基于三阶统计特征的被动图像拼接检测

Passive image splicing detection based on third order statistical features

针对被动图像拼接检测问题,提出了一种基于三阶统计特征的检测算法。该算法把图像状态矩阵中三个相邻状态之间的依赖关系建模为条件共生概率矩阵,然后将其作为识别特征输入到支持向量机(SVM)进行分类。由于高阶统计特征维数随着统计阶数的增加而呈指数级增加,为了降低高维特征在分类阶段所引入的高计算复杂度以及避免可能出现的过拟合现象,引入了主成分分析法(PCA)对提取的特征进行降维处理。实验结果显示,条件共生概率矩阵特征在空间域和8×8分块DCT域的检测结果均优于传统的马尔可夫特征和共生矩阵特征;PCA是图像拼接检测的一个有力分析工具,在大幅度降低特征维数的同时能够保持识别率不降低。

This paper proposed a third order statistical features based method to detect image splicing operation passively.The dependences among neighboring three states in the state-array were modeled as CCPM which was treated as discriminative features for SVM classification.Since the dimensionality of higher order statistical features grows exponentially with the order,it introduced PCA to decrease the complexity for classification and to overcome the potential over-fitting problem.Experimental results show that the conditional co-occurrence probability matrix features outperform traditional Markov features and gray level co-occurrence matrix features in both 8×8 block DCT domain and spatial domain.PCA is an effective tool for image splicing detection and new features with much fewer dimensionalities after PCA perform as good as original features.