A Lightweight Network-Based Approach for Cross-View Drone Image Matching

Aiming at the existing problem of cross view drone image matching, that is, the amount of parameters is too large when using the large backbone network (ResNet-50) to extract features, and it is not easy to deploy on embedded devices, this paper proposed a lightweight network cross view drone image matching method. This method introduced the multi head attention mechanism, which enables LNet to effectively capture the interaction between features. At the same time, the triple loss function is introduced to ensure that the model can effectively distinguish the feature representation of positive and negative samples. Through the verification experiment on the dataset university-1652, the LNet network contains only 16.8M parameters. Experimental results show that the recall rate of this method in drone positioning task (Recall@K, R@K) and average precision (AP) reached 87.02% and 88.93%; The recall rate and average accuracy of drone navigation mission reached 89.91% and 84.79%.