Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (8): 289-296.DOI: 10.3778/j.issn.1002-8331.2010-0192

• Engineering and Applications • Previous Articles     Next Articles

Defect Detection in Transmission Line Based on Scale-Invariant Feature Pyramid Networks

ZHAO Jielun, ZHANG Xingzhong, DONG Hongyue   

  1. College of Software, Taiyuan University of Technology, Taiyuan 030024, China
  • Online:2022-04-15 Published:2022-04-15

基于尺度不变特征金字塔的输电线路缺陷检测

赵杰伦,张兴忠,董红月   

  1. 太原理工大学 软件学院,太原 030024

Abstract: As the existing object detection algorithms have the problem of low detection accuracy of power components and inspection defects in complex inspection scenarios of high-voltage power, defect detection in transmission line based on scale-invariant feature pyramid networks is proposed. Firstly, by applying current baseline detection methods to the problem in this article, the RepPoints v2 networks has the highest accuracy. Secondly, as FPN structure cannot effectively extract cross-level semantic information and ignoring the scale normalization in the corner point verification process of RepPoints v2, the scale-invariant feature pyramid networks(SI-FPN) structure is proposed by combining the efficient channel attention(ECA) mechanism and the scale-equalizing pyramid convolution(SEPC). In SI-FPN, the ECA attention module enhances the features of FPN at the channel level, and then SEPC extracts scale-invariant features from FPN and fuses cross-level pyramid features. Through training and testing on the self-built data set of six objects including insulator, shock hammer, suspension clamp, insulator self explosion, shock hammer falling off and bird’s nest, the proposed method improves 1.9 percentage points on the baseline of RepPoints v2, and the mAP reaches 96.3%. The detection accuracy is far beyond the current baseline detection models. Moreover, the SI-FPN module designed in this paper can be used as an independent structure to improve other detection models, which has certain universality.

Key words: scale-invariant feature pyramid networks(SI-FPN), RepPoints v2, object detection, scale-equalizing pyramid convolution(SEPC), power inspection, defects detection

摘要: 针对现有目标检测算法在高压电力复杂巡检场景下电力部件与巡检缺陷检测精度较低的问题,提出一种基于尺度不变特征金字塔的输电线路缺陷检测方法。将主流目标检测方法用于该场景,对比得出RepPoints v2网络模型的检测精度最高。针对RepPoints v2中FPN结构不能有效提取跨层次间语义信息及角点验证过程中忽略尺度归一化的问题,结合高效通道注意力模块(ECA)与尺度均衡金字塔卷积(SEPC)提出了一种尺度不变特征金字塔结构SI-FPN(scale-invariant feature pyramid networks)。其中ECA注意力模块对FPN的特征进行通道级别的增强,之后SEPC从FPN中提取尺度不变特征并对跨层次的金字塔特征进行融合。通过在自建的包含绝缘子、防震锤、悬垂线夹、绝缘子自爆、防震锤脱落与鸟巢六类对象的数据集上进行训练测试表明,该方法在RepPoints v2的基准上提升1.9个百分点,mAP达到96.3%,检测精度远超当前基准检测模型,且所设计的SI-FPN模块可作为一种独立的结构改善其他检测模型,具有一定的通用性。

关键词: 尺度不变特征金字塔结构(SI-FPN), RepPoints v2, 目标检测, 尺度均衡金字塔卷积(SEPC), 电力巡检, 缺陷检测