Analysis and Comparison of Single-Class Classifiers for Extracting the Feature Patterns of Different Cameras for Tamper Detection
No Thumbnail Available
Date
2012-09
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Image forgery or tampering is the art of manipulating or modifying an
image in either an intentional or unintentional manner. Tampering of images
is widely used by tabloid magazine, fashion industry, scientific joumals, court
rooms, main media outlet and photo hoaxes received in email. In most of the
cases, these manipulations resuh in copyright violation and hence, have to be
detected.
This research work uses the camera properties to detect tampered
images and is working under the fact that different cameras have dissimilar
properties and any tampering operation will reflect in a change in these
features. The proposed automatic algorithm performs tamper detection in
three steps, feature selection, feature extraction and one-class classification.
The system extracts four types of features, namely, color feature, image
quality features, wavelet features and bicoherance features, to form feature
database. To improve the detection process, dimensionality reduction is
performed using Fisher classifier algorithm. The third stage, uses two oneclass
classification algorithms. Radial Basis Function and Support Vector
Machine, to detect tampered and genuine images. The tamper detection
process is performed in two steps, training and testing. During training, the
training images are segmented into 128 x 128 blocks and the features are
extracted for each block. These features are used for training the classifiers.
During testing, the same features are extracted and the trained classifier is
used to detect the presence or absence of manipulafion.
The CASIA TIDE dataset was used during the evaluation steps of the
research work. Three different cameras. Canon, Nikon and Sony, were
selected and the splice tampering was performed using three main variations.
The first is to splice images in different manners, using different camera
combinations, using different image transformations and using different shape
methods. Nine different camera combinations, namely, Canon-Canon, CanonNikon,
Canon-Sony, Nikon-Nikon, Nikon-Canon, Nikon-Sony, Sony-Sony,
Sony-Canon and Sony- Nikon were formed. During photomontage creation,
four different types of transformations (Resize, Deform, Rotate, Do Nothing)
were used with four different shapes (Circle, Rectangle, Triangle and
Arbitrary). Thus, a total of 1725 (800 authentic and 925 spliced color images)
were used during experimentation.
From the experimental results, it can be concluded that the maximum
efficiency was produced by SVM one-class classifier with both tamper and
authentic image identification. The system is efficient in identifying tampered
images formed using rectangle shape followed by triangle. Poor performance
was obtained when arbitrary shape was used during splice creation.
Considering transformations, best performance was achieved when no
transformation was used during the creation of tampered images. The system
found rotate transformation tampering easily followed by resizing
modifications. Thus, it could be concluded that the four features, color, image
quality, wavelet and bicoherance, when combined with SVM one class
classifier, is efficient in identifying tampered and untampered images and can
be used in important applications like legal, media and World Wide Web.