一种基于对象意识多编码生成器网络的遥感场景分类方法
A Remote Sensing Scene Classification Method Based on Object-aware Multi-code Generator
刘卓 1边小勇 1杨博1
作者信息
- 1. 武汉科技大学计算机科学与技术学院 武汉 430065;武汉科技大学智能信息处理与实时工业系统湖北省重点实验室 武汉 430065
- 折叠
摘要
论文引入了一种对象意识多编码生成器网络的遥感场景方法,从多示例对象定位以及生成式对抗网络图像生成两个方面研究遥感场景分类方法.首先利用原始图像以及生成式对抗网络生成的伪图输入联合定位模块,利用两者特征上的差异性和互补性以定位多个信息性对象部件.然后提出一种多编码生成器网络,每个潜在编码用以恢复图像的特定区域,同时在生成器的中间层通过通道重要性权重组合这些潜在编码,合成足够细节的伪图.最后,论文联合对象部件级分类以及图像全局级分类作为最终的分类结果,用于遥感场景分类.该方法主要在两个公开的大规模场景数据集AID和NW-PU-RESISC45上进行了验证.与现有的方法相比,该方法有效定位到了多个信息性对象部件并且提高了遥感场景分类性能.
Abstract
This paper introduces an object-aware multi-code generator network approach for remote sensing scenes from two aspects,which are multiple instance object localization and generative adversarial networks.First,the original image as well as the fake image generated by the generative adversarial network are used as input to the joint localization module to exploit the differenc-es and complementarities in their features in order to localize multiple informative object parts.Then,a multi-coding generator net-work is proposed,where each latent code is used to recover a specific region of the image while combining these latent codes by channel importance weights in the middle layer of the generator to synthesize a fake image with sufficient detail.Finally,this paper combines object part-level classification and image global-level classification as the final classification results for remote sensing scene classification.The method in this paper is mainly validated on two publicly available large-scale scene datasets,AID and NW-PU-RESISC45.Compared with existing methods,this method effectively locates multiple informative object parts and improves the performance of remote sensing scene classification.
关键词
遥感场景分类/多示例定位/生成式对抗网络Key words
remote sensing scene classification/multiple instance parts localization/generative adversarial network引用本文复制引用
基金项目
湖北省自然科学基金项目(2018CFB575)
国家自然科学基金面上项目(62071456)
出版年
2024
国家科技期刊平台
NSTL
万方数据