Scholarlink

<p>The rapid proliferation of edge-computing applications, including autonomous vehicles, wearable electronics and mobile robotics, is driving demand for compact vision systems capable of real-time intelligent processing under strict energy and latency constraints. Conventional imaging architectures, however, separate sensing from computation, producing large data streams that increase power consumption and system complexity. Here we report electrochromic hyperspectral embedding, an in-sensor computing framework that adaptively compresses spectral information at the pixel level. Our approach exploits electrically tunable photocurrent responses in electrochromic photodetectors, enabling each pixel to selectively encode its most task-relevant spectral components before readout. The resulting low-dimensional outputs interface directly with lightweight memristor-based analogue computing hardware for efficient inference. Compared with conventional artificial intelligence vision systems, electrochromic hyperspectral embedding reduces data transmission by more than an order of magnitude while maintaining high classification accuracy, offering a materials-to-system pathway towards compact, high-speed and energy-efficient intelligent vision for scalable edge applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Electrochromic hyperspectral embedding for ultracompact intelligent vision

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Ran Li,
  • Chaoyi He,
  • Yi Huang,
  • Enzi Zhai,
  • Vinod K. Sangwan,
  • Jingyi Zou,
  • Kevin St. Luce,
  • Jianzhou Cui,
  • Rex Kim,
  • Roland Liu,
  • Zhixing Wang,
  • Xi Ling,
  • Helen Xie,
  • Xu Zhang,
  • Mark C. Hersam,
  • Qiangfei Xia,
  • Linda Katehi,
  • Yuxuan Cosmi Lin

摘要

The rapid proliferation of edge-computing applications, including autonomous vehicles, wearable electronics and mobile robotics, is driving demand for compact vision systems capable of real-time intelligent processing under strict energy and latency constraints. Conventional imaging architectures, however, separate sensing from computation, producing large data streams that increase power consumption and system complexity. Here we report electrochromic hyperspectral embedding, an in-sensor computing framework that adaptively compresses spectral information at the pixel level. Our approach exploits electrically tunable photocurrent responses in electrochromic photodetectors, enabling each pixel to selectively encode its most task-relevant spectral components before readout. The resulting low-dimensional outputs interface directly with lightweight memristor-based analogue computing hardware for efficient inference. Compared with conventional artificial intelligence vision systems, electrochromic hyperspectral embedding reduces data transmission by more than an order of magnitude while maintaining high classification accuracy, offering a materials-to-system pathway towards compact, high-speed and energy-efficient intelligent vision for scalable edge applications.