计算机与现代化 ›› 2024, Vol. 0 ›› Issue (01): 47-52.doi: 10.3969/j.issn.1006-2475.2024.01.008

• 人工智能 • 上一篇    下一篇

改进RetinaNet的电力设备目标检测方法

  

  1. (云南大学信息学院,云南 昆明  650500)
  • 出版日期:2024-01-23 发布日期:2024-02-23
  • 作者简介:王秋忆(1997—),女,云南昆明人,硕士研究生,研究方向:图像处理,目标检测,E-mail: 459821599@qq.com; 通信作者:周浩(1972—),男,云南玉溪人,副教授,博士,研究方向:数字图像处理,计算机视觉,智能视频监控,E-mail: zhouhao@ynu.edu.cn; 郑婷婷(1994—),女,河南周口人,硕士研究生,研究方向:图像处理,目标检测,E-mail: 958052264@qq.com。
  • 基金资助:
    国家自然科学基金资助项目(12263008); 云南省重大科技专项(202202AD080004)

Improved RetinaNet Target Detection Method for Power Equipment

  1. (School of Information, Yunnan University, Kunming 650500, China)
  • Online:2024-01-23 Published:2024-02-23

摘要: 摘要:针对电力设备检测中小目标识别精度低的问题,提出一种基于RetinaNet的电力设备目标检测方法。通过K-means聚类方法优化原始网络的锚点框尺寸。然后在特征融合中加入拥有更高分辨率的浅层特征图,解决通过多层卷积后特征图包含信息过少的问题。在此基础上,引入ECA(Efficient Channel Attention)注意力机制,使网络定位电力设备的有效特征,抑制无用特征信息。实验结果表明,相比原始方法,本文方法对于电塔、销钉、工程车、绝缘子、电杆5种电力设备的平均识别精度提升了18.1个百分点,表明改进后的方法能显著提高电力设备的检测水平。

关键词: 关键词:目标检测, 电力设备, 聚类分析, 特征提取, 注意力机制

Abstract: Abstract: A RetinaNet-based target detection method for power equipment detection is proposed for the problem of low accuracy of small target recognition in power equipment detection. The anchor box size of original network is optimized by K-means clustering method firstly. Then shallow feature maps with higher resolution are added to feature fusion to solve the problem that the feature maps contain too little information after convolution through multiple layers. Based on this, ECA (Efficient Channel Attention) attention mechanism is introduced to enable the network to locate the effective features of power devices and suppress the useless feature information. The experimental results show that compared with the original method, the average recognition accuracy of the method in the paper is improved by 18.1 percentage points for five types of power equipment: electric towers, pins, construction vehicles, insulators and poles, which indicates that the improved method can significantly improve the detection level of power equipment.

Key words: Key words: object detection, power equipment, cluster analysis, feature extraction, attention mechanism

中图分类号: