Copy-Move Forgery Detection Using Cellular Automata - Cellular Automata in Image Processing and Geometry - page 115

Table 6.1 shows detection results for 40 plain CMF images from the CoMoFoD
database [24]. Block size and threshold T d are chosen according to the image con-
tent in order to maximise the F-measure, threshold T s is set to zero, and no post-
processing is applied. Results show that 82.5% of the images have a F-measure
score higher than 0.6, indicating that detection is very accurate. The main problem
for detection represents homogeneous regions because in that case many blocks are
falsely detected.
Furthermore, results achieved with this method are comparable with results of
other CMFD methods. Lower accuracy and weaker performances are noticeable
only in the case of large homogeneous regions with many pixels having similar
values. In that case, better results are achieved with methods that do not require rep-
resentation of the image in binary, such as DCT [9] or Zernike moments [19]. Better
performance in this case can be accomplished by changing the binary representa-
tion to produce a more discriminative description of such similar areas. However,
the advantage of the CA method is the very small number of false detected regions
in all other tested cases, for example images with more complex textures.

6.4.3 Application on Post-processed Images

Post-processing of forged images introduces differences in pixel values resulting
in differences in the binary representation and generated set of rules. One of the
common problems is detection of forgery in the image after the addition of noise.
Figure 6.9 shows degradation of performance with adding of Gaussian noise on
plain CMF image (simple 1D CA is used for detection). The result is accomplished
using image with added noise of zero mean and variance equal to 0.00001 (image
intensities were normalized to range [0,1] prior to addition of noise). Block size is
set to 32 32 pixels and neighbourhood size of 7 pixels is used. Values of thresholds

are T s =6and T d =1.5

×

b = 48 pixels. Conversion of the RGB image into a greyscale
image is used before noise adding, and no post-processing is applied (for example,
removing of false detected areas). Even when the amount of noise is very small,
only part of the copied area is successfully detected (Fig. 6.9c). A larger part of the
copied area can be detected by adjusting the similarity threshold for detection of
similar blocks, but that would also introduce much false detected area.
Coping with noise can be done by using simple pre-processing of image in the
form of filtering [23]. Figure 6.10 shows example of CMF detection on image with
Gaussian zero mean noise with variance equal to 0.0001. In the case when no pre-
processing is applied, detection is not possible at all (Fig. 6.10c). After application
of an averaging filter of size 3

×

× 3 on noise image, detection becomes very accurate

(Fig. 6.10d).
Another very common post-processing method is image blurring. Figure 6.11
shows one example of detection of blurred image accomplished by applying of 3 3

filter. It is possible to notice that detection is quite accurate because blurring does
not affect pixels values in a way that changes properties of copied regions.

×