一种遥感重标注的像素级人机交互神经网络
A deep pixel-level guided neural network based on human-computer interaction for remote sensing data re-labeling
梁桂明 1肖明虹 1余凡 2谢俊威2
作者信息
- 1. 广西地理信息测绘院,广西柳州 545006
- 2. 北京建筑大学测绘与城市空间信息学院,北京 102627
- 折叠
摘要
针对遥感图像标注产品的自动化像素标注难题,本文提出一种遥感重标注的像素级人机交互深度神经网络(DGN)方法,可自动生成图像标注,并允许标注者在发现错误后通过简单的指导信息自适应地纠正先前的标注,且采用了一种新的训练方法和度量标准,用以衡量重新标注的效率.在Vaihingen数据集上使用不同的基础架构和骨干网络对该算法进行了实验验证,结果表明,DGN能有效地引导指导模块利用指导信息,将重新标注效率提高至2.52倍,并一定程度上提高了分类的精度.
Abstract
This paper proposes a pixel level human-machine interactive deep neural network(DGN)method for remote sensing image annotation products,which can automatically generate image annotations and allow annotators to adaptively correct previous annotations through simple guidance information after discovering errors.A new training method and measurement standard are adopted to measure the efficiency of re-annotation.The algorithm was experimentally validated using different infrastructure and backbone networks on the Vaihingen dataset,and the results showed that DGN can effectively guide the guidance module to utilize guidance information,increasing the efficiency of re labeling by 2.52 times and improving the accuracy of classification to a certain extent.
关键词
图像标注/机器学习/语义分割Key words
image annotation/machine learning/semantic segmentation引用本文复制引用
出版年
2024
NETL
NSTL
万方数据