Available online at www.sciencedirect.com
ScienceDirect Available online at www.sciencedirect.com Procedia Computer Science 00 (20 (2017) 000–000
ScienceDirect
www.elsevier.com/locate/procedia
Procedia Computer Science 115 (2017) 501–508
7th International Conference on Advances in Computing & Communications, ICACC ICACC--2017, 2224 August 2017, Cochin, India
Image Inconsistency Detection Using Local Binary Pattern (LBP) Vivek H. Mahalea,*,MouadM.H.Alib,Pravin L.Yannawarc, Ashok T. Gaikwadd a,b c
Research Scholar at Department of CS & IT, Dr.BabasahebAmbedkarMarathwada University, Aurangabad, (M.S),431001, India Vison and Intelligent System Lab, Department of CS & IT, Dr.BabasahebAmbedkarMarathwada University, Aurangabad,(MS) India d Institute of Management Studies and Information Technology, Aurangabad, (M.S), (M.S),,431001, India
Abstract Day to day Digital Image has widely increased popularity in Human life. People edit ima image ge with the help of editing tools or software for malicious intent. This work is to identify inconsistency in an image. The paper contains different steps suchas such preprocessing, feature extraction, and matching process, which is highlights effective use of local binary pattern method for feature extraction mechanism. Euclidean distance is exploited for matching measures. The result obtained exhibits that LBP with 2x2 block size gives the best result with accuracy reach to 98.58 % for automatic detection off inconsistencies in an image. © 2017 The Authors. Published by Elsevier B.V. review under responsibility of the scientific committee of the 7th International Conference on Advances in Computing & Peer-review Communications. Keywords:Inconsistency ;LBP; FeatureExtraction; Matching atching;
1. Introduction Recent days Digital Image has decreased its pureness due to inception of massive operational flexibility incorporated in modification of digital images through image editing software’s such as sophisticated photo editing editin tools, this manipulation of image hasbecome more common. Manipulation of digital image in image analysis deals with many questions [1] like: render. • Is the image is true, digitally enhance or computer render • If the image is true, actual details of an image.
* VivekH. Mahale Tel.: +919028042502 E-mail address:
[email protected] 1877-0509© © 2017 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the scientific committee of the 7th International Conference on Advances in Computing & Communications. 1877-0509 © 2017 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 7th International Conference on Advances in Computing & Communications 10.1016/j.procs.2017.09.097
Vivek H. Mahale et al. / Procedia Computer Science 115 (2017) 501–508 Mahale V. H et al/ Procedia Computer Science 00 (2017) 000–000
502 2
• •
If the image is digitally raised, what was the manipulated and how was manipulation executed. If the image is computer render, how was the image produce.
In order to address aforesaid questions researchers have put their efforts towards design of different mechanism towards examination of image inconsistencies and these method were broadly classified under three classes as, • • •
Observation: Sometimes inconsistent image can be identified through direct observation i.e. specular highlights and shadows, colour tones anti-aliasing, reflections, scale. Basic image enhancement: Through an algorithms i.e. scaling, sharpening, re colouring, blurring, attributes with image can be made discrete. Advance Image analysis : This involves PCA , wavelet and , LBP, and light direction (i.e. gradient)
Copy-Move image inconsistency is evil and normally usedbylarge population of users for various purposes, it may be for authorized or unauthorized. In case such operations using Copy-Move operation the forgery in image datatakes place where some part of image information is copied and pasted to another place in same image and human eyes could not identify these changes incorporated in image and this mechanism itself makes this task challenging and critical.Farid[2] proposed somemethods about detect image forgery. There are two types method for image forgery detection, first is active [3], [4], [5]and second is passive or blind [5], [6], [7], [8]. In active forgery detection method there is need of prior information of image. Where as in case of passive forgery detection method, there is no need for prior information. Figure 1 shows methods involved image forgery detection and its types.
Fig.1 Types of Image forgery detection.
Active image forgery method is categorized in the two type [9] i.e. digital signatures and digital watermarking. Where as in case of passive method it iscategorised in two types i.e.forgery dependent and forgery independent. The proposed detection method discussed in this research paper is categorizedunder forgery dependent type i.e. Copy-Move forgery detection [10]. Forgery type independent further categories again into two typesi.e. retouchingdetection andlightingconditions.This paper is organised into five sections, where section 1 provided introduction of the problem beingintegratedSection 2 gives the related work of Image forgery detection; Section 3 introduces the methodology of the system; Section 4 gives experimental results. Finally, Section 5 gives the conclusions and future work.
