Research On Remote Sensing Image Retrieval Based On Graph Convolutional Neural Network

With the enhancement of remote sensing data acquisition means,the amount of remote sensing information that needs to be processed has increased dramatically,and how to improve the accuracy and retrieval efficiency of remote sensing image retrieval has become a key issue.The traditional management methods based on manual classification or annotation are inefficient,time-consuming and costly,which are difficult to adapt to the processing needs of massive data.Convolutional neural network has powerful feature extraction ability and has been widely used in the field of remote sensing image retrieval,but convolutional neural network can only deal with the information of image itself,ignoring the semantic relationship between images,while graph convolutional neural network can better deal with the semantic relationship between images.Therefore,this paper investigates the remote sensing image retrieval method based on graph convolutional neural network,and the specific work is as follows:(1)For the problem that convolutional neural networks ignore the semantic relationships between images,this paper proposes a remote sensing image retrieval feature extraction method based on GraphSAGE(Graph Sample and Aggregate).The method proposes a neighbor node sampling method based on the node attention mechanism of node similarity,and the sampled neighbor node vectors are aggregated by a weighted aggregation function to generate a new embedding representation of graph nodes.In addition,the Image Rank Similarity(IRS)method is proposed,which combines node connection similarity with node feature similarity.Finally,a category weighting method is proposed to further improve the retrieval accuracy.The experimental results on the UCMD(UC Merced)dataset show that the mean Average Precision(mAP)of this method reaches 96.53%,which has good retrieval performance.(2)In order to improve the accuracy of node graph structure and further enhance the retrieval speed,we first propose a method that fuses the GraphSAGE model and IDGL(Iterative Deep Graph Learning)model.The inaccurate graph structure is continuously optimized using the IDGL model to obtain a better graph structure in order to improve the effectiveness of the embedding representation of nodes.Secondly,by establishing a positive and negative sample library,a suitable loss function is constructed to further improve the effectiveness of node embedding representation.Finally,in order to improve the efficiency of remote sensing image retrieval,K-D Tree is used to reduce the time spent on constructing the graph structure in order to improve the retrieval speed.The experimental results on UCMD and PatternNet datasets show that the method in this paper can improve the retrieval speed by 8%,and the average search accuracy rate reaches 96.59%,which is a significant improvement compared with other methods.