Learning Adaptability Analysis for Remote Sensing Image Matching via Multi-Scale Saliency and Uniqueness Feature Fusion

Image matching navigation technology, known for its exceptional precision and anti-interference capabilities, has been widely applied in aircraft navigation. Adaptability analysis, a core component of this technology, is critical for enhancing matching performance. Existing deep learning-based adaptability analysis methods lack feature extraction networks specifically designed for adaptability analysis, which hinders their ability to fully capture the saliency and uniqueness features within adaptability regions. This limitation consequently restricts their performance in the adaptability analysis of remote sensing images. A novel remote sensing image adaptability network architecture named US-SwinT is proposed. The network consists of two core modules: saliency analysis and uniqueness analysis, designed to effectively utilize regional scene perception to accurately predict matchability of reference images. The multi-scale saliency analysis module adaptively fuses global contextual information with local detail features through a dynamic weighting mechanism, enhancing the model's ability to discern key regions. The uniqueness module employs cosine similarity to analyze feature similarity between regions, effectively assessing redundancy of regional features. By suppressing highly similar regional features, it enhances capture of regional differences. Experimental results demonstrate that US-SwinT outperforms traditional methods and other deep learning models across multiple key metrics, significantly improving accuracy and robustness of remote sensing image adaptability analysis.