GLoCNet: Robust Feature Matching With Global–Local Consistency Network for Remote Sensing Image Registration

Feature matching is a fundamental and critical task for remote sensing image registration (RSIR). However, numerous outliers (false matches) harm the feature point neighborhood structure due to the view transformation from the camera. Meanwhile, unknown local distortion obscures the distinction between inliers (correct matches) and outliers. To solve these problems, we propose a global–local consistency network (GLoCNet) for feature matching to exclude the interference of outliers under various transformation patterns and provide stable neighborhood support for the similarity metric of feature points. Specifically, a global transformation consistency (GTC) module is proposed to obtain a neighbor pool by exploiting the compact nature of the inlier distribution under different transformation patterns. In addition, feature points interact with information through the local neighborhood consistency (LNC) module in center-based graph construction. Finally, the difference between inliers and outliers is increased by dynamically adjusting the upper limitation distance of the outlier and suppressing its effect in the neighborhood. We conducted rich experiments on extensive datasets to verify the effectiveness of the proposed method. The experimental results illustrate that the proposed GLoCNet can effectively handle numerous outliers and achieve satisfactory registration results under various transformation patterns.