Cross-view Geographic Image Retrieval Based On Information Bottleneck

With the rapid development of science and technology,image data is exploding,and therefore the task of image retrieval has become important.Cross-domain image retrieval is a difficult area of research,and researchers have combined their efforts to promote the development of this field.Cross-view geographic image retrieval is one of the research hotspots in this field.Cross-view geographic image retrieval is to search and match the same geographic targets in images from multiple different perspectives,and it has a wide range of applications,such as drone navigation and assisting GPS to achieve precise positioning.Cross-view geographic image retrieval is an extremely challenging task because the change of viewpoint can cause a huge change in appearance,which can lead to a huge domain gap between them.The key to this task lies in learning discriminative features with viewpoint invariance.Most of the present work utilizes deep neural networks to extract and learn image features while incorporating attention mechanisms and orientation information into the models to improve retrieval performance.Although these methods have made some progress in cross-view geographic image retrieval tasks,they still have the following problems:(1)previous methods directly use the extracted features as retrieval features after metric learning,and they ignore the negative impact of redundant information in the extracted features on the retrieval performance?(2)as the retrieval features contain too much redundant information,it makes the dimensionality of the retrieval features large.The high-dimensional feature vector leads to the problems of difficulty in similarity measures and space complexity in retrieval.To address these two problems,two solutions are proposed in this thesis as follows.(1)To address the problem of how to remove redundant information from features,this thesis proposes a cross-view geographic image retrieval method based on a variational distillation of information bottleneck.Based on the information bottleneck theory,this method uses a variational self-distillation strategy to compress the features to remove redundant information.It ensures that the prediction information is not lost during the feature compression process,and only the task-irrelevant redundant information is discarded.Finally,discriminative features with viewpoint invariance are obtained as retrieval features,which effectively improve the metrics of cross-view geographic image retrieval.(2)To obtain the minimum and sufficient representation of retrieval features,this thesis also proposes a cross-view geographic image retrieval method.The method is based on information bottleneck,aiming to learn a sufficient and low-dimensional feature as retrieval features.The method first obtains an adequate representation of the image by filtering out the noise in the image features through a variational information bottleneck module and retaining only the most relevant information to the task.It then reduces the dimensionality of the features and discards the redundant information contained in them by the feature compression module.Finally,sufficient and low-dimensional discriminative features with viewpoint invariance are obtained as retrieval features,which significantly improve the retrieval performance.In summary,the proposed methods in this thesis have been richly experimented on the publicly available University-1652 dataset and CVACT dataset,respectively,and the results are compared with other state-of-the-art methods.The experimental results show that the proposed methods in this thesis achieve competitive results on several metrics.