A forgery detection algorithm for exemplar-based inpainting images using multi-region relation

https://doi.org/10.1016/j.imavis.2012.09.002Get rights and content

Abstract

The identification of image authenticity has received much attention because of the increasing power of image editing methods. This paper proposes a novel forgery detection algorithm to recognize tampered inpainting images, which is one of the effective approaches for image manipulation. The proposed algorithm contains two major processes: suspicious region detection and forged region identification. Suspicious region detection searches the similarity blocks in an image to find the suspicious regions and uses a similarity vector field to remove the false positives caused by uniform area. Forged region identification applies a new method, multi-region relation (MRR), to identify the forged regions from the suspicious regions. The proposed approach can effectively recognize if an image is a forged one and identify the forged regions, even for the images containing the uniform background. Moreover, we propose a two-stage searching algorithm based on weight transformation to speed up the computation speed. The experimental results show that the proposed approach has good performance with fast speed under different kinds of inpainting images.

Graphical abstract

Highlights

► We propose a novel forgery detection algorithm for inpainting images. ► The proposed method can detect the forged images containing uniform area. ► We speed up the searching process by a two-stage searching algorithm. ► The proposed method can also detect the copy-move images.

Introduction

The detection of image authenticity [1], [2], [3], [4] is becoming an important research topic because advanced image processing tools can help people make forged images easily. The detection of digital forgery can be classified into active-based [5], [6], [7] and passive-based [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24] approaches. Active-based approaches extract prior inserted information from a digital image (e.g., digital watermarks or signatures) to identify the authenticity. If the embedded information is detected to be changed, the image is recognized as a tampered image.

In contrast to active-based approaches, passive-based approaches detect the feature consistency in a digital image without the embedded information. Region duplication is a common type of forgery. Fridrich et al. [8] detected duplicated regions by matching the discrete cosine transform (DCT) coefficients of overlapping blocks. Langille and Gong [9] used zero-normalized cross correlation (ZNCC) as the similarity measurement. Mahdian and Saic [10] applied blur moment invariants to detect duplicated regions. Huang et al. [11] used a scale invariant feature transform (SIFT) algorithm to extract SIFT descriptors, which describe the correlations between the copy region and the paste region.

When an image is forged using the copy-paste method or the matting technique [25], lighting inconsistency is a good indication of a forged image because the lighting is not consistent over the image. Johnson and Farid [14] described how to estimate the 3-D direction of a light source from the lighting reflection in a human eye. Another one of their methods [15] estimated a low-parameter model for complex lighting environments. Lee et al. [16] proposed a detection method by computing the lighting direction from a segmented image.

The corresponding device parameters are not the same when the images are acquired from different acquisition devices. A good indicator for a normal image is the comparison of these parameters because they should be the same between two different areas in an image. Lukas et al. [19] estimated the sensor's pattern noise of the target region and the referred sensor's pattern noise. Then the statistical analysis of regional correlations was used to identify forged regions. Lin et al. [20] employed a mismatch camera response function to detect image composition. Johnson and Farid [21] used lateral aberration to identify tampered region. When the image was tampered with, the local lateral aberration of the tampered region is inconsistent with the global aberration. Popescu and Farid [22] employed an expectation–maximization (EM) algorithm to quantify specific correlations between the pixels of the image.

JPEG images can be re-saved in JPEG format after being manipulated, and the characteristics of double JPEG compression are a good clue for the detection of forged images. Li et al. [12] examined the mismatch information of block artifact grids (BAGs) as evidence of copy-paste forgery. Qu et al. [13] proposed a novel independent component analysis (ICA)-based algorithm to search for areas of double JPEG compression to find image splicing forgeries.

Re-sampling is usually used in image composites, such as image resizing, rotation and translation. The interpolation is adopted to generate irregular periodic relations between the pixels of the image. Popescu and Farid [17] introduced an EM algorithm to examine whether or not the image is re-sampled. Mahdian and Saic [18] analyzed specific periodic properties for interpolated signals that can be used to detect any re-sampling operation in the image.

The technique of image inpainting can be applied to produce a type of forged image. Consequently, the detection of these forged images is becoming an important issue. Wu et al. [23] proposed a detection algorithm that finds the abnormal similarity for all blocks in the image, and then employs a fuzzy membership function to identify inpainting regions in the image. Lin et al. [24] presented a fast algorithm that analyzes the DCT coefficients of the double-quantization effect. The algorithm is able to detect fine-drained forged regions in the JPEG image.

