Scholarlink

<p>Current computer vision is data-intensive and faces bottlenecks in shrinking computational costs. Incorporating physics into a bioinspired visual system is promising to offer unprecedented energy efficiency, while the mismatch between physical dynamics and bioinspired algorithms makes the processing of real-world samples rather challenging. Here, we report a bioinspired in-materia analogue photoelectronic reservoir computing for dynamic vision processing. Such system is built based on InGaZnO photoelectronic synaptic transistors as the reservoir and a TaO<sub>X</sub>-based memristor array as the output layer. A receptive field inspired encoding scheme is implemented, simplifying the feature extraction process. High recognition accuracies (&gt;90%) on four motion recognition datasets are achieved based on such system. Furthermore, falling behaviors recognition is also verified by our system with low energy consumption for processing per action (~45.78 μJ) which outperforms most previous reports on human action processing. Our results are of profound potential for advancing computer vision based on neuromorphic electronics.</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

A bioinspired in-materia analog photoelectronic reservoir computing for human action processing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年3月6日

查看全文
  • Hangyuan Cui,
  • Yu Xiao,
  • Yang Yang,
  • Mengjiao Pei,
  • Shuo Ke,
  • Xiao Fang,
  • Lesheng Qiao,
  • Kailu Shi,
  • Haotian Long,
  • Weigao Xu,
  • Pingqiang Cai,
  • Peng Lin,
  • Yi Shi,
  • Qing Wan,
  • Changjin Wan

摘要

Current computer vision is data-intensive and faces bottlenecks in shrinking computational costs. Incorporating physics into a bioinspired visual system is promising to offer unprecedented energy efficiency, while the mismatch between physical dynamics and bioinspired algorithms makes the processing of real-world samples rather challenging. Here, we report a bioinspired in-materia analogue photoelectronic reservoir computing for dynamic vision processing. Such system is built based on InGaZnO photoelectronic synaptic transistors as the reservoir and a TaOX-based memristor array as the output layer. A receptive field inspired encoding scheme is implemented, simplifying the feature extraction process. High recognition accuracies (>90%) on four motion recognition datasets are achieved based on such system. Furthermore, falling behaviors recognition is also verified by our system with low energy consumption for processing per action (~45.78 μJ) which outperforms most previous reports on human action processing. Our results are of profound potential for advancing computer vision based on neuromorphic electronics.