Research On Remote Sensing Image Retrieval Method Based On Deep Learning

Remote sensing data has the characteristics of mass,complexity and diversity,which proposes a higher request of remote sensing image retrieval.Strategy for content-based remote sensing image retrieval is a research hotspot,and the method of feature extraction is the key point.Traditional feature extraction methods usually extract the low-level visual feature of remote sensing images,and has the drawback that low-level visual feature can’t represent the semantic information.This paper presents different kinds of methods based on deep learning for remote sensing image retrieval.By training neural network,mapping between low-level visual feature and semantic information is founded.And in this paper,two different deep learning methods are presented,which are described below:(1)This paper proposed a semi-supervised remote sensing image retrieval method based on deep learning,which can be divided into four steps: first,the remote sensing images are pretreated,using the method of ZCA Whitening.Second,the feature dictionary was extracted utilizing Sparse Autoencoder,an algorithm deals with data that has not been annotated.Then the feature dictionary is used in the image convolution process following the principle of Convolutional Neural Networks,and average pooling is conducted after convolution.Third,Softmax classifier is used to classify remote sensing images.At last,the remote sensing image retrieval is sorted based on the Euclidean distance between the query image and database in the same category as the query image,and the retrieval result is accomplished.It turns out that this method achieves good accuracy and efficiency.(2)This paper proposed another remote sensing image retrieval method based on Convolutional Neural Network(CNN),which is realized in a deep learning frame named keras.And this method can be divided into CNN feature extraction network and Softmax classifier network.This method intends to get high-level semantic feature through deep CNN network.At the same time,this method introduces the Dropout layer to improve the generalization ability of this model.Finally,the efficiency of remote sensing image retrieval is improved.In the same data sat,method(1)achieves the retrieval accuracy of 90.6%,and the retrieval time is 7.1844 s,method(2)achieves the retrieval accuracy of 98.8% and the retrieval time is 9.138 s.Drawback in this paper is as follows: when the query image is incorrectly classified,it achieves low retrieval accuracy.