Vivek H. Mahale et al. / Procedia Computer Science 115 (2017) 501–508 Mahale V. H et al / Procedia Computer Science 00 (2017) 000–000
503 3
2. Related Work Now a day, there has been so much study going on in image inconsistency detection, Motasemet. al.[11]has put forwarded copy-move image forgery detection using Local Binary Pattern and Neighbourhood clustering. Forgery detection using LBP involves on different Color information of an image to detect forgery. They acquired dataset from CASIA TIDE V2.0 for evaluation purpose. To detect forgery using LBP they first input image is segmented into Red, Green and Blue Color components. Then components were divided into 20x20 overlapping block and calculate LBP of each block. Then calculate distance between each block pair according to low distance, and then they do neighbourhood clustering. In experiment author wrote that to bring forth accuracy original image is subtract from forged. Muhhamad et al. [12] proposed Passive copy move image forgery detection using undecimated dyadic wavelet transform, in this first apply DyWT on all images. Then divide into LL1 and HH1 sub bands. Then LL1 and HH1 further divided into overlapping blocks. Then calculate Euclidean distance between every pair of blocks, and then sort it in ascending order. Then with the threshold value match the blocks. If matching, copy move forgery takes place. They compare result with li et al. [13] which use DWT and LL and another with Mahadian and saic [14] that use DWT and HH1. There experiment results with False Positive rate as 4.02 and False Negative rate as 6.35 in copy move forgery without rotation image and with rotation False Positive rate is 3.52 and False Negative rate is 6.92. AmaniAlahmadi et.al. [15] Presented passive detection of image forgery using DCT and local binary pattern. They work on CASIA-1, CASIA-2 and Columbia datasets. In their method they did pre-processing, feature extraction, classification and evaluation. In pre-processing, then convert RGB to YCbCr image. In feature extraction YCbCr image divide into block and then apply LBP and DCT then convert set of data in training and testing set. SVM is used for classification. They evaluate their method on three dataset. They calculate True positive rate (TPR) and True Negative rate (TNR) and depend on TPR and TNR calculate accuracy. They got accuracy as 97.00%, 97.50% and 97.77% onCASIA-1,CASIA-2 and Columbia dataset respectively . 3. Methodology The methodology of this system is utilized for detecting image inconsistency which is followed by different mechanism related to proposing and feature extraction techniques. In this proposed study, the preprocessing steps are done on COMOFOD dataset. Figure 2 presents the block diagram of methodology adopted in this proposed research work and process view of represented in Algorithm 1.
Fig.2 Diagram of methodology system
Input:Image for determination of inconsistencies check
Vivek H. Mahale et al. / Procedia Computer Science 115 (2017) 501–508 Mahale V. H et al/ Procedia Computer Science 00 (2017) 000–000
504 4
Output: Either Original or Forge image Begin: Step 1: Read Image Step 2: Convert Colour Image in to Gray scale image Step 3: Divide image into overlapping blocks Step 4: Compute LBP Features from each block Step 5: Apply Lexicographical sorting and match similar pairs of blocks. Step 6: Matched blocks are mapped to indicate forgery detection. Step 7: Draw the part which containing the forge End. Algorithm 1: Automatic Forgery detection
3.1. Preprocessing In pre-processing steps the colour image converting gray scale image using following formula: I = 0 .2 9 9 R + 0 .5 8 7 G + 0 .1 1 4 B
(1 )
Where, R, G and B refer to red, green, and blue respectively.In addition RGB separated in to Red, Green and Blue. After that image divide into overlap blocks with fixed size. To identify block which have inconsistency. This case image is divided into 2 x 2 block size. 3.2. Feature extraction The feature extraction process takes the output from pre-processing which is of different block size. In this case the local binary pattern (LBP) process was applied to each block by performing block division and preprocessing steps in first section and extracts the histogram of each block. Finally all histogram of all blocks concatenated into one histogram which represent as a feature factor. Then sorting it by lexicographical sorting for which is use for matching block by block and detect which part of image is inconsistency. The example show how the LBP work as in fig.3
Fig.3. Example of LBP Process
3.3. Matching process The matching process incorporated block wise manner to determine whether block has Inconsistency,with the help of lexicographical sorting of feature fact. Similar feature are located in different blocks. The mechanism using for matching is Euclidean distance measure. The pair wise Euclidean distance between blocks were calculated using Eq. (2) d ( b lo c k v1, b lo c k v 2 ) =
∑ nN=1 b lo c k n 1 − b lo c k n 2
(
)
2
(2 )
Eq. (2) finds the distance between two blocks from training vs. Test dataset. The result obtained through this process presents the distance equal to zero or near, it means the image does not contain any inconsistencies and represents that the image under observation is original; otherwise it is considered to be inconsistent.
