An Adaptive Remote Sensing Image-Matching Network Based on Cross Attention and Deformable Convolution

There are significant background changes and complex spatial correspondences between multi-modal remote sensing images, it is difficult for existing methods to extract common features images effectively, leading poor matching results. In order improve the effect, with high robustness extracted; this paper proposes a multi-temporal algorithm CMRM (CNN matching) based on deformable convolution cross-attention. First, VGG16 backbone network, Deformable (DeVgg) constructed by introducing convolutions adapt geometric distortions in of different shapes scales; second, extracted from DeVgg input cross-attention module better capture correspondence changes; finally, key points corresponding descriptors output feature map. stage, solve problem quality points, BFMatcher used rough registration, then RANSAC adaptive threshold constraint. The proposed performs well public dataset HPatches, MMA values 0.672, 0.710, 0.785 when selected as 3–5. results show that compared methods, our method improves accuracy images.