Abstract: We describe a set of natural image statistics that are built upon two multi-scale image decompositions, the quadrature mirror filter pyramid decomposition and the local angular harmonic decomposition. These image statistics consist of first- and higher-order statistics that capture certain statistical regularities of natural images. We propose to apply these image statistics, together with classification techniques, to three problems in digital image forensics: (1) differentiating photographic images from computer-generated photorealistic images, (2) generic steganalysis; (3) rebroadcast image detection. We also apply these image statistics to the traditional art authentication for forgery detection and identification of artists in an art work. For each application we show the effectiveness of these image statistics and analyze their sensitivity and robustness.

@phdthesis{lyu05,
  school       = {Department of Computer Science, Dartmouth College},
  url          = {http://www.cs.dartmouth.edu/farid/publications/slthesis05.pdf},
  author       = {Siwei Lyu},
  year         = {2005},
  title        = {Natural image statistics for digital image forensics},
  address      = {Hanover, NH},
}