Remote Sensing Image Retrieval Based On Convolutional Neural Network

With the development of remote sensing technology,the number of remote sensing images shows the trend of massive growth.How to retrieve the interested image quickly in the big-scale remote sensing image database has become a research hotspot.The extensive application of deep learning in recent years has made remote sensing image retrieval more and more mature,especially in the use of convolutional neural network to extract image features,and the performance is relatively excellent.However,the image features based on convolution neural network have different characteristics.Specifically,image features from the convolution layers contain rich details,while image features from the fully connected layer contain abstract semantic information.Therefore,how to use the advantages of both features to extract more effective feature representation is the research focus of remote sensing image retrieval based on convolutional neural network.In addition,as remote sensing images contain many ground objects and complex scenes,single scene labels are difficult to describe the complex images.Therefore,how to use a variety of ground objects in remote sensing images for multi-label remote sensing image retrieval is another problem that needs to be studied.In view of the above problems,this paper studies the remote sensing image retrieval methods based on convolutional neural network,and proposes multi-layer feature integration retrieval method and multi-label remote sensing image retrieval method.The main research contents can be listed as follows:(1)This paper makes a general summary and analysis of current remote sensing image retrieval techniques,mainly focusing on difficulties in remote sensing image retrieval based on convolutional neural network and the corresponding solutions.(2)Aiming at the problem of different features extracted from the convolutional layers and the fully connected layers in convolutional neural network contain different image information and single layer feature expresses limited image content,a remote sensing image retrieval method based on multi-layer feature integration of convolutional neural network is proposed.The method first extract the features from the convolutional layers and the fully connected layers of the pre-trained convolution neural network layer,and features from the convolutional layers are as local features,then according to the mapping relationship between the convolutional layers and the fully connected layers,which means the change of the receptive field area,we integrate the two kinds of features and extract more powerful feature which contains local and global information.The experimental results show that the multi-layer feature combination method improves the retrieval accuracy comparing with methods of the convolutional layers or the fully connected layers singly or conventional methods.(3)Aiming at the problem that single label retrieval method is difficult to accurately describe and distinguish remote sensing image with complex scenes,a multi-labels remote sensing image retrieval method based on convolutional neural network is proposed.For the existing public multi-labels remote sensing image datasets are relatively lack,which limits the research of remote sensing image retrieval,so we build a dense labeling remote sensing dataset(DLRSD,dense labeling remote sensing dataset)and test,which is suitable for image retrieval based on pixels.We extract features of different regions for multi-labels remote sensing image retrieval,and compare with single label remote sensing image retrieval.We also compare the result of features from convolution neural network with the conventional manual design features at the same time.The experimental results show that multi-labels remote sensing image retrieval has better retrieval accuracy to some extent,which also indicates that the retrieval method of multi-labels feature is able to narrow down the semantic gap between low-level features and high-level semantic concepts present in remote sensing images to some extent and further improve the retrieval performance.The innovations of the paper are as follows:(1)The remote sensing image retrieval method based on multi-layer feature integration of convolutional neural network is proposed,which integrates the features of the fully connection layer and the convolutional layer to extract more effective feature representation.(2)The dense labeling remote sensing dataset built makes up for the problem that there are few such open data sets at present.The multi-label remote sensing image retrieval method based on convolutional neural network is proposed,which takes into account the information of various objects in the complex scenes of remote sensing images and further improves the retrieval accuracy.