利用风格迁移和特征点的多模态图像配准算法

Multimodal Image Registration Algorithm Using Style Transfer and Feature Points

由于多模态遥感图像在光谱成份上存在巨大的差异,传统图像配准算法在该类图像的配准中正确率非常低。针对这一难题,提出了一种利用风格迁移和特征点的图像配准算法。首先,利用卷积神经网络对基准图像的风格特征以及待配准图像的内容特征进行抽取并重新组合,得到一幅与基准图像差异性较小的生成图像;其次,通过图像分割的方法分离出待配准图像中没有明显纹理信息的部分,清除生成图像中多余的纹理;最后,使用加速鲁棒性特征(speed up robust features,SURF)算法提取特征点,进行图像配准。实验结果表明,与传统图像配准算法相比,该方法有效提高了多模态遥感图像配准的正确率和鲁棒性。

Due to the huge difference in spectral composition of multimodal remote sensing images,the accuracy of traditional image registration algorithms of such images is very low.To address this problem,an image registration algorithm using style transfer and feature points is proposed.Firstly,use the convolutional neural network to extract and recombine the style features of the reference image and the content features of the image to be registered,and obtain a generated image with a small difference from the reference image.Secondly,the image segmentation method is used to separate the parts of the image to be registered that have no obvious texture information,and to remove the redundant texture in the generated image.Finally,use speed up robust features(SURF)algorithm to extract feature points and perform image registration.Experimental results show that,compared with traditional image registration algorithms,this method effectively improves the accuracy and robustness of multimodal remote sensing image registration.