Frequency based edge-texture feature using Otsu’s based enhanced local ternary pattern technique for digital image splicing detection

Received Jun 30, 2021 Revised Sep 10, 2021 Accepted Oct 14, 2021 Sharing information through images is a trend nowadays. Advancements in the technology and user-friendly image editing tool make easy to edit the image and spread fake news through different social networking platforms. Forged image has been generated through an advanced image editing tool, so it is very challenging for image forensics to detect the micro discrepancy which distorted the micro pattern. This paper proposes an image forensic detection technique, which implies multi-level discrete wavelet transform to implement digital image filtering. Canny edge detection technique is implemented to detect the edge of the image to implement Otsu’s based enhanced local ternary pattern (OELTP), which can detect forgery-related artifact. DWT is implemented over Cb and Cr components of the image and using edge texture to improve the Otsu global threshold, which is used to extract features using ELTP technique. Support vector machine (SVM) is used for classification to find the image is forged or not. The performance of the work evaluated on three different open available data sets CASIA v1, CASIA v2, and Columbia. Our proposed work gives better results with some of the previous states of the work in terms of detection accuracy.


INTRODUCTION
Image forgery is a widespread activity in different fields by different modes like generating fake evidence, spread fake news. The digital image is a huge source of information in various areas such as disease detection in the medical field, forensic detection, electronic media, e-commerce, and social networking sites. Millions of images are uploaded daily on WhatsApp, Facebook, and other networking sites, and many are fake. They tampered with the help of advanced image editing tools and techniques. These counterfeit images negatively affect our society. So it's required to develop such forensic detection tools with the forgery detection technique to identify image forgery over the different formats of the image.
There are two types of image forensic techniques, active and passive. Digital signature and watermarking are subpart of active, and copy-move and splicing are passive subpart. We are discussing the passive image forensic technique which is applicable on both copy-move and splicing, which is trendy but not easy to implement. It's also known as the blind image forgery detection technique. In this technique, the source of the image is unknown, and only the image is present. Due to image forgery, internal statistical information of the image gets disturbed. Detecting the forgery needs to extract disturbed internal statistics as a feature extraction, which helps to detect forgery. In this paper, we used frequency based edge texture for feature extraction technique. Through edge the performance of the optimal global threshold has been improved and the performance of the enhanced local ternary pattern (ELTP) has been improves show in result.
The tampering detection technique plays a significant role in identifying the image splicing. So many techniques were discussed last year. Some of them are discussed here, based on extraction of the various feature of the image illumination, JPEG compression, camera based property or limitations, and so on. To identify the image forgery [1], [2] the property and restrictions of the camera play an important role, it's work like a fingerprint because every camera has its own noise. They identify the differences generated by camera noise as a feature to identify the image splicing. Image forgery detected by feature fusion technique discuss in [3], where camera features and image quality are used to train them in color, quality, and frequency. The concept of color filter array (CFA) is used in [4], [5], to generate RGB through interpolation. Image tampering detected by CFA and interpolation. The best approach for pixel based feature technique is local binary pattern (LBP) [6]- [8], LBP is used for feature extraction with the help of discrete cosine transform (DCT) over the chrominance component of the image after converting the RGB image into YCbCr and derived the image into non-overlapping block. Frequency based technique is used to classified the image for forgery detection. In paper [9], [10], DCT transition probability metrics implements Markov based approach for forgery detection. Frequency-based domain generated by discrete wavelet transform (DWT) to characterize the wavelet coefficient across position, scale and orientations. Support vector machine (SVM) is used for classification. To detect blind image splicing multi-scale local binary pattern [11] is implemented over discrete cosine transform coefficient for feature extraction. K-fold cross validation technique is used to train the system. Steerable pyramid transform (SPT) [12] and LBP is implemented over the chrominance component of the image. Every texture of SPT is described by LBP histogram. Which are concatenate to generate features, and SVM is used for classification. Entropy filter and local phase quantization (LPQ) texture operator [13] is used for forgery detection. Image boundary highlighted by entropy filter and internal statistic of the image obtained by LPQ. SVM is used to classify the image. In paper [14], a traditional gray level co-occurrence material (GLCM) and LBP for texture descriptor of the image. Convolutional neural network (CNN) and multi-scale patch based recognition is used as a fisher vector. A modern hybrid frame work [15] has been proposed for existing all possible features of the image tampering to develop the best feature set. Classify the feature with the classification technique to get better result. Rotation invariant technique [16] is used to detect uniform pattern on LBP. Generalized gray scale and rotation invariant operator is used to detect the uniform reason for any quantization of angular space. It generates multiple operators for multiplex resolution. Standard deviation filter is used to highlight the image information [17]- [19], internal statistic extracted by rotation-invariant co-occurrence among adjacent local binary pattern (RIC-LBP) operator and classify to detected forgery. The concept of the ternary pattern is firstly discuss in paper [20], binary pattern (0, 1) is used for LBP and ternary pattern (0, -1, 1) used for local ternary pattern (LTP) to remove the limitations of basic LBP technique. LTP is more reliable then LBP. In paper [21], ELTP is used for enhanced the performance of the LTP technique. Features are extracted with of standard deviation filter (STD) and ELTP over chrominance component of the image. STD filter highlighted the abrupt changes of the image. SVM is used for classification to find the forgery. In paper [22] image texture features are considered to detect the edge of the image. This technique is implemented on an image where multiple faces are present. CASIA v1.0 data set for tampering detection in this data set image are stored in JPEG format. JPEG format-based technique images are compressed multiple times for tampering, quantization, and blocking artifacts. It is implemented on both frequency and special domain.

