摘要
由于遥感影像中建筑物的纹理的复杂性和背景的多样性,从高分辨率遥感图像中实现建筑物自动提取是一个具有挑战性的任务。现有的最先进的深度学习框架使用重复的池化操作,导致详细信息的丢失。在LinkNet框架上结合空洞卷积、注意力机制和多尺度预测,提出了ADLinkNet,并在WHU Building数据集上进行了测试。实验结果表明,ADLinkNet优于DeepLabv3、UNet、SegNet和D-LinkNet等。
Due to the complexity of texture and diversity of background of buildings in remote sensing images,automatically extracting buildings from high-resolution sensing images is a challenging task.The most advanced deep learning frameworks use repeated pooling operations,resulting in the loss of detailed information.Therefore,we combine the dilated convolution,attention module and multi-scale prediction based on the LinkNet to propose ADLinkNet.And then we test it on WHU Building dataset.The experimental results show that ADLinkNet is superior to DeepLabv3,UNet,SegNet and D-LinkNet,etc.
作者
吴目宇
胡翔云
荣子豪
WU Muyu;HU Xiangyun;RONG Zihao(School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,China;Institute of Artificial Intelligence in Geomatics,Wuhan University,Wuhan 430079,China)
出处
《测绘地理信息》
CSCD
2021年第S01期237-241,共5页
Journal of Geomatics
