Image copy-move forgery detection based on polar cosine transform and approximate nearest neighbor searching

https://doi.org/10.1016/j.forsciint.2012.10.031Get rights and content

Abstract

Copy-move is one of the most commonly used image tampering operation, where a part of image content is copied and then pasted to another part of the same image. In order to make the forgery visually convincing and conceal its trace, the copied part may subject to post-processing operations such as rotation and blur. In this paper, we propose a polar cosine transform and approximate nearest neighbor searching based copy-move forgery detection algorithm. The algorithm starts by dividing the image into overlapping patches. Robust and compact features are extracted from patches by taking advantage of the rotationally-invariant and orthogonal properties of the polar cosine transform. Potential copy-move pairs are then detected by identifying the patches with similar features, which is formulated as approximate nearest neighbor searching and accomplished by means of locality-sensitive hashing (LSH). Finally, post-verifications are performed on potential pairs to filter out false matches and improve the accuracy of forgery detection. To sum up, the LSH based similar patch identification and the post-verification methods are two major novelties of the proposed work. Experimental results reveal that the proposed work can produce accurate detection results, and it exhibits high robustness to various post-processing operations. In addition, the LSH based similar patch detection scheme is much more effective than the widely used lexicographical sorting.

Introduction

Digital imaging technologies have experienced tremendous advances during recent decades, and the conventional film camera is being overshadowed by its successor, the digital camera. By virtue of the proliferation of low-cost and high-resolution digital cameras, the ubiquitous digital images have become the dominant source of visual information. Meanwhile, a number of powerful image editing softwares have been developed, amidst which Adobe Photoshop might be the most popular one. The content of a digital image can be easily doctored with the assistance of image editing softwares. Sometimes it is very difficult, or even impossible, to detect the trace of elaborate forgeries by human visual inspections. As a result, the credibility of digital images have been greatly undermined. In the past few years, the cases of image forgeries on publications have been frequently reported, ranging from news magazines to scientific journals. Therefore, great efforts have been dedicated to tackle image forgeries, and the topic of image forensic has attracted increasing attentions [1]. Image forensic aims at identifying the evidence of forgeries, whose primary mission is to reinforce the credibility of digital images. In contrast to the watermarking- [2] and signature-based [3] authentication approaches, image forensic can accomplish blind authentications without referring to any auxiliary information such as watermark or signature, and it is therefore of more practical significance. After nearly a decade of developments, image forensic has grown from infancy to maturity, and a series of algorithms have been proposed to cope with diverse forms of forgeries.

In this paper, we focus on the detection of copy-move that is one of the most common forgeries. In copy-move forgery, a specific region of the image is copied and then pasted to another part of the same image. The most famous and widely spread fake image involving copy-move forgery should be the news photo of Iran's missile launching. On July 9, 2008, a photo with four missiles firing into the sky, as shown below in Fig. 1(a), appeared on the front pages of several major news websites, including The Los Angeles Times, The Chicago Tribune, The Financial Times, The New York Times, etc. The news photo was initially obtained from the website of Iran's Sepah News. However, shortly after its release, the photo raised considerable public attention. It was suspected that the photo had been digitally altered, since it contains some repeating patterns that may be the result of copy-move forgery. As marked in Fig. 1(a), the two missiles and their smoke trails, as well as the four folds of smoke plumes on the ground are highly similar. One day later, the Associated Press news agency published the original photo with only three missiles, as Fig. 1(b) shows, which further proved that the controversial photo was fake.

The copy-move forgery can be identified by detecting duplicated regions in the image. However, the duplicated region may not exactly match the source one since it could undergo some post-processing operations, such as blur, filtering and rotation. Thus, an effective forgery detection algorithm should be robust to post-processing operations. In what follows, several state-of-the-art copy-move forgery detection (CMFD) algorithms will be briefly introduced. Most of the existing algorithms are patch-matching based, where the image is divided into overlapping patches and the copy-move forgery is detected by searching the patches with similar features. In [4], the quantized discrete cosine transform (DCT) coefficients were calculated as the robust representation of each patch. The feature vectors were then lexicographically sorted to identify matching pairs, and it was assumed that the feature vectors of matching patches in a copy-move pair should come consecutively in the sorted list. In order to filter out false matches, the displacement between the positions of every two patches whose feature vectors rank in neighboring rows of the list was examined. Since all the copy-move pairs should yield the same position displacement, two matching patches can be labeled as a copy-move pair if the occurrence frequency of their position displacement is higher than a pre-defined threshold. Compared with exhaustive searching, the lexicographical sorting based detection scheme can make a reasonable tradeoff between computation complexity and detection accuracy, thus it has been widely adopted by various CMFD algorithms. In Huang et al.'s work, a similarity metric was incorporated in the lexicographical sorting based scheme to improve its detection accuracy [5]. In addition, some alternative features have also been exploited. For example, the principal component analysis (PCA) was applied on the patches to obtain compact and robust features for CMFD in [6]. Similarly, the singular value decomposition (SVD) was employed in [7] to decompose the low-frequency wavelet sub-band of the image. Inspired by the research in digital watermarking, Bayram et al. proposed to extract features from patches using the Fourier–Mellin Transform (FMT) [8]. The rotationally-invariant Zernike moments were computed in [9] as features, so that the duplicated regions can be detected even after rotations. In [10], a number of blur-invariant features were developed for CMFD, and the kd tree was adopted to detect similar patches.

