An Enhanced Statistical Approach to Identifying Photorealistic Images

Abstract

Computer graphics identification has gained importance in digital era as it relates to image forgery detection and enhancement of high photorealistic rendering software. In this paper, statistical moments of 1-D and 2-D characteristic functions are employed to derive image features that can well capture the statistical differences between computer graphics and photographic images. YCbCr color system is selected because it has shown better performance in computer graphics classification than RGB color system and it has been adopted by the most popularly used JPEG images. Furthermore, only Y and Cb color channels are used in feature extraction due to our study showing features derived from Cb and Cr are so highly correlated that no need to use features extracted from both Cb and Cr components, which substantially reduces computational complexity. Concretely, in each selected color component, features are extracted from each image in both image pixel 2-D array and JPEG 2-D array (an 2-D array consisting of the magnitude of JPEG coefficients), their prediction-error 2-D arrays, and all of their three-level wavelet subbands, referred to as various 2-D arrays generated from a given image in this paper. The rationale behind using prediction-error image is to reduce the influence caused by image content. To generate image features from 1-D characteristic functions, the various 2-D arrays of a given image are the inputs, yielding 156 features in total. For the feature generated from 2-D characteristic functions, only JPEG 2-D array and its prediction-error 2-D array are the inputs, one-unit-apart 2-D histograms of the JPEG 2-D array along the horizontal, vertical and diagonal directions are utilized to generate 2-D characteristic functions, from which the marginal moments are generated to form 234 features. Together, the process then results in 390 features per color channel, and 780 features in total Finally, Boosting Feature Selection (BFS) is used to greatly reduce the dimensionality of features while boosts the machine learning based classification performance to fairly high.