基于注意力和特征融合的光学遥感图像配准

Optical remote sensing image registration based on attention and feature fusion

针对部分基于深度学习的光学遥感图像配准模型配准精度较低、用时较长的问题,提出了一种结合了注意力块和残差网络的配准方法。首先使用随机颜色抖动增强数据,然后采用改进的双注意力块过滤图像中的无关信息和干扰信息,紧接着采用改进的残差网络提取过滤后图像的高、低维特征信息并进行融合。最后利用双向相关特征匹配得到两个匹配关系,并计算合成图像的匹配参数,再通过仿射变换完成图像配准。在四个数据集上的实验结果表明,平均配准精度提升了6.87%,而平均配准时间仅为1.23秒,提高了光学遥感图像配准的精度与效率。

In order to deal with the problems of low registration accuracy and long application time of some optical remote sensing image registration based on deep learning model,we propose a registration method that combines attention blocks and residual networks.Data were first augmented using random color jitter,then a modified double attention block was applied to filter out irrelevant and interfering information from the images,followed by a modified residual network to extract and fuse the high and low dimensional feature information from the filtered images.Two matching relationships were finally obtained using bi-directional correlation feature matching,and matching parameters were calculated for the resultant images,followed by complete image registration via affine transformation.Experimental results on the four datasets show that,under the"proportion of key points correctly estimated"metric,the accuracy improves by 6.87%on average,while the average registration time is only 1.23 s,improving the accuracy and efficiency of optical remote sensing image registration.