A Subpixel Accurate Image-Matching Method Integrating SuperGlue and Local Optimization

Abstract

Subpixel accurate image matching is critical for high-precision applications across diverse scenarios, including multi-temporal satellite image registration, underwater inspection, and rigorous photogrammetric processing. Deep learning–based matching methods demonstrate strong robustness. However, applying them to out–of–dataset scenes makes it difficult to achieve subpixel accuracy. They require constructing a new scene-specific dataset and retraining the model, which demands significant human and material resources. Therefore, this paper presents a coarse-to-fine matching strategy that requires no additional scene-specific training data or fine-tuning. SuperGlue first establishes reliable initial correspondences under challenging conditions, which are then refined to subpixel accuracy through iterative gray-level least squares optimization with an adaptive step-size strategy and geometric constraints. This hybrid strategy retains the adaptability of deep learning to diverse imaging conditions while preserving the mathematical rigor of photogrammetric optimization. Experimental validation across four distinct application scenarios demonstrates the effectiveness and cross-scenario applicability of the proposed strategy. On remote sensing imagery, the proposed method achieves a mean root mean square error (RMSE) of 0.83 pixels under the dataset–specific window configuration, compared with 1.36 pixels obtained by SuperGlue. For underwater imagery, RMSE is reduced from 1.30 to 0.90 pixels. On indoor ScanNet, pose estimation accuracy improves consistently across all thresholds. In addition, an image-stitching application experiment shows that the refined correspondences improve geometric alignment and reduce black boundary gaps in stitched images. These results indicate that the proposed hybrid refinement strategy provides practical utility for precision-demanding photogrammetric and remote sensing applications.