Insulator Defect Detection via Attention-Driven Multi-Scale Feature Fusion

Abstract

Insulator defect detection is a critical task for ensuring the safe operation of power transmission systems. With the increasing adoption of unmanned aerial vehicles and intelligent inspection robots, defect detection models are required to achieve high accuracy under complex outdoor environments while maintaining low computational complexity for deployment on mobile devices. However, existing object detection models often suffer from insufficient feature extraction for small-scale defects, limited robustness to complex backgrounds, and suboptimal regression performance. To address these challenges, this paper proposes an improved YOLOv8-based insulator defect detection model that integrates an attention mechanism, multi-scale feature fusion, and an optimized bounding box regression strategy. First, a Convolutional Block Attention Module (CBAM) is embedded into the backbone network to enhance spatial and channel feature representation, improving the detection of small, occluded, and low-contrast defects. Second, the original PANet structure in YOLOv8 is replaced with a Bidirectional Feature Pyramid Network (BiFPN), enabling more effective multi-scale feature fusion while reducing redundant computation. Finally, the SIoU loss function is introduced to account for angular differences between predicted and ground truth bounding boxes, thereby improving regression convergence and localization accuracy. Experimental results on a public insulator defect dataset demonstrate that the proposed method achieves superior detection performance compared with Faster R-CNN, RetinaNet, YOLOv5, YOLOv6, and the original YOLOv8 model. The improved model attains an [email protected] of 77.8%, outperforming the baseline YOLOv8 by 2.3%, while exhibiting stronger robustness in complex inspection scenarios. These results indicate that the proposed approach provides an effective and practical solution for insulator defect detection in intelligent power line inspection systems.