摘要
由于农村道路具有位置特殊、结构和材质多样且差异大等特点,使得高精度遥感图像在提取农村公路信息中混入大量噪声,造成影像特征聚类和分割模糊,难以保证路网提取精度。因此,该文提出一种EDDNN模型,在基础层次结构下,交叉采用池化层与卷积层作为网络隐藏层主要结构,提取遥感影像全局图像特征;在ED-DNN网络架构内部拟采用跳跃连接方式,结合深度神经网络架构内深层和浅层分辨率信息,提高路面影像分割的精细度。实例表明:采用ED-DNN模型提取道路信息具有良好的泛化能力,训练集准确率达到0.9982,损失值降低为0.0137,能准确地提取农村公路特征参数。
The rural roads have the characteristics of special location,diverse structures and materials,and large differences,which makes the high-precision remote sensing images mixed with a large amount of noise in extracting the rural highway information,resulting in image feature clustering and fuzzy segmentation,which is difficult to ensure the accuracy of road network extraction.Because of this problem,this paper proposes a ED-DNN model.Under the basic hierarchy,the pool layer and convolution layer are used as the main structure of the network hidden layer to extract the global image features of remote sensing images.In the ED-DNN network architecture,it proposes to adopt the jump connection mode,combined with the deep and shallow resolution information in the deep neural network architecture,so as to improve the precision of pavement image segmentation.The example shows that the ED-DNN model has good generalization ability for extracting road information,the accuracy of the training set is 0.9982,the loss value is reduced to 0.0137,and the characteristic parameters of rural highway can be extracted accurately.
作者
昌宏哲
杜红静
袁洋
Chang Hongzhe;Du Hongjing;Yuan Yang(Highway Administration Bureau of Henan Provincial Transportation Department;Henan Transportation Development Research Institute Co.,Ltd.)
出处
《勘察科学技术》
2022年第3期11-15,共5页
Site Investigation Science and Technology
基金
河南省交通运输厅项目(河南省农村公路“百县通村入组工程”建设目标任务智能化核查关键技术研究)(编号2020G2)