FCD-YOLO:一种高效检测绝缘子缺陷的模型  

FCD-YOLO:an efficient algorithm for detecting insulator defects

作　　者：邓豪 林国军 秦海洋 蔡大洋 DENG Hao;LIN Guojun;QIN Haiyang;CAI Dayang(Sichuan University of Science and Engineering,Yibin 644000 China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,四川宜宾644000

出　　处：《新余学院学报》2025年第2期48-55,共8页Journal of Xinyu University

基　　金：四川省科技计划项目“基于深度字典学习的人脸识别算法研究”(2024YFZ0026)。

摘　　要：在充分考虑检测速度与精度的双重需求上,提出一种基于YOLOV8n的改进模型——FCD-YOLO。在YOLOV8n的网络模型中,用FasterNet网络结构替代主干特征提取网络,以加强网络模型的空间特征提取能力;在金字塔池化层添加了通道先验卷积注意力机制,可以更有效地提取空间信息,降低检测中的漏检率;采用轻量化的动态上采样模块替换了颈部网络中的上采样模块,加强网络模型聚合上下文信息能力。实验结果表明:FCD-YOLO算法的平均精度均值达到90.7%,较原始的网络模型提升2.2%;FPS为238帧/秒。该算法具有较强的实时性、准确性和鲁棒性,可快速、精确识别绝缘子缺陷。In order to fully consider the dual requirements of detection speed and accuracy,an improved model based on YOLOV8n-FCD-YOLO is proposed.the FasterNet structure is used to replace the backbone feature extraction network to enhance the spatial feature extraction capability of the network model;A channel prior convolutional attention mechanism was added in the pyramid pooling layer for more effective spatial information extraction to reduce miss detection,and a lightweight dynamic up-sampling module in the neck network was adopted to enhance contextual information aggregation.Experimental results showcase FCD-YOLO's superiority,achieving a mean average precision(mAP)of 90.7%,a 2.2%improvement over the baseline,with a frames per second(FPS)of 238,ensuring real-time performance,high accuracy,and robust defect detection for insulators.

关 键 词：绝缘子缺陷 FCD-YOLO FasterNet 通道先验卷积注意力 动态上采样结构 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]