摘要
为实现高分辨率卫星图像的道路自动提取,设计一种编码器-解码器结构的图像分割方法。针对卫星图像中乡村地区的道路提取结果不佳,以及不能对阴影区域、被遮挡区域的道路进行有效提取的问题,以不含全连接层的VGG13作为编码器的骨干网络,对解码器部分进行设计,达到对道路区域进行有效提取的目的,并对模型训练使用的损失函数进行介绍。在开始训练之前,对DeepGlobe道路提取数据集进行预处理。使用PaddlePaddle深度学习框架展开实验,改进后的方法在验证集上的IoU,acc,Kappa分别可以达到0.6194,0.9811,0.7551,对比实验结果显示,与使用DeepLabv3+和U-Net的道路提取方法相比,可以有效提升道路提取结果的准确性和完整性。
In order to achieve the automatic road extraction of high-resolution satellite images,an image segmentation method based on the encoder-decoder structure is designed.In view of the poor results of road extraction in rural areas in satellite images,and the inability to effectively extract roads in shadow areas and sheltered areas,VGG13 without full connection layers is used as the backbone network of the encoder,the decoder is designed to effectively extract road areas,and the loss functions used in the training of model are introduced.Before starting the process of training,the DeepGlobe road extraction data set is preprocessed.The PaddlePaddle deep learning framework is used to carry out experiments.On the validation set,the value of IoU,acc and Kappa of the improved method can reach 0.6194,0.9811 and0.7551 respectively.The experimental results show that this method can effectively improve the accuracy and integrity of road extraction results compared with the road extraction methods using DeepLabv3+and U-Net.
作者
晏美娟
魏敏
文武
YAN Meijuan;WEI Min;WEN Wu(College of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China)
出处
《成都信息工程大学学报》
2022年第1期46-50,共5页
Journal of Chengdu University of Information Technology
基金
四川省科技计划资助项目(2020YFG0442、2020YFG0453)。
关键词
深度学习
卫星图像
道路提取
图像分割
飞桨
deep learning
satellite imagery
road extraction
image segmentation
PaddlePaddle
