[VOLUME 6 I ISSUE 2 I APRIL – JUNE 2019] e ISSN 2348 –1269, Print ISSN 2349-5138

http://ijrar.com/ Cosmos Impact Factor 4.236

Enriching Fake Colorized Image Detection

1Neetu Pillai, 2Dr.Ashok Kanthe, 3Srijita Bhattacharjee
1,2,3Department of Computer Engineering, Pillai HOC College of Engineering and Technology,
Maharashtra, India

Received: March 05, 2019 Accepted: April 07, 2019

ABSTRACT: : As the social media grows, the use of the fake images are at their highest, more than 50% of
the images stored at the data centers or in the cloud are being morphed or fake. So, this eventually
increases the risk of storage and also processing complexity of the images. Very few systems exists to tell
whether images are fake or not with not so great accuracy. The Neural network being the most advanced
of all, identifies fake images by using different features of the images and by using the learning process of
the pattern of faking an image. Proposed methodology uses features like Hue, alpha channel and RGB
channel. Edges are evaluated using the Gaussian distribution to identify the features of the fake colorized
images. Convolution neural network perform deep forensic analysis along with the fuzzy classification
process to enhance the probability of identifying fake images.

Key Words: Gaussian Distribution model, Image Features, Convolution neural network.

I.INTRODUCTION
People say that a picture is worth a thousand words, which holds truer than ever, owing to the
digital age and the enormous influx of cameras and Smartphone that are equipped with the latest sensor
hardware. These factors contribute to an exponential growth in the number of photographs. As the number
of images keeps increasing, there is also an influx of an almost equal number of forged images. We live in an
age where anything can be altered or manipulated with assistance from technology. The forged images are
designed with nefarious purposes to discredit the original owner by persons with malicious intents.
Due to a large number of forged images, it is necessary to be able to discern an original image from
the forged ones. The researchers infer that the two most widely used techniques of forgery are Copy-Move
forgery and Image Splicing. The copy-move method is more common in comparison to the Image splicing
technique. With the introduction of innovative image editing applications such as Photoshop, GIMP etc. the
process of forgery has been simplified and introduce modifications discreetly. Therefore, the discovery of
forged images is paramount to maintain the integrity and authenticity of the original image.
Machine Learning imitates human-like response, this is necessary for certain applications where
reasoning is heavily required for the process to be completed. This is usually achieved with the help of
certain computational networks known as neural networks. Machine learning could be fed data about the
types of image forgery techniques, as it learns, with every picture, the algorithm gets faster and a lot more
accurate. Once the relationship in-between all the system components is completely established, the
machine learning algorithm is stopped. Its addition into the forged image detection can be very useful as a
machine would not be biased and a person cannot go through a large set of images efficiently. Machine
learning can help identify the patterns in the image to verify its validity. Application of machine learnin g into
the image forgery speeds up the process and generally out-performs most of the conventional methods.
Colorization of grayscale images is not a complex task for the creative human brain. A person is only
required to recall that grass is green and the sky is blue, for various other objects, the mind is free to
imagine several possible colors. Deep CNN can assist as tools to consolidate localization and semantic
parsing into a colorization system. The researchers have done an extensive study on the image forgery
techniques and their detections and provide us with an exhaustive evaluation of the detection methods
utilized in determining the type of forgery, copy-move, re-sampling, splicing, retouching etc. For the purpose
of identification of fake images, it is imperative that we look for the sources of the

