医学图像的不同工作域下高性能及鲁棒可逆水印的深入研究

Aiming at the remote diagnosis aid of medical images, this project studies the techniques for the increase of the rate-distortion (or embedding) performance, the decrease of computation complexity, and the robustness against attacks in different processing domains. The main contents are as follows. .For reversible watermarking (RW) methods in the spatial domain, the distortion model based on gradient magnitude and gradient direction is built. We identify the features which can be used for estimating most accurately local complexity by the edge-detection operator and the invariability of the average of an image block based local complexity analysis. By the function impacting algorithm’s performance determined by distortion model in the spatial domain and GA (Genetic Algorithm), when the embedding capacity is given, the locations are determined, which minimize the global distortion. In other words, the induced distortion is least when the pixels at these locations are modified in watermark embedding. Based on this reason, the computational complexity is deceased largely. .For RW methods in the transform-domain, we build the distortion model by the following two aspects. One is that we try our best to determine the range over which the integer DCT (discrete cosine transform) coefficients of a 8×8 block vary. When the coefficients are restricted in this range, there are no pixel overflows (> 255 for a 256 gray-scale image) or underflows (< 0) in the spatial domain. The other one is to estimate the distribution of integer DCT coefficients so that those small amplitudes are selected for embedding. Thus, the low distortion is obtained..For RW methods in encrypted image, our main method is divided into the following two parts. One is that sparse representation processing is used before encryption, which is favorable to provide enough embedding capacity required by the following RW method. The other one is performed after encryption phase, in which the reserved redundancy among pixels is utilized to get the invariability to the embedding process, and then we employ this invariability to design the corresponding RW scheme..For research on improving the robustness of reversible watermarking used in medical images, the bijective integer transform (or an statistical invariant and its corresponding embedding method) resistant to the attacks is proposed. .The research results of this project can provide rigorous theoretical support for RW of medical images, and have important applications.

本项目以医学图像远程医疗为应用背景，研究医学图像不同工作域下算法性能提升和抵抗攻击的能力。具体内容包括: 针对空域可逆水印算法，建立空域失真模型，利用边缘检测算子和均值不变性提取最能描述局部复杂性的特征，利用从理论上得到的空域失真模型对算法性能的影响函数和遗传算法，获得指定容量下全局失真最小的最优嵌入位置，从而有效降低算法复杂度；针对变换域算法，从估算不会引起像素值溢出的系数取值范围和预测系数幅值分布两个方面建立失真模型；针对加密域算法，在加密之前，通过稀疏表示理论事先预留出可逆水印算法所需的嵌入空间，或在加密之后，利用像素间仍存在的冗余提取嵌入不变量并设计相应的水印嵌入算法；针对抵抗攻击的能力，提出对某些攻击具有不变性的双射的整数不变域（或空域统计不变量及其嵌入方式），提高医学图像的可逆水印的鲁棒性。研究成果能为医学图像的可逆水印研究提供重要的理论支持，同时具有重要的应用价值。

本项目自从2016年1月1号获得资助以来，这四年主要从以下几个方面展开了研究: .⑴. 空域可逆水印算法：.围绕低容量下的最小失真模型（即通过限定嵌入容量来追求最小嵌入失真）展开研究：⑴ 构建一个全新的多直方图平移的可逆信息框架，通过聚类算法划分出图像纹理复杂度不同的区域，利用改进的纵横交叉算法进行多直方图快速选点的可逆信息隐藏算法；⑵ 设计更加精确的自适应局部复杂度评估算法来提高局部复杂度的评估性能，优先选取位于平滑区域的载体进行信息隐藏从而来降低嵌入失真；通过改进现有预测器，产生一个信息熵较小的预测误差载体序列，即载体序列拥有一个陡峭的直方图，同时尽可能增加预测误差的数量。.高容量可逆信息隐藏算法：将占了所有像素的四分之三的像素都能被环绕它的像素预测，从而提高了预测精度，值为0和为1的预测误差的数量大大增加，从而提高了嵌入容量；像素位于的区域越平滑，携带的水印信息就越多，反之，像素位于的区域纹理越高，携带的信息就越少，从而保证了图像的视觉效果。.⑵. 整数域可逆水印算法：.在给定容量下尽可能降低嵌入失真的思想指引下，实现整数域可逆信息隐藏算法。具体从以下两个方面展开研究：充分利用Alattar 整数变换的具有均值不变性的特点，将均值用于局部复杂度的评估中从而极大地提高评估性能；通过设计不规则的块分割算法，使得每一块都是由一个中心像素和环绕它的像素组成，用中心像素去预测环绕它的每一个像素，从而产生出更尖锐的预测误差直方图。.⑶. 加密域可逆信息算法：.加密域可逆信息隐藏算法的难点在于：加密会消弱相邻像素之间的相关性，而信息隐藏却是利用相邻像素之间的相关性进行信息隐藏的。在加密之前，设计出16种重构算法，每一种重构算法的最终目标是使得相关性的像素排在一起，然后通过设计压缩算法对排序后的5个高位平面进行压缩，将空出的嵌入空间携带最低3个位平面，低3个位平面用来存放水印，最后采用加密和重排的方式对含水印图像进行加密生成加密水印。