Forensics of image blurring and sharpening history based on NSCT domain

Abstract

Detection of multi-manipulated image has always been a more realistic direction for digital image forensic technologies, which extremely attracts interests of researchers. However, mutual affects of manipulations make it difficult to identify the process using existing single-manipulated detection methods. In this paper, a novel algorithm for detecting image manipulation history of blurring and sharpening is proposed based on non-subsampled contourlet transform (NSCT) domain. Two main sets of features are extracted from the NSCT domain: extremum feature and local directional similarity vector. Extremum feature includes multiple maximums and minimums of NSCT coefficients through every scale. Under the influence of blurring or sharpening manipulation, the extremum feature tends to gain ideal discrimination. Directional similarity feature represents the correlation of a pixel and its neighbors, which can also be altered by blurring or sharpening. For one pixel, the directional vector is composed of the coefficients from every directional subband at a certain scale. Local directional similarity vector is obtained through similarity calculation between the directional vector of one random selected pixel and the directional vectors of its 8-neighborhood pixels. With the proposed features, we are able to detect two particular operations and determine the processing order at the same time. Experiment results manifest that the proposed algorithm is effective and accurate.