基于改进RCF和无人机影像的电力线检测  

Power line detection based on an improved RCF and the UAV images

作　　者：郭家 江洪[1] 张雍 GUO Jia;JIANG Hong;ZHANG Yong(Key Laboratory of Spatial Data Mining&Information Sharing of MOE,National&Local Joint Engineering Research Center of Satellite Geospatial Information Technology,Academy of Digital China(Fujian),Fuzhou University,Fuzhou,Fujian 350108,China)

机构地区：[1]福州大学空间数据挖掘与信息共享教育部重点实验室,卫星空间信息技术综合应用国家地方联合工程研究中心,数字中国研究院(福建),福建福州350108

出　　处：《福州大学学报（自然科学版）》2024年第2期168-175,共8页Journal of Fuzhou University(Natural Science Edition)

基　　金：福建省科技计划引导性资助项目(2021Y0005)。

摘　　要：针对更丰富卷积特征(RCF)算法检测电力线时存在边缘模糊、特征图包含太多噪声、在融合特征图时丢失多尺度信息等问题,对RCF算法进行改进.首先,使用具有平移不变性的下采样技术增强模型的鲁棒性;然后,在RCF主干网络中引入卷积块注意力模块(CBAM)机制,提高模型对电力线特征的表达能力;最后,在RCF的侧输出网络中加入级联网络,借助基于通道注意力机制的多尺度特征融合模块对特征图进行融合,从而获得更优异的细节保持效果.实验结果表明,改进模型的最优数据集规模、最佳图像比例和平均精度可分别提高0.7%、 1.3%和1.7%,检测结果噪声数量少,电力线更加清晰准确.In order to solve the problems such as edge blur when richer convolutional features(RCF)algorithm detects power lines,feature maps contain too much noise,and multi-scale information is lost when fusing feature maps,RCF algorithm was improved in this paper.Firstly,the down-sampling technique with translation invariance was used to enhance model robustness.Secondly,convolutional block attention module(CBAM)mechanism was introduced into the convolutional block attention of RCF trunk network to enhance the power lines characteristics.Thirdly,the cascade network is added into the side output network of RCF,and the feature map is fused with the multi-scale feature fusion module using the channel attention mechanism,to obtain better details.The results showed that the optimal dataset scale,the optimal image scale and average precision of the improved RCF increased 0.7%,1.3%and 1.7%,respectively.The detection results of the improved model are less noise,and the power line is more clear and accurate.

关 键 词：电力线 边缘检测 更丰富卷积特征(RCF) 无人机 注意力机制 多尺度融合 

分 类 号：P237[天文地球—摄影测量与遥感]