Lunar Remote Sensing Image Matching under a Limited Computational Budget

Abstract

Lunar remote sensing image matching faces strict computational constraints in practical applications. This work investigates the trade-off among matching success rate, geometric accuracy, and computational cost under limited budgets. A unified evaluation framework is constructed using 60 multi-illumination Chang’e-3 image pairs. Input conditions, candidate matching, geometric verification, and success criteria are kept consistent throughout the experiments. The influence of RANSAC sampling limits and reprojection thresholds is first evaluated. The results indicate that increasing the sampling limit alone does not significantly improve the success rate. In contrast, the reprojection threshold directly affects the success rate, inlier ratio, and geometric error. ORB, SIFT, fixed-parameter ALIKE, and two adaptive variants are then compared within the same framework. On this basis, a budget-adaptive multi-scale matching method, termed BAMM-ALIKE, is proposed. The method allocates computational resources by jointly adjusting image scale and the number of keypoints. Feasible configurations are selected by using geometric verification feedback under different time budgets. Experimental results demonstrate that BAMM-ALIKE maintains stable performance across budget levels. It further achieves a more favourable balance between accuracy and efficiency, especially under medium-budget conditions, which supports robust image matching in resource-constrained lunar remote sensing tasks.