03-01-2013, 01:23 PM
Detecting Digital Tampering by Blur Estimation
1Detecting Digital.pdf (Size: 855.73 KB / Downloads: 23)
Abstract
With powerful computer and mighty software, seasoned users could turn digital media into
what they want. The detection of digital tampering has become a crucial problem. In most of
the time, digital tampering is not perceptible by human; however, some traces of digital
tampering may be left in the media during the process. Based on this idea, we present a method
in this paper that could reveal blurred regions that indicate possible tampering without any
embedded information such as watermarking technique. Effectiveness and results will be
presented, robustness would also be discussed.
Introduction
Considering the making of digital image frauds, for example, face replacement, in order to
produce a seamless doctored face, applying blurring is inevitable. Other than face replacement,
skin smoothing and panorama are also blurring involved digital image tampering. Blurring is a
very common process in digital image manipulation; it could be used to reduce the degree of
discontinuity or to remove unwanted defects, ultimately, it is used to generate plausible digital
image forensics. Hence if an additional blurring process applied on an image is detectable,
possible fraud can be exposed even in a credible image.
We will first introduce our blur estimation method. From the fundamentals of frequency
domain knowledge about the mathematical model, we develop a scheme to detect blur regions
for general images then modify it for identifying digital forensics.
To illustrate the general effectiveness of this technique, we show results on naturally blurred
images first to prove the effectiveness of our blur estimation method and proceed on
perceptually credible forgeries applied manually smoothing, and then demonstrate the
possibility to detect computer-generated objects in movie scenes by our blur estimation scheme.
Review on digital forgeries detection
A study about how to verify digitally altered contents without the implementation of
watermark becomes a novel and emerging subject. However, several researchers have carried
out some representing results in different approaches, but these approaches are different from
ours.
Blur estimation
In the making of digital image frauds, considering the face replacement process, imaginable,
the size of the two faces are not likely to match, in most of the time the author would
commence a resizing procedure. Resizing will cause the two regions become similar in their
dimension, then the author have to adjust the angle that the face towards by rotating it.
Moreover, the dissimilarity of the joining regions, intensity or color, which is resulted from
light source direction (such as from left or right, makes the shadows have non-consistent
positions) or type (a tungsten light or fluorescent tube, which will cause identical object
exhibits different colors), adjusting brightness or contrast will preserve global illumination
consistency, and correcting color can remove biased color effect resulted from different light
source, adding or erasing shades would gain shading homogeneity.
Till now, the forged region is quite similar to the original, but it won’t be exactly
identical to the covered one; consequently, removing the seamy regions is ineluctable. To
achieve this goal, a ubiquitous way is to apply a Gaussian blur on the seamy areas. After
that, a believable face replacement fraud is done, such as Fig. 1: the face of the center
image is cropped, resized, rotated, re-lighted, smoothed and pasted onto the left one to
obtain the result image.
Digital image tampering detection using blur estimation
From the knowledge of blurring characterized by point spread function, we know that if we
restrict the spectrum for only a certain kind of signal (or texture), the relationships (sharp and
blurred ones) will match the illustration as Fig. 5, which means we could estimate a small or a
large amount of blur, resulted from a small or a large deviation from the focus plane of the
subject. Seeing that the blur estimation will give us some clue of depth, and with the result of
blur estimation applied on normal images using various thresholds, it becomes possible to
construct a valid tool for blurring involved digital image fraud detection.
The main idea we use to expose digital image blurring is by examining the consistency
between defocus knowledge (depth of field) and blur estimation results (blurrier regions). For
example, in Fig. 14, we know that the focus is on the boy thus he should be a sharp object, and
so is his entire face because they are all in the focus plane, this is called the defocus knowledge.
If the blur estimation results show that some part of his face disobey the defocus knowledge,
there must be some other procedures like blurring is taken on the image; in this example, the
blurring applied on his cheeks and chin is revealed.
Note that other than examining the inconsistency between depth of field and results of blur
estimation, we could define some region is highly potential for digital tampering. Considering
the face replacement case, the cheeks are usually the location smoothing will be applied on. If
we could observe odd blur appearance on such region, we can say the image is suspicious for
being manipulated.
Detection results
We have test on several kinds of images, the following results are grouped by the manner of
test images. We will first present the effectiveness of blur estimation on unaltered images, then
the demonstration about the frequency domain approach of blur estimation to identify the
existence of possible tampering.
In the first group, results are the of normal photography image. We test on the image which
has partial sharp and blurred regions. As seen in Fig. 11 (a) is the input image, which has a
quality factor of 78.07%; this number is generated from the first phase, indicates the degree of
global sharpness. Fig. 11 (b) is the normalized image of Fig. 11 (a), which is the result of the
second phase. It could be observed that sharper regions are brighter and blurrier regions are
darker. After threshold applied, final result is shown in Fig. 11 ©, black blobs indicate sharp
regions and white blobs indicate blurred ones. This algorithm does indicate the sharp regions
(keypad on the phone) and blurred ones (backgrounds). Blurred regions are locally estimated,
blur percentage is 88.3848 %. The final results and are quite close to our expectation.
Conclusions
We proposed a method that is able to detect blurred regions in an image. And the results of
the detection show the method is reliable in various kinds of images even there are no blur in
the image at all or the blur is made artificially. From the idea that blending is inevitable to hide
unwanted traces in the making of digital forensics, we applied modified approach of blur
estimation based on frequency domain knowledge.
The effectiveness of blur estimation on plausible digital image frauds, even on those
non-perceptible ones for human are demonstrated. By combining additional information, such
as depth relationships of the objects or unnatural sharpness, both help identify digital image
frauds. All tests of this method are operated on a computer environment of 1.3G CPU with
640MB RAM. For each run of estimation, the calculation interval depends on the size of the
target image and varies from couples of seconds (5 mega pixels image) to instant. Experiment
results show that this scheme is powerful in blurred regions detection for defocused images as
well as digital image tampering involving blurring process, with the advantage of simplicity.