Research On Visual Distortion And Image Information Hiding

Data is ubiquitous in information society. Secure distribution and access of data become increasingly important. Information hiding, which can embed messages into cover works without arousing the attention from eavesdroppers or attackers, dose not increase the payload of communication, and hence is used for digital rights management, covert communication, error resilient and etc. Information hiding must satisfy the distortion constraint and seek a tradeoff between the invisibility and the robustness of watermark. Therefore, it is crucial for image-based-watermarking to determine a reasonable visual distortion constraint.Starting from the laws of visual perception, this thesis proposes a hierarchical, multi-channel model of visual distortion. The visual distortion is layered into signal distortion and content distortion, and the latter is decomposed into textural distortion and structural distortion. Under the models of all the distortion categories, image distortion can be controlled during watermarking. To this end, this thesis carries out research in four aspects.At first, signal distortion. The previous studies show that, visual distortion is not perceivable when the response of a complex cell is kept invariant between the stimuli of two images. Therefore, characterizing neural responses by independent subspace analysis, this thesis proposes to embed watermarks into ISA coefficients of the cover image by energy-conserving hiding code of angle quantization index modulation. This approaches presents intrinsically low distortion, and is robust against valumetric attack because of angle quantization.Secondly, image decomposition into structure and texture. The structure and the texture are the particular components of visual data. The structure, as the sketch of scene and the shape of object, includes large-scale, continuous edges and contours, while texture, as the surface character or illumination effects, contains fragmentary, regulated or unregulated oscillatory regions as well as smooth regions. Discrimination between the structure and the texture is the foundation of discussion on content distortion. According to principles of sparse coding, this thesis improves the image decomposition approach which is based on completive representation by corresponding bases, so as to meet the requirement of distortion analysis and watermarking schemes. Then, texture distortion. Texture has been regarded as suitable region for information hiding. Psychology research on texture perception and graphical study about texture synthesis show that high order statistics of feature are key clues for texture recognition. However, due to its nonlinearity, high order statistics is hard to guide the distortion metric for watermarking. Therefore, this thesis design a linear proximate distortion metric based on statistical feature, under the guidance of which watermark intensity can be increased efficiently, and watermark capacity too.At last, structure distortion. As the geometrical attack shown, image warping may incurs no perceptible distortion, which proves the existence of structurally undistorted image space. Within the space, watermark can be hidden and endowed with robustness against collusion attack. The theory on structure perception shows that human is prefer to perceive or more sensible to large-scale structure rather than local one. According to the large-scale priority principle, this thesis defines the structure distortion, and embedding watermarking by image warping.Based on the proposed theory of visual distortion, this thesis focuses on the improvement of watermarking schemes. Watermarking is united into a framework of energy minimization, and posed as a biobjective optimization problem for a tradeoff between the invisibility and the robustness. In the framework, constraint about texture distortion and structure distortion are both considered, meanwhile, spread spectrum modulation watermark and quantization index modulation watermark are unified. The improved watermarking is both imperceptible and robust against compression, filtering valumetric scaling and collusion attack, and thus offers high capacity. Moverover, the proposed theory of visual distortion is also useful for other visual creative application, including image completion, blending, inpainting and etc.