Vivek H. Mahale et al. / Procedia Computer Science 115 (2017) 501–508 Mahale V. H et al / Procedia Computer Science 00 (2017) 000–000
505 5
4. Result and discussion The proposed methodis evaluated on database COMOFOD, which include different types of images forgery. The database is taken from Video Communication Laboratory (VLC), University of Zagreb, Croatia,and Departmentof Wireless Communication. The proposed method was evaluated on laptop Intel Core i3, with 4GB Ram andHardware infrastructure with MATLAB 2013, necessary toolbox. The figure 4 shows the some sample taken from COMOFOD database in to cases original and forge.
Fig.4. Show some sample taken from COMOFOD dataset
The study of image forgery detection method is evaluated using training and testing which is selected randomly from the COMOFODdatabase with the size of images as 512x512 colour image in PNG format. The forged images are already created by the owner of the database in different case either single or multiple forge operations. The advantage of this system it can work in any types of image format with different size. The implementation of pre-processing steps are resulting as shows in figure 5 with the sequence from input image up to divided image into blocks.
. Fig5. Pre-processing steps
From the experimental work the LBP feature extraction are applied for different block size namely 2x2, 4x4, 8x8, 16x16. It was get different results for each block with LBP feature. The visualization result of LBP is shows in figure 6 Example table shown below.
Fig. 6 The visualization result of LBP feature extraction
After extracting LBP feature from the image and sorting the feature by lexicographical sort. This feature is passed to matching step where comparison is done block by block which provides guideline about the portion of test
Vivek H. Mahale et al. / Procedia Computer Science 115 (2017) 501–508 Mahale V. H et al/ Procedia Computer Science 00 (2017) 000–000
506 6
image was being modified or forged based on the inconsistent image information marked by proposed algorithm. And Constructing shift vectors for all the matching block pairs and Looking for common shift vectors among matching blocks. By counting the number of shift vector in particular image, we can determine how many part as forge in the image. If the number of shift vector equal to zero that means the image is original otherwise the image contains forge part. The figure 7 shows the LBP histogram and forgery image detection with the exactly forged parts.
Fig.7. LBP histogram and forgery image detection with the exactly forge parts
For evaluation of the system the 100 images are taken and applied all the steps above and store the feature in template. The 100 images are divided randomly into training and testing set and extract the LBP feature from each image then calculate the evaluation matrix TPR and FPR with the help of threshold values. Table 1 shows the results of the system with help of TPR and FPR measure shows by formula in Eq. (3) and (4). True positiveTP =
Number of Images Detected as Forged being Forge Number of Forge Image
False PositiveFP =
Number of Images Detected as Forged being Original Number of Original Images
3 4
Table 1. Show the results evaluation of LBP for different block size Sr.No. 1 2 3 4
Database size 100 100 100 100
Block size 2x2 4x4 8x8 16x16
TPR 0.0142 0.0301 0.0517 0.0800
FPR 0.0995 0.0992 0.0994 0.0997
The evaluation system based on True positive (TP) and False Positive (FP) Rate with the help of threshold value which is generated from distance matrix from training and testing sets True positive (TP) and False Positive (FP) rate are calculated with the help of ROC (Receiver Operating Characteristics) curve which is shown in fig.8
Vivek H. Mahale et al. / Procedia Computer 115000–000 (2017) 501–508 Mahale V. H et al / Procedia Computer ScienceScience 00 (2017)
5077
ROC curve for image forgery detection 1 0.9
True Positive Rate (TPR)
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
0.1
0.2
0.3 0.4 0.5 0.6 0.7 False Positive Rate (FPR)
0.8
0.9
1
Fig.8. ROC curve of the system show the FPR and TPR
Table 2 shows comparison result of proposed method andother methods. It is clearly seen that the proposed algorithmperform as compared with Ghulam M. Et.al.[16]and Amani A.[17]. Table 2.Comparison results of between the proposed and other methods. Sr. No. 1 2 3 4 5 6 7
Methods STP & LBP[16] STP& LBP[16] STP& LBP[16] DCT&BP[17] DCT_LBP[17] DCT_LBP[17] Proposed
Dataset CASIA1 CASIA2 Columbia CASIA1 CASIA2 Columbia COMOFOD
Accuracy (%) 94.89 97.33 96.39 97.00 97.50 96.60 98.58
5. Conclusion The paper present a unique method of determination of image inconsistency and detection method based on Local Binary Pattern (LBP) technique. The method was evaluated on COMOFOD dataset. The results were obtained from the proposed method on different block base. The efficiency of the system were calculated using True positive Rate (TPR) and False Positive Rate (FPR) with database size 100 images from original and forge. The evaluation from the proposed work shows True positive Rate (TPR) and False Positive Rate (FPR) for different block size 2 x 2, 4x4, 8x8, 16x16 the TPR are 0.0142, 0.0301, 0.0517, 0.0800 respectively, while False Positive Rate (FPR) for the same blocks are 0.0995, 0.0992, 0.0994, and 0.0997 respectively. It was understood from the result that the LBP of 2x2 block size is achieved best result compare with another blocks due to its window size. The future work may be extending to combining LBP with DCT and DWT for improving the performance of system.
