A Very Fast Copy-Move Forgery Detection Method for 4K Ultra HD Images

Abstract

Copy-move forgery detection is a challenging task in digital image forensics. Keypoint-based detection methods have proven to be very efficient to detect copied-moved forged areas in images. Although these methods are effective, the keypoint matching phase has a high complexity, which takes a long time to detect forgeries, especially for very large images such as 4K Ultra HD images. In this paper, we propose a new keypoint-based method with a new fast feature matching algorithm, based on the generalized two nearest-neighbor (g2NN) algorithm allowing us to greatly reduce the complexity and thus the computation time. First, we extract keypoints from the input image. After ordering them, we perform a match search restricted to a window around the current keypoint. To detect the keypoints, we propose not to use a threshold, which allows low intensity keypoint matching and a very efficient detection of copy-move forgery, even in very uniform or weakly textured areas. Then, we apply a new matching algorithm, and finally we compute the cluster thanks to the DBSCAN algorithm. Our experimental results show that the method we propose can detect copied-moved areas in forged images very accurately and with a very short computation time which allows for the fast detection of forgeries on 4K images.

A GPU-accelerated algorithm for copy move detection in large-scale satellite images

“…Very high resolution satellite images, which convey a lot of extremely fine spatial information in the form of patterns, shapes, objects and textures, can easily contain millions of keypoints. Consequently, as the resolution of the satellite image increases, off-the-shelf SIFT detectors implementations such as those used in, 1,17 can typically run into one of these problems: (i) slow processing; (ii) the exhaustion of system resources when allocating the memory required to build and store the Gaussian pyramid (recall that, depending on the settings, the SIFT detector might also compute the pyramid level at double image resolution); and (iii) the exhaustion of system resources when computing the keypoints or the descriptors among millions of potential candidates.…”

“…Inevitably, because of the extremely large number of keypoints in satellite images, the g2NN test rapidly becomes computationally prohibitive, thus slowing or, in the worst case, crashing the detector. To tackle with this problem, Bertojo et al 1 proposed their optimised version of the detector, called VFCMD (Very Fast Copy-Move Detector), relying on a test with lower complexity as follows. First, they sort all the keypoints according to the value of their angles.…”

“…To avoid the problems caused by the warping and the ZNCC map computation, Bertojo et al 1 proposed an alternative solution to generate the binary mask. Given two clusters R 0 and R 1 , the method first computes their convex hulls H 0 and H 1 , which identify the heart of the manipulation, and their respective borders B 0 , B 1 ; then, the borders are iteratively expanded as follows: for each each pair i = 1, ..., N of points (p 0,i , p 1,i ) ∈ B 0 × B 1 , the PSNR between respective pixel neighborhoods in the input image is compared against a threshold and if it is higher, the coordinates of the two neighbourhoods are assigned to the binary map and the new borders are recomputed.…”