UAV-Satellite Cross-View Image Matching Based on Siamese Network

In recent years, UAVs have played an important role in the military and civilian fields, and are widely used in intelligence reconnaissance, fire strikes, urban management, video shooting, and other fields. However, UAVs have a limited field of view and sometimes need to be paired with satellites to determine their geographic location. Therefore, this paper uses the matching of images taken by drones and satellites to solve the problem of geo-localization. Based on the Siamese neural network, the shared weight convolutional neural network is applied to extract the feature vectors of the UAV image and the satellite image respectively. After measuring the distance between the two feature vectors, the matching loss is calculated to train the matching model. This paper compares the classification performance of various loss functions and introduces a weight factor to solve the problem of unbalanced positive and negative samples. Besides, a hard example mining strategy is applied to improve the efficiency of model training. Experiments are applied with the dataset containing 127,064 UAV images and 2,353 satellite images. A variety of evaluation metrics are used to evaluate the performance of the proposed model and show competitive matching results.