Research On Image Feature Extraction And Its Privacy Preserving

The rapid development of the mobile Internet has caused a surge in the number of images in people's lives.How to better manage and use these images becomes an important issue.As a key step in image processing tasks,image feature extraction can extract the main characteristics from redundant image information as the criteria for discrimination and analysis,so extracting features from images to mine and utilize important information contained in image data is an effective method to solve this problem.Besides,people are more willing to store a large number of images to the remote cloud server to solve the storage and computing burden.However,while enjoying convenience,storing image data containing a large amount of sensitive information directly to the cloud platform makes image data out of the direct physical control of the image owner.For the image owner,privacy and security have become the most important issues.In order to solve privacy and security issues,encryptionbefore-outsourcing is the fundamental solution to protect the image privacy,but this makes traditional image processing solutions extremely difficult.Specifically,it becomes almost impossible to extract features over encrypted image.In addition,in the case of storing image data on a third-party cloud platform,the image owner does not only use the image data stored in the cloud alone,but may also need to share the image data with other authorized parties.Therefore,it is a challenging problem to achieve secure image retrieval using image features while ensuring the security of image data and its features.Aiming at the above problems,this dissertation focuses on image feature extraction and its privacy preserving.The main contributions of this dissertation are as follows:1.In view of the problem of poor performance of existing quaternion moments in color image reconstruction and color object recognition,this dissertation proposes a new quaternion circularly orthogonal moment named as Quaternion weighted Spherical Bessel-Fourier Moment(QSBFM)and construct a set of invariants to rotation,scaling and flipping transformation.The relationship between QSBFM and the conventional weighted spherical Bessel-Fourier moment has been established so that the computational cost of QSBFM can be efficiently reduced.The performance of QSBFM and its invariant are widely studied.And this dissertation conducts experiments on popular color image to examine our theoretical analysis.This dissertation compares QSBFM with frequently-used quaternion moments,including quaternion Bessel-Fourier moment,quaternion radial Harmonic-Fourier moment,quaternion Chebyshev-Fourier moment and quaternion orthogonal Fourier-Mellin moment,with respect to color image reconstruction,color object recognition and computational complexity performance to provide useful information for potential users.2.In order to solve the privacy-preserving feature extraction problem of the circularly orthogonal moment,this dissertation first proposes a novel Legendre Circularly Orthogonal Moment(LCOM)defined in whole polar coordinates domain.In comparison to the typical circularly orthogonal moments,the performance of LCOM is more prominent as it does not have the geometric error.This dissertation further presents an effective scheme to achieve Privacy-preserving Legendre Circularly Orthogonal Moment(PLCOM)by combining the LCOM with somewhat homomorphic encryption to solve the problem of privacy-preserving LCOM feature extraction in cloud computing environment.Moreover,this dissertation provides the message space analysis about the expansion of LCOM plaintext data,so that the correct LCOM can be acquired after decrypting PLCOM.The formal security analysis proves that the PLCOM scheme can guarantee the image content security.Besides,the empirical results show that PLCOM can provide performance(in terms of image reconstruction capability and image recognition accuracy)close to directly calculating LCOM in the plaintext domain when the parameters are correctly set.3.In order to solve the privacy-preserving feature extraction problem of discrete orthogonal moment,this dissertation takes Krawtchouk orthogonal moment as an example to present an effective scheme to achieve Privacy-preserving Krawtchouk Moment(PPKM)by utilizing Paillier cryptosystem.In addition,this dissertation provides an upper bound analysis about Krawtchouk moment's plaintext data expansion so that the correct Krawtchouk moment can be obtained by decrypting PPKM.Furthermore,this dissertation proposes a block-based parallel algorithm to reduce the computation complexity of PPKM,and the proposed parallel algorithm has excellent acceleration effect and high scalability.Sufficient security analysis shows PPKM can guarantee the image content security.Besides,the experimental results show that PPKM has almost the same performance as plaintext Krawtchouk moment in terms of image reconstruction capability and image recognition accuracy when the parameters are set correctly.4.In order to solve the problems of key sharing caused by single key mechanism,low retrieval precision and untraceability to malicious users in encrypted image retrieval schemes,this dissertation proposes a traceable encrypted image retrieval in the multi-user setting using VGG16 convolutional neural network,called MU-TEIR.Compared with the traditional single-key mechanism that incurs the problem of key-sharing with users,MU-TEIR uses the distributed two trapdoors public-key cryptosystem technology to provide a multi-key mechanism,which settles the authorization problem in the multi-key environment.Secondly,compared with the traditional invalid feature extraction method that suffers from high image retrieval overhead and poor precision,MU-TEIR utilizes the lower bound on the squared Euclidean distance technology and VGG16 convolutional neural network to extract image features to achieve higher retrieval efficiency and precision,respectively.Furthermore,compared with the traditional schemes that result in an inability to punish malicious users publishing pirated images,MU-TEIR ingeniously makes use of an encrypted image watermarking method that enables tracking of malicious users.