PROPOSED METHOD
Training and testing two stages are used to detect image forgery show in Figure 1. In training, pristine and forged images are used to train the classifier to detect image forgery. In testing same classifier is used to test the image is forged or not. Many techniques are used for image forgery detection. If two are more images are combined to generate a forged image is a spliced technique. A single image is used to copy and paste somewhere in the same image, known as the copy-move forgery detection technique. Some approaches are used to find the image forgery, like image compression, camera noise, and edge detection. In this paper, we proposed a technique for image splicing detection. We implement wavelet transform for digital filtering to excess the time scale depiction. The edge detection technique is implemented over wavelet transform for detecting meaningful discontinuities in intensity value. In this proposed work, texture-based image splicing detection is used because it divides into the region of interest and classifies the region. It also helps to generate a global threshold by Otsu. The ELTP technique is implemented on the image. Forgery can be detected with the help of a support vector machine, gives awe-inspiring results and easy to implement [23], over three different types of data set, CASIA v1, CASIA v2 [24], and Columbia. We implement DWT to implement frequency-based digital image filtering and excess the image edge texture to produce batter global-threshold to implement ELTP and get batter feature extraction for forgery detection. Algorithm a. Convert RGB image into the YCbCr image component. b. Chrominance component of the image is used for feature extraction. c. Frequency-based model is implemented to extract edge texture. d. Otsu-based enhanced local ternary pattern (OELTP) is used for feature extraction. e. SVM classifier is used for forgery detection.

RESEARCH METHOD 3.1. Proposed splicing detection system
Texture based feature extraction technique is very effective due to the highly discriminative property at the local level. In this paper, we proposed an edge-based feature extraction technique. We detect the edges to find the meaning of discontinuities in the intensity of pixels [25]. A threshold becomes better if it is tall, narrow, and separated by deep values. A most straightforward way to improve the shape of the histogram is to consider those pixels near the edge between the object and the background. We implement Otsu-based global threshold on ELTP technique for feature extraction [26], [27]. SVM is used for classification to find the image is fake are pristine. We are considering three image databases for implementing our proposed work. These are CASIA v1, CASIAv2, Columbia. Chrominance components of the image are considered. A multi-level wavelet transform is implemented over the chrominance component to get the important details in the frequency domain, the architecture of the proposed work represented by Figure 2.

Preprocessing
Image forgery artifacts are embedded in the form of edge discrepancy. An effective color component plays an essential role in feature extraction. We choose the YCbCr color component after converting RGB into YCbCr, RGB and luminance and chrominance relation shown in (1)

Feature extraction technique
To detect image forgery artifact, need to extract features of the image as per application and dataset. It is found that the operation which is used for forged the image is predicted, help to find the features.

Discrete wavelet transform
DWTis a frequency-based model. It's reliable for texture feature classification. A digital filtering technique is implemented by discrete wavelet transform for time scale depiction of the digital signal. In DWT, the image decomposed into four frequency components, represented by LL (approximation coefficient), low-high (LH), high-high (HH), high-low (HL), shown in Figure 3 (a). The energy store in the approximation coefficient block. The effect of DWT on the image is represented by Figure 3 (b). In this proposed work, we implement DWT on the image. We again decomposed approximation coefficient into four sub-bands until the fourth level after converting RGB into YCbCr. Features are calculated through the mean of approximation coefficient and standard deviation.

