Automatic fast feature-level image registration for high-resolution remote sensing images

Abstract

The size and amount of remote sensing images constantly increase with the improving resolution of remote sensing images. Meanwhile, the development of remote sensing applications also requires high image registration performance. Therefore, an automatic fast feature-level image registration method for high-resolution remote sensing images is proposed. The method includes five steps. First, the reference image and the image to be registered are processed by Haar wavelet transform to obtain the low-frequency approximate images to match. Then, the original images are registered according to the matching result of the approximate images, thereby potentially effectively reducing calculation and improving registration speed. Second, edges in the optical image are extracted by the Canny operator, and edges in the SAR image are extracted by the Ratio Of Averages (ROA) operator. Then, the edge line features are transformed into point features. The use of edge point features can achieve positioning accuracy and acquire stable features. Third, in the feature matching session, the main and auxiliary directions of the point features are considered such that each point feature has multiple directions to enhance the robustness of image registration. Then, the initial matching points are determined by the ratio of the minimum angle to the second minimum angle, which is less than a threshold. Fourth, in the matching point pair filtering session, the random sample consensus is enhanced to improve registration accuracy by adding the constraint condition. The high-quality matching point pairs are selected to fit the model parameters. Finally, in the affine session, the block thought is used to uniformly select matching point pairs to be evenly distributed in the image to avoid the local optimal problem on the registration and further improve image registration accuracy. To verify the efficiency of the method, experiments are conducted under the following conditions: the same sensor optical image registration and sensor SAR image registration, image registration among different bands, image registration with different resolutions, and image registration of different satellite sensors with large size. High resolution WorldView-2, Pleiades, and TerraSAR images are used to perform the experiments. The proposed method is compared with the typical SIFT and SURF algorithms. Four quantitative evaluation indexes, namely, Matching Ratio (MR), Matching Efficiency (ME), Root Mean Square Error (RMSE), and time consumed are used for the registration result evaluation. Experimental results show that the proposed method effectively achieves high registration accuracy under the different conditions. An automatic fast feature-level image registration method for high-resolution remote sensing images is proposed. Multiple datasets of registration experimental results for high-resolution remote sensing images indicate that the proposed method can be rapidly implemented and has high accuracy and strong robustness.