Remote Sensing Image Retrieval Based On Deep Feature

With the rapid development of aerospace and sensor technology,the earth obser-vation techniques grow up.which results in an increase of a large number of remote sensing images with good quality.It is a research hotspot and difficulty in the field of remote sensing image retrieval to design a feature extraction method and similarity mea-sure method that can achieve accurate and efficient retrieval based on characteristics of large remote sensing images.The existing approach of feature extraction is not good enough to represent the remote sensing images because of the characteristics of mass,complexity and diversity of the remote sensing images,which results in a semantic gap between the feature rep-resentation and high-level semantics.In addition,the feature dimension from existing approach is too high to decrease the performance of large-scale remote sensing im-age retrieval.To address the above issues,we conducted research on feature extraction and similarity measure method for remote sensing image retrieval.The specific research work is as follows.(1)Based on imaging characteristics of remote sensing images,considering the multi-angle imaging characteristics of remote sensing images,Rotation Invariance Fea-ture Transformer Network(RI-FT-NET)and Rotation Invariance Spatial Trans-former Network(RI-ST-NET)are proposed to extract the target rotation invariance features in the image.RI-FT-NET rotates the last convolutional layer at different angles and summarizes all loss to optimize the network parameters,aiming to make the same classification result for different rotation patterns.RI-ST-NET combines the Spatial Transformer Networks(STN)and is trained by means of Siamese net-work,which aims to make the image features of different rotation patterns of the same target more similar.The experimental results show that both RI-FT-NET and RI-ST-NET can describe the rotation invariant features of the target,and effectively improve the accuracy of remote sensing image retrieval.(2)We proposed Hybrid Triplet Remote Sensing Network(TRI-RS-NET)based on hard negative feedback.TRI-RS-NET is based on the triplet mixing loss,and considers the association information and category information of the image to extract more distinguishing features.At the same time,our method introduces the sample se-lection probability matrix,and selects the difficult sample pairs according to the probability matrix,thus improve the difficult sample selection effect from a global perspective.The experimental results show that TRI-RS-NET is superior to the tra-ditional triplet method and can effectively improve the accuracy of remote sensing image retrieval.(3)We proposed a similarity measure method base on Locality Sensitive Hashing(LSH)using the in-memory database.The feature index is established by LSH and the features extracted by the model are hashed by random binary projection.And we persist feature indexes through an in-memory database,which achieving efficient retrieval.Based on the above work,we construct a remote sensing image retrieval test system which visualizes retrieval results.Finally we test and evaluate the re-trieval performance on the simulation dataset.