All-Sputtered Oxide n-p Heterojunction Anti-Ambipolar Transistors for Semantic Segmentation Networks

초록

Neuromorphic vision provides an efficient route for extracting meaningful information from visual inputs and supporting intelligent perception in dynamic environments. However, visual scenes acquired under excessive or insufficient illumination often contain distorted contrast, weakened boundary information, and shifted input distributions, which degrade semantic segmentation accuracy before network computation. Conventional synaptic or neuromorphic hardware systems usually require additional normalization or preprocessing steps to compensate for such input distortion, limiting the direct integration of signal regulation at the device level. In this work, an all-sputtered oxide anti-ambipolar transistor (AAT) based on an IGZO/TeSeOx heterojunction is introduced as a device-level preprocessing element for illumination-robust neuromorphic vision. IGZO/TeSeOx AAT exhibits a well-defined Gaussian-like anti-ambipolar transfer characteristic, with a peak-to-valley current ratio of 1.17 & times; 103 and a maximum negative transconductance of -240.8 nS, enabling effective attenuation of excessive input signals. The resulting Gaussian-like anti-ambipolar response is used to construct Gaussian smoothing and Laplacian of Gaussian filtering kernels. Applied to semantic segmentation under overexposed and underexposed conditions, the proposed preprocessing scheme improves pixel-wise classification consistency and achieves 86.1% pixel accuracy.

키워드