weight storage in a single device. The memristor exhibits stable bipolar switching, long retention, endurance over 500 cycles, and self-powered state-programmable photoresponse, enabling in-sensor convolution without electrical multiply accumulate (MAC) energy. We further incorporate experimental in-sensor convolution into a binarized spiking neural network (BSNN), thereby achieving 98.09% accuracy, compared to 98.42% for the software-based BSNN, while reducing the total network energy to 1.4 μJ, which corresponds to a 97.10% reduction versus a full-precision convolutional neural network (CNN). This work presents a self-powered, physics-aware framework integrating material-level signal transduction with low-power neuromorphic computing, enabling scalable, noise-resilient edge AI vision.