基于改进YOLOv8的架空线路危物辨识方法

A Detection Method for Dangerous Objects on Overhead Power Lines Based on Improved YOLOv8

中文关键词:架空线路  异物入侵  YOLOv8  注意力机制  损失函数

keywords:overhead lines  foreign object invasion  YOLOv8  attention mechanism  loss function

中文摘要:架空线路所处环境复杂，外力破坏、异物入侵均会不同程度损坏线路。针对无人机巡检存在的漏检、误检问题，提出一种基于改进YOLOv8（you only look once version 8，YOLOv8）的架空线路危物辨识方法。设计C2f-ODConv（omni-dimensional dynamic convolution，ODConv）结构来捕获更丰富的上下文信息，有效增强了模型的特征提取能力；提出SPPF（spatial pyramid pooling fast，SPPF）-MHSA（multi-head self-attention，MHSA）模块，提高了模型在复杂背景下遮挡小目标的检测率；在颈部网络添加SimAM（similarity-based attention mechanism，SimAM）无参数注意力机制，增强了模型对重要信息的关注程度；引入基于辅助边框的Inner-IoU（intersection over union，Inner-IoU）损失函数，使模型的均值平均精度达到94%，提高了3.8%。通过消融实验和对比实验，验证了所提方法的有效性、优越性。

Abstract:Overhead transmission lines operate in complex environments， where external damage and foreign object intrusion can cause varying degrees of damage to the lines. To address the issues of missed detection and false detection in unmanned aerial vehicle （UAV） inspection， this paper proposes an overhead line hazard identification method based on the improved YOLOv8 （You Only Look Once Version 8， YOLOv8）. A C2f-ODConv （Omni-Dimensional Dynamic Convolution， ODConv） structure is designed to capture richer contextual information， effectively enhancing the feature extraction capability of the model； an SPPF （Spatial Pyramid Pooling Fast， SPPF）-MHSA （Multi-Head Self-Attention， MHSA） module is proposed to improve the detection rate of the model for small occluded targets in complex backgrounds； a parameter-free SimAM （Similarity-based Attention Mechanism， SimAM） attention mechanism is added to the neck network to strengthen the model's focus on important information； an Inner-IoU （Intersection over Union， Inner-IoU） loss function based on auxiliary bounding boxes is introduced， enabling the model to achieve a mean average precision （mAP） of 94%， representing an increase of 3.8%. The effectiveness and superiority of the proposed method are verified through ablation experiments and comparative experiments.