摘要
针对绝缘子表面缺陷检测中存在的小尺度目标特征微弱、复杂背景干扰等技术难题,提出一种基于YOLOv8s架构的轻量化与细粒度特征增强相结合的检测框架YOLOv8-CEDA。通过设计GCA模块,融合Ghost模块的轻量化优势与Coordinate Attention的空间信息建模能力,在显著降低模型参数和计算复杂度的同时,有效增强对微小缺陷的空间特征表征能力;在检测头中部署轻量级ECA通道注意力模块,实现关键特征通道的自适应权重分配,从而提升微弱缺陷信号的响应强度;此外,采用DySample动态上采样模块替代传统上采样方法,使多尺度特征融合过程具备内容感知能力,进一步保持细节特征并增强边缘信息。实验结果表明,所提方法在保持模型轻量化特性的基础上,检测精度较基线模型显著提升,尤其在小目标缺陷检测任务中表现优异,展现出良好的工程应用价值。
Detecting small-scale insulator surface defects remains challenging due to weak feature representation and complex background interference.To address this issue,this study proposes YOLOv8-CEDA,an enhanced lightweight object detection framework based on YOLOv8s,integrating fine-grained feature enhancement.First,a Ghost and Coordinate Attention(GCA)module is designed,combining the parameter efficiency of Ghost convolution with the spatial modeling capability of Coordinate Attention to improve small defect localization while reducing computational overhead.Second,an efficient channel attention(ECA)mechanism is incorporated into the detection head to adaptively amplify discriminative features,strengthening the response to subtle defect patterns.Additionally,DySample,a dynamic upsampling operator,replaces conventional interpolation to preserve structural details and enhance edge recovery during multi-scale feature fusion.Extensive experiments demonstrate that YOLOv8-CEDA achieves superior detection accuracy over baseline models while maintaining computational efficiency,particularly in small defect detection.The proposed method exhibits strong potential for real-world industrial applications.
作者
潘学华
梁炜皓
李金瑾
张震
龚宇平
尹显贵
刘益畅
PAN Xuehua;LIANG Weihao;LI Jinjin;ZHANG Zhen;GONG Yuping;YIN Xiangui;LIU Yichang(Measurement Center,Guangxi Power Grid Co.,Ltd.,Nanning 530000,China;Inspur Communication Information System Co.,Ltd.,Jinan 250014,China)
出处
《机械与电子》
2025年第12期10-17,23,共9页
Machinery & Electronics
基金
广西电网公司科技项目资助(GXKJXM20230003)。
