Abstract
In this article, we propose an effective method for remote sensing image registration. Point features are robust to remote sensing images with low quality, small overlapping area, and local deformation. Therefore, we extract point features from remote sensing images and convert the problem of remote sensing image registration into the problem of feature point matching. A correspondence set constructed solely on the similarity of features often contains many false correspondences or outliers, so our key idea is to remove the mismatches in the initial correspondence set and obtain a stable correspondence through a two-step strategy. First, we use two constraints to construct the optimization model which can be solved in linear time. The first constraint is that the topology of the points and their neighbors can be maintained after the spatial transformation. Another constraint is that the feature distance of the correct matches are similar to the neighbors. Then, we design a strategy to increase the number of inliers and raise the precision by a global constraint calculated from the solution in the previous step. Experiments on a variety of remote sensing image datasets demonstrate that our method is more robust and accurate than state-of-the-art methods.
Introduction
Image registration—the process of aligning two or more images of the same scene—is a cornerstone of remote sensing tasks like change detection and image fusion. However, remote sensing data is notoriously difficult to align due to differences in sensors, viewpoints, and lighting, as well as local distortions caused by terrain.
While algorithms like SIFT (Scale-Invariant Feature Transform) are effective at finding potential matching points, the initial "correspondence set" they produce is often riddled with outliers (false matches). Traditional methods to clean this data, such as RANSAC, can become computationally expensive or fail entirely when the ratio of outliers is high. This research introduces a method that is both highly accurate and computationally efficient, operating with linear time complexity to clean these complex datasets.
Methodology: A Two-Step Framework
The proposed framework eliminates mismatches through a sequential application of local and global constraints.
Step 1: Local Neighborhood Constraints
The algorithm first filters the initial correspondence set by examining the local neighborhood of each point. This step relies on two primary assumptions:
- Topology Preservation: If a match is correct, its neighbors in the first image should remain its neighbors in the second image after transformation.
- Feature Similarity: The feature distances between a point and its neighbors should remain consistent across images.
By combining these into a mathematical optimization model, the algorithm identifies a "cleaner" set of matches in linear time ($O(N \log N)$), significantly improving upon the existing Locality-Preserving Matching (LPM) framework.
Step 2: Global Constraints
Using the filtered matches from Step 1, the algorithm calculates a global affine transformation matrix. This matrix acts as a geometric guide: any point from the original set that aligns well with this global model is re-integrated as an "inlier," while remaining outliers are discarded. This step boosts the final count of correct matches (NCM) and ensures global consistency.
Datasets
The method was evaluated on 102 pairs of images across four diverse datasets:
- CIAP: 30 pairs of orthorectified images (700x700 pixels) for image mosaicking.
- UAV: 22 pairs of drone-captured images with very small overlapping areas.
- SAR (Synthetic Aperture Radar): 10 pairs of satellite images corrupted with heavy speckle noise.
- SUIRD: 40 pairs featuring complex perspective variations and severe outliers.
Key Results and Analysis
The proposed method consistently outperformed state-of-the-art algorithms like RANSAC, LPM, and GLPM across all test cases. In the CIAP and UAV datasets, the method achieved near 100% precision and recall, effectively finding almost all correct matches while eliminating every outlier.
On the SAR dataset, where speckle noise makes feature matching extremely difficult, the method maintained high accuracy where others faltered. The synergy between local and global constraints allows the model to handle local distortions while maintaining overall geometric integrity.
Conclusion and Future Work
This two-step strategy provides a robust and efficient solution for remote sensing image registration. By using local constraints to "prime" a global affine model, the method achieves high precision even in high-outlier scenarios.
One primary limitation remains: Non-rigid transformations. The second step relies on an affine transformation, which may struggle with complex, non-rigid deformations where terrain warps inconsistently. Future work will focus on developing more flexible global transformation models to handle these complex cases while maintaining linear efficiency.
Summary updated