Scholarlink

The paper addresses the task of constructing models of the underlying surface using sensors of various spectral ranges by employing edge detection methods. To further improve the obtained models, a neural network structure in the form of an acyclic graph, consisting of 58 layers and including convolutional, deconvolutional, fully connected, and segmentation layers, is proposed. The trained model improves the accuracy of the detected boundaries by up to 26%. An approach to integrating data from sensors of various spectral ranges and training the model in real-time is also proposed. The developed model can be used for determining or correcting the determination of the location onboard an unmanned aerial vehicle.

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

Using Segmentation Neural Networks for Correcting Underlying Surface Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kirill V. Parfentiev,
  • Maria B. Pimenova,
  • Tatiana L. Anisova

摘要

The paper addresses the task of constructing models of the underlying surface using sensors of various spectral ranges by employing edge detection methods. To further improve the obtained models, a neural network structure in the form of an acyclic graph, consisting of 58 layers and including convolutional, deconvolutional, fully connected, and segmentation layers, is proposed. The trained model improves the accuracy of the detected boundaries by up to 26%. An approach to integrating data from sensors of various spectral ranges and training the model in real-time is also proposed. The developed model can be used for determining or correcting the determination of the location onboard an unmanned aerial vehicle.