Research On Remote Sensing Image Retrieval Method Based On Multi-scale Self-attention Feature Fusion

Remote sensing image retrieval is one of the key research directions in the field of remote sensing.At present,there are still many problems in remote sensing image retrieval tasks that need to be solved,such as different target directions,large changes in target size,complex target background,and large differences between similar targets.Remote sensing image retrieval requires efficient image feature extraction technology.Therefore,in response to the above problems,the specific work of this paper includes the following two parts:(1)Aiming at the problems of large changes in target scale,different target directions,and complex backgrounds in remote sensing images,this paper proposes a multi-scale selfattention feature fusion remote sensing image retrieval algorithm.This method uses variable convolution to solve the problem of target angle transformation and enhance the ability of the network model to adapt to geometric transformation.The self-attention mechanism can capture long-distance contextual information,and can extract features of salient regions in complex backgrounds,and improve the model’s robustness under complex remote sensing image backgrounds.At the same time,in the case of large changes in the target scale,multiscale self-attention features are integrated to enhance the expressive ability of features and improve the accuracy of remote sensing image retrieval.In the experiments of this algorithm on the three classic data sets of UCM,SATREM,and NWPU,the average retrieval accuracy of remote sensing image retrieval can be improved by 1%,which shows the effectiveness of the two modules.(2)The large difference between the targets of the same type will cause the problem that the extracted remote sensing image features cannot accurately reflect their true category information.This paper proposes a remote sensing image feature metric learning algorithm based on multiple similarities.By comprehensively considering the three similarities between features(self-similarity,positive similarity and negative similarity),it effectively reduces the distance of similar features and increases The distance of different types of features enhances the ability to distinguish features,so that features can accurately display their true category information,thereby improving the accuracy of remote sensing image retrieval.In the experiments on the three classic data sets of UCM,SATREM,and NWPU,the method using the multiple similarity loss function has higher average retrieval accuracy than the other two methods(comparative loss function and triple loss function).It can be seen from the feature distribution graph that the multiple similarity loss function can greatly enhance the discriminative ability of features.At the same time,this paper also proposes a non-metric similarity network for remote sensing image retrieval to replace the traditional similarity measurement method.This method also improves the average retrieval accuracy of remote sensing image retrieval.The above two algorithms are combined with the multi-scale selfattention feature fusion remote sensing image retrieval algorithm.The average retrieval accuracy on the three classic data sets is higher than the existing remote sensing retrieval algorithms with better performance,both are above 90%.Especially on the NWPU data set with a large number of remote sensing images,the performance of the method proposed in this paper far exceeds the existing methods,showing the superiority of the method proposed in this paper on large-scale data sets,and is suitable for practical application scenarios.Demand.In summary,experiments show that the algorithm proposed in this paper can well solve the problems of remote sensing image retrieval,enhance the generalization of network models,improve the performance of remote sensing retrieval and retrieval,and meet the needs of massive remote sensing image data in practical applications.High performance requirements for retrieval tasks.