II.LITERATURE REVIEW
J. Li[1] introduces a method of the image with CMF consists of no less than a few areas in which
contents are indistinguishable. CMF maybe conducted by a forger pursuing to enhance the visual effect of
the image or to cover the truth. In this light, the researchers propose to segment the test image into a
number of non-overlapped patches. Therefore, by coinciding these patches CMFD can be executed , as long
Research Paper IJRAR- International Journal of Research and Analytical Reviews 255𝗑
[ VOLUME 6 I ISSUE 2 I APRIL– JUNE 2019] E ISSN 2348 –1269, PRINT ISSN 2349-5138
as the copying source regions and the pasting target does not exist in the same patch. This paper mainly
focuses on the detection of the segment copied from one image into another image for the purpose of faking
an image. Here a source image is needed to identify the forgery. This paper proposed a CMFD strategy
established on image segmentation. The proposed method is very computationally taxing and could be done
in a better way with the integration of parallel processing.
G. Larsson[2] narrates about Colorization of grayscale images. The technique of automatic
colorization transforms two impulsive observations into a design philosophy. Firstly, semantic information
matters. In order to colorize random images, a system must perceive the semantic configuration of the scene
in addition to localizing objects. Deep CNNs can assist as tools to consolidate localization and semantic
parsing into a colorization system. The article presents a system that demonstrates the conventional ability
to automatically color grayscale images. Two innovative offerings enable this development: a color
histogram prediction framework and a deep neural architecture that is trained to integrate semantically
significant features of diverse complexity into colorization.
M.A. Qureshi[3] explores the advent of digital images and their proliferation in this new internet
age. The extensive use of smartphones and cameras has led to an abundance of images which in turn has
increased the probability of persons with mischievous intents to tamper and forge images by leaving very
little or no trace. The researchers have done an extensive study on the image forgery techniques and their
detections and provide us with an exhaustive evaluation of the detection methods utilized in determining
the type of forgery, copy-move, resampling, splicing, retouching etc. The researchers after having done an
extensive evaluation of the techniques have come across an observation that the researchers who have been
actively developing methods and algorithms for tampering detection have not been paying attention to the
research that goes into counter-forensics or anti-forensics. This area of research focuses on the explicit
removal of tampering traces been sought after by these forgery detection methods.
D. Gragnaniello [4] explains that due to the inflation of social media as it turns into a quite powerful
communication tool. Due to the boom of social media, there has been an enormous influx of manipulated
images. This has led to an epidemic and there are researchers that have been trying to combat this problem
with the inclusion of CNN's in their detectors. As recent studies have shown that CNN based detection
methods are vulnerable to attacks. Therefore, to ameliorate this effect the researchers have analyzed the
adversarial attacks done on various different CNN based detection systems and tested their performance
adequately. The study concentrates only on the attacks that take place on a CNN based network, which is
very limiting and have not proposed any remedial systems that can be used in the absence of a CNN based
approach.
Y. Guo [5] introduces the different types of forgeries that can be done on an image and the various
ways that can be used to detect the fakes. But this does not take into account the colorization of grayscale
images. This is difficult to detect as there aren’t many techniques that have been developed for the detection
of fake colorization of the images. Therefore, to detect fake colorized image the authors proposed two
different methods, the first one is the FCID-HIST Histogram based Fake Colorized Image Detection, and the
other one is the FCID-FE which stands for Feature Encoding based Fake Colorized Image Detection. These
techniques have been proven to be better than the traditional detection methods by a big margin. As with
any learning method, most of the detection depends on the kind of training given to the system and is
limited by that perspective, as it cannot detect forgery even when the image is forged, but the system is
limited to its understanding.

III.PROPOSED METHODOLOGY
The proposed idea of finding fake colorized images is depicted in the below proposed methodology
figure 1. And the steps involved in this process are elaborated deeply with the below mentioned steps.

256𝗑 IJRAR- International Journal of Research and Analytical Reviews Research Paper
[VOLUME 6 I ISSUE 2 I APRIL – JUNE 2019] e ISSN 2348 –1269, Print ISSN 2349-5138
http://ijrar.com/ Cosmos Impact Factor 4.236

Figure 1: Over view of the Proposed Idea

Step 1: Feature Extraction - This is the very first and very important step of the proposed model.
Here the base features are like Hue, alpha channel and RGB values are extracted from the input image and
process in a list.
To fetch these features proposed system take input to traverse each and every pixel to get a signed integer
(PSIGN) . Then on right shifting 24 bits and then AND one hexadecimal value yields Alpha channel value. Right
shifting 16,8 and 0 bits and then AND one hexadecimal value yields the Red, Green and blue channels
respectively. This can be shown in the equation (1,2,3,4)

α =PSIGN >>24 AND HEX _______(1)

R =PSIGN >>16 AND HEX ______(2)

G=PSIGN >>8 AND HEX ________(3)

B =PSIGN >>0 AND HEX _______(4)

Where
PSIGN - A Singed Integer
HEX -Hexadecimal Number

Hue Feature- It indicates the frequency of the light that eventually is represented by a single
variable. Hue can be estimated using the ratio of difference between the other two colors for maximum of
RGB to the difference between max and min.
Histogram feature - The image Intensity is adjusted based on the histogram of the pixels, which can
be evaluated using the equation 5.

PN=No of pixels with intensity n / Total no of Pixels __ (5)

Where
PN - Image with N pixels

Step 2: Edge Estimation - If an image is morphed or colorized, then the edges of the objects in the
image are always tend behave out of the natural values. These edges are eventually helping us to identify the
border area of the objects in the image.

Edge Detection is used for this purpose, which works based on the 5 following steps.
i) Smoothing - Here image is blurred using the Gaussian blur technique to widen the edges of the objects in
the image. For this Gaussian distribution technique is used as shown in equation no 6.
Research Paper IJRAR- International Journal of Research and Analytical Reviews 257𝗑
[ VOLUME 6 I ISSUE 2 I APRIL– JUNE 2019] E ISSN 2348 –1269, PRINT ISSN 2349-5138
ii) Finding the gradients- Here the edges with the maximum width is identified.
iii) Non- maximum Suppression - Edges with the maximum width is labeled.
iv) Double Thresholding - Here the edges which are already labeled are doubled by their width.
v) Edge Tracking- Here all the threshold edges are traversed recursively to mark them in white color and
rests of the pixels are marked with the black color.
2
x
1 2 2
f (x | , ) e _________ (6)
2

Where µ is mean, xi instance pixel value,𝛿 is the Standard deviation

Step 3: First Layer of the CNN ( Edge Traversing) - Here this is the first layer of CNN wherethe
edge image is taken into consideration to traverse on only edges and record their luminance with respect to
the orignal image. This is because if the image is natural then the luminance of the image is undisturbed by
the external colors. If it is edited or morphed to create its fake version, then it is disturbed in many locations.
This process can be depicted in below algorithm.
______________________________________________
ALGORITHM: Edge Traversing
______________________________________________
// Input: Edge Image I E, Original Image OI
// Output: Pixel List PL
//TL :Temporary List
// ISIGN: Signed Integer
1: Start
2: For i = 0 to size of Width of IE
3: For j=0 to size of Height of IE
4: ISIGN =IE(ij) RGB
5: R= ISIGN >> 16 & HD
6: G= ISIGN >> 8 & HD
7: B= ISIGN >> 0 & HD
8: IF( R==255 && G==255 && B==255),THEN
9: LU→luminance(ROI,GOI,BOI)
10: TL → I,J,LU
11: ADD TL to PL
12: End IF
13 :End for
14:End for
15: returnPL
16: Stop
____________________________________________

Step 4: Second Layer of the CNN ( Edge Traversing) - Here in this step once the Edge image pixel
position for white pixel is recorded as I,J. then, with respect to this I and J RGB of the original image is
extracted to evaluate the luminance based on the equation no 6.

Luminance= 0.299 * R + 0.587 * G + 0.114* B _____(7)

Luminance is the component that represents the amount of white light on the pixel, For any natural
image luminance of the edges of the images are always remain intact. For any forgery in the image ,
luminance keep changing on the edges.
Step 5: Third Layer of the CNN (Luminance Correlation) - Here in this step a correlation array is
created in the edge pixels based on the values of luminance and the histogram. Then another correlation
array is created with respect to the trained images stored in the database for the luminance and histogram.
Then these two arrays are fed to Pearson correlation estimation technique, which eventually
generates a decimal value in between the 0 and 1 as shown in equation 8. This correlation value is subjected
to classify the image into fake or original based on the Fuzzy classification model.

258𝗑 IJRAR- International Journal of Research and Analytical Reviews Research Paper
[VOLUME 6 I ISSUE 2 I APRIL – JUNE 2019] e ISSN 2348 –1269, Print ISSN 2349-5138
http://ijrar.com/ Cosmos Impact Factor 4.236

______________(8)

Where
x is the entities of Query image
y is the entities of trained images from database
n is the array Size
Step 6: Fuzzy Classification- Here in this step a correlation array result which is in the range of 0 to
1 is divided into 5 parts with respect to the Fuzzy crisp values. Fuzzy Crisp values are VERY LOW, LOW,
MEDIUM , HIGH AND VERY HIGH.
A fuzzyfier engine is used to classify the value for the fake or original based on the classification
label of crisp values. Any value which is nearer to 1 is rated as thevery highly fake . Where as the any value
which is nearer to 0 is rated as non fake or original.
IV. RESULTS AND DISCUSSIONS
In this section the proposed system of fake image detection using the features like histogram and
CNN, which is boosted using fuzzy classification is subjected to some experiments to evaluate its
performance with comparison of same existed methodology. The proposed model is deployed in real time
scenario using Java programming language and Netbeans as the IDE in Core i5 Environment with the
primary memory of 6GB.
Histogram Equalization Result (HTER)-Proposed model is put under hammer for HTER with the
comparison of the two different methods are used to identify the fake colorized images as FCID-HIST and
FCID - FE[5].
FCID-HIST: Here in this step four features are being extracted from the images like Hue, Saturation,
Dark Channel, and Bright Channel. Once these features are being extracted, then they are aggregated to
normalize the histogram. And then the proposed methodology deploys the features into Support vector
machine (SVM) to get the classification values which are being used to classify the fake colored images.
FCID - FE: Here in this step again four features are being extracted as mentioned in FCID-HIST and
then these features are subject to identify the distribution of data using Gaussian mixture model and
exploring the divergence inside the different moments of the distribution. And finally Gaussian mixture
model is blended with the SVM training Classifier. Where LIBSVM Classification protocol is being used to
identify the Fake images.
Cross Validations are performed on the basis of histogram role to identify the fake images on FCID-
HIST, FCID-FE of [5] and with the proposed model of FCID- CNN on various image folders belong different
categories like sports, nature, sky etc. The table 1 represents the cross validation result between the FCID-
HIST, FCID -FE and FCID-CNN.
Table 1 : HTER Results of Cross Validation in Percentage

On observing the values and the plot in figure 2 and 3 it clearly indicates the fact that FCID -CNN provides
better performance than that of FCID-HIST and FCID-FE. Where average HTER of FCID -HIST is around
18.423%, Average HTER of FCID-FE is around 16.994 % and average FCID-CCN is about 14.47%.
Research Paper IJRAR- International Journal of Research and Analytical Reviews 259𝗑
[ VOLUME 6 I ISSUE 2 I APRIL– JUNE 2019] E ISSN 2348 –1269, PRINT ISSN 2349-5138

Figure 2: HTER Comparisons of Cross Validation between FCID- HIST and FCID-CNN

Figure 3: HTER Comparisons of Cross Validation between FCID- FE and FCID-CNN

On observing some facts on FCID-HIST in [5] it indicates that the using SVM for all different types of images
creates different bins of distinctive values so it is not performing well than that of FCID-FE. On the other
hand FCID-FE out performs than that of FCID-HIST because of the moments based feature extraction using
Gaussian mixture model. But on the other side FCID-CNN shows more improved results than that of FCID-FE
due to using of Gaussian kernel model and CNN, which enhances the process with better average result of
14.47%.
V. CONCLUSION
This research paper properly scrutinizes the different features of the images like hue, saturation, dark
channel, bright channel and an alpha channel. Then Distribution factors of all these features are being
estimated using a Gaussian distribution model for the edges of the images, which indeed the plays a vital
role in identifying morphed and fake colorized images.
Then based on this distribution factor the convolution neural network efficiently handles the fake image
detection process based on the luminance factor of the image edges. The Correlation between the luminance
of the real edges and the fake edges are being estimated using Pearson correlation model. To efficiently
classify the Pearson correlation fuzzy classification model is being used to enrich the model.
The Histogram equalization results are being evaluated in between FCID-FE and FCID-HIST of [5], where the
proposed model's concept of FCID- CNN performs better by 4% compared to FCID-HIST and 2.5% compared
to FCID_FE. In the future this fake image detection process can be enhanced to more accurately by
considering more vigorous features like DCT and wavelet transform techniques to handle more deep High
definition images of space and etc.

260𝗑 IJRAR- International Journal of Research and Analytical Reviews Research Paper
[VOLUME 6 I ISSUE 2 I APRIL – JUNE 2019] e ISSN 2348 –1269, Print ISSN 2349-5138
http://ijrar.com/ Cosmos Impact Factor 4.236
VI. REFERENCES
1. J. Li, X. Li, B. Yang,and X. Sun, “Segmentation-Based Image Copy-MoveForgery Detection Scheme,” IEEE Trans.
Inf. Forensics and Security, vol.10, no. 3, pp. 507-518, 2015.
2. G. Larsson, M. Maire,and G. Shakhnarovich, “Learning representations for automatic colorization,” in Proc.
European Conf. Comp. Vision (ECCV), pp. 577-593, 2016.
3. M.A. Qureshi and M. Deriche, “A bibliography of pixel-based blind image forgery detection techniques,” Signal
Process.: Image Communication,vol. 39, pp. 46-74, 2015.
4. Diego Gragnaniello, Francesco Marra, Giovanni Poggi, Luisa Verdoliva, “Analysis of Adversarial Attacks against
CNN-based Image Forgery Detectors”, 26th European Signal Processing Conference. 2018.
5. Y. Guo, X. Cao, W. Zhang,and Rui Wang, “Fake Colorized Image Detection”, IEEE TRANSACTIONS ON
INFORMATION FORENSICS AND SECURITY, 2018
6. Y. Zhao, X. Zhang, M. Xu, Z. Sun and G. Liu, “Web Identification Image Recognition Based on Deep Learning”, 3rd
International Conference on Information Science and Control Engineering, 2018.
7. M. Bachtiar, D. Gusti, M. Hidajat and I. Wijaya, “Web-Based Application Development for False Images Detection
for Multi Images Through Demosaicing Detection”, International Conference on Information Management and
Technology, 2018.
8. Tushar D. Gadhiya, Anil K. Roy, Suman K. Mitra,and Vinod Mall, “Use of Discrete Wavelet Transform Method for
Detection and Localization of Tampering in a Digital Medical Image”, IEEE Region 10 Symposium, 2017.

Research Paper IJRAR- International Journal of Research and Analytical Reviews 261𝗑