Insulator Defect Detection Based on Feature Fusion and Attention Mechanism

Yue Zhang, Baoguo Wei, Lina Zhao, Jinwei Liu, Zhilang Hao, Lixin Li, Xu Li

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

The performance of insulator defect detection model is not satisfactory due to the small object size, imbalanced and insufficient data. In this paper, based on YOLOv5 model, we propose an insulator defect detection method incorporating feature fusion and attention mechanism. Firstly, multi-scale feature fusion is introduced to strengthen the ability to extract minute features from images. Secondly, an attention mechanism based on SE-C module is proposed to improve the detection of defective objects. In addition, K-means++ is used to customize anchor boxes to meet the actual requirements and avoid mismatches. The experimental results show that the proposed model achieves 92.4% precision on the public insulator dataset, which demonstrates the applicability of the auto-detection system for insulator defects significantly.

Original languageEnglish
Title of host publication2022 IEEE International Conference on Signal Processing, Communications and Computing, ICSPCC 2022
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781665469722
DOIs
StatePublished - 2022
Event2022 IEEE International Conference on Signal Processing, Communications and Computing, ICSPCC 2022 - Xi'an, China
Duration: 25 Oct 202227 Oct 2022

Publication series

Name2022 IEEE International Conference on Signal Processing, Communications and Computing, ICSPCC 2022

Conference

Conference2022 IEEE International Conference on Signal Processing, Communications and Computing, ICSPCC 2022
Country/TerritoryChina
CityXi'an
Period25/10/2227/10/22

Keywords

  • attention mechanism
  • defect detection
  • insulator
  • object detection

Fingerprint

Dive into the research topics of 'Insulator Defect Detection Based on Feature Fusion and Attention Mechanism'. Together they form a unique fingerprint.

Cite this