摘要
为了解决复杂图像背景条件下高压塔上鸟巢检测的问题,提出一种基于Faster R-CNN的多尺度高压塔鸟巢检测方法.在特征提取方面,通过改进基于卷积神经网络的ResNet-50特征提取网络分别完成图像中高压塔和鸟巢的特征提取;在建议区域提取方面,提出在两种不同深度的卷积特征图上采用合理尺度的不同大小滑动窗口方式完成建议区域的提取,增强了对鸟巢的检测能力;在目标检测方面,提出在高分辨卷积特征图上进行上采样操作,并额外增加SENet特征增强模块,有效提高了目标检测效果.最后利用同时检测出的高压塔区域条件剔除了此区域之外的鸟巢检测结果,进一步提升了最终的鸟巢检测精度.该方法在2 000幅无人机实际巡检航拍的含有鸟巢的图像数据集上进行了测试.实验结果表明:本文方法的平均检测精度达到了84.55%.对比已有的基于HOG+SVM的检测方法和基于经典的Faster R-CNN ResNet-50检测方法,本文所提出的方法平均检测精度分别提高了43.5%和15.2%,并加快了检测速度.该方法为无人机电力智能巡检提供了一种新的解决办法.
In order to solve the problem of bird’s nest detection on high-voltage towers using ima-ges with complex backgrounds,ageneral method of bird’s nest detection in multi-scale on high-voltage towers based on Faster R-CNN is proposed in this paper.In the aspect of feature extrac-tion,the feature of high-voltage towers and bird’s nests in the image are extracted by the im-proved ResNet-50network.In the aspect of region proposal,it is proposed to use sliding windows of different sizes and reasonable scales to gain region proposals on two convolution feature maps with different depths,which enhances the detection ability of bird’s nest.For the object detec-tion,with the help of an upsampling layer is added to the feature maps with high-resolution and SENet feature enhancement module is added to the ResNet network,the proposed method a-chieves excellent results.Finally,the bird’s nest detection precision is further improved by sim-ultaneous detection of high-voltage towers.The method is evaluated on a data set composed of 2000images containing bird’s nests which are sampled by unmanned aerial vehicle during actual inspection.The experimental results show that the average detection precision of this method rea-ches 84.55%.Compared with the existing detection methods based on HOG+SVM or classical Faster R-CNN ResNet-50,the average detection precision of the proposed method is improved by 43.5%and 15.2%respectively,and the detection speed is accelerated.This method provides a no-vel solution for intelligent inspection of UAV power system.
作者
王纪武
罗海保
鱼鹏飞
郑乐乐
胡方全
WANG Jiwu;LUO Haibao;YU Pengfei;ZHENG Lele;HU Fangquan(School of Mechanical,Electronic and Control Engineering,Beijing Jiaotong University,Beijing 100044,China)
出处
《北京交通大学学报》
CAS
CSCD
北大核心
2019年第5期37-43,共7页
JOURNAL OF BEIJING JIAOTONG UNIVERSITY
基金
中央高校基本科研业务费专项资金(KMGY318002531)~~
作者简介
第一作者:王纪武(1970—),男,辽宁辽阳人,副教授,博士.研究方向为深度学习、图像处理.email:jwwang@bjtu.edu.cn.
