Insulator Detection and Defect Identification Method Based on Dual Attention and Multi-Scale Feature Fusion

To address the challenges in power transmission and distribution inspection images captured by unmanned aerial vehicles (UAVs), specifically severe scale variations of insulators and their defects, complex background textures, and difficulty in modeling long-range dependencies caused by slender structures, this paper proposes YOLOv-DGI, an insulator object detection method based on dual-attention feature enhancement and feature fusion. Based on the YOLOv11 framework, this method introduces systematic improvements across three dimensions: feature extraction, feature fusion, and bounding box regression. In the feature extraction stage, a Dynamic Convolution with Large-kernel and Spatial attention (DCLSConv) module is designed. By leveraging a collaborative mechanism of multi-scale large-kernel weight generation and dynamic weighted convolution, DCLSConv adaptively models local details and global contextual dependencies, effectively mitigating the loss of detail for small targets during downsampling. In the feature fusion stage, a Grouped-shuffle Half-Convolutional Attention Module (GSHalfConv) is introduced. It utilizes a group-shuffling mechanism to break cross-channel information barriers and leverages the sparse processing characteristics of half-convolution to achieve semantic alignment and discriminative reorganization of cross-level features, thereby significantly enhancing recall sensitivity for occluded and small-scale insulators. In the bounding box regression stage, an improved ICIoU loss function is proposed. It enhances the gradient contribution of small targets via dynamic scale weighting and exponentially weighted center distances, while incorporating an aspect ratio smoothing term to optimize localization constraints for extreme-aspect-ratio targets. Experimental results on a power transmission and distribution dataset featuring complex scenarios demonstrate that, compared with the baseline YOLOv11, the proposed method achieves consistent and significant improvements in both detection accuracy and generalization performance. These results validate its effectiveness and robustness in insulator detection and defect identification tasks, while indicating that further optimization is required for strict onboard real-time deployment.

Authors

Institutions

Publication Details

Journal
Mathematics
Published
2026-11-17
DOI
https://doi.org/10.3390/math14224225
Primary Topic
Power Line Inspection Robots
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Insulator Detection and Defect Identification Method Based on Dual Attention and Multi-Scale Feature Fusion

Feng Shen, Chenghao Zhang, Meng Su, Shuai Zhou et al.
4 links
Mathematics
Power Line Inspection Robots
article

Insulator Detection and Defect Identification Method Based on Dual Attention and Multi-Scale Feature Fusion

Feng Shen, Chenghao Zhang, Meng Su, Shuai Zhou, Hong Yu, Shuailun Geng
article en

Abstract

To address the challenges in power transmission and distribution inspection images captured by unmanned aerial vehicles (UAVs), specifically severe scale variations of insulators and their defects, complex background textures, and difficulty in modeling long-range dependencies caused by slender structures, this paper proposes YOLOv-DGI, an insulator object detection method based on dual-attention feature enhancement and feature fusion. Based on the YOLOv11 framework, this method introduces systematic improvements across three dimensions: feature extraction, feature fusion, and bounding box regression. In the feature extraction stage, a Dynamic Convolution with Large-kernel and Spatial attention (DCLSConv) module is designed. By leveraging a collaborative mechanism of multi-scale large-kernel weight generation and dynamic weighted convolution, DCLSConv adaptively models local details and global contextual dependencies, effectively mitigating the loss of detail for small targets during downsampling. In the feature fusion stage, a Grouped-shuffle Half-Convolutional Attention Module (GSHalfConv) is introduced. It utilizes a group-shuffling mechanism to break cross-channel information barriers and leverages the sparse processing characteristics of half-convolution to achieve semantic alignment and discriminative reorganization of cross-level features, thereby significantly enhancing recall sensitivity for occluded and small-scale insulators. In the bounding box regression stage, an improved ICIoU loss function is proposed. It enhances the gradient contribution of small targets via dynamic scale weighting and exponentially weighted center distances, while incorporating an aspect ratio smoothing term to optimize localization constraints for extreme-aspect-ratio targets. Experimental results on a power transmission and distribution dataset featuring complex scenarios demonstrate that, compared with the baseline YOLOv11, the proposed method achieves consistent and significant improvements in both detection accuracy and generalization performance. These results validate its effectiveness and robustness in insulator detection and defect identification tasks, while indicating that further optimization is required for strict onboard real-time deployment.

MathematicsVol. 14(22)
Harbin Engineering University (CN), Harbin Institute of Technology (CN), Yunnan Power Grid Co., Ltd. (China) (CN)
Yunnan Power Grid Company
Industry, innovation and infrastructure
Power Line Inspection Robots
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon