High‐Performance Neuromorphic Visual System With Image Pre‐Processing via Normally‐Off GaN MOS‐HEMT

ABSTRACT

This study presents a novel image processing and recognition system enabled by an enhancement‐mode   recessed‐gate GaN MOSHEMT artificial synapse. By leveraging trap modulation near the AlGaN/GaN heterojunction, long‐term potentiation and long‐term depression characteristics were well demonstrated, ensuring high linearity and symmetry required for neural computing. Upon being implemented in an artificial neural network, this approach achieves a recognition accuracy of 98.27% on the MNIST dataset of handwritten digits. Additionally, an image compression preprocessing scheme capable of ultra‐high compression ratio is implemented based on the photoelectric conversion and memory capabilities, thereby significantly enhancing transmission efficiency. This work demonstrates a promising approach for developing comprehensive signal‐processing systems that include sensing, storage, and computing, paving the way for advanced human‐machine interaction and visual neuromorphic technologies.