A Secure Spectral Imagery Retrieval System Based On Deep Featrues

With the rapid development of remote-sensing earth observation technology,the number of spectral imagery has shown an exponential growth.Furthermore,the accelerating rise of artificial intelligence technology and high-performance computing has promoted the arrival of big data era for spectral imagery.Therefore,how to efficiently organize and manage large-scale spectral imagery has become a practical application problem to be solved urgently.As basic data for the major construction projects in digital earth,spectral imagery has been widely applied in the field of urban planning,environmental monitoring and military reconnaissance,etc.However,the openness and sharing of the network age have made the network information security increasingly prominent.Especially for spectral imagery containing important information,it should have strict confidentiality to ensure that no confidentiality event occurs during retrieval.Traditional image retrieval methods based on hand-crafted features are still difficult to form robust feature representations for spectral imagery as a result of the semantic gap between low-level features and high-level concepts,which restricts the further development of spectral imagery retrieval techniques.Moreover,the existing spectral imagery retrieval technology focuses on designing an effective mechanism to improve the retrieval performance as well as ignores the security of the image content.Therefore,it has important practical significance and extensive engineering application value how to ensure the security of large-scale spectral imagery with satisfied retrieval accuracy.In this thesis,a spectral imagery secure retrieval system based on deep features is designed and implemented,which jointly associate image retrieval with security.The mainly works in this thesis include:(1)Considering the powerful feature learning ability of deep learning technology,deep spectral-spatial features extraction method of spectral imagery is proposed.Firstly,the pure pixels of spectral imagery are selected to obtain spectral and spatial vectors.Then,the Deep Convolutional Generative Adversarial Networks(DCGAN)model is trained by combining spectral and spatial vector as training set.Finally,the deep spectral-spatial features are extracted with the trained DCGAN model.The experimental results show that the accuracy of image retrieval can be improved by using deep spectral-spatial features to represent spectral imagery.(2)For high-dimensional deep features,a dimensionality reduction method of deep spectral-spatial features is designed for spectral imagery.Firstly,the representative data are selected from deep spectral-spatial features in spectral imagery database by using fuzzy C-Means(FCM)method.Then,dimensionality reduction for representative data are made by using t-Distributed Stochastic Neighbor Embedding(tSNE)based nonlinear manifold method.Finally,dimensionality reduction of all spectral imagery are made by using the t-SNE based nonlinear manifold hashing method.The experimental results show that t-SNE based nonlinear manifold hashing method can effectively reduce the dimensionality of deep spectral-spatial features as well as greatly improve the retrieval efficiency owning to maintaining the feature representation ability.(3)To ensure the security in retrieval process,a spectral imagery secure retrieval system based on feature randomization encryption is designed and implemented.Firstly,the hash codes of deep spectral-spatial features after dimensionality reduction are encrypted by using the feature randomization encryption technology under preserving the Hamming distance.Then,the multi-index hash is utilized to compute the Hamming distance for similarity matching in the encryption domain.Finally,the retrieval results are optimized with relevance feedback mechanism based on the feature weight adjustment.The experimental results show that the proposed spectral imagery retrieval system can effectively protect the security of image content under proper image retrieval accuracy.