Defects Localization and Classification Method of Power Transmission Line Insulators Aerial Images Based on YOLOv5 EfficientNet and SVM
Accurate localization and classification of defects in insulator images captured by UAVs (Unmanned Aerial Vehicles) are critical for the maintenance of high-voltage power transmission grids. However, existing methods face challenges in achieving satisfactory accuracy, particularly in complex environ...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10962140/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Accurate localization and classification of defects in insulator images captured by UAVs (Unmanned Aerial Vehicles) are critical for the maintenance of high-voltage power transmission grids. However, existing methods face challenges in achieving satisfactory accuracy, particularly in complex environments. Moreover, previous approaches primarily focus on detecting a single type of insulator defect without further categorizing them. To address these limitations, we propose a novel deep-learning framework that integrates YOLOv5, EfficientNet and Support Vector Machines (SVM) to improve both the localization and classification of insulator defects in aerial images. The framework specifically targets overcoming the challenges associated with low signal-to-noise ratios in defect detection. The proposed approach divides the task into two primary modules: 1) YOLOv5-based object detection for accurate defect localization, and 2) defect classification using EfficientNet and SVM. Specifically, YOLOv5 enhances feature extraction to improve the signal-to-noise ratio, while EfficientNet and SVM enable precise defect classification. The effectiveness of the proposed method is validated through experiments conducted on a range of insulator image datasets collected from the Guangdong power grid. Comparative analysis of experimental results shows that the proposed method outperforms existing approaches in terms of both accuracy and practical applicability, achieving a top-1 accuracy of 98.22%, an F1-score of 98.53%, and an accuracy improvement of 3.89% over conventional backbone-based networks. |
|---|---|
| ISSN: | 2169-3536 |