目的 图像篡改检测主要分为图像区域复制篡改、图像拼接和对象移除三个方向，其中图像复制粘贴篡改是图像篡改检测的重要研究方向之一。针对目前大多数复制粘贴篡改检测方法难以检测平滑和小的篡改区域，并且虚警率较高等问题，本文提出了一种基于匹配对的密度聚类MP-DBSCAN (matched pairs - density based spatial clustering of applications with noise)和点密度过滤策略的图像复制粘贴篡改检测方法。方法 首先，在图像中提取大量的关键点，根据关键点的灰度值分组后进行匹配。其次，本文提出了一种改进的密度聚类算法MP-DBSCAN，聚类对象为匹配对的一侧，并利用匹配对的另一侧约束聚类过程，即使篡改区域在空间上距离较近，或者篡改区域存在多个的情况，也能把不同的篡改区域较好地区分开来。本文还提出了一种点密度过滤策略，通过删除低密度簇，降低了检测结果的虚警率。最后，通过估计仿射矩阵并使用ZNCC (zero-mean normalized cross-correlation) 算法定位篡改区域。结果 消融实验证明了MP-DBSCAN算法和点密度过滤策略的有效性。在FAU、MICC-F600、GRIP和CASIA v2.0四个数据集上与几个经典的和最新的检测方法进行了对比实验，本文方法的F1在四个数据集上像素层的实验结果分别是0.9143、0.8906、0.9391和0.8568。结论 本文提出的MP-DBSCAN聚类算法和点密度过滤策略能有效提高检测算法的性能，即使篡改区域经过旋转、缩放、压缩和添加噪声等处理，本文方法依旧能够检测出大部分的篡改区域，性能优于当前的检测算法。

Objective In recent years, with the development of the Internet and computer technology, it is a trivial task to manipulate images and change their content. Therefore, it is becoming important to develop robust image tampering detection methods. As a passive forensic method, the image forgery methods can be categorized into: copy-move, splicing and inpainting. Copy-move means copying a part of the original image to another part of the same image. In recent years, many excellent copy-move forgery detection (CMFD) methods have been proposed. All the proposed copy-move forgery detection methods can be divided into three types: block-based methods, keypoint-based methods and deep learning methods. However, the existing methods still have the following draw-backs: 1) It is difficult to detect small or smooth tampered regions. 2) A massive number of features lead to the high computational cost. 3) The false alarm rates are high when the tampered images involve self-similar regions. To solve the above issues, in this paper, a novel CMFD methods based on matched pairs - density based spatial clustering of applications with noise (MP-DBSCAN) and the filtering of the point density is proposed. Method First, a large number of scale-invariant feature transform (SIFT) keypoints are extracted from the input image by lowering the contrast threshold and normalizing the image scale. Therefore, sufficient keypoints can be detected in small and smooth regions. Second, in order to manage multiple keypoint matching, the generalized two nearest neighbor (G2NN) matching strategy is employed. Therefore, the detection algorithm can perform better when tampered region has been copied multiple times. In order to solve the keypoint matching problems over a massive number of keypoints, a hierarchical matching strategy is employed. To speed up the matching process, keypoints are first grouped by their grayscale values, then the G2NN matching strategy is applied to each group, rather than the keypoints detected from the entire image. The efficiency and accuracy of the matching procedure are improved without deleting correct matched pairs. Third, an improved clustering algorithm called MP-DBSCAN is proposed. The matched pairs are grouped into different tampered regions accurately. During the clustering process, it is necessary to adjust the direction of the matched pairs. The cluster objects are only one side of the matched pairs, not all the extracted keypoints, and the keypoints from the other side are used as a constraint of the clustering process. A satisfying detection result is obtained even the tampered regions are close to each other. Compared with the traditional copy-move forgery detection methods, the proposed method obtains the best F1 measures. Fourth, the prior regions are constructed based on the clustering results. The prior regions can be regarded as the approximate tampered regions. Furthermore，a filtering policy of the point density is proposed. Each point density of the region is calculated. Then, the region with lower point density is deleted. The false alarm rate is reduced according to the filtering policy of the point density. Finally, the tampered regions are located accurately using the affine transforms and the zero-mean normalized cross-correlation (ZNCC) algorithm. Result The proposed method is compared with the state-of-the-art CMFD methods on four standard datasets, including FAU, MICC-F600, GRIP and CASIA v2.0. The FAU dataset is provided by Christlein. The average resolution of the FAU dataset is about 3000 × 2300 pixels. The FAU dataset involves the tampered images under post-processing operations (such as additional noise, JPEG compression, etc.) and various geometrical attacks (such as scaling, rotation, etc.). The total number of plain copy-move is 48, the total number of scaling is 480, the total number of rotation is 384, the total number of JPEG compression is 432 and the total number of noise addition is 240. The MICC-F600 dataset includes images in which a region is duplicated once or more times. The resolutions of the MICC-F600 dataset are from 800×533 to 3888×2592 pixels. There are 600 images on the MICC-F600 dataset, 440 original images, taken from the MICC-F2000 dataset and 160 forged images, taken from the SATS-130 and 28 new ones. The GRIP dataset involves the tampered images with low resolution, all the resolution of the GRIP dataset is 1024×768 pixels. There are 80 original images on the GRIP dataset. Some tampered regions on the GRIP dataset are very smooth, which are challenging for those features based on sparse sampling. The size of the tampered regions is from about 4000 pixel to about 50000 pixels. The CASIA v2.0 dataset contains 7491 authentic and 5123 forged images. There are 1313 copy-move forged images on the CASIA v2.0 dataset. Precision, recall and F1 scores are used as the assessment criteria in the experiments. The F1 score of the proposed method on the FAU dataset at the pixel level is 0.9143. The F1 score of the proposed method on the MICC-F600 dataset at the pixel level is 0.8906. The F1 score of the proposed method on the GRIP dataset at the pixel level is 0.9391. The F1 score of the proposed method on the CASIA v2.0 dataset at the pixel level is 0.8568. Extensive experimental results demonstrate the superior performance of the proposed method compared with the existing state-of-the-art methods. Furthermore, the effectiveness of the MP-DBSCAN algorithm and the filtering policy of the point density are demonstrated by the ablation studies. Conclusion To detect the tampered regions accurately, a novel CMFD method based on MP-DBSCAN algorithm and the filtering policy of the point density is proposed in this paper. The matched pairs of the image can be divided into different tampered regions effectively by the MP-DBSCAN algorithm, and the tampered regions can be detected accurately. The mismatched pairs are discarded by the filtering policy of the point density, thus, the false alarm rate of the detection result is reduced. Extensive experimental results demonstrate that the proposed method exhibits quite satisfactory accuracy and robustness compared with the existing state-of-the-art methods.