Research On HVS Model Based Image Perceptual Hashing

Images, as an essential carrier of vision information, are indispensable in the process of information transfer and knowledge accumulation. However, many major fields such as public security have been suffered threaten from the security of images brought by the development of digitial image processing techniques, which made image authentication techniques to an urgently issue to be solved.The key issues of the security of images lie on the authenticity and integrity during the use. These requirments are no longer satisfied by the conventional methods of authentication and encryption. Perceptual hashing also refers as robust hashing and fingerprinting, maps digital multimedia data into a compact digital digest based on their perceptual contents. Multimedia with different contents would be mapped into distinct hashing sequence, while multimedia with same contents would be mapped into same hashing sequence values. Robustness to content-preserving manipulations, sensitivity to malicious tampering and security are the basic criteria of perceptual hashing. Perceptual hashing supplies an efficient solution to image identification and authentication. It has a wide application in content-based identification, retrieval, tampering localization and etc. However, feature extraction of existing hashing algorithms concentrates on the data expression instead of the perception of human vision system, which degrades their robustness to content-preserving manipulations, sensibility to content changes, and tampering localization in distorted images.In this dissertation, the reasons why the structural statistics based human man vision (HVS) model could not correspond well with the judgments of human observers are analyzed. A novel HVS model which fuses the classical HVS model and the structural statistics based HVS model is proposed. Then, in the guidelines of the fusion models of human vision system, hashing algorithms based on pixel domain, DCT and DWT domain are propsed. In addition, a universal tampering localization model to improve the performance of existing hashing algorithms on tampering localization in distorted images is brought out.The major works of this dissertation are described as follow:Fistly, based on the classical HVS models which imitate the information processing from human eyes to human brains and the model which measures the human observers' feeling by structural statistics, the HVS fusion model is proposed. Experimental results show that, this fusion model is consistent with the judgements of human observers, serving as a feasible theroretical guideline for the extraction of robust features.Secondly, image hashing algorithms in the pixel domain are proposed. By Gabor filter, whose response is in accordance with HVS, the two dimensional (2D) edge features are extracted. Then, by Radon transform, which has high robustness to rotation, 2D edge features are transformed into one dimensional (1D) features. Last, by quantization of 1D features, hashing sequence is generated. Experimental results show that the above proposed algorithm is very robust to common content-preserving manipulations, sensitive to content changes, and has the performance on tampering localization. Since features are extracted in the pixel domain, the prposed hashing algorithms could be applied to images regardless of their storing formats. In addition, an hashing algorithms based on the fuzzy distance matching is proposed to improve the robustness of rotation for existing hashing algorithms with low level features.Thirdly, image hashing algorithms in DCT and DWT domains are proposed respectively. For the hashing algorithm in DCT domain, robust features are extracted from the bit stream which is entropy decoded for JPEG compressed images. Experimental results show that the proposed algorithm is not only robust to JPEG recompression, histogram equalization and some other content-preserving manipulations, but also robust to mild rotation attack, which is scarce for most existing hashing algorithms in DCT domain. For the hashing algorithm in DWT domain, significant bits in intra scales instead of inter scales are extracted, which is in accordance with the scalable property of JPEG2000 bits steam. Experimental results demonstrate the proposed algorithm excellant performance.Fourthly, a novel image hashing model for existing hashing algorithms to improve their performance on tampering localization in distorted images is proposed in this dissertation. Many image hashing algorithms have been proposed to detect the malicious tampering for content authentication. However, their tampering localization performance degrades dramatically on images with content-preserving distortion, as these algorithms cannot distinguish the malicious tampering from content-preserving distortion. In this framwork, high precision of tampering localization in distorted images is achieved by controlling the robustness of extracted features. By experimenting with classical image hashing algorithms, the correctness of the proposed model is proved.