The present work is devoted to the development and investigation of the algorithm for detection intentional distortions (forgeries) of a single digital image in an image time series (time sequences) of one scene. The proposed algorithm consists of three stages. At the rst stage, a set of errors that were calculated during reconstruction the fragments of the analyzed image by the 'neighboring' ones are estimated. After errors were calculated throughout image, we analyze their distribution on the second stage. At the nal stage, fragments of the analyzed image that are anomalies are selected as 'suspicious'. The proposed algorithm, unlike existing algorithms, will allow uni ed detection of such attacks as intra-image copy-move and inter-image copy-move. Also, copy-move fragments may fall under geometric transformation, linear enhancement and other distortions. The investigation results of intra-image copy-move and inter-image copy-move detection using the proposed solution are presented.
Nowadays the image forgery methods complexity increases in relation to their detection complexity. This is due to the number spheres increasing using digital images in their work, as well as their processing tools availability and popularization. Image time series show the scene dynamics and allow it to be compared over time. So, having an image time series of some scene, with some admission, you can model an image that will be next in the scene, or an image that may have been distorted.
This algorithm will allow detection spatial tampering attacks. At the
moment, several detection techniques were development. They are technique based
on camera's ngerprint, technique based on coding artifacts, and techniques that
use temporal and spatial correlation [
We will mean forgery in the sense of an anomaly to develop an algorithm for
image forgery detection. In the general sense, an anomaly is a data fragment that
does not correspond to the precisely de ned concept of normal behavior [
The work consists of two parts, namely, description of the proposed algorithm and analysis of experiments results. The description of the proposed algorithm is subdivided into the three sections. The rst section contains a characteristic of the image fragments description method. The second section de nes the statistic construction method. The third section explains the rule for assigning fragments of the analyzed image to anomalies. 2
Let It (n1; n2), t = 0; T is an image series (time sequence) of one scene. Every image in the series has the same size N1 N2, ni 2 0; Ni 1 (i = 1; 2; T 1).
For de niteness, we assume that the image I0(n1; n2) is checked, although it can be located in the sequence anywhere. We analyze a certain square region of the image D(n1; n2) 0; N1 1 0; N2 1 in a sliding window with a position (n1; n2). For certain region D(n1; n2), we have image fragments It(m1; m2), where (m1; m2) 2 D(n1; n2). To simplify the exposition, the arguments of the region in the record D(n1; n2) can be omitted. Experimentally, we have spotted the best in terms of detection quality and runtime window has 15 15 size. 2.1
For each possible position of the window D in the image plane, the corresponding fragments It(m1; m2) are successively divided into k fragments, k = 2p, by clustering by brightness (a k-means clustering algorithm is used). An example of such splitting under k = 22 is shown in Figure 1:
The new fragments of the image obtained in this way on step k can be denoted by Itj (n1; n2), j = 0; k 1. Next, we solve the problem of image I0 fragment representation for this region by means of corresponding by the window D position fragments I10; I11; :::; I1k 1; :::; IT0 ; IT1 ; :::; ITk 1 linear combination, that is:
I0
T k 1
X X
1 A comprehensive coverage of the eld of outlier analysis from a computer science
point of view can be found in [
0
We perform this procedure for k = 4; 8; 16. The rst stage result for each position of the analysis region is a normalized mean squared deviations set of the analyzed image fragment representation. For convenience of further use, we designate them as a vector: We represent the obtained vectors x(n1; n2) set in the coordinate system "~24"~28"~216. This set is located in the three-dimensional cube with sides equal to 1 as shown in Figure 2: (2) (3) (4) 2.3
There are no absolute unchanging objects on images obtained in real conditions. This is due both to real cameras properties that have their own noises and the information transfer path properties of from the camera to the processing system. In this path, the image is compressed before shipment and then decoded, which often leads to additional system distortions. Moreover, there are often objects on the scene that have certain dynamic characteristics although they are static in our understanding. For example, it may be trees swaying in the wind.
In accordance with this fact, it is impossible to obtain an errors vector with coordinates (0; 0; 0) after authentic image fragment representation by means of neighboring images fragments linear combination. So we can conclude the errors vector with coordinates (0; 0; 0) corresponds to the image region, which is a duplicate inserted from one of neighboring images of the image time series.
On the other hand, the errors "~ , "~ , "~126 of an authentic fragment represen2 2 4 8 tation must have values that do not exceed a certain threshold. It is obvious the error value of the same fragment representation decreases with the clusters number increasing. Therefore, it is justi ed to use di erent thresholds for "~42, "~8 2 and "~126. So the following relation should be observed:
After the statistic construction stage, we analyze the distribution histograms and select the thresholds according to the relation (5) as shown in Figure (3). We choose rst local minimum and consider its value a threshold.
Then the cube with the errors vectors x(n1; n2) set is splitted into three areas: 1) Origin of the coordinate system; 2) A parallelepiped that is adjacent to the origin; 3) Rest area of the cube.
In accordance with the above, vectors from the rst area correspond image region which is a duplicate inserted from one of neighboring images of the image time series (inter-image copy-move). Vectors from second area refer to authentic image regions and vectors from third area correspond to fragments which is the duplicate within one image (intra-image copy-move). This splitting is shown in Figure 4.
After extraction of errors vectors from the relevant area and labeling them as suspicious, we create corresponding binary mask. After this, we process it with a noise lter that removes regions from the binary mask that have a square less than some value. So after this, we keep only errors vectors which corresponding forgery regions in the set of suspicious errors vectors. 3
The experiments were carried out on a desktop PC with Intel Core i5-4460 processor and 16 GB RAM using MATLAB R2016b software. c
Five image time series were obtained using the same camera. The camera was still all the time. It has captured the scene and token image every 10 sec. As result of this procedure, we have got ve image time series with six images in every series. Next, we transform all images to gray-scale. Obtained images have 920 1380 dimension. These time series were chosen as the objects of experiments. We developed a copy-move generation procedure which enables to add distortions and control their parameters.
The experiment results that were carried out on all image series with duplicate taken from another image of the image series are showed in the Table 1. Example of this detection is shown in Figure 5.
The experiment results with duplicate taken from the same image are shown in the Table 2 and in the Figure 6