基于广义视觉失真模型的高保真可逆信息隐藏

Considering that there exist significant differences between performance evaluation results and human visual system (HVS) in the conventional reversible data hiding (RDH) algorithms especially for high capacity, we will explore a new research field to achieve the high-fidelity RDH based on generalized visual distortion model. The research is composed of the following three parts: ① the multi-level histogram shifting (HS) based RDH that. is combined with one new visual valuation standard is proposed so as to achieve both high-fidelity and high capacity RDH, in the proposed method, the characteristics of the visual valuation standard are extracted to construct the multi-level histogram and rate-distortion model, and then the local iterative optimization method is proposed to quickly solve the rate-distortion model so that the embedding performance is largely increased, and the computational complexity is highly decreased; ② adaptive reversible data hiding is proposed, in which the just noticeable difference (JND) is used to adjust the strength of difference expansion so that the high visual quality is maintained for high-capacity, besides, the optimization method having the characteristic of self-feedback is presented to quickly obtain the optimal embedding performance; ③ the integer transform based on adaptive distortion model is proposed to achieve high-capacity RDH, considering that the adaptive distortion model has the characteristic of knapsack problem, we utilize the methods that solve the knapsack problem or a simple method to obtain the optimal solution, so that the embedding performance is largely increased and the distortion is significantly decreased. From the description above, we focus on achieving high-fidelity with HVS for high capacity, reconstruct the RDH distortion model considering HVS, provide the theoretical significance and practical value for high-capacity and high-fidelity under the big data background, and also provide a new research direction for the future RDH.

针对传统可逆隐藏算法中普遍存在性能评价结果与人的主观感觉不一致的情况，本项目拟开辟一条与传统思路反其道而行的新研究思路来实现基于广义视觉失真模型的高保真可逆信息隐藏。具体研究内容如下：① 构建一个面向全新视觉评价标准的多直方图平移的可逆信息框架，并在此基础上给出基于新评价标准的特征选择和多直方图构建方法，研究局部迭代的寻优策略来简化求解空间；② 建立以最小可视觉差（JND）为评价指标的动态差扩展可逆信息隐藏框架，提出具有自反馈机制的优化算法来快速求解率（容量）/失真(嵌入失真)问题；③ 建立基于整数变换的自适应失真模型，基于模型所呈现出的背包问题特性，本项目拟借助具有启发式特征的演化算法来实现快速寻优。本项目旨在从上述三个方面构建符合人眼视觉特性的优化模型，并基于其特点设计相应寻优方案，为实现大数据环境下高容量、视觉高保真提供理论及技术支持，也为可逆信息隐藏发展提供了一种新的研究视角。

本课题旨在研究基于广义视觉失真模型的高保真可逆信息隐藏，推进该领域的理论创新，提升嵌入性能，扩展可逆算法的应用面。课题以面向全新视觉评价标准的多直方图平移的可逆信息框架、建立以最小可视觉差为评价指标的动态差扩展可逆信息隐藏框架和建立基于整数变换的自适应失真模型，四年来提出了融合多特征和聚类算法的直方图生成、率失真模型建立及嵌入点寻优策略的多直方图修改一般性框架、人眼视觉评价标准、基于人眼视觉特性的对比度增强的动态差扩展算法、结合了自适应高维映射生成策略和自适应系数选择策略的JPEG图像的可逆信息隐藏算法、融合深度学习技术的高性能预测器。从实验结果来看，所得研究成果能够有效提升率失真性能，所提理论可为可逆信息隐藏自适应研究提供支撑。此外，新的算法框架展现出与先进人工智能研究有机结合的强大潜力，为进一步研究指明了方向。