Large-Model-Enabled UAV Inspection for Transmission Lines: A Vision-Language Reasoning and Lightweight Detection Framework

Abstract

Unmanned Aerial Vehicle (UAV) inspection has become an important technical approach for improving the safety, efficiency, and traceability of transmission-line operation and maintenance. However, automatic interpretation of UAV images remains difficult because transmission-line defects are often small, partially occluded, visually similar to normal components, and affected by complex outdoor backgrounds. To address these problems, this paper proposes a large-model-enabled UAV inspection framework that integrates vision-language reasoning, lightweight You Only Look Once (YOLO)-based detection, and maintenance-oriented risk evaluation. The vision-language model is used as an auxiliary semantic module for weak annotation, hard-sample mining, semantic consistency checking, and inspection-report generation, while the lightweight detector performs real-time defect localization through multi-scale feature fusion, Convolutional Block Attention Module (CBAM) attention, focal loss, and semantic consistency regularization. A structured UAV inspection dataset containing typical transmission-line components and defects is used to evaluate the proposed method. Experimental results show that the proposed framework achieves 95.2% precision, 93.4% recall, 94.3% F1-score, and 96.1% mAP@0.5, outperforming mainstream detector-only baselines while maintaining practical inference efficiency. The results indicate that large-model assistance can improve small-defect recall, reduce ambiguous false alarms, and provide interpretable decision support for intelligent transmission-line maintenance.