Research On Deep Semantic Retrieval Of High-Resolution Optical Remote Sensing Images

Recent years,launch of a series of satellites has provided massive amount of data for image retrieval.Large-scale image data sets have three main features e.g.,large data volume,high feature dimension and short response time.Therefore,how to achieve quick and efficient image retrieval has become an increasingly challenging problem.The hashing method is one of the key technologies to solve these problems.Because the hashing method has the advantages of simple structure,efficient retrieval,low space cost,simplicity to expand,and not affected by curse of dimensionality,it has become an important technology for large-scale image retrieval.his paper combines the idea of image hashing and deep learning to study the retrieval of large-scale remote sensing images.The main work of this paper is as follows:(1)The basic theories and methods involved in image retrieval,hashing algorithm and deep learning are introduced in detail,including the development history of image retrieval,the basic concepts and processes of hashing algorithm,the description of image features,the method of similarity measurement,Classification of hashing algorithms,convolutional neural networks and back propagation algorithms.(2)This study proposed an image retrieval method based on deep semantic hashing for mining semantic information of remote sensing images,which is called deep semantic hashing(DSH),for RS images with tags or other semantic annotations.The contribution of this paper consists of introducing the hashing methods for RS images which encodes the high-dimensional image feature vector to binary bits by exploiting limited number of labeled(annotated)images.Furthermore,DSH directly learns the discrete hash codes without relaxation which deteriorates the accuracy of the learned hash codes.Hence,DSH provides high time-efficient(in terms of both storage and speed)and accurate search capability within huge data archives.Experiments carried out on our RS dataset demonstrate the effectiveness of the proposed DSH methods.We have conducted experiments on three different archives.On GF-2 satellite and Google Earth remote sensing image dataset,when the hash bit is 64,the mAP value can be improved by about 2 % unlike DPSH.On the CIFAR-10 dataset,the proposed method attains the improvements by 6% to 7% than DPSH for the mAP evaluation when hash bit is 64.On the FLICKR-25 K dataset,the proposed method attains the improvements by about 0.6% than DPSH for the mAP evaluation when hash bit is 64.(3)Due to the low storage cost and fast query speed,the hashing retrieval algorithm has been widely used for large-scale image retrieval.Aiming at the inefficiency of large-scale remote sensing image dataset training,we proposed a query point-oriented hashing retrieval method.First,the image features are extracted from the remote sensing image data training set with multiple semantic tags by using the deep convolution network.Then,the hashing function is learned from the query points and the hashing codes of the query points are generated by using the learned hashing function,and finally the binary hashing codes of the whole image database are obtained through iterative learning,which is helpful to improve the retrieval accuracy.The feature extraction of the entire database is avoided in the process of image retrieval,and thus the supervised information in the large-scale database is more effectively utilized for image retrieval.We have conducted experiments on three different archives.On GF-2 satellite and Google Earth remote sensing image dataset,when the hash bit is 64,the mAP value can be improved by about 11% unlike DPSH.On the CIFAR-10 dataset,the proposed method attains the improvements by 15% than DPSH for the mAP evaluation when hash bit is 64.On the FLICKR-25 K dataset,the proposed method attains the improvements by about 7% than DPSH for the mAP evaluation when hash bit is 64.