Distinguishing Photographic Images and Photorealistic Computer Graphics Using Visual Vocabulary on Local Image Edges

摘要

Differentiating computer graphics from natural images remains a representative problem of digital image forensics because the two categories of images reflect typical different aspects of generation and forgery of digital images. This paper aims to address this problem through analyzing the statistical property of local edge patches in digital images. First, we preprocess image edge patches and project them into a 7-dimensional sphere as in [7]. Then, a visual vocabulary is constructed via determining the key sampling points in accordance with Voronoi cells. The proposed approach to constructing visual vocabulary avoids troubles in traditional partitioning algorithms such as k-means, And then, a given image is represented as a binned histogram of visual words and the corresponding feature vector is formed by the bins. Finally, we employ an SVM classifier for image classification. Our experimental results demonstrate the efficient discrimination of the proposed features.