Hierarchical image resampling detection based on blind deconvolution

Highlights

• We propose a novel image resampling detection algorithm based on blind deconvolution.

• To the best of our knowledge, this is the first study to classify different types of resampling algorithms.

• Our method fundamentally avoids those drawbacks generated by JPEG block artifacts and other intractable interference.

Abstract

Resampling detection is a helpful tool in multimedia forensics; however, it is a challenge task in cases with compression and noisy. In this paper, by modeling the recovery of edited images using an inverse filtering process, we propose a novel resampling detection framework based on blind deconvolution. Different interpolation types in the resampling process can be distinguished by our algorithm, which is significant for practical forensics scenarios. Furthermore, in contrast to traditional resampling detection algorithms, our method can effectively avoid interference caused by JPEG block artifacts. As the experimental results show, our method is more robust than other state-of-the-art approaches in the case of strong JPEG compression and substantial Gaussian noise.

Introduction

Geometric transformations are helpful preprocessing in many multimedia analysis [1] and social network tasks [2], [3], [4], [5]. They are also big challenges in some computer vision tasks, such as image retrieval [6], [7]. Moreover, to create high-quality and consistent image forgeries, geometric transformations are an inevitable process in image splicing. As interpreted in [8], during geometric transformations, resampling operations are a typical technique widely used by forgers.

Image resampling operations, namely resizing images by interpolating pixels, are common operations in the image editing and tampering process. It is essential to conduct a resampling operation when a natural photo taken by a camera is presented on different sizes of screens, such as those of mobile phones or televisions, as illustrated in Fig. 1. The detection of resampling operations can effectively recover the image editing history so that one can justify the authenticity of an image.

Many researchers have made great efforts concerning the blind detection of digital image resampling operations. Previous studies have mostly focused on the periodical trace of resampling operations. Popescu et al. [8] noted that resampling operations produce new statistical correlations, despite a few obvious visual differences. They built a probability correlation diagram to uncover the latent periodical traces, and resampled images can be easily distinguished using this diagram. An exact formulation of how transformation parameters influence the appearance of periodic artifacts was given by Kirchner [9], who presented a reliable and efficient detection scheme.

The second difference of an image is a powerful tool for detecting resampling operations. Gallagher et al. [10] presented a detection method for linear and cubic interpolation in images subject to JPEG compression, therein exploiting the second derivative of each row for forensic tasks on image resampling operations. Prasad et al. [11] noted the periodicity of the zero-crossings of the second difference in a resampled image and proved that their detection scheme can be effectively applied to image copy-move forgery detection. Mahdian et al. [12] introduced the second derivative of the covariance feature for reflecting the signal correlation. Cao et al. [13] noticed that the correlation coefficient of the adjacent rows (or columns) varied periodically in resized images, and the proposed algorithm performed well in an experimental scenario. Ryu et al. [14] exploited the periodic properties of the second derivative of the transformed image in both the row and column directions. Both the magnitude and phase information of the derived signals were analyzed to estimate the transformation matrix accurately. Birajdar et al. [15] detected a resampled image using the autocovariance sequence of the zero-crossings of the second difference.

In recent years, Singular Value Decomposition (SVD), which reflects the correlation of data, has been promisingly applied to detect resampled images. Wang et al. [16] indicated that resampling operations increase the number of zero singular values significantly, as the dependence of the image pixels changes. Gul et al. [17] extracted SVD-based features from the image sub-blocks and then merged them together to represent the whole image for detection. In [18], the authors detected resampled images based on the fact that the magnitudes of non-zero singular values vanish more sharply than usual.

Moreover, some works have used other characteristics of resampled images for detection. Vázquez-Padín et al. [19] proposed to use a resampling-based method to provide an accurate way to distinguish original and tampered regions by analyzing the resampling factor of each area. In the paper [20], an improved scheme for resampling factor estimation was presented using prefilters based on cyclostationarity theory. Feng et al. [21] used the normalized energy density to detect traces of resampled images. Their proposed method achieved relatively robust results in JPEG compression cases. Hou et al. [22] considered resampling detection as a texture classification problem because the resampling operation often changes the texture pattern of the original image.

Additionally, some researchers have extended their study to detecting resampling operations on JPEG images. Because of block artifacts, as explained above, few resampling detectors have shown competitive performance when Q is lower than 85 in the case of JPEG compression. Nataraj et al. [23] added noise to suppress JPEG artifacts, but the periodic patterns caused by resampling remained partially intact. In the work [24], a more suitable detection method under the condition of double JPEG compression was proposed and analyzed how the affine transformation of JPEG compression affected resampling detection.

Although image resampling blind detection has been explored from diverse perspectives, this detection method still must address many challenges. Essentially, most previous approaches have explored intrinsic periodic traces of resampling operations for constructing detectable features. Because most image resampling detectors are designed in the frequency domain, it is difficult to detect aperiodic interpolations, e.g., nearest neighbor interpolation [14]. In addition, there are few reliable detection algorithms applicable to strong JPEG compression. When the compression quality is low, the interference is quite severe [8], [12]. However, in practical scenarios, the suspected images are often saved in JPEG format or are subject to JPEG compression during capture or transmission.

To overcome the above obstacles, this paper proposes a general algorithm for detecting resampled images and recognizing interpolation methods. We extend the application of blind deconvolution to image resampling detection. Furthermore, a content-adaptive decision framework is formulated via hierarchical multi-region fusion. Two tasks in this area are considered: (1) the classification of resampled images and (2) the recognition of interpolation types. Both tasks are challenging when subject to severe lossy compression or the addition of substantial noise. We conduct our experiments on the UCID dataset [25], therein obtaining better performance compared with other methods. The main contributions of this paper can be summarized as follows:

• We propose a novel image resampling detection algorithm based on blind deconvolution that can recover the image editing history. To the best of our knowledge, this is the first study to classify different types of resampling algorithms, which can be used for copy-move detection and splicing detection.

• Our method fundamentally avoids those drawbacks generated by JPEG block artifacts and other intractable interference. The proposed method is extensively evaluated on the publicly available UCID dataset. We perform exhaustive tests over various parameters, therein obtaining more promising results compared with traditional algorithms.

• Images resampled by zeroth-order interpolation (e.g., nearest neighbor interpolation), second-order interpolation (e.g., bilinear interpolation), and third-order interpolation (e.g., bicubic interpolation) are considered. Both aperiodic interpolation (e.g., nearest neighbor interpolation) and periodic interpolation (e.g., bilinear interpolation and bicubic interpolation) can be detected accurately.

The remainder of this paper is organized as follows. We review background knowledge in Section 2, and then, we illustrate our detection algorithm in Section 3. The experimental results are presented in Section 4. Finally, Section 5 summarizes the highlights and discusses future work.