A robust and high-accuracy registration method for large format aerial images with adaptive space constraints

Geometric distortions and repetitive or weak textures lead to uneven geometric control across large-format aerial images, posing challenges to precise registration and fusion. To overcome these issues, we propose an adaptive spatially constrained registration and fusion framework. An improved Adaptive Adaptive Threshold RANSAC (At-RANSAC) integrates spatial geometric similarity (SGS) and neighborhood gray continuity (NGC) for robust mismatch elimination. The spatial distribution of matched points is optimized using Non-Maximal Suppression with Square Covering (NSSC) and Local Self-Similarity Correlation (LSCC) to ensure uniform geometric control. A multi-threaded TIN-based differential rectification achieves subpixel-accurate registration and fusion of PAN and multispectral images. Experiments on six aerial datasets show that our method reduces registration error from 1.1 to 0.3 pixels, improves efficiency nearly fivefold, and enhances image sharpness by over 2%, outperforming existing methods in accuracy, robustness, and computational efficiency.