使用自适应阈值的图像篡改检测与定位算法

Image tampering detection and localization algorithm using adaptive thresholding

针对利用相机传感器模式噪声的篡改检测在待测图像纹理复杂区域存在较高的虚警,提出了一种考虑纹理复杂度的自适应阈值检测算法。根据Nyman-Pierson(N-P)准则,确定不同纹理复杂度对应的相关性匹配判定阈值,而得到相关性阈值与纹理复杂度的关系拟合函数。在不重叠分块计算待测图像噪声残差和其来源相机传感器模式噪声对应块相关性的基础之上,根据待测图像块不同的纹理复杂度进行相关性匹配,确定大致篡改位置;再用快速零均值归一化互相关(ZNCC)算法计算两噪声图像中大致篡改区域对应点的相关性,实现精确定位。在手机图像库上的实验表明,与现有的固定阈值方法相比,本文算法的检测率达到了98.8%,而虚警率仅为1.897%,有效地降低纹理复杂区域的虚警率,并实现对篡改区域的精确定位;同时,与传统的滑动窗口方法相比,本文算法检测效率平均提高了26倍。

As the existing image tampering detection algorithms based on camera s ensor pattern noise have high false alarm rate in strong texture areas,a novel detection algorithm using adap tive thresholding adaptive to the texture complexity is proposed.According to Nyman Pierson criteria,the correla tion decision thresholds for different texture complexity are determined. Furthe rmore,the relationship between the correlation threshold and the texture complexity is obtained. While detecting,the correlati on coefficients of each pair of non-overlapping blocks from the noise residual of the test image and the refere nce sensor pattern noise of the inspected camera are computed first.The tampered areas are roughly localized ba sed on the correlation matching determined by the adaptive thresholds.Then,the fast zero mean normalized cross c orrelation （ZNCC） algorithm is used to calculate the correlation of the corresponding pixels in the roughly localized regions between the reference sens or pattern noise and the noise residual of the test image,and to realize the accurate localization.The experi mental results on the mobile phone image database show that the detection rate of the proposed algorithm is 98.8%,while the false alarm rate is only 1.897%.Compared with the existing methods using a f ixed threshold,the proposed algorithm using adaptive thresholds can effectively reduce the false alarm rate in the complex texture re gion ,and can accurately locate the tampered region.At the same time,the proposed algorithm can improve the effici ency by 26times compared with the conventional sliding-window-based algorithms.