DEF-YOLO: A Dual-Branch Framework for UAV-Based Insulator Defect Detection With Feature Enhancement and Adaptive Feature Fusion

Abstract

Accurate and real-time detection of insulator defects in UAV inspection images is crucial for ensuring the safe and reliable operation of power transmission systems. However, existing single-backbone detection networks are often susceptible to issues such as missed detections, false positives, and poor localization in UAV scenarios. To address these challenges, this study proposes an improved insulator defect detection network, namely DEF-YOLO, based on YOLOv8. By constructing a dual-branch backbone composed of a structural branch and a texture branch, the proposed network separately extracts the overall structural features of the insulator body and the local texture features of defect regions, thereby reducing the interference between different types of information in shared representations. To enhance the representation capability for weak-texture defects, a texture feature enhancement module (TFEM) is introduced into the texture branch to strengthen edge features and local detail information. Meanwhile, an adaptive feature fusion module (AFFM) is designed to achieve adaptive fusion of structural and textural information, thereby suppressing complex background interference and improving detection stability. In addition, a Scale-Aware Regression Loss (SAR Loss) is proposed to enhance the model’s localization capability for small-scale defects. Considering the requirements of UAV inspection scenarios, channel pruning and fine-tuning strategies are further adopted to compress the model in a lightweight manner. Experimental results on a self-built UAV insulator defect dataset show that the final lightweight DEF-YOLO achieves mAP0.5 and mAP0.5:0.95 values of 91.4% and 72.5%, respectively, with AP0.5 values of 92.0%, 90.6%, and 91.6% for Flashover, Breakage, and Insulator, respectively. With only 9.5 M parameters and an inference speed of 92 FPS, DEF-YOLO achieves a favorable balance between detection accuracy and real-time performance.