Automated detection and localisation of duplicated regions affected by reflection, rotation and scaling in image forensics

Although the detection of duplicated regions plays an important role in image forensics, most of the existing methods aimed at detecting duplicates are too sensitive to geometric changes in the replicated areas. As a result, a slight rotation can be used not only for the copied region to fit better the scene in the image, but also to hinder the detection of the tampering. In this paper, a novel forensic method is presented to detect duplicated regions, even when the copied portions have undergone reflection, rotation and/or scaling. To achieve this, overlapping blocks of pixels are mapped to log-polar coordinates, and then summed along the angle axis, to produce a one-dimensional (1-D) descriptor invariant to reflection and rotation. Besides, scaling in rectangular coordinates results in a simple translation of the descriptor. The dimension-reduced representation of each block has a favourable impact in the computational cost of the search of similar regions. Extensive experimental results, including a comparative evaluation with two existing methods, are presented to demonstrate the effectiveness of the proposed method.