| Color neighbor histograms are investigated as an approach to detecting the manipulation of JPEG images. Color histograms have previously been used as a basis for the study of image manipulations. The generalization of previously explored color neighbor approaches is found to be a convolution of the multi-dimensional color histogram with a neighborhood defining filter. Various parameter settings and their implications are explored, including choice of color space, choice of counting methods, and choice of methods for selecting regions under analysis. Algorithm performance was analyzed on the Nimble Challenge 2016 image set, both by categories of the images and overall. A discussion of Fisher's linear discriminant, the classification scheme used, is also provided.;The analysis revealed color neighbor histograms to have overall poor performance, with better performance seen in certain cases, particularly for spliced images. Methods for selecting regions for analysis can produce results that can be misrepresentative of color neighbor histogram performance. |
