基于注意力网络尺度特征融合的遥感场景分类
Remote sensing scene classification based on attention network scale feature fusion
投稿时间:2022-01-26  修订日期:2022-01-26
DOI:
中文关键词: 遥感图像  场景分类  多尺度特征  多选框注意力模型  LBP特征融合
英文关键词: Remote sensing image  Scene classification  Multi-scale features  Multi-check box attention model  LBP feature fusion
基金项目:湖北省技术创新专项重大项目(2019ABA101);湖北省科技重大专项(2020AEA011);武汉市科技计划应用基础前沿项目(2020020601012267)
作者单位邮编
帖军 中南民族大学计算机科学学院 
肖鹏飞 中南民族大学计算机科学学院 
郑禄 中南民族大学计算机科学学院 430074
马海荣 湖北省农业科学院农业经济技术研究所 
彭丹 中南民族大学计算机科学学院 
摘要点击次数: 173
全文下载次数: 0
中文摘要:
      针对遥感数据集存在的类内差异性大与类间相似性高的特点导致遥感场景分类准确率不高的问题,提出一种基于注意力网络尺度特征融合(MSA-CNN)的遥感影像场景分类模型。首先,将遥感图像经过尺度变换和预处理操作输入到VGG-16网络提取遥感影像的多尺度特征,使用多选框注意力模型提取图像多尺度下的目标区域,对目标区域进行剪切和放大并输入到三层网络结构中。然后,融合原始影像的多尺度特征和目标区域的特征,并且利用LBP对全局特征表达,克服遥感图像因拍摄角度不同带来的差异性。最后,将融合的多尺度特征输入到网络全连接层来完成最终的分类预测任务。实验结果显示,MSA-CNN平均分类精度较注意循环卷积网络(ARCNet)和传统细粒度循环注意力网络(RA-CNN)在NWPU-RESISC45公开数据集上分别提升1.63%和2.66%,在UC Merced Land-Use公开数据集上较RA-CNN提升0.64%。结果表明,提出的MSA-CNN能够有效提高遥感图像场景分类的准确率。
英文摘要:
      Aiming at the existence of remote sensing data sets within the class difference between large and class high similarity features in remote sensing scene classification accuracy is not high, this paper proposes a remote sensing image scene classification model based on attention network scale feature fusion (MSA-CNN). Firstly, the remote sensing image is input to the VGG-16 network through scaling transformation to extract the multi-scale features of the remote sensing image. The multi-scale target region of the image is extracted by using the multi-scale attention model, and the target region is cut and enlarged and input into the three-layer network structure. Then, the multi-scale features of the original image and the features of the target region are integrated, and LBP is used to express the global features to overcome the difference of the remote sensing image caused by different shooting angles, Finally, the fused multi-scale features are input to the full connection layer of the network to complete the final classification and prediction task. The experimental results show that the average classification accuracy of MSA-CNN is 1.63% and 2.66% higher than that of the attention circular convolutional network (ARCNet) and the traditional fine-grained circular attention network (RA-CNN) on the NWPU-Resisc45 public dataset, respectively. In the open data set of UC Merced Land-Use, it is 0.64% higher than that of RA-CNN. The results show that the proposed MSA-CNN can effectively improve the accuracy of scene classification of remote sensing images.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