关键词: 输电线路缺陷检测, 分层可分离视觉Transformer, 双向递归特征金字塔网络, 双重注意力模块, 可变形卷积

Abstract: As the core transmission and distribution carrier of the power system, high-voltage transmission lines were closely related to the safety of the power grid. To address the problems of low efficiency and high omission rate in traditional manual inspection, a lightweight defect detection method based on a two-stage multimodal attention mechanism and dynamic feature decoupling was proposed. In the first stage, accurate localization of key components was achieved based on an improved lightweight detection network, Light-YOLO. In the second stage, precise classification and identification of defects were performed using the difference-enhanced feature network, Dual-DifferNet. In the design of Light-YOLO, a hybrid structure of hierarchical Separable Visual Transformer (SepViT) and deep Deformable Convolution Network (DCN) was introduced. By alternately stacking local perception convolution layers and global attention Transformer blocks, the modeling capability of long-range dependencies was enhanced while reducing computational cost, which effectively improved the detection accuracy of small targets such as insulators and conductor joints. For the defect classification task, Dual-DifferNet adopted a dual-branch structure, in which a Spatial-Channel Dual Attention (SCDA) module was embedded in each branch. Cross-modal feature interaction was promoted by a cross-attention mechanism, thereby improving the robustness and generalization capability of defect identification. Experimental results showed that the proposed method achieved a mean Average Precision (mAP@50) of 96.9%, which was 16.1 percentage points higher than the baseline model (YOLOv8), while the computational cost (GFLOPs) was reduced by 56.73%, fully verifying the method’s high detection accuracy, excellent computational efficiency, and deployment potential.

Key words: transmission line defect detection, hierarchical separable vision transformer, bidirectional recursive feature pyramid network, dual attention module, deformable convolution