High-Resolution Remote Sensing Image Registration Leveraging Locally Invariant Feature Points

In order to solve the problem of low registration accuracy in traditional point feature matching algorithms, this paper proposes a high-resolution remote sensing image registration algorithm based on locally invariant feature points. Firstly, a bilateral filter is used to preprocess the reference image and the image to be registered separately, preserving their edge information and removing noise. Secondly, extract local invariant feature points from the preprocessed two images. Then, bidirectional matching is performed using the Euclidean distance ratio method of nearest and second nearest neighbors to obtain matching feature points[1]. Finally, perform affine transformation on the registered images. To validate the method's effectiveness, we propose the affine transformation root mean square error (AT-RMSE) as an evaluation metric for accuracy. PLEIADES images from two time phases in Shenzhen are selected, and the results demonstrate that although the bilateral filter increases the running time of the algorithm, it can reduce AT-RMSE by more than 55%; In addition, registration methods based on SURF features have better performance and significantly improved accuracy compared to SIFT based methods.