| 摘要: |
| 针对现有遥感图像场景分类算法处理数据类型单一的问题，以及缺乏星上部署研究的现状，将域泛化算法应用于多模态的遥感图像场景分类任务中，提出一种任务相关特征加权的场景分类算法，并在商用货架(Commercial Off-The-Shelf，COTS)器件上进行部署和测试。该算法基于通道和空间注意力机制对骨干网络提取到的特征进行加权，以增强任务相关特征；引入解码器模块重建图像的高频特征，该模块在训练过程中辅助训练，加强模型对任务相关特征的学习。以上两方面改进能够增强模型对跨域不变特征的学习，提升模型的泛化能力。实验结果表明，该算法具有82.95%的跨域平均分类准确率，相较于ERM、MixStyle、V-Rex、POEM和SNSC等5种算法至少提升3.09%。在COTS器件上进行算法部署，测试发现使用TensorRT部署方式能够在器件整体功耗7 W内实现每秒超过200帧的执行速度，满足在轨应用需求。 |
| 关键词: 遥感图像 场景分类 域泛化 在轨部署 COTS器件 |
| DOI：10.20079/j.issn.1001-893x.240401004 |
|
| 基金项目:中国科学院空间科学战略性科技先导专项基金（XDA04060300） |
|
| A Satellite-oriented Domain Generalization Algorithm for Remote Sensing Image Scene Classification |
| DU An揳n,ZHOU Qing |
| (1.National Space Science Center,Chinese Academy of Sciences,Beijing 100190,China;2.University of Chinese Academy of Sciences,Beijing 100049,China) |
| Abstract: |
| To address the problems of existing remote sensing image scene classification algorithms handling single data types and the lack of on-board deployment studies,the domain generalization algorithm is applied to multimodal remote sensing image scene classification tasks.A scene classification algorithm with task-relevant feature weighting is proposed and deployment and testing on Commercial Off-The-Shelf(COTS)devices is carried out.The algorithm weights the features extracted from the backbone network based on the channel and spatial attention mechanism to enhance the task-relevant features.Additionally,the decoder module is introduced to reconstruct the high-frequency features of the image,which assists in the training process and enhances the learning of task-related features.The above improvements can enhance the model’s learning of cross-domain invariant features and improve the model’s generalization ability.Experimental results demonstrate that the proposed algorithm achieves a cross-domain average classification accuracy of 82.95 |
| Key words: remote sensing image scene classification domain generalization on-orbit deployment COTS device |
