UAV-to-satellite image geo-localization via dual-branch deep learning: a comparative study of ResNet-18 and MLP-Mixer

Abstract

Unmanned Aerial Vehicle (UAV)-satellite cross-view geolocalization is a core technology for autonomous navigation in GNSS-denied environments, which achieves precise pose estimation by matching low-altitude oblique UAV images with high-altitude orthophoto satellite images. However, the huge visual differences in perspective, scale and distortion between the two types of images significantly increase the difficulty of feature matching. Current mainstream methods are mostly based on CNN or Transformer architectures, while systematic comparative studies of pure MLP architectures in this task are still insufficient, and their performance differences and efficiency trade-offs remain unclear. To this end, this paper compares and analyzes the accuracy, robustness and computational efficiency of CNN and pure MLP architectures in UAV-satellite cross-view geolocalization on a unified benchmark dataset. The experimental results show that CNN has advantages in local feature extraction and is more robust to perspective distortion, while pure MLP architectures show potential in global feature modeling and inference speed, but their performance needs to be improved in small-sample and extreme perspective change scenarios. The comparative analysis in this paper provides an empirical basis for the architecture selection and optimization of cross-view geolocalization models.