508 8
Vivek H. Mahale et al. / Procedia Computer Science 115 (2017) 501–508 Mahale V. H et al/ Procedia Computer Science 00 (2017) 000–000
References [1]. Krawetz, Neal. A pictures worth digital image analysis andforensics. Black Hat Briefings, 2007: 1-31. [2]. Farid, Hany. , Image forgery detection. IEEE Signal processing magazine, 2009; 26(2): 16-25. [3]. Cheddad, Abbas, Joan Condell, Kevin Curran, and Paul McKevitt., Digital image steganography: Survey and analysis of current Methods. Signal processing, 2010; 90(3): 727-752. [4]. Rey, Christian, and Jean-Luc Dugelay. A survey of watermarking algorithms for image authentication. EURASIP Journal on Advances in Signal Processing 2002; 6: 218932. [5]. Yeung, Minerva M., Digital watermarking: marking the valuable while probing the invisible. Communications of the ACM, 1998; 41(7): 3135. [6]. Lee, Jen-Chun, Chien-Ping Chang, and Wei-Kuei Chen., Detection of copy–move image forgery using histogram of orientated gradients. Information Sciences, 2015;321: 250-262. [7]. Mahdian, Babak, and StanislavSaic. Blind authentication using periodic properties of interpolation. IEEE Transactions on Information Forensics and Security, 2008; 3(3): 529-538. [8]. Zhang, Zhen, Yuan Ren, Xi-Jian Ping, Zhi-Yong He, and Shan-Zhong Zhang., A survey on passive-blind image forgery by doctor method detection.International Conference in Machine Learning and Cybernetics,2008, 6: 3463-3467. [9]. Birajdar, Gajanan K., and Vijay H. Mankar. , Digital image forgery detection using passive techniques: A survey ,Digital Investigation, 2013,10(3): 226-245. [10]. Mushtaq, Saba, and AjazHussain Mir., Digital image forgeries and passive image authentication techniques: A survey., International Journal of Advanced Science and Technology 2014; 73: 15-32. [11]. AlSawadi, Motasem, Ghulam Muhammad, Muhammad Hussain, and George Bebis., Copy-Move Image Forgery Detection Using LocalBinary Pattern and Neighborhood Clustering. In Modelling Symposium (EMS), 2013: 249-254. [12]. Muhammad, Ghulam, Muhammad Hussain, and George Bebis., Passive copy move image forgery detection using undecimated dyadic wavelet transform. Digital Investigation, 2012; 9(1): 49-57. [13]. Li, Guohui, Qiong Wu, Dan Tu, and Shaojie Sun., A sorted neighborhood approach for detecting duplicated regions in image forgeries based on DWT and SVD. IEEE International Conference onMultimedia and Expo, 2007: 1750-1753. [14]. Mahdian, Babak, and StanislavSaic.,Using noise inconsistencies for blind image forensics. Image and Vision Computing, 2009; 27(10): 1497-1503. [15]. Alahmadi, Amani, Muhammad Hussain, HatimAboalsamh, Ghulam Muhammad, George Bebis, and Hassan Mathkour., Passive detectionof image forgery using DCT and local binary pattern., Signal, Image and Video Processing , 2017;11(1): 81-88. [16]. Muhammad, Ghulam, Munner H. Al-Hammadi, Muhammad Hussain, and George Bebis., Image forgery detection using steerable pyramid transform and local binary pattern., Machine Vision and Applications, 2014; 25(4): 985-995. [17]. Alahmadi, Amani A., Muhammad Hussain, HatimAboalsamh, Ghulam Muhammad, and George Bebis., Splicing image forgery detection based on DCT and Local Binary Pattern., In Global Conference on Signal and Information Processing (GlobalSIP),2013: 253-256.
