系统工程与电子技术 ›› 2025, Vol. 47 ›› Issue (8): 2475-2485.doi: 10.12305/j.issn.1001-506X.2025.08.06

• 电子技术 • 上一篇    

PE-Net:一种优化剪枝的实时山体滑坡检测网络

于营1,2(), 王春平1, 徐金辉2, 吕述杭3, 付强1,*, 陈明2   

  1. 1. 中国人民解放军陆军工程大学石家庄校区,河北 石家庄 050003
    2. 三亚学院容淳铭院士工作站,海南 三亚 572022
    3. 清华大学集成电路学院,北京 100084
  • 收稿日期:2024-08-12 出版日期:2025-08-25 发布日期:2025-09-04
  • 通讯作者: 付强 E-mail:yingyu@sanyau.edu.cn
  • 作者简介:于 营(1990—),女,副教授,博士,主要研究方向为计算机视觉、语义分割、目标检测
    王春平(1965—),男,教授,博士,主要研究方向为智能火控、目标检测
    徐金辉(2003—),男,本科,主要研究方向为多智能体增强学习、视觉语言理解
    吕述杭(2002—),男,本科,主要研究方向为电子设计自动化插件开发、人工智能
    陈 明(1998—),男,硕士研究生,主要研究方向为目标检测
  • 基金资助:
    海南省院士创新平台专项(YSPTZX202145);海南省自然科学基金(622RC733)资助课题

PE-Net: a real-time landslide detection network with optimized pruning

Ying YU1,2(), Chunping WANG1, Jinhui XU2, Shuhang LYU3, Qiang FU1,*, Ming CHEN2   

  1. 1. Shijiazhuang Campus,Army Engineering University of PLA,Shijiazhuang 050003,China
    2. Academician Workstation of Chunming Rong,University of Sanya,Sanya 572022,China
    3. School of Integrated Circuits,Tsinghua University,Beijing 100084,China
  • Received:2024-08-12 Online:2025-08-25 Published:2025-09-04
  • Contact: Qiang FU E-mail:yingyu@sanyau.edu.cn

摘要:

山体滑坡实时检测对于减少人员伤亡和财产损失至关重要。为了解决传统图像识别方法在滑坡监测中的时间滞后和误判问题,构建了一个多域数据集,以增强对山体滑坡和沙尘暴视觉特征的理解,并提出一种剪枝和增强的山体滑坡自动检测模型。该模型基于改进的VanillaNet网络,结合动态多头注意力检测模块,显著提高了山体滑坡场景的视觉感知能力。此外,采用基于性能感知近似的全局通道剪枝(performance-aware approximation of global channel pruning,PAGCP)算法对该模型进行了压缩,以适应嵌入式部署。实验结果表明,所提出的模型在达到实时检测的前提下,显著提高了山体滑坡场景检测的准确性,对山体滑坡自然灾害监测与预警具有参考价值。

关键词: 山体滑坡, 目标检测, VanillaNet, 动态检测头, 全局通道剪枝

Abstract:

Real-time detection of slope landslides is crucial for reducing casualties and property damage. To address the issues of time lag and misjudgment in traditional object recognition methods for landslide monitoring, a multi-domain dataset is constructed to enhance the understanding of the visual features of slope landslides and sandstorms, and proposes a pruned slope and enhanced model named PE-Net for landslide automatic detection. This model is based on an improved VanillaNet network and incorporates a dynamic multi-head attention detection block, significantly enhancing the visual perception capability for landslide scenarios. Additionally, the model is compressed using the performance-aware approximation of global channel pruning (PAGCP) algorithm to facilitate embedded deployment. Experimental results demonstrate that the proposed model significantly improves the accuracy of slope landslide detection in real-time scenarios, providing valuable insights for natural disaster monitoring and warning of slope landslides.

Key words: landslide, object detection, VanillaNet, dynamic detection head, global channel pruning

中图分类号: