IRJET- Digital Image Forgery Detection using Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG)

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 02 | Feb 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 1844
Digital Image Forgery Detection using Local Binary Patterns (LBP) and
Histogram of Oriented Gradients (HOG)
Manish Shankar Kaushik1, Rabul Saikia2, Dr. Aditya Bihar Kandali3
1Department of Electrical Engineering, Jorhat Engineering College, Jorhat, Assam, India
2Department of Electrical Engineering, Jorhat Engineering College, Jorhat, Assam, India
3Associate Professor, Dept. of Electrical Engineering, Jorhat Engineering College, Jorhat, Assam, India
-----------------------------------------------------------------------***--------------------------------------------------------------------
Abstract - The detection of accurate forgery in digital
images plays an exceedingly substantial role in the field of
forensics and medical forgery. As forgery in images is a
sensitive issue, utmost security and caution must be
maintained while safeguarding it. This paper proffers an
approach in order to detectauthenticandtampered imagesby
incipiently implementing the LBP (Local Binary Pattern)
descriptor on the image and then the HOG (Histogram of
Oriented Gradients) descriptor is applied ontheextractedLBP
features and finally they are classified into the two different
classes: Authentic Images and TamperedImages, adoptingthe
Support Vector Machine (SVM). The structured model
implemented on the CASIA 1 and CASIA 2 databases signifies
92.3% and 96.1% rate of detection respectively. The time
complexity is also considerably reduced and the method is
found to be functioning well under diverse illuminating
variation conditions.
Key Words: Digital Image Forgery, Digital Image
Authentication, Local Binary Patterns, Histogram of
Oriented Gradients, Support Vector Machine, CASIA.
1. INTRODUCTION
In the recent times, the interest in digital imageforensicshas
grown exponentially with the ingress of new sophisticated
cameras, smartphones and tablets. Social media such as
WhatsApp, Facebook have further contributed to their
extensive applications. Consequently, the editing tools like
Photoshop, Picasa, and smartphone beautify apps for
digitally manipulating digital images have also increased
greatly, hence making it very trivial to easily manipulate the
images. Several methods have been avowed to discern
diverse manipulations in digital images, that encompasses
resampling detection, detection of copy-move, object
removal and splicing, machine learning, JPEG artifacts
differentiating and deep learningtechniques[1].In[2],Farid
and Popescu proposed a statistical method to determine
fragments of image resampling and discrete cosine
transform (DCT). Further, convolutional neural network
(CNN), that shows excellent results in computer vision
problems, has also been applied to detect various image
distortions. Stamm and Bayar in [3] displayed excellent
performance with CNN in detecting various image
manipulation. Cozzolino et al. in [4], subsequently put
forwarded a system network that provide significant
detection performance in image forensics while making use
of a small image training set through a local descriptor.
These techniques, however, only apply to certain image
manipulations and file formats and are limited to detecting
manipulations in uncompressed images and do not reflect
the real-world image forgery, where various image
manipulations and superimposed compression occur. No
previous study, as far as our knowledge and studies is
concerned, has considered fusion of LBP and HOG
descriptors in digital image forgeries even though they are
widely used in Recognition of Facial Expression (FER) task
and Face Recognition (FR) task [5]. In this paper, we
contemplate fusing the LBP-HOG descriptors followed by
Linear SVM [5] and performed on the CASIA 1 and CASIA 2
datasets [6] whichoutperformsmanyexistingmethods[7,8].
The system in [7] works on an image noise map, by applying
a regression filter on the image noise map, and then feeds
the output to the fusion of support-vector-machine (SVM)
based and extreme-learning (ELM) based classifiers. Kim
et.al in [9] proposed a two-stream neural network approach
for image forensics, based on Convolutional Neural Network
(CNN) and Markov characteristics.
The rest of the paper being systematized as follows. Section
2 illustrates our proposed method. Results and their
Discussions are narrated in Section 3. Conclusion being
asserted in Section 4.
2. PROPOSED METHODOLOGY
The system weproposed, encompassthreefunctionalblocks.
The elemental function is to select the input frame from the
database. The succeeding function block is to acquire the
features using LBP descriptor and then applying HOG on the
LBP extracted features. The SVM classifier training and
classification constitute the endmost function step. Figure.1
shows schematic overview of our proposed system for
detection of forgery in digital images.

Figure -1: The Proposed Method of Forgery Detection
2.1 LOCAL BINARY PATTERN
Local Binary Pattern (LBP) [10] is primarily a texture
measure operator which is gray-scale invariant. Initially, it
was found to be a powerful feature for texture classification,
but its application became widespread also in the fields of
facial expression recognition, face recognition, etc. In LBP
based descriptions, features can be efficiently derived and
then combining the textures, the shape and all the dynamics
in a featurevector. The first step of LBP descriptor modelling
is converting the image to the grayscale. For every grayscale
image pixel, the LBP operator, thresholding the 3×3
neighborhood with the center value, forms certain labels for
the image pixels and considering a binary number as the
result. The LBP coding of a pixel is done as following
equation:
.
The binary encoding function s(x) is defined as follows:
When the center pixel’s value is larger than the
neighbor’s value, "0" is assigned, otherwise, a "1" is being
assigned to it. For calculating the center pixel’sLBPvalue,we
can proceed in clockwise or counter-clockwise from any
arbitrary neighboring pixel. This results in a binary number
of 8-digit, converted to decimal. The endmost step is
histogram computing over the LBP array. In our structured
method, weused a 3 × 3 LBPoperator on the inputimageand
we obtain a feature vector.
2.2 HISTOGRAM OF ORIENTED GRADIENTS
Since HOG features are quite sensitive to object
deformations, we propose to apply it. In 2005, Dalal and
Triggs proposed “Histogram of Oriented Gradients” (HOG)
[11]. HOG [11], is widely used in various object detection
fields, predominantly in detection task of pedestrian.
HOG computes in an image the emergence of the gradient
orientation in a local patch. The idea is to evaluate the local
intensity and orientation distribution that could depict the
local object shape and appearance [11]. The HOG can be
computed in the basic five steps as follows:
1. Gradient Computation: At first, the LBP encodedimage
is being convoluted with the 1-D Sobel filters [-1, 0, 1] and
for forming respectively the horizontal gradient
map and vertical gradient map.
2. Magnitude and Orientation Computation: Using the
horizontal and vertical gradient maps, we can compute the
magnitude and orientation map.
3. Image division: The features are divided into cells of
size 10×10 pixels.
4. Quantization: Each cell orientation value is being
quantized in histogram form using 8 orientation bins, where
the values of magnitude represent voted weights.
5. Normalization: A block forms from four adjacent cells
with each block having 50% overlapping with the adjacent
one. The orientation histograms of each of the blocks are
being normalizedlocallyandthenconcatenatedintoafeature
vector of 1-D. We resize the LBP feature vector to 50% of its
original size and repeat the HOG steps on them. After thatwe
concatenatethe both the obtained HOG features and thenwe
normalize them.
2.3 SUPPORT VECTOR MACHINE
Support Vector Machine [12] finds wide applications in the
domain of pattern recognitions. In separating the various
classes, the SVM achieves a separation level which is near
optimum. In our study, by using the extracted features we
train the SVM to perform the classification of the images i.e.
Authentic or Tampered. The SVM, in general, separates the
high-dimensional space by building a hyperplane. As the
interspace between the data of training-set of any of the
classes and the hyper plane is the largest, it is termed as the
ideal separation. Given labelled samples of a training set:
The SVM tries to acquire a hyper plane that distinguishes
the samples having the smallest errors.
For an input vector, the classification process is achieved
by evaluating the distance from , the input vector to the
hyper plane. The actual SVM is a binary classifier. However,
forperforming the multi-class classification,herewetakethe
one-against-rest strategy.
3. EXPERIMENTAL RESULTS
In order to appraise our proposed approach, we make
utilization of the two customarilyuseddatasets:TheCASIA1
and CASIA 2 databases [8].