An inpainting image is constructed by filling the target area using the surrounding regions in the same image; therefore, the key problem in recognizing the fake region is to search for similar regions in the faked image. However, a non-faked image may have some very similar parts in the background, especially uniform regions such as the sky or sea, which will interfere with the detection process. This paper proposes an effective forgery detection algorithm for inpainting forged images. Suspicious region detection searches for similar blocks in an image to find the suspicious regions and uses vector filtering to filter out non-forged regions. Forged region identification applies a multi-region relation (MRR) technique to recognize the forged region. MRR can effectively remove some suspicious regions from a uniform region. The technique has improved the recognition accuracy for the images containing a uniform background. Moreover, we propose a two-stage searching algorithm based on weight transformation to speed up the similarity computation. To improve the searching speed of similar blocks, we propose a two-stage searching algorithm based on a weight transformation. The method is able to reduce the searching time of the best matched block by examining key values for similarity detection. The experimental results show that the proposed algorithm can locate the forged region in forged images and performs well in terms of speed. In addition to the inpainting forgeries, our algorithm is also capable of detecting copy-move forgeries.

The rest of this paper is organized as follows. Section 2 summarizes the exemplar-based image inpainting technique. Section 3 describes the proposed algorithm and the high-speed searching method. Section 4 illustrates the experimental results and analysis. Finally, we present our conclusions in Section 5.

Section snippets

Background knowledge

Image inpainting [26], [27], [28], [29], [30], [31], [32] is a very active research field because it can effectively repair damaged or removed regions in a visually plausible way. Previously, pixel-based approaches were used to repair little blemishes or scratches, but this technique generates obvious blur when filling a larger region in the image. Criminisi's algorithm [27], [28], [29], [31] is the most popular exemplar-based inpainting algorithm among all the similar approaches.

In Criminisi's

Proposed forgery detection algorithm

Our detection algorithm consists of two major modules: suspicious region detection and forged region detection. The first module searches for similar blocks in an image, selects candidates, and then groups them into several suspicious regions. The second part analyzes the multi-region relation of these suspicious regions and identifies the forged regions.

Searching the most similar blocks is a time-consuming process. Instead of an exhaustive search, we propose a two-stage searching strategy,

Experimental results and discussion

This section illustrates the experimental results and analysis for four kinds of forged images, namely (1) images with a single forged region, (2) images with a single forged region and a uniform region, (3) images with multiple forged regions, and (4) images with copy-and-paste forgery. To quantitatively evaluate the performance, we define the recall and precision rates in the following way:Recall=DcDc+DmPrecision=DcDc+Df,where Dc is the number of correct pixels, Dm is the number of miss

Conclusion and feature work

In this paper, we propose a fast and effective forgery detection algorithm for exemplar-based inpainting forged image. We first search the suspicious regions using similar block computation, and then identify forged regions using the MRR technique. In addition to the inpainting type, our algorithm is also able to detect copy-move forgeries. Moreover, to speed up the execution time of similarity computation, we propose a fast searching algorithm based on weight transformation. The experimental

Acknowledgement

This work was supported by the Ministry of Economic Affairs, Taiwan, ROC, under Grant No. 100-EC-17-A-02-S1-032 and National Science Council, Taiwan, ROC, under Grant No. NSC 101-2221-E-259 -013.

References (33)

  • J. Fridrich et al.

    Detection of copy-move forgery in digital images

  • A. Langille et al.

    An efficient match-based duplication detection algorithm

  • H.L. Huang et al.

    Detection of copy-move forgery in digital images using SIFT algorithm

  • W. Li et al.

    Doctored JPEG image detection

  • Z. Qu et al.

    A convolutive mixing model for shifted double JPEG compression with application to passive image authentication

  • M.K. Johnson et al.

    Exposing digital forgeries through “”specular highlights on the eye

  • Cited by (86)

    • Transformer-Based Image Inpainting Detection via Label Decoupling and Constrained Adversarial Training

      2024, IEEE Transactions on Circuits and Systems for Video Technology
    View all citing articles on Scopus

    This paper has been recommended for acceptance by Etienne Memin.

    View full text