Based on the conclusion that the frequency residual DF(u, v) of filtered images present a band-pass shape, while that of original images present a high-pass shape, a bandwidth feature is defined to discriminate the filtered images from original images. Then, a low-pass Gaussian filter bank with a selection of different normalized cutoff frequencies is applied as the detection filter to build a robust detection scheme.A. Bandwidth EstimationBefore feature extraction, DF(u,v) is converted from Cartesian to polar coordinate system and a 1-D radial function R(ρ) is found by the integration over [0 2π], denoted by the expression