摘要
图像处理中应用广泛的中值滤波可用于图像反取证技术、隐秘通信技术,因此中值滤波取证受到了研究者的关注.文中提出了一种基于中值滤波残差及其差分(Median Filtering Residual Difference,MFRD)的鲁棒中值滤波取证技术.首先根据方向性和对称性将多方向MFRD分组,然后分别建立自回归模型(Auto-Regressive model,AR)并提取其模型参数和直方图特征,最后将所有分组特征组合成中值滤波检测特征.作者证明了在一定条件下p阶AR模型等价于一个p阶马尔科夫模型,因而采用AR参数代替转移概率可以大大降低信息取证特征的维数.MFRD减少了来自图像内容和JPEG压缩块效应痕迹的干扰,因而增强了提出方法的鲁棒性.在多个图像数据库组成的混合图像库上的测试结果和多个图像库之间的泛化测试结果都表明作者提出的算法能有效地辨别JPEG压缩图像及小尺寸图像是否经过中值滤波操作,其检测错误率远低于现有的一些中值滤波取证技术,并且检测结果不依赖于训练图像.
Median filtering is widely used in image editing and image anti-forensic techniques. The forensics of median filtering has recently drawn much attention since it can be used to reveal image＇s processing history or used as auxiliary clues for image tampering. To achieve a robust detector against JPEG compression and small-size image block, we first explored the multi-directional differences from median filtering residual （MFRD）, then fit each difference into an Auto-Regressive （AR） model and extracted the AR parameters and histograms of the difference as the feature subset. It is theoretically proved that under some assumptions, a p-order AR model is equivalent to a p-order Markov model. Due to the low dimension of the AR coefficients and histogram, the final feature concatenated from multi-directional differences has low dimension. Experimental results on a large composite image database indicate that the proposed detector performs better than the-state-of-art methods, especially in the detection of median filtering under heavy JPEG compression. The proposed method possesses excellent generalization ability, and outperforms other methods.
出处
《计算机学报》
EI
CSCD
北大核心
2016年第3期503-515,共13页
Chinese Journal of Computers
基金
国家自然科学基金(61379155
U1536204)
教育部博士点基金项目(20110171110042)
广东省自然科学基金重点项目(s2013020012788)资助~~
关键词
数字图像取证
中值滤波
中值滤波残差多方向差分
自回归模型
digital image forensics
median filtering
median filtering residual difference
auto-regressive model
