High-Performance Flexible ITZO-Based Deep-Ultraviolet Phototransistors for Retina-Inspired Vision Sensors

Retina-inspired vision sensors can extract input visual information through preprocessing operations to reduce system power consumption and latency. Metal–oxide thin-film transistor (MOTFT)-based phototransistors are attractive for in-sensor computing, but are often limited by sluggish temporal response, constraining high-frequency and time-dependent operation. To address this constraint, a flexible InSnMgO (ITMO)/InSnZnO (ITZO)/ITMO heterojunction deep-ultraviolet phototransistor (HDPT) was developed as an active vision sensor that exhibited stable photosynaptic behavior and emulated retinal light adaptation. Under 5-ms deep-ultraviolet (DUV) pulses, the device sustained 100-Hz operation with the 26-fJ/spike single-event energy and enabled 4-bit temporal sequence recognition. By combining blockwise superpixel aggregation and the HDPT-enabled 4-bit temporal encoding, the network input dimension was reduced to 25% of the original size while achieving over 95% accuracy in static image recognition. For dynamic perception, a 95% accuracy was achieved on dynamic gesture recognition. These results indicate hardware-level feature encoding and dimensionality reduction at the point of capture, enabling low-power, dynamic in-sensor vision.