利用SAM大模型对SAR影像配准的深度学习方法

Deep Learning Approach for SAR Image Registration Using Segment Anything Model

在SAR影像的配准中,由于SAR影像固有的斑点噪声,导致配准稳定性差、配准效率低。针对这个问题,文章提出了利用SAM大模型对SAR影像配准的深度学习方法。该方法使用SAM分割大模型对参考图像进行分割,得到参考图像的掩码,然后利用基于编码器和解码器的深度学习网络寻找该掩码在主图像中的位置,以此来寻找参考图像和主图像之间的对应关系,并把这个对应关系转成点与点的对应关系,进而达到配准的目的。实验结果表明:该方法能准确找到参考图像的掩码在主图像中的位置,均值误差在5%以内;在大尺寸SAR影像配准中,该方法具有很好的稳定性,且配准效率较高。

In the registration of SAR images,the inherent speckle noise of SAR images results in poor registration stability and low registration efficiency.To address this problem,this paper proposes a deep learning method that uses segment anything model to register SAR images.This method uses the segment anything model to segment the reference image to obtain the masks of the reference image.Then,it utilizes encoder-decoder-based deep learning networks to locate this mask in the master image,thereby establishing the correspondences between the reference image and the master image.Next,convert this correspondence into point-to-point relationships to achieve the purpose of SAR image registration.Experimental results show that this method can accurately find the position of the reference image mask in the master image,with a mean error within 5%.And in the registration of large-scale SAR images,this method has good stability and high registration efficiency.