Abstract: In view of the complex and changeable geographical information of remote sensing images, the difficulty in fully extracting local details and contextual information, and the low accuracy and long time of partial registration models, this paper proposes a registration that combines attention and multi-scale features. The model first introduces a lightweight convolutional network combining Transformer and inverse residual structure in the feature extraction stage, and deepens the focus on channel spatial information by embedding hybrid attention blocks. Furthermore, in order to capture contextual feature information more effectively, the attention-enhanced multi-scale dilated convolution module is used to perform deep filtering extraction and obtain more refined and rich feature semantic maps. Secondly, in the matching stage, the cross-correlation optimal weighted bidirectional matching method is used to calculate the dense correspondence to obtain bidirectional parameters, and the final transformation parameters are synthesized by weighting the parameter regression network, and finally the affine transformation completes the image registration. Experimental results show that when the proportion coefficients of correctly estimated key points are 0.03, 0.05 and 0.1, the registration accuracy on the three datasets reaches 61.9%, 86.2% and 93.6%, while the average registration time is only 1.05 seconds, which proves that this model effectively improves the accuracy and efficiency of remote sensing image registration.

Key words: remote sensing image registration, contextual features, enhanced attention, bidirectional matching

摘要： 针对遥感图像地理信息复杂多变、局部细节与上下文信息难以被充分提取，以及部分配准模型精度较低、用时较长等问题，提出了一种融合注意力与多阶尺度特征的配准模型，在特征提取阶段引入Transformer与逆残差结构结合的轻量级卷积网络，通过嵌入混合注意力块加深对通道空间信息的关注，进一步地，为了更有效地捕获上下文特征信息，使用增强注意力的多尺度扩张卷积模块进行深层次过滤提取，以获取更精细和丰富的特征语义图。在匹配阶段采用互相关最优加权的双向匹配方法，计算密集对应关系得到双向参数，并通过参数回归网络加权合成最终变换参数，仿射变换完成图像配准。实验结果表明，关键点正确估计的比例系数为0.03、0.05和0.1的情况下，在三个数据集上的配准精度达到61.9%、86.2%、93.6%，而平均配准时间仅为1.05?s，证明了该模型有效提升遥感图像配准的精度和效率。

关键词: 遥感图像配准, 上下文特征, 增强注意力, 双向匹配