Enhance local ternary pattern
Local binary pattern (LBP) is reliable for local feature extraction in texture. Its descriptor has some restrictions due to dependency on the central pixel. It does not perform well on random and quantization noise over uniform resigns. LBP used binary values 0 and 1 to encode the intensity between central and surrounding pixels shown by (4). Ternary value 0, 1, -1 are consider to evaluate the LTP. It's represented by (5). LTP used a fixed threshold value to find the intensity of the central pixel. The LTP is updated by an ELTP. In ELTP, the dynamic threshold value concept is used [27]. The intensity relationship between the central pixel (Ic E ) and threshold (Is E ) of ELTP is represented by (6). M represents the image matrix of 3x3, neighbor pixels of the central pixel represented by Ic E , and the threshold is enhanced mean absolute deviation (MAD) in the ELTP technique. The ternary matrix of ELTP is generating after implementing M, Ic E , and T E in (5).
where; Ii=Intensity value of the surrounding pixel. Ic=Intensity value of the central pixel.
where; E=ELTP ternary matrix. Ts=Static threshold value Ii E= Intensity of the surrounding pixel M=Ii E | i=0, 1, 2, 3….8 | Ic E =Mean(M) Ts E =MAD(M) In ELTP, the intensity of the central pixel is obtained by the upper case and lower case ELTP matrix. We take a 3x3 matrix of the image and generate a new ternary matrix. The value assigns to surrounding pixels is based on (6). Further, it was divided into positive and negative matrices in positive matrix 0 and -1intensity represented by 0 and in negative matrix 0 and 1 intensity represented by 0 and -1 intensity represented by 1, shown in Figure 4 [27]. Central pixel intensity is evaluated by (7) of the coordinate (m, n).
ELTPm,n=ELTP_Pm,n * (m + 2) -(ELTP_Pm,n * (ELTP_Pm,n + 1)) / 2 + ELTP_Nm,n Where N-1 ELTP_Pm,n=∑ e(S E (Ii E , Ic E , T E ), 1) i=0 N-1 ELTP_Nm,n=∑ e(S E (Ii E , Ic E , T E ), -1) i=0 enhanced local ternary pattern. We implement Canny edge detection technique, which is the most popular edge detector. Due to the implementation of the edge detection technique, the performance of the global threshold has been improved after considering the pixels which lie on or near the object and background. Before implementing the predefine ELTP technique, we implement the Canny edge detection technique to improve the performance of the global threshold by which the performance of ELTP has been improved the performance of the edge, the effect of Otsu technique represented by Figure 5.

Database and classifier
The proposed work is evaluated on three open source available image databases, CASIA v1, CASIA v2, and Columbia. CASIA v1 has 1721 RGB images of different sizes 384 x 256 and 256 x 384. In this database, 800 images are pristine, and 921 are tampered. CASIA v2 has 12614 different image datasets like JPEG, BMP, and TIFF. In this, 7491 images are pristine, and 5123 are fake. In Columbia, different camera types are used for data collection, Canon G3, Canon 350D RebelXT, NikonD70, and Kodak DCS330, generating an uncompressed image of size 757 x 569 and 1152 x 768 sizes. To classify these images SVM is used, gives awe-inspiring results over the small and big datasets, and easy to implement.

Experimental results and comparison
The result of the proposed work is based on the accuracy of the forgery detection on different image datasets, CASIA v1, CASIA v2, and Columbia. We compare our work with the previous technique and previous state of work in terms of accuracy. The accuracy of the proposed work is 99.98% on CASIA v1.0, 99.03% on CASIA v2, and 98.56% on the Columbia image dataset is shown in Figure 6 and compares the proposed work accuracy with different states of work [6], [10], [12], [15], [26] is shown in Table 1.  Table 1. Comparison of proposed work accuracy with previous state of work Techniques CASIA v1.0 CASIA v2.0 Columbia He et al. [10] 89.76 89.76 − Muhammad et al. [12] 94.89 97.33 96.39 Goh and Thing [15] 90.18 96.21 − Alahmadi et al. [6] 97.5 − − Kanwal, N. et al [26] 98

CONCLUSION
In this paper, we discuss edge texture feature extraction using OELTP techniques for forgery detection. A multi-level wavelet transform is implemented over the chrominance component of the image to get the import details in the form of the frequency domain. The Canny edge detection technique is implemented over the frequency domain to obtain a better global threshold with the help of the Otsu technique. Extracted feature by enhanced local ternary pattern classified by SVM to get the image is forged or not. Our proposed approach obtains progress with other states of artwork in terms of accuracy. The accuracy of the proposed work is 99.98% on CASIA v1.0, 99.03% on CASIA v2, and 98.56% on the Columbia image dataset. We do not get better results when the background illumination is highly nonuniform. In the future, we need to develop such system which is reliable for highly non-uniform surface or background.