A novel copy-move forgery detection algorithm via two-stage filtering

SMDAF: A novel keypoint based method for copy‐move forgery detection

“…In the past few years, many copy-move forgery detection methods have been proposed, which can be roughly divided into block-based methods [11][12][13][14][15], deep learning-based methods [16][17][18][19][20][21], and feature point-based methods [22][23][24][25][26][27][28][29][30]. The block-based algorithms are inefficient and do not perform well in detecting the copy-move regions with different geometric and post-processing attack operations.…”

“…Although deep learningbased methods have gained state-of-the-art success in many other fields, one interesting fact here is that these methods need many training data, and they cannot achieve good performance in CMFD. Compared with block-based and deep learning-based methods, the keypoint-based methods have achieved relatively superior performance concerning detection consequences and running speed [29,31]. However, in the existing keypoint-based methods, the descriptors extracted from keypoints may not all meet the needs.…”

“…The results contain a large number of false matching and reduce the matching accuracy. To overcome this weakness, researchers proposed some algorithms for filtering the wrong matching keypoint pairs [27][28][29]. Chen et al [27] aggregated the feature and colour of the extracted keypoints, then used reverse-G2NN and outpoint filtering algorithm to filter the keypoints.…”

“…Wang et al [28] used candidate clustering based on offset information to remove wrong matching pairs and used RANSAC and ZNCC to locate duplicate regions. And Yang et al [29] proposed a two-stage(mesh-based and clusterbased) filter to select most of the keypoint pairs of wrong matching and then applied the Delaunay triangulation algorithm for image matting.…”

“…The results contain a large number of false matching and reduce the matching accuracy. To overcome this weakness, researchers proposed some algorithms for filtering the wrong matching keypoint pairs [27–29]. Chen et al.…”