Figure -2: Accuracy Plot for CASIA-1 Dataset
The images here are resized to 256×256. When the input
image is fed, at first, we apply the 5×5 median filter.
Subsequently the techniques of the proposed method are
applied for acquiring the features. The CASIA 1 and CASIA 2
datasets contains images of different subjects like animal,
art, archaeology, nature, etc. of two classes: Authentic and
Tampered images. Of all the various subjects of the two
datasets, from each one of the datasets of Authentic and
Tampered images, we selected randomly the images of each
of the two categories to constitute the training set and also
the test dataset. Finally, in the training set, there were 321
images and the test data set have 58 images. Then, we train
the SVM on the train dataset and subsequently check the
result by enquiring an image from the test datasettocheck if
the enquired image is correctly recognized. The plot of the
confusion matrix of CASIA 1 and CASIA 2 are shown in Fig.2
and Fig.3 respectively.
Table-1: Comparison Table
Table-1 gives a comparative study of our structured
proposed method to existing methods. Our method worked
siginificantly and has a high accurate rate of detection
irrespective of rotation and scaling.
Figure -3: Accuracy Plot for CASIA-2 Dataset
4. CONCLUSION
We propose an effective method in this paper to handle the
problem of Digital Image Forgery Detection. Forgery in
digital images like copy-move forgery, leaves little traces of
spikes or edges. These subtle spikes or changes are
described by the LBP and HOG features, that are very
sensitive to the shape of objects. The features are then used
to train an SVM classifier. Experimental results ontheCASIA
1 and CASIA 2 database signifies that the proposed method
consummates a significantly good accuracy, the accuracy
being respectively 92.3% and 96.1%. Also, the method we
proposed here consumes much lesser time. The task of
forgery detection is very exacting. More efforts to improve
the recognition accuracy is to be necessitated for vital
applications. Our future work objective is improving the
performance accuracy of the method in the medical images
and in multimedia (video) applications.
REFERENCES
[1] M. A. Qureshi and M. Deriche, “A Bibliography of Pixel-
Based Blind Image Forgery Detection Techniques,”
Signal Processing: ImageCommunication,vol.39,PartA,
Nov. 2015, pp. 46–74
METHOD RECOGNITION RATE
SVM + ELM [7] 84.3%
FCID [8] 85.37%
CNN + Markov [9] 86.11%
PROPOSED METHOD 96.1%

[2] Popescu, A.C., and Farid, H.: ‘Exposing digital forgeries
by detecting traces of resampling’, Trans. Signal
Process., 2005, 53, pp. 758–767
[3] Bayar, B., and Stamm, M. C.: ‘A deeplearningapproachto
universal image manipulation detection using a new
convolutional layer’. Proc. of the 4th ACM Workshop on
Information Hiding and MultimediaSecurity, June2016,
pp. 5–10
[4] Cozzolino, D., Poggi, G., and Verdoliva, L.: ‘Recasting
residual-based local descriptors asconvolutional neural
networks: an application to image forgery detection’,
arXiv preprint arXiv:1703.04615, 2017
[5] Kaushik, Manish Shankar, and Aditya Bihar Kandali.
"Recognition of facial expressions extracting salient
features using local binary patterns and histogram of
oriented gradients." In 2017 International Conference
on Energy, Communication, Data Analytics and Soft
Computing (ICECDS), pp. 1201-1205. IEEE, 2017.
[6] J. Dong and W. Wang, “CASIATamperedImageDetection
Evaluation (TIDE) Database, v1.0, and v2.0,” 2011;
http:// forensics.idealtest. org/
[7] Ghoneim, A., Muhammad, G., Amin, S. U., & Gupta, B.
(2018). Medical Image Forgery Detection for Smart
Healthcare. IEEE Communications Magazine, 56(4), 33-
37.
[8] Guo, Yuanfang, Xiaochun Cao, Wei Zhang, andRuiWang.
"Fake Colorized Image Detection." IEEETransactionson
Information Forensics and Security 13, no. 8 (2018):
1932-1944.
[9] Kim, H-G., J-S. Park, D-G. Kim, and H-K.Lee."Two-stream
neural networks to detect manipulation of JPEG
compressed images." Electronics Letters 54, no. 6
(2018): 354-355.
[10] T. Ahonen, A. Hadid, and M. Pietikainen, “Face
Description with Local Binary Patterns: Application to
Face Recognition,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 28, 2006, pp. 2037–41
[11] N. Dalal and B. Triggs, Histograms of oriented gradients
for human detection," in Computer Vision and Pattern
Recognition, 2005. IEEE Computer Society Conference
on, vol. 1, pp. 886{893, IEEE, 2005.
[12] C.-W. Hsu, C.-C. Chang, C.-J. Lin, et al., A practical guide
to support vector classification," 2003. “Title of paper if
known,” unpublished.
BIOGRAPHIES
Manish Shankar Kaushik
Research Scholar, Department of
Electrical Engineering, Jorhat
Engineering College, Jorhat
Research area includes: Digital
ImageProcessing,ImageForensics,
Computer Vision Application, Face
Recognition, Facial Expression
Recognition.
Rabul Saikia
Research Scholar, Department of
Research area includes: Digital
imageprocessing,Medical Imaging,
Computer Vision Application, Face
Recognition, Facial Expression
Recognition
Dr. Aditya Bihar Kandali
Associate Professor, Dept. of
Research area includes: Speech
Processing, Digital Image
Processing, Computer Vision,
Artificial Intelligence, Neural
Network, Machine Learning

IRJET- Digital Image Forgery Detection using Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG)

Recommended

More Related Content

What's hot (20)

Similar to IRJET- Digital Image Forgery Detection using Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG) (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Digital Image Forgery Detection using Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG)