东北大学学报(自然科学版) ›› 2024, Vol. 45 ›› Issue (6): 769-775.DOI: 10.12068/j.issn.1005-3026.2024.06.002

• 信息与控制 • 上一篇    

基于改进YOLOv4轻量化网络的机械手状态检测算法

郭立新1(), 毕素涛1,2, 赵明扬2   

  1. 1.东北大学 机械工程与自动化学院,辽宁 沈阳 110819
    2.季华实验室,广东 佛山 528200
  • 收稿日期:2023-02-02 出版日期:2024-06-15 发布日期:2024-09-18
  • 通讯作者: 郭立新
  • 作者简介:郭立新(1968-),男,辽宁沈阳人,东北大学教授,博士生导师.
  • 基金资助:
    国家自然科学基金资助项目(52275283)

State Detection Algorithm of Manipulator Based on Improved YOLOv4 Lightweight Network

Li-xin GUO1(), Su-tao BI1,2, Ming-yang ZHAO2   

  1. 1.School of Mechanical Engineering & Automation,Northeastern University,Shenyang 110819,China
    2.Jihua Laboratory,Foshan 528200,China.
  • Received:2023-02-02 Online:2024-06-15 Published:2024-09-18
  • Contact: Li-xin GUO
  • About author:GUO Li-xin, E-mail: lxguo@mail.neu.edu.cn

摘要:

YOLOv4网络结构复杂、参数较多、模型较大,因此极大地限制了其在工业上的应用.针对这一问题,提出一种改进YOLOv4的轻量化网络.首先,采用GhostNet代替YOLOv4主干网络,简化网络结构,降低模型参数量;其次,为了弥补网络简化后带来的精度损失,在其余两个输出特征层后加入Spatial Pyramid Pooling结构,加强特征提取;再次,加入Squeeze and Excitation Network通道注意力机制,增强网络重要信息提取能力;最后,将损失函数CIOU替换为SIOU,加快模型收敛,进而产生更好的模型.实验结果表明,在满足工业要求的前提下,改进后的轻量化网络相比于YOLOv4网络,在牺牲较小检测精度的情况下,模型参数量和计算量大幅下降,同时检测速度得到了提升,从而证明了改进算法在光纤插拔任务中机械手夹持状态识别检测的有效性.

关键词: YOLOv4, GhostNet, 深度可分离卷积, 注意力机制, 损失函数

Abstract:

The YOLOv4 network is difficult to be widely used in industry due to its complex structure, many parameters, and large model size. In view of this problem, an improved lightweight network based on YOLOv4 is proposed.Firstly, GhostNet is used to replace the YOLOv4 backbone network, simplifying the network structure and reducing the number of model parameters; Secondly, in order to make up for the accuracy loss caused by network simplification, Spatial Pyramid Pooling structure is added after the other two output feature layers to enhance feature extraction. Thirdly, the attention mechanism of channel, which is Squeeze and Excitation Network, is added to improve the network’s ability to extract important information. Finally, the loss function CIOU is replaced by SIOU to accelerate the convergence of the model and thus produce a better model. Experimental results show that, on the premise of meeting industrial requirements, compared with YOLOv4 network, the improved lightweight network significantly reduces the number of model parameters and the amount of computation, while improving the detection speed, at the same time, at the expense of less detection accuracy, thus proving the effectiveness of the improved algorithm in the identification and detection of the clamping state of the manipulator in the optical fiber plugging task.

Key words: YOLOv4, GhostNet, deepthwise separable convolution, attention mechanism, loss function

中图分类号: