Scholarlink

<p>Unpredictable light pollution such as glare, laser beams and light-emitting diode sources poses a major challenge to machine vision systems in autonomous driving and humanoid robotics, where reliability is essential for safety and efficiency. Here we introduce a machine learning-based tunable band-stop photodetector that combines a bio-inspired visual perception strategy with a band-stop centre wavelength dynamically defined by a bias-controlled operating point (<i>V</i><sub>t</sub>). A deep learning model within an incremental learning framework maps incident spectral features to optimal voltage settings, enabling continuous self-calibration under rapidly varying illumination. The machine learning-based tunable band-stop photodetector spans the visible-to-infrared range and achieves an extinction ratio of ~43 dB between target and interference signals. In simulated autonomous driving scenarios with severe light contamination, the system improves multitarget recognition accuracy from about 60% for conventional broadband photodetectors to more than 92%, enabling robust operation in uncontrolled and light-polluted environments.</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

Tunable band-stop photodetection with machine learning-enabled broadband spectral adaptation

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • He Yu,
  • Yunlu Lian,
  • Shengwang Jia,
  • Wenzhuang Ma,
  • Jintao Fu,
  • Xingsi Liu,
  • Xinyun Zhu,
  • Li Zhang,
  • Tiancheng Han,
  • Changyong Lan,
  • Kewei You,
  • Guangliang Qin,
  • Moufu Kong,
  • Hanyu Zheng,
  • Yadong Jiang,
  • Jun Wang,
  • Cheng-Wei Qiu

摘要

Unpredictable light pollution such as glare, laser beams and light-emitting diode sources poses a major challenge to machine vision systems in autonomous driving and humanoid robotics, where reliability is essential for safety and efficiency. Here we introduce a machine learning-based tunable band-stop photodetector that combines a bio-inspired visual perception strategy with a band-stop centre wavelength dynamically defined by a bias-controlled operating point (Vt). A deep learning model within an incremental learning framework maps incident spectral features to optimal voltage settings, enabling continuous self-calibration under rapidly varying illumination. The machine learning-based tunable band-stop photodetector spans the visible-to-infrared range and achieves an extinction ratio of ~43 dB between target and interference signals. In simulated autonomous driving scenarios with severe light contamination, the system improves multitarget recognition accuracy from about 60% for conventional broadband photodetectors to more than 92%, enabling robust operation in uncontrolled and light-polluted environments.