基于GAN-Attention-UNet的网络地理遥感图像配准

TGeo-Remote Sensing Image Registration Based on GAN-Attention-UNet Network

基于地理遥感图像UNet配准网络,改进了一种半监督对抗性与注意力机制相结合的UNet配准网络。该网络由提取特征的编码器和精确定位的解码器组成,解码部分通过相加操作融合来自编码器的信息,卷积层与池化层之间引入了空间注意力与通道注意力相结合,有效抑制地理遥感图像中无关的区域,突出显著地理超遥感特征,使用对抗性相似优化与空间变换相结合。利用地理遥感数据集对该方法进行实验,实验结果表明,该方法在配准精度与速度上都有较大的提升。

A UNet alignment network combining semi-supervised adversarial and attention mechanisms is improved based on the UNet alignment network for geographic remote sensing images.The network firstly consists of an encoder for extracting features and a decoder for precise positioning,and the decoding part fuses the information from the encoder by summation operation;secondly,spatial attention combined with channel attention is introduced between the convolution and pooling layers,which can suppress irrelevant regions in geographic remote sensing images and highlight significant geographic hyper-remote features;finally,adversarial similarity optimization combined with spatial transformation is used.The method is experimented using geographic remote sensing dataset,and the experimental results show that the method has a large improvement in alignment accuracy and speed.