A Comprehensive Study on Third Order Statistical Features for Image Splicing Detection

Abstract

Second order statistical features (e.g. Markov transposition probability matrix and gray level co-occurrence matrix) have been proved to be effective for passive image forgery detection in the past few years. In this paper, third order statistical features are proposed for image splicing detection. We model the thresholded adjacent difference block DCT coefficient array of an image as conditional co-occurrence probability matrix, second order Markov transition probability matrix and second order co-occurrence matrix. Since the dimensionality exponentially depends on the order, dimensionality of the third order features is much larger than that of second order features, principal component analysis (PCA) is therefore introduced in our work to overcome the high dimensionality introduced computational complexity and the possible overfitting for a kernel based supervised classifier. Experimental results show that conditional co-occurrence probability matrix outperforms second order features and PCA is proved to be an effective dimensionality reduction tool for image splicing detection. We also test the robustness of third order statistical features, despite higher dimensionality, third order statistical features demonstrate the same robustness as that of second order features.