针对目前基于信道脉冲响应(Channel Impulse Response,CIR)的非视距(None Line of Sight,NLoS)/视距(Line of Sight,LoS)识别方法精度低、泛化能力差的问题,提出了一种多层卷积神经网络(Convolutional Neural Network,CNN)与通道注意力...针对目前基于信道脉冲响应(Channel Impulse Response,CIR)的非视距(None Line of Sight,NLoS)/视距(Line of Sight,LoS)识别方法精度低、泛化能力差的问题,提出了一种多层卷积神经网络(Convolutional Neural Network,CNN)与通道注意力模块(Channel Attention Module,CAM)相结合的NLoS/LoS识别方法。在多层CNN中嵌入CAM提取原始CIR的时域数据特征,利用全局平均池化层代替全连接层进行特征整合并分类输出。使用欧洲地平线2020计划项目eWINE公开的数据集进行不同结构模型和不同识别方法的对比实验,结果表明,所提出的CNN-CAM模型LoS和NLoS召回率分别达到了92.29%与87.71%,准确率达到了90.00%,F1分数达到了90.22%。与现有多种传统识别方法相比,均具有更好的识别效果。展开更多
This study combines ground penetrating radar(GPR)and convolutional neural networks for the intelligent detection of underground road targets.The target location was realized using a gradient-class activation map(Grad-...This study combines ground penetrating radar(GPR)and convolutional neural networks for the intelligent detection of underground road targets.The target location was realized using a gradient-class activation map(Grad-CAM).First,GPR technology was used to detect roads and obtain radar images.This study constructs a radar image dataset containing 3000 underground road radar targets,such as underground pipelines and holes.Based on the dataset,a ResNet50 network was used to classify and train different underground targets.During training,the accuracy of the training set gradually increases and finally fluctuates approximately 85%.The loss function gradually decreases and falls between 0.2 and 0.3.Finally,targets were located using Grad-CAM.The positioning results of single and multiple targets are consistent with the actual position,indicating that the method can eff ectively realize the intelligent detection of underground targets in GPR.展开更多
文摘针对目前基于信道脉冲响应(Channel Impulse Response,CIR)的非视距(None Line of Sight,NLoS)/视距(Line of Sight,LoS)识别方法精度低、泛化能力差的问题,提出了一种多层卷积神经网络(Convolutional Neural Network,CNN)与通道注意力模块(Channel Attention Module,CAM)相结合的NLoS/LoS识别方法。在多层CNN中嵌入CAM提取原始CIR的时域数据特征,利用全局平均池化层代替全连接层进行特征整合并分类输出。使用欧洲地平线2020计划项目eWINE公开的数据集进行不同结构模型和不同识别方法的对比实验,结果表明,所提出的CNN-CAM模型LoS和NLoS召回率分别达到了92.29%与87.71%,准确率达到了90.00%,F1分数达到了90.22%。与现有多种传统识别方法相比,均具有更好的识别效果。
基金supported in part by the National Natural Science Fund of China under Grant 52074306in part by the National Key Research and Development Program of China under Grant 2019YFC1805504in part by the Fundamental Research Funds for the Central Universities under Grant 2023JCCXHH02。
文摘This study combines ground penetrating radar(GPR)and convolutional neural networks for the intelligent detection of underground road targets.The target location was realized using a gradient-class activation map(Grad-CAM).First,GPR technology was used to detect roads and obtain radar images.This study constructs a radar image dataset containing 3000 underground road radar targets,such as underground pipelines and holes.Based on the dataset,a ResNet50 network was used to classify and train different underground targets.During training,the accuracy of the training set gradually increases and finally fluctuates approximately 85%.The loss function gradually decreases and falls between 0.2 and 0.3.Finally,targets were located using Grad-CAM.The positioning results of single and multiple targets are consistent with the actual position,indicating that the method can eff ectively realize the intelligent detection of underground targets in GPR.
