Research On High Resolution Remote Sensing Image Retrieval Technology Based On Deep Learning

Along with the rapid development of high-resolution remote sensing for earth observation technology,remote sensing image data volume exponential growth,how to go from vast amounts of quick browsing and efficient retrieval in remote sensing image data to the desired image,has become the bottleneck and difficult problem of remote sensing image information extraction and sharing,is the key to give play to effect of remote sensing data.Content based Image Retrieval（CBIR）framework is widely used in high-resolution remote sensing image retrieval.CBIR includes two core parts,image content is represented by features,and images with similar features are used as retrieval results.So,how to extract image features with low dimension,strong expression ability and good discrimination is the key of CBIR.In view of the weak feature expression of high-resolution remote sensing image,the dispersion of the earth’s surface objects,and the problem of weak information interaction between low-level convolution features and high-level semantic features of the Convolutional Neural Network（CNN）,CNN is used for feature learning,the method of feature extraction and dimension reduction of high-resolution remote sensing image is focused on,and the accuracy and fast retrieval of high-resolution remote sensing image is realized.The specific work is as follows:（1）Aiming at the weak feature expression of high-resolution remote sensing image,a high-resolution remote sensing image retrieval method based on Inception V4 deep feature coding and Large Vis dimension reduction is proposed.In this method,the convolution layer feature map output from the first Reduction Block is extracted based on the Inception V4 network.After reorganizing the data,VLAD（Vector Locally Aggregated Descriptors）is used to encode the data to obtain a compact feature representation vector.It is combined with the feature from full connection layer to form the image feature representation vector.In order to avoid the problem of "dimension disaster",the extended Large Vis algorithm is used to reduce the dimension of feature representation vector,and the feature representation with low dimension and strong distinguishing ability is obtained.Finally,L2 distance measurement method is used to compare the similarity of the reduced features to achieve high-resolution remote sensing image retrieval.Experimental results on RS19,UCM and RSSCN7 datasets show that the proposed method can achieve higher retrieval accuracy than the existing methods,the mAP reached 98.26%、97.57%and 91.34%,respectively.That’s an average increase of 6.24%over the then-best approach.（2）According to the characteristics of object dispersion of high-resolution remote sensing image,a high-resolution remote sensing image retrieval method based on disperse attention mechanism and ResNeSt is proposed.In this method,Disperse Attention（DA）is introduced into ResNeSt network,and then combined with DNN（Deep Neural Network）feature transformation network,a DA-ResNetSt network structure is builded,which is used as the backbone network to extract the deep feature of image.The network pays more attention to the information integration of global features,which can improve the expression and discrimination ability of remote sensing image deep features.Next,the extended Large Vis algorithm is used to reduce the dimension of deep features,and L2 distance metric is used to match the reduced features.Experimental results on RS19,UCM,RSSCN7 and AID datasets show that the proposed method can achieve higher retrieval accuracy than the existing methods,the mAP reached 99.99%、100%、97.54%and 96.58%,respectively.That’s an average increase of 0.84%over the then-best approach.（3）Aiming at the problem of weak information interaction between low-level convolution features and high-level semantic features of CNN,a high-resolution remote sensing image retrieval method based on ResNetSt multi feature fusion is proposed.In this method,ResNetSt is used as the backbone network.The ability of feature expression and discrimination is improved through the effective fusion of high-level semantic features and low-level convolution features.Then,a Multilayer Perceptron（MLP）is designed to map high-dimensional features to low dimensional features with strong discrimination,and L2 distance metric is used for similarity matching to realize high-resolution remote sensing image retrieval.Experimental results on RS19,UCM,RSSCN7 and AID datasets show that the proposed method can achieve higher retrieval accuracy than the existing methods,the mAP reached 99.72%、99.38%、97.42%和 95.42%,respectively.That’s an average increase of 0.30%over the then-best approach.