融合YOLOX和ASFF的高原山地灾害检测模型
作者:
作者单位:

作者简介:

何明杰(1973—),男,教授级高工,学士。主要从事水工结构设计和边坡灾害治理研究。E-mail:he_mj@hdec.com

通讯作者:

中图分类号:

TU443

基金项目:


A Plateau Mountain Disaster Detection Model by Integrating YOLOX and ASFF
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    YOLOX 是全球首个一阶无锚框目标检测模型,超越了 YOLO‐V(3‐5)和 SSD 等传统锚框模型,极大提高了山地灾害检测识别精度。然而,该模型存在不同尺度之间特征不一致的问题,融合后的特征图质量有待提升。以西藏高原山地灾害重灾区为试验区,在建立了西藏高原山地灾害数据集的基础上,通过融合 Adaptively Spatial Fea‐ ture Fusion(ASFF)注意力机制和骨干网络尺寸调整机制,设计了一个优化的、不同适用性的高原山地灾害检测模型(Plateau mountain disaster detection model,PMDDM)。为验证 PMDDM 模型的优越性,将其与传统 YOLOX 模型、不同注意力机制、不同目标检测模型进行了对比分析,并且对不同尺度模型的检测性能和可视化结果也进行了对比分析研究。结果表明:ASFF 注意力机制可以有效的解决传统 YOLOX 模型中存在的不同尺度特征间的特征不一致问题,且对模型检测性能提升明显优于 SE、CBAM、ECA、GAM 和 Coord 等注意力机制;PMDDM 模型对山地灾害的检测精度优于 Faster‐RCNN、SSD 和 YOLO‐V3 模型,可以满足不同工作场景对硬件配置、检测速度和精度的需求,且模型尺度越大,识别目标的准确率越高。

    Abstract:

    YOLOX is the world's earliest first-order anchor-free frame target detection model, which surpasses traditional anchor frame models such as YOLO-V (3-5) and SSD, and greatly improves the accuracy of mountain disaster detection and recognition. However, the model suffers from feature inconsistency between features at different scales, and the quality of the fused feature map needs to be improved. This paper presents the mountain disaster-stricken areas on the Tibetan plateau as the experimental site, based on the basis of the established Tibetan plateau mountain disaster dataset, introduces the Adaptively Spatial Feature Fusion (ASFF) attention mechanism and backbone network size adjustment mechanism to design an optimized system with different applicability. The new plateau mountain disaster detection model (PMDDM) improves the detection accuracy of plateau mountain hazards and enhances the convenience of plateau mountain hazard detection. Meanwhile, the PMDDM model was also compared with traditional YOLOX and other models with different attention mechanisms and different object detection to verify it superiority. The detection performance and visualization results of different scale models are also comparatively analyzed and studied in this paper. The results show that the ASFF attention mechanism can effectively solve the problem of feature inconsistency among different scale features in the traditional YOLOX model; the ASFF attention mechanism improves the model detection performance much better than the attention mechanisms of SE, CBAM, ECA, GAM and Coord; the detection accuracy of the PMDDM model for mountain hazards is better than that of Faster-RCNN, SSD and YOLO-V3 models; the PMDDM model can meet requirements of hardware configuration, detection speed and accuracy for different working scenarios; the larger the scale of PMDDM model, the higher the accuracy rate of identifying targets.

    参考文献
    相似文献
    引证文献
引用本文

何明杰,刘德方,张猛,李高会.融合YOLOX和ASFF的高原山地灾害检测模型[J].防灾减灾工程学报,2023,43(6):1215-1223

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2023-01-05
  • 最后修改日期:2023-03-13
  • 录用日期:
  • 在线发布日期:2024-01-11
  • 出版日期: