Frequency-Enhanced Network for cross-view geolocalization

Abstract

Cross-view geolocalization is a crucial task that aims to match remote sensing images captured from different perspectives within the same scene, particularly important for applications involving drone and satellite imagery, where varying viewpoints can significantly impact analysis accuracy. To address these challenges, we propose FENet (Frequency-Enhanced Network), a novel multiple frequency scale framework that integrates frequency decomposition with adaptive feature fusion for hierarchical feature extraction. FENet uniquely decomposes images into different frequency components, enabling the extraction of fine-grained local details and global structural information at various scales, enhancing the network’s ability to capture complementary features. Leveraging the robust visual encoding capabilities of the CLIP model without modifying its structure, FENet ensures adaptability and efficiency across diverse geolocalization tasks. Its adaptive feature fusion module dynamically integrates features from multiple frequency components, while the InfoNCE loss function supervises cross-scale feature alignment, ensuring consistent and precise matching between views. Extensive experiments on SUES-200, University-1652, and DenseUAV datasets show that FENet outperforms the state-of-the-art methods in cross-view image matching.