BLIND DETECTION OF PHOTOMONTAGE USING HIGHER ORDER STATISTICS
Tian-Tsong Ng, Shih-Fu Chang
Department of Electrical Engineering
Columbia University, New York, NY
{ttng,sfchang}@ee.columbia.edu
Qibin Sun
Institute of Infocomm Research
Singapore
ABSTRACT
In this paper, we investigate the prospect of using bicoherence
features for blind image splicing detection. Image splicing is an
essential operation for digital photomontaging, which in turn is a
technique for creating image forgery. We examine the properties
of bicoherence features on a data set, which contains image
blocks of diverse image properties. We then demonstrate the
limitation of the baseline bicoherence features for image splicing
detection. Our investigation has led to two suggestions for
improving the performance of the bicoherence features, i.e.,
estimating the bicoherence features of the authentic counterpart
and incorporating features that characterize the variance of the
feature performance. The features derived from the suggestions
are evaluated with Support Vector Machine (SVM) classification
and shown to improve the image splicing detection accuracy
from 62% to about 70%.
1. INTRODUCTION
Photomontage refers to a paste-up produced by sticking together
photographic images. In olden days, creating a good composite
photograph required sophisticated skills of darkroom masking or
multiple exposures of a photograph negative. In today’s digital
age, however, the creation of photomontage is made simple by
the cut-and-paste tools of the popular image processing software
such as Photoshop. With such an ease of creating good digital
photomontages, we could no longer take image authenticity for
granted especially when it comes to legal photographic evidence
[1] and electronic financial documents. Therefore, we need a
reliable and objective way to examine image authenticity.
Lack of internal consistency, such as inconsistencies in
object perspective, in an image is sometimes a telltale sign of
photomontage [1]. However, unless the inconsistencies are
obvious, this technique can be subjective. Furthermore, forgers
can always take heed of any possible internal inconsistencies.
Although image acquisition device with digital watermarking
features could be a boon for image authentication, presently
there still is not a fully secured authentication watermarking
algorithm, which can defy all forms of hacking, and the
hardware system has to secure from unauthorized watermark
embedding. Equally important are the issues such as the need for
both the watermark embedder and detector to use a common
algorithm and the consequence of digital watermarks degrading
image quality.
On the premise that human speech signal is highly Gaussian
in nature [2], a passive approach was proposed [3] to detect the
high level of non-gaussianity in spliced human speech using
bicoherence features. Unlike human speech signal, the premise
of high guassianity does not hold for image signal. It was shown
[4] that bispectrum and trispectrum of natural images have a
concentration of high values in regions where frequency
components are aligned in orientation, due to image features of
zero or one intrinsic dimensionality such as uniform planes or
straight edges. As images originally have high value in higher
order spectrum, detecting image splicing based on the same
principle of increased non-gaussianity would be a very low
signal-to-noise problem, not to mention the possible complex
interaction between splicing and the non-linear image features.
Recently, a new system for detecting image manipulation
based on a statistical model for ‘natural’ images in the wavelet
domain is reported [5]. Image splicing is one kind of image
tampering the system takes on; however, no further detail about
the technical approach is provided in the article.
Image splicing is defined as a simple joining of image
regions. We currently do not address the combined effects of
image splicing and other post-processing operations. Creation of
digital photomontage always involves image splicing although
users could apply post-processing such as airbrush style edge
softening, which can potentially be detected by other techniques
[5]. In fact, photomontages with merely image splicing, as in
Figure 1, can look deceivingly authentic and each of them only
took a professional graphic designer 10-15 minutes to produce.
Figure 1: Spliced images that look authentic subjectively
In this paper, we pursue the prospect of grayscale image
splicing detection using bicoherence features. We first examine
the properties of the proposed bicoherence features [3] in
relation to image splicing and demonstrate the insufficiency of
the features. We then propose two new methods on improving
the performance of the bicoherence features for image splicing
detection. Lastly, we evaluate the methods using SVM
classification experiments over a diverse data set of 1845 image
blocks. More details about this work are included in [6].
2. BICOHERENCE
Bicoherence is a normalized bispectrum, i.e., the third order
correlation of three harmonically related Fourier frequencies of a
signal, X(Ȧ) [7]:
9;,((( ,6&$6
Authorized licensed use limited to: NORTHERN JIAOTONG UNIVERSITY. Downloaded on April 10,2010 at 09:43:04 UTC from IEEE Xplore. Restrictions apply.