摘要
针对基于深度学习的绝缘子缺陷检测方法存在泛化性能低、难以识别复杂背景下的绝缘子等问题,从特征提取和融合角度出发,提出一种结合特征重用与重建的YOLO(YOLO-RR)模型的绝缘子缺陷检测方法。首先,在特征提取阶段,以DenseNet为基础构建dense35网络作为主干网络,通过特征的重用增加对特征细节的感知能力,提升了模型在低饱和度和低对比度成像情况下的检测精度,并降低了网络参数量。其次,在特征融合阶段,提出基于沙漏模块的双向特征金字塔网络(H-BiFPN)结构进行不同尺度特征间的双向融合,通过特征重建和融合丰富了不同尺度的特征信息,解决了连续卷积下小目标信息丢失的问题,提升了对小目标的检测精度。最后,使用Wise-交并比(WIoU)损失函数优化模型,通过重点关注普通锚框使预测更加精准。在扩充后的中国输电线路绝缘子数据集(CPLID)上的实验结果表明,YOLO-RR模型识别率达到93.6%,网络参数量压缩至5.16×10^(6),优于对比模型,能够满足绝缘子缺陷定位的准确性和实时性要求,同时在背景干扰较大、受光照影响的成像上也有很好的检测效果。
To overcome challenges such as low generalization performance and difficulty in identifying insulators amidst complex backgrounds in deep learning-based insulator defect detection methods,this study introduces a novel method based on the You Only Look Once(YOLO)with feature Reuse and Reconstruction(YOLO-RR)model,focusing on feature extraction and fusion.Firstly,in the feature extraction stage,a dense35 network is constructed based on DenseNet as the backbone network.By reusing features,the model enhances its perception of feature details,thereby improving detection accuracy under low saturation and low contrast imaging while reducing the number of network parameters.Secondly,in the feature fusion stage,an Hourglass-based Bidirectional Feature Pyramid Network(H-BiFPN)structure is introduced for bidirectional fusion of features at different scales.Through feature reconstruction and fusion,this method enriches feature information of varying scales,addressing the issue of information loss of small targets under continuous convolution and enhancing small target detection accuracy.Finally,the Wise Intersection of Union(WIoU)loss function is employed to optimize the model,enhancing prediction accuracy by focusing on common anchor boxes.Experimental results on the expanded Chinese Power Line Insulator Dataset(CPLID)demonstrate that the YOLO-RR model achieves a recognition rate of 93.6% with network parameters compressed to 5.16×10^(6),outperforming comparative models.The proposed model meets the requirements of accurate localization and real-time performance for insulator defect detection,exhibiting robust detection performance even in scenarios with significant background interference and lighting effects.
作者
杨露露
马萍
王聪
李新凯
孟月
张宏立
YANG Lulu;MA Ping;WANG Cong;LI Xinkai;MENG Yue;ZHANG Hongli(College of Electrical Engineering,Xinjiang University,Urumqi 830017,Xinjiang,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2024年第7期303-313,共11页
Computer Engineering
基金
国家自然科学基金(52065064,52267010,62263030)
新疆维吾尔自治区自然科学基金(2022D01E33)。
关键词
绝缘子检测
YOLO模型
特征重用
特征重建
轻量化
智能巡检
insulator detection
YOLO model
feature reuse
feature reconstruction
lightweight
intelligent inspection
