基于多方向互交叉模式算子与极坐标变换的鲁棒图像哈希算法

为了提高哈希序列对几何攻击的鲁棒性与正确识别率,设计了基于多方向互交叉模式算子与极坐标变换的鲁棒哈希算法。引入插值运算与Gaussian滤波器,完成图像的尺寸规范化与去噪处理,使其对于任意的可疑目标均可输出一个固定长度的哈希序列。基于极坐标变换(log-polar transform,LPT),对滤波规范图像实施处理,输出抗旋转攻击的二次图像。随后,利用多方向互交叉模式算子,从8个方向将二次图像变换为两个编码映射。将两个编码映射分割为非重叠子块,通过提取这些子块的直方图,将其视为纹理特征,作为第一个哈希序列。利用强度概率密度梯度代替强度梯度,对SURF (speeded up robust features)方法予以改进,充分提取图像中的稳定角点,形成角点图像;将角点图像分割为一系列的非重叠子块,通过计算每个子块所含的角点数量,将含有结构信息最丰富的子块予以标记,输出其在图像中对应的位置信息;并借助离散小波变换(discrete wave transform,DWT)来分解这些标记子块,获取每个子块对应的低频系数。联合位置信息与低频系数,形成结构特征,作为第二个哈希序列。设计加密机制,分别对两个哈希序列完成扩散,从而形成最终的哈希序列。通过计算源图像与可疑图像之间的l_2范数距离,将其与用户识别阈值的大小对比,对目标的真实性完成判别。试验数据显示:较已有的哈希生成机制而言,所提方法拥有更高的鲁棒性,对各类几何攻击均有更高的识别准确率。

Robust Image Hash Algorithm Based on Multi-Directional Mutual Cross Pattern Operator and Polar Coordinate Transformation

In order to improve the robustness and correct recognition rate of hash sequences against geometric attacks, a robust hash algorithm based on multi-directional mutual cross pattern operator and polar coordinate transformation is designed in this paper. By introducing interpolation operation and gaussian filter, the size normalization and denoising of the image are completed so that it can output a fixed-length hash sequence for any suspicious target. And the filtered standard image is processed based on polar coordinate transformation to output the secondary image against rotation attack. Subsequently, the secondary image is transformed into two coding maps from eight directions by using the multi-directional mutual cross pattern operator. Two coding maps are segmented into non-overlapping sub-blocks, and the histograms of these sub-blocks are extracted to treat as texture features as the first hash sequence. The SURF method is improved by replacing the intensity gradient with the intensity probability density gradient to fully extract the stable corners in the image for forming the corner image. The corner image is divided into a series of non-overlapping sub-blocks, and the sub-blocks with the richest structural information are marked by calculating the number of corners contained in each sub-block to output their corresponding position information in the image. And the low frequency coefficients corresponding to each sub-block are obtained by decomposing these marker sub-blocks with the help of discrete wavelet transform. The structural features are formed by combining location information with low frequency coefficients, which as the second hash sequence. The encryption mechanism is designed to diffuse the two hash sequences for forming the final hash sequence. By calculating the l2 norm distance between the source image and the suspicious image, the authenticity of the target is judged based on comparing it with the user recognition threshold. The experimental data show that the proposed method has higher robustness and recognition accuracy for various geometric attacks than the existing hash generation mechanism.