Scholarlink

<p>Remote sensing image pairs of different time phases have complex and changeable semantic contents, and traditional convolutional registration methods are challenging in modeling subtle local changes and global large-scale deformation differences in detail. This results in poor registration performance and poor feature representation. To address these problems, a semantically enhanced dual-graph perception framework is proposed. This framework aims to gradually achieve semantic alignment and precise registration of remote sensing image pairs of different time phases via coarse to fine stages. On the one hand, a newly designed large-selection kernel convolution attention module is used to learn affine transformation parameters. Attention to global semantics perceives the large pixel displacement deviation caused by large-scale deformation, and the association relationship is established between remote sensing image pairs of different time phases. At the same time, dual-graph perception modules are embedded in multiple subspace structures, and the subtle local changes of remote sensing image pairs are modeled through the dynamic aggregation ability of graph perception nodes to achieve coarse registration of remote sensing images. On the other hand, a U-shaped module guided by global attention with deformable convolution is used to refine the local spatial structural features and global contextual semantic information of the rough registration, establish dependencies between channels, and correct the pixel displacement deviation of remote sensing image pairs of different phases through position encoding. It is worth noting that the newly designed weighted loss function supervises the learning of each module and the entire network structure from the perspective of inverse consistency, promoting the network’s optimal performance. Finally, the experimental results on the AerialData and GFRS datasets show that the proposed framework has good registration performance, with mean absolute error (MAE) of 3.64 and 3.81, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SDGANets: a semantically enhanced dual graph-aware network for affine and registration of remote sensing images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年2月28日

查看全文
  • Xie Zhuli,
  • Wan Gang,
  • Liu Jia,
  • Bu Dongdong

摘要

Remote sensing image pairs of different time phases have complex and changeable semantic contents, and traditional convolutional registration methods are challenging in modeling subtle local changes and global large-scale deformation differences in detail. This results in poor registration performance and poor feature representation. To address these problems, a semantically enhanced dual-graph perception framework is proposed. This framework aims to gradually achieve semantic alignment and precise registration of remote sensing image pairs of different time phases via coarse to fine stages. On the one hand, a newly designed large-selection kernel convolution attention module is used to learn affine transformation parameters. Attention to global semantics perceives the large pixel displacement deviation caused by large-scale deformation, and the association relationship is established between remote sensing image pairs of different time phases. At the same time, dual-graph perception modules are embedded in multiple subspace structures, and the subtle local changes of remote sensing image pairs are modeled through the dynamic aggregation ability of graph perception nodes to achieve coarse registration of remote sensing images. On the other hand, a U-shaped module guided by global attention with deformable convolution is used to refine the local spatial structural features and global contextual semantic information of the rough registration, establish dependencies between channels, and correct the pixel displacement deviation of remote sensing image pairs of different phases through position encoding. It is worth noting that the newly designed weighted loss function supervises the learning of each module and the entire network structure from the perspective of inverse consistency, promoting the network’s optimal performance. Finally, the experimental results on the AerialData and GFRS datasets show that the proposed framework has good registration performance, with mean absolute error (MAE) of 3.64 and 3.81, respectively.