A statistical feature based approach to distinguish PRCG from photographs

Abstract

We present a passive forensics method to distinguish photorealistic computer graphics (PRCG) from natural images (photographs). The goals of our work are to improve the detection accuracy and the robustness to content-preserving image manipulations. In the proposed method, Homomorphic filtering is used to highlight the detail information of image. We find that the texture changes are different between photographs and PRCG images under same Homomorphic filtering transformation, and then we use the difference matrixes to describe the differences of texture changes. We define a customized statistical feature, named texture similarity, and combine it with the statistical features extracted from the co-occurrence matrixes of differential matrixes to construct forensics features. Then we develop a statistical model and use SVM as classifier to distinguish PRCG from photographs. Experimental results show that the proposed method enjoys following advantages: (1) Proposed method reaches higher detection accuracy, synchronously, it is robust to tolerate content-preserving manipulations such as JPEG compression, adding noise, histogram equalization, and filtering. (2) Proposed method is provided with satisfactory generalization capability, it will be available when the training samples and the testing samples come from different sources.