基于方向金字塔分解与稳定几何失真校正的鲁棒图像水印算法

目的为了提高水印算法的抗几何攻击能力,并兼顾较高的鲁棒性与不可感知性,设计一种基于方向金字塔分解与稳定几何失真校正的鲁棒图像水印算法。方法首先,引入方向金子塔,对载体图像完成分解,输出对应的低通与高通子带;将低通子带分割为一系列的非重叠块;根据载体的亮度、纹理与边缘掩码,计算水印嵌入强度,最大程度地平衡水印图像的不可感知性与鲁棒性;设计水印嵌入方法,将经过Arnold映射加密后的水印嵌入到非重叠子块中,通过修改载体的方向金子塔分解系数,获取水印图像;将不同的攻击类型作用于水印图像,建立训练样本;再利用方向金子塔分解训练样本,计算高通子带的高斯-厄米矩能量,将其视为特征矢量;再利用特征矢量对模糊支持向量机完成训练,以预测几何失真参数,准确校正受攻击的水印图像;设计水印检测机制,从水印图像中复原水印。结果实验数据表明,与当前图像水印方案相比,所提算法具有更高的抗几何变换能力,以及较好的不可感知性与鲁棒性,其提取的水印失真度最小,对应峰值信噪比保持在40dB以上。结论所提水印算法具有较高的鲁棒性和视觉隐秘性,在版权保护、信息防伪等领域具有一定的参考价值。

Robust Image Watermarking Algorithm Based on Directional Pyramid Decomposition and Stable Geometric Distortion Correction

The work aims to design a robust image watermarking algorithm based on directional pyramid decomposition and stable geometric distortion correction, in order to improve the watermarking algorithm''s ability of resisting geometric attacks and give consideration to both higher robustness and insensitivity. Firstly, the directional pyramid was introduced to decompose the carrier image for outputting the corresponding low pass and high pass subbands. Subsequently, the low pass subbands were divided into a series of non-overlapping blocks. According to the brightness, texture and edge mask of the carrier, the embedding strength of the watermark was calculated to balance the imperceptibility and robustness of the watermark image to the maximum extent. Then, the watermark embedding method was designed to embed the watermark encrypted by Arnold map into the non-overlapping subblocks, and the watermark image was obtained by modifying the decomposition coefficient of directional pyramid in the carrier. The training samples were established by applying different attack types to watermark image. The directional pyramid was then used to decompose the training samples for calculating the Gauss Hermite moment energy of the high pass subbands, which was regarded as the feature vector. Then, the feature vector was used to train the fuzzy support vector machine for predicting the geometric distortion parameters and accurately correcting the attacked watermark image. Finally, the watermark detection mechanism was designed to restore the watermark from the watermark image. The experimental data showed that, compared with the current image watermarking scheme, the proposed algorithm had a higher anti-geometric transformation ability, as well as better imperceptibility and robustness with the least distortion degree of extracted watermark, with the corresponding peak signal to noise ratio (PSNR) maintained above 40 dB. The proposed watermarking algorithm has higher robustness and visual insensitivity, which has certain reference value in copyright protection, information security and other fields.