解耦与引导协同的输电线路螺栓缺陷检测

PDF下载HTML

Title:

Synergistic decoupling and guidance for transmission line bolt defect detection

作者:

赵文清^1,2, 张乐¹

1. 华北电力大学 控制与计算机工程学院, 河北 保定 071003;
2. 华北电力大学 河北省能源电力知识计算重点实验室, 河北 保定 071003

Author(s):

ZHAO Wenqing^1,2, ZHANG Le¹

1. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China;
2. Hebei Key Laboratory of Knowledge Computing for Energy & Power, North China Electric Power University, Baoding 071003, China

关键词:

螺栓; 缺陷识别; 输电线路巡检; 小目标检测; 特征融合; 解耦金字塔池化; 深度学习; 一阶段检测器

Keywords:

bolt; defect detection; power line inspection; small object detection; feature fusion; decoupled pyramid pooling; deep learning; one-stage detector

分类号:

TP391

DOI:

10.11992/tis.202509002

摘要:

针对输电线路巡检图像中螺栓等紧固件目标微小、细节易丢失、上下文信息利用不充分以及多尺度特征融合存在语义冲突的问题，本文提出一种融合解耦与引导思想的螺栓缺陷检测方法。该方法主要由以下3部分构成：1)高频细节引导的特征增强模块利用高频分量驱动门控机制增强微弱目标的细节表达，2)解耦上下文与显著性金字塔池化模块通过并行分支分别捕获峰值响应与区域统计特征以实现信息互补，3)非对称引导融合金字塔网络以非对称路径分流和细节引导融合策略缓解多尺度特征的语义冲突。在自建螺栓缺陷数据集上的实验结果表明，所提方法性能优越，在mAP@50和mAP@.5:.95指标上达到93.9%和63.4%，较基线模型YOLOv11s分别提升了3.1和2.9百分点，并在公开数据集RSOD与DIOR上表现出良好的泛化能力。

Abstract:

This paper proposes a novel detection method integrating decoupling and guidance strategies to address the challenges of detecting small fasteners in UAV imagery, including detail loss, insufficient contextual information, and conflicts in multi-scale feature fusion. The method mainly consists of the following three modules: a high-frequency detail-guided feature enhancement (HFDG) module that leverages high-frequency components to drive a gating mechanism, thereby enhancing the representation of faint targets; a decoupled context and salient pyramid pooling (DCSPP) module that captures peak responses and regional statistics in parallel to achieve complementary information; and an asymmetric guidance fusion pyramid network (AGF-PN) that mitigates semantic conflicts via asymmetric path shunting and a detail-guided fusion strategy. Experiments on a self-built bolt defect dataset demonstrate superior performance, achieving 93.9% mAP@50 and 63.4% mAP@.5:.95, surpassing the YOLOv11s baseline by 3.1 and 2.9 percentage points, respectively. The proposed method also demonstrates strong generalization on the public RSOD and DIOR datasets.