卷积神经网络特征在遥感图像配准中的应用

Application of convolutional neural network feature to remote sensing image registration

遥感图像配准是许多遥感应用的重要步骤之一。卷积神经网络(convolutional neural network,CNN)提取的图像高层特征在图像分类和检索问题上表现优异,能够克服低层配准特征的表达能力有限、容易受到干扰等问题。因此对利用CNN特征进行遥感图像配准开展研究。首先,针对遥感图像配准问题,对CNN中的全连接层特征和不同聚合大小的卷积层特征进行了研究;然后,对利用CNN特征进行图像配准的方法进行了分析;最后,将CNN特征与尺度不变特征变换(scale-invariant feature transform,SIFT)特征在图像的旋转角度、缩放倍数和亮度依次变换时的配准性能进行了对比分析。实验结果表明,在匹配精度和正确对应点的数量方面,CNN特征比SIFT方法具有更好的匹配性能;对变换后的图像而言,微调后的CNN特征比SIFT特征具有更强的鲁棒性。

Successful remote sensing image registration is one of the foundations of many remote sensing applications.Image high-lever features extracted by convolutional neural network (CNN) have achieved excellent performance in image classification and retrieval,and can be used to solve some problems of low-lever image registration features,such as the limitation of expression capability and easily being interfered.Hence,in this paper,the authors investigated the problem as to how to use CNN feature for remote sensing image registration.First,the authors investigated different CNN features from fully connected layers and aggregating convolutional features with different sizes from convolutional layer to register remote sensing image.Then the authors introduced the procedure by using CNN feature for image registration.Finally,the authors compared the registration performances of CNN features and scale-invariant feature transform (SIFT) features after the transformation of the image's perspective,brightness and scale,respectively.The experimental results show that the CNN feature has better matching performance than the SIFT method in terms of matching accuracy and correct number of corresponding points.The finely tuned CNN feature has stronger robustness to the transformed image than the SIFT feature.