As discussed above, robust feature extraction and similar patch identification are two primary concerns in CMFD; however, both of them are fraught with technical challenges. Although most of the algorithms perform well in the robustness against alignment-preserving post-processing operations such as blur, rotation is still very difficult to deal with in CMFD and most algorithms cannot detect the forgeries with (large degree) rotations. However, rotation is commonly used in the copy-move forgery. The duplicated region may be rotated to better fit the background region, so that the forgery can be more visually convincing and even harder to detect. Take the forged image shown in Fig. 2(b) for example, the duplicated region is rotated by 2° in clockwise to synchronize with its neighbors. Regarding similar patch identification, the research on this issue is rather rare, and the lexicographical sorting based scheme has been used by a vast majority of CMFD algorithms. Nevertheless, it has been observed in simulations that the feature vectors of two patches in a copy-move pair do not always rank consecutively after lexicographically sorting if post-processing has been imposed on the duplicated region, and such pairs could be missed in CMFD. In order to tackle these challenges, we propose a polar cosine transform and approximate nearest neighbor searching based CMFD algorithm. The main novelties of the proposed work are twofold. First, the problem of similar patch identification is studied from the perspective of approximate nearest neighbor searching, and an LSH based identification scheme is developed. Second, post-verification methods are devised by exploiting the unique characteristics of copy-move pairs. The post-verification procedure can benefit CMFD by enhancing the accuracy of forgery detection. Experimental results reveal that the proposed work can exhibit higher accuracies than the comparative algorithm in detecting the copy-move forgeries with various post-processing operations. Moreover, it has been observed that the proposed similar patch identification scheme outperforms lexicographical sorting, and it can significantly reduce the rate of miss detection.

The organization of this paper is as follows. In Section 2, the proposed CMFD algorithm is described in detail, including feature extraction, similar patch identification and post-verification. Experimental results are presented in Section 3 to demonstrate the efficacy of the proposed work. Finally, the paper is concluded in Section 4.

Section snippets

Robust feature extraction

Feature extraction is a prerequisite step for CMFD and crucial to detection accuracy. It is desired that the patches in a copy-move pair can be mapped to similar features even in the presence of post-processing. At the same time, the features should correctly distinguish distinct patches in the image. To accomplish these goals, the polar cosine transform is investigated to construct the features for CMFD. Polar cosine transform (PCT) [11] is a recently proposed orthogonal transform and it

Experimental results and discussions

Experimental results are presented in this section to demonstrate the effectiveness of the proposed algorithm. We start by describing the experimental setup, including the testing data, metrics for performance evaluation, as well as the parameter settings of the proposed and comparative algorithms.

Conclusions

In this paper, we have presented a robust copy-move forgery detection algorithm using the PCT and approximate nearest neighbor searching. The feature extraction scheme have been developed using the orthogonal PCT. The rotation invariance in PCT has been exploited to enable rotation-resistant CMFD. In addition, the problem of similar patch identification has been formulated as approximate nearest neighbor searching and solved by utilizing LSH. To further enhance the accuracy of CMFD, a set of

Acknowledgement

This work was supported by National Science Foundation of China under grant 61202164.

References (16)

  • Y.P. Huang et al.

    Improved DCT-based detection of copy-move forgery in images

    Forensic Sci. Int.

    (2011)
  • B. Mahdian et al.

    Detection of copy-move forgery using a method based on blur moment invariants

    Forensic Sci. Int.

    (2007)
  • H. Farid

    Image forgery detection: a survey

    IEEE Signal Process. Mag.

    (2009)
  • D. Kundur et al.

    Digital watermarking for telltale tamper proofing and authentication

    Proc. IEEE

    (1999)
  • S. Roy et al.

    Robust hash for detecting and localizing image tampering

  • J. Fridrich et al.

    Detection of copy-move forgery in digital images

  • A.C. Popescu, H. Farid, Exposing digital forgeries by detecting duplicated image regions, Technical Report, TR2004-515,...
  • G.H. Li et al.

    A sorted neighborhood approach for detecting duplicated regions in image forgeries based on DWT and SVD

There are more references available in the full text version of this article.

Cited by (144)

  • Image copy-move forgery detection based on dynamic threshold with dense points

    2022, Journal of Visual Communication and Image Representation
    Citation Excerpt :

    use frequency-domain features to describe the region of interest, and [12–14] and [17–21] use texture features to describe the region of interest. [17–22] uses 2NN matching, [3–7] and [12–13] uses dictionary sorting, etc. [3–13] is a block-based approach, which obviously has disadvantages such as computationally intensive, lack of robustness to geometric attacks, and difficulty in detecting multiple tampering. [14–22] is a point-based algorithm, and even though it overcomes some of the disadvantages of block-based methods, it has a number of problems of its own.

View all citing articles on Scopus
View full text