Research On Image Geo-localization Based On Cross-view Matching

The task of image geo-localization based on cross-view matching aims to determine the geo-location(GPS coordinates)of a query ground-view image by matching it with the GPStagged aerial(satellite)images in a reference dataset.However,due to the dramatic changes in the viewpoints between ground and aerial images,there exist huge difference between groundview and aerial-view images,which makes matching the cross-view images very challenging.This paper summarizes several research hotspots of cross-view image matching at the current and studies the representation,distance measurement and perspective transformation of crossperspective images.The main research work is as follows:1)Cross-view image geo-localization algorithm based on representation learningIn this paper,we propose an end-to-end network architecture,namely GeoNet,for crossview image-based geo-localization,by incorporating the multiple convolutional layers and the capsule layers to capture the feature representation and their relationships,which can further model the spatial feature hierarchies and enhance the representation power.Moreover,we also introduce a simple and effective weighted soft-margin loss with online batch hard sample mining,which cannot only avoid the selection of the parameter margin when using margin triplet and quadruple loss,but also improve the image retrieval accuracy.Extensive experiments on two publicly available datasets demonstrate that our GeoNet significantly outperforms the state-of-the-art approaches for cross-view image geo-localization.2)Cross-view image geo-localization algorithm based on perspective transformationIn this paper,we also propose a multi-stage learning algorithm to solve the problem of cross-view image geo-localization,i.e.,the ground-view query images are converted into aerialview images,and then similarity learning is carried out from the aerial-view images.In this paper,pix2pix-gan network is firstly used and improved to realize the conversion from aerial(satellite)images to ground images.The ground-view query image is converted into an aerialview image by the GAN,so that the query image and the reference dataset image belong to the same view.This makes cross-view problem into the same perspective problem,reducing the complexity of cross-view matching.Next,we design a Siamese architecture double branch network and adopt weighted soft-margin triplet loss function to learn similarity.It makes that the matching image pairs are close and unmatched image pairs are far apart.Experiments on two public data sets demonstrate the effectiveness and advancement of the proposed method.