基于改进Faster R-CNN的输电线路断股识别算法研究
Modified-faster-R-CNN-based Identification Algorithm for Transmission Line Broken Strand
张吉庆 1姚攀 2宋坤 3谢蓉 3廖红华3
作者信息
- 1. 湖北民族大学信息工程学院,湖北 恩施 445000;国网湖北省电力有限公司恩施供电公司,湖北 恩施 445000
- 2. 国网湖北省电力有限公司恩施供电公司,湖北 恩施 445000
- 3. 湖北民族大学信息工程学院,湖北 恩施 445000
- 折叠
摘要
为解决输电线路断股图像识别不精确的问题,设计了一种基于改进 Faster R-CNN算法的输电线路断股识别检测方法.该方法利用 DenseNet121 代替 VGG16,并融合各密集块的输出特征;同时,采用 Soft-NMS 算法取代NMS算法来缓解重叠目标漏检的问题.实验结果表明,更换特征提取网络后,mAP 达 85.1%,且特征融合后的mAP 达到了 90.2%,相比改进前提升 5.1%.采用Soft-NMS算法后,mAP 提升了 1.6%.最终改进后的模型mAP从 80.8%提高到了 92.5%,证明了改进后的算法能有效提高检测能力.
Abstract
In order to solve the problem of inaccurate image recognition of power transmission line broken strands,an i-dentification method for power transmission line broken strands was designed based on modified Faster R-CNN algorithm.This method uses DenseNet121 instead of VGG16 and integrates the output features of each dense block.At the same time,NMS algorithm is replaced by Soft-NMS algorithm to alleviate false negative identification of overlapping targets.The experimental results showed that after replacing feature extraction network,the mAP reaches 85.1%.Meanwhile,the mAP after feature fusion reached 90.2%,achieving an increase of 5.1%compared to before improvement.By adopting the Soft-NMS algorithm,mAP was increased by 1.6%.The final improved model had anmAP increment from 80.8%to 92.5%,demonstrating its effectiveness in improving detection performance.
关键词
输电线路/Faster/R-CNN/断股/特征融合Key words
transmission line/Faster R-CNN/broken strand/feature fusion引用本文复制引用
基金项目
国家自然科学基金项目(62163013)
湖北省自然科学基金项目(2021CFB542)
出版年
2024
NETL
NSTL
维普
万方数据