摘要
高分辨率遥感图像的语义分割问题是目前遥感图像处理领域中的研究热点之一。传统的有监督分割方法需要大量的标记数据,而标记过程又较为困难和耗时。针对这一问题,提出一种基于生成式对抗网络的半监督高分辨率遥感图像语义分割方法,只需要少量样本标签即可得到较好的分割结果。该方法为分割网络添加全卷积形式的辅助对抗网络,以助于保持高分辨率遥感图像分割结果中的标签连续性;更进一步,提出一种新颖的能够进行注意力选择的对抗损失,以解决分割结果较好时判别器约束的分割网络更新过程中存在的难易样本不均衡问题。在ISPRS Vaihingen 2D语义标记挑战数据集上的实验结果表明,与现有其它语义分割方法相比,所提出方法能够较大幅度地提高遥感图像的语义分割精度。
Semantic segmentation of very high resolution(VHR)remote sensing images is one of the hot topics in the field of remote sensing image processing.Traditional supervised segmentation methods demand a huge mass of labeled data while the labeling process is very consuming.To solve this problem,a semi-supervised semantic segmentation method for VHR remote sensing images based on Generative Adversarial Networks(GANs)is proposed,and only a few labeled samples are needed to obtain pretty good segmentation results.A fully convolutional auxiliary adversarial network is added to the segmentation network,conducing to keeping the consistency of labels in the segmentation results of VHR remote sensing images.Furthermore,a novel adversarial loss with attention mechanism is proposed in the paper in order to solve the problem of easy sample over-whelming during the updating process of the segmentation network constrained by the discriminator when the segmentation results can confuse the discriminator.The experimental results on ISPRS Vaihingen 2D Semantic Labeling Challenge Dataset show that the proposed method can greatly improve the segmentation accuracy of remote sensing images compared with other state-of-the-art methods.
作者
刘雨溪
张铂
王斌
LIU Yu-Xi;ZHANG Bo;WANG Bin(Key Laboratory for Information Science of Electromagnetic Waves(MoE),Fudan University,Shanghai 200433,China;Research Center of Smart Networks and Systems,School of Information Science and Technology,Fudan University,Shanghai 200433,China)
出处
《红外与毫米波学报》
SCIE
EI
CAS
CSCD
北大核心
2020年第4期473-482,共10页
Journal of Infrared and Millimeter Waves
基金
国家自然科学基金(61971141,61731021)。
关键词
高分辨率遥感图像
语义分割
深度学习
生成式对抗网络
损失函数
very high resolution remote sensing images
semantic segmentation
deep learning
generative adversarial networks
loss function
