面向绝缘子缺陷检测的多模块协同轻量化YOLO网络

A Multi-module Coordinated Lightweight YOLO Network for Insulator Defect Detection

摘要:

针对无人机电力巡检中绝缘子缺陷检测面临的小目标占比高、细微裂纹特征易丢失及复杂背景干扰等特定挑战,提出一种基于改进YOLOv11n的轻量化检测模型YOLO-SCD；通过多维度结构优化提升对绝缘子缺陷的检测精度：设计基于星形运算的C3Star模块,利用隐式高维映射增强主干网络对缺陷细微特征的捕获能力；构建融合上下文感知注意力的CAA-HSFPN颈部网络,以强化多尺度特征的语义融合并抑制背景干扰；采用解耦增强型DE-LSCD检测头,结合可变形卷积与EMA注意力机制提升对几何形变目标的检测适应性；引入Inner-WIoU损失函数优化边界框回归过程；在自制绝缘子缺陷数据集上的实验结果表明,YOLO-SCD的准确率P达到89.79%,mAP@0.5达到87.06%,较基准模型YOLOv11n分别提升2.72%和2.47%,而参数量仅增加0.37×10?,实现了检测精度与模型轻量化之间的有效平衡,可满足实际巡检工程中对高效率、高准确度绝缘子缺陷检测的需求。

Abstract:

Aiming at the specific challenges in UAV power inspection, such as the high proportion of small targets, the easy loss of fine crack features, and complex background interference in insulator defect detection, a lightweight detection model named YOLO-SCD is proposed based on an improved YOLOv11n architecture. The method enhances detection accuracy through multi-dimensional structural optimizations: a C3Star module based on star operation is designed to strengthen the backbone network’s ability to capture subtle defect features via implicit high-dimensional mapping; a context-aware attention fused CAA-HSFPN neck network is constructed to reinforce multi-scale feature fusion and suppress background interference; a decoupled enhanced DE-LSCD detection head is adopted, combining deformable convolution and an EMA attention mechanism to improve adaptability to geometrically distorted targets; and the Inner-WIoU loss function is introduced to optimize the bounding-box regression process. Experimental results on a self-built insulator defect dataset show that YOLO-SCD achieves a precision P of 89.79% and an mAP@0.5 of 87.06%, representing improvements of 2.72% and 2.47%, respectively, over the baseline YOLOv11n, while the parameter increase is only 0.37×10?, effectively balancing detection accuracy and model lightweightness, thereby meeting the practical requirements for high-efficiency and high-accuracy insulator defect detection in inspection engineering.