Research On Benchmark And Algorithm Of Image Perceptual Hashing

With the rapid expansion of digital media in the quantity and innovative applica-tions, a number of new challenges have been introduced to the traditional managementand application modes of digital media. Perceptual hashing is one of prominent new tech-niques that may conquer these challenges. It provides a more ?exible, effective and safersolution for the identification and authentication of media content by mapping the percep-tual information of the content to a brief have value and matching based on the hash value.However, perception hash is still far from mature since there are still many fundamentalproblems to be solved, such as its essential properties.The most essential property of perceptual hashing is robustness. According robust-ness, it is required to compute the hash value base on perceptual content of image, ratherthan the binary data. However, the existing perceptual hashing algorithms hardly con-sider the subjective visual factors. Therefore, the robustness of such perceptual hashingalgorithm deviates the"perception", the original intention. In addition, there is no rea-sonable basis of performance comparison for the same reason. The digest, one-way andanti-collision features of the perceptual hashing are enough to meet the security require-ments of many application environments. But due to the insufficient understanding onrelation between goals of security applications and performances requirements of percep-tual hashing algorithms, the efforts of improving randomicity of hash value cannot meetthe expectation.Aiming to these issues above, firstly, this thesis summarizes the origin, develop-ment, application and state-of-art of perceptual hashing. The application models andperformance requirements of each model are summed up. The evaluation methods ofalgorithms are also reviewed, and their shortcomings are pointed out. Subsequently, thethesis studies the testing and evaluation benchmark of perceptual hashing algorithms,Human Visual System based perceptual hashing algorithms and base on structural infor-mation based perceptual hashing algorithms. The main contributions of the thesis aresummarized as follows:The concept of media space division by perceptual hashing is proposed. It is clarifiedthat the core objective of perceptual hash algorithm is to determine the perceptual or physical boundary in the media space, and provide perceptual distance metric within theboundary.The benchmark of robustness and discriminability based on subjective identifica-tion experiment is proposed. The corresponding methods of performance testing andcomparing is also presented. The empirical data about the subjective feelings to dis-torted/degraded images are obtained by a elaborately designed subjective identificationexperiment.Aiming to approaching the perceptual boundary, four perceptual hashing algorithmsbased on Human Visual System are proposed. The cyclic coding based perceptual hash-ing algorithm studies the functions of different perceptual channels in perceptual featureextraction. The other 3 algorithms are based on perceptual error measurement, weight-ing by constrast masking threshold, and multi-channel decomposition and error pooling,respectively. The benchmark testing results demonstrate some value of research achieve-ment of HVS in desgining perceputual hashing algorithm, and reveal the limitations ofexisting HVS models and the problems when applying them in perceptual hashing.Aiming to approaching the the physical boundary, a perceptual hashing algorithmbased on structure information coding is proposed. Its robustness and discriminability isfar better than existing algorithms with lower computational complexity.A secure perceptual hashing algorithm based on higher-order luminance differenceand random confusion operation is proposed. This method not only improves the ro-bustness and discriminability, but also improves the randomness of the hash value sig-nificantly. Through the theoretical analysis and experiments on the diffusion and con-fusion capability of several algorithms, it is clarified that there are misunderstanding inthe existing research on the security of perceptual hashing. When concerning about theauthentication of integral image and security applications of key identification, securityrequirement can be met based on discriminability of algorithms and randomicity of hashvalue. However, it is not sufficient to support the authentication of local content only relyon randomicity of hash value.To overcome the contradiction between robustness of algorithm, a tampering detec-tion and localizing algorithm based on perceptual hashing is proposed. The sensitivity ofthe local tampering can be achieved by pursuing the robustness and discriminability oflocal hash.