改进金字塔和跳跃连接的YOLOv5目标检测网络
DOI:
作者:
作者单位:

重庆理工大学两江人工智能学院

作者简介:

通讯作者:

中图分类号:

TP391

基金项目:

国家重点研发计划“智能机器人”重点专项项目(2018YFB1308602); 国家自然科学基金面上项目(61173184); 重庆市自然科学基金(cstc2018jcyjA2328, cstc2018jcyjAX0694)


YOLOv5 Object Detection Network based on Improved Pyramid and Skip Connection
Author:
Affiliation:

School of Artificial Intelligence,Liangjiang,Chongqing,China.

Fund Project:

National Key R&D Plan "Intelligent Robots" Key Project(2018YFB1308602); the National Natural Science Foundation Project(61173184); the Chongqing Natural Science Foundation Project(cstc2018jcyjA2328, cstc2018jcyjAX0694)

  • 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:

    针对YOLOv5在进行目标检测时因金字塔结构未有效利用跨尺度信息以及梯度消失而导致精确率不高的问题,本文通过引入FPT(Feature Pyramid Transformer)结构改进原YOLOv5网络模型中的FPN(Feature Pyramid Networks)结构和PAN(Path Aggregation Network)结构,利用注意力机制有效提取网络跨尺度特征,提升目标检测精确度和鲁棒性;针对网络模型加深后的梯度消失问题,在FPT结构两端加入跳跃连接结构,提升网络目标检测能力同时传递显著性特征;引入Mish激活函数,提升了目标检测精确度,结合以上结构,从而提出了改进金字塔和跳跃连接的YOLOv5目标检测网络模型。在PASCAL VOC数据集和MSCOCO数据集上的实验结果表明,本文网络的目标检测精确度相较于YOLOv5有所提升。

    Abstract:

    Aiming at the problem that the accuracy rate is not high due to the ineffective use of cross-scale information in the pyramid structure and the disappearance of gradient in the object detection of YOLOv5. This paper introduces the FPT (Feature Pyramid Transformer) structure to improve the FPN (Feature Pyramid Networks) structure and PAN(Path Aggregation Network) structure in the original YOLOv5 network model. The attention mechanism is used to effectively extract the cross-scale features of the network to improve the accuracy of object detection. Aiming at the gradient disappearance problem after the deepening of the network model, a skip connection structure is added at both ends of the FPT structure to improve the network object detection ability and transfer the salient features at the same time. The Mish activation function is introduced to improve the accuracy of object detection. Combined with the above structure, the fs-yolov5 network model is proposed. Experimental results on Pascal VOC datasets and MSCOCO datasets show that the detection accuracy of this algorithm is improved compared with YOLOv5.

    参考文献
    相似文献
    引证文献
引用本文
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:2021-08-11
  • 最后修改日期:2022-02-12
  • 录用日期:2022-02-25
  • 在线发布日期:
  • 出版日期: