CN103310236A - Mosaic image detection method and system based on local two-dimensional characteristics - Google Patents

Abstract

A mosaic image detection method and system based on local two-dimensional features in the field of image processing and information security technology. The image is divided into squares with different side lengths and then subjected to block DCT transformation, and the obtained block DCT coefficients are localized. After the two-dimensional features are described and combined into complete detection features, a classifier is used for classification. The invention can take both detection accuracy and detection complexity into consideration, and the detection accuracy can reach 89.9%.

Description

Method and system for stitching image detection based on local two-dimensional features

technical field

The present invention relates to a method and system in the technical field of image processing and information security, in particular to a spliced image detection method and system based on local two-dimensional features for spliced images without prior knowledge.

Background technique

Digital image technology has become very popular in today's society, and the threshold for using it is also decreasing day by day. Many powerful digital image processing software are also accessible to the public, and ordinary people without professional training can also create fake images that cannot be directly identified by the naked eye. When forged images appear in hot posts such as forums and microblogs, they can have a great negative impact on the government, enterprises or individuals. Therefore, the use of computer technology to detect forged images has become a research hotspot in the field of information content security.

Generally speaking, the current mainstream counterfeit image detection technologies are divided into two categories: active methods and passive methods. The active method is mainly to embed digital signatures or digital watermarks in the process of generating pictures, and to ensure that pictures are not tampered with by detecting these marks. Passive methods mainly use the statistical properties of the image itself, rather than relying on pre-implanted identifiable marks. Marks such as digital signatures and digital watermarks implanted in an active way are destructive to the image itself and should not be used in some cases. On the contrary, the adaptability of the passive method is stronger, so it has become the main research direction at present.

Image stitching is the most basic step in forging images. The complete image tampering process usually includes stitching, scaling, rotation and subsequent processing, and the detection of stitching operations is the basis of most anti-counterfeiting identification methods.

The classic image mosaic detection method includes the bispectral feature detection method proposed by Ng et al., see Ng TT, Chang SF., Adataset of authentic and spliced image blocks (dataset of real and spliced image blocks), (ADVENT TechnicalReport, #203- 2004-3, Columbia University.) This document also provides a general spliced image detection data set to compare the pros and cons of various algorithms and is widely cited. Ng et al achieved a detection accuracy of 72% on this dataset. Fu et al. used the Hilbert-Huang transform and moment feature of the characteristic function in the wavelet transform domain for detection, and achieved a detection accuracy of 80.15%.

Different from the above statistical method, Johnson et al. used the inconsistency of illumination features in different stitching areas of stitched images to detect, see Johnson MK, Farid H. Exposing digital forgeries by detecting inconsistencies inlighting (digital forgery identification based on illumination inconsistency detection) .(In Proceedings of ACM Multimedia and Security Workshop, New York, USA, 2005; 1–9.) The detection method based on statistical features has a mature method framework, including the steps of feature selection, feature extraction and classification learning. Commonly used detection methods. However, the detection accuracy of existing detection methods based on statistical features needs to be improved, and a large number of statistical features have not been used for detection of this problem.

After searching the existing technology, it is found that Chinese Patent Document No. CN102855496A, with a publication date of 2013-01-02, discloses a method and system for face masking authentication. The technology includes: S1, collecting face video images; S2, The collected face video image is preprocessed; S3, detect and calculate the occluded face, according to the motion information of the video sequence, use the three-frame difference method to estimate the position of the face image, and then perform further face detection through the Adaboost algorithm. Confirmation of the position; S4. Identify and calculate the occluded face, divide the face sample into several blocks, and use the SVM dichotomy algorithm combined with the supervised 1-NN nearest neighbor method to perform occlusion discrimination on the face blocks. , then discard it directly. If the block is not occluded, extract the corresponding LBP texture feature vector for weighted recognition, and then use the classifier based on the orthogonal projection method to reduce the number of feature matching. This technology divides the image mechanically into 6 areas. Although it can be used to solve the problem of relatively fixed positions of the face and important facial organs, it cannot solve the problem of passive image forgery identification without any prior knowledge of the image.

Contents of the invention

Aiming at the above-mentioned deficiencies in the prior art, the present invention proposes a mosaic image detection method and system based on local two-dimensional features, which can take both detection accuracy and detection complexity into consideration.

The local two-dimensional feature refers to a local binary pattern (Local Binary Pattern, LBP) feature customized for splicing detection. The LBP feature was first proposed by Ojala et al., see Ojala T, Pietikainen M, Maenpaa T.Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. (IEEE Transactions on Pattern Analysis and Machine Intelligence2002;24(7):971–987 .)

The LBP feature is widely used as an image texture feature in face recognition, background modeling and steganalysis. The core idea of LBP is to compare an image pixel with its surrounding pixels, and then get a 0/1 sequence after thresholding. The 0/1 sequence can be regarded as a binary number, so it can be represented by a decimal integer. Each pixel in the image matrix corresponds to a decimal integer. These integers can form a histogram, which reflects statistical laws such as the edges of the image.

The present invention is achieved through the following technical solutions:

The invention relates to a splicing image detection method based on local two-dimensional features, which divides the image into squares with different side lengths and performs block DCT transformation, and describes the obtained block DCT coefficients in the form of local two-dimensional features And after merging into a complete detection feature, a classifier is used for classification.

Described method specifically comprises the following steps:

Step 1: Perform multi-scale block DCT (discrete cosine) transform on any image in the data set to be processed to obtain the block DCT coefficient matrix, and take the absolute value of all block DCT coefficients to obtain the absolute value of block DCT coefficients matrix.

The block-by-block DCT transformation refers to: select a side length b, divide the image to be processed into small square blocks of the same size with side length b, and then perform DCT transformation in each small square block.

The side length b of the small square block is preferably a multiple of 8, such as usually 8, 16, 32, . . . , but can also be any value according to application requirements.

When the size of the image to be processed cannot be divisible by b, add 0 to the rightmost and bottom sides of the image to be processed until it meets an integer multiple of b; the block DCT coefficient matrix obtained after the block DCT transformation is a For an M×N matrix, the size of the image to be processed is M′×N′, usually the two are the same size. When 0 columns and 0 rows are added to the rightmost and bottommost sides of the image to be processed, the resulting coefficient matrix will be larger than the image to be processed.

Step 2: Use local two-dimensional features to represent the changes in statistical features caused by image stitching, and convert the block DCT coefficient absolute value matrix into a local two-dimensional feature histogram.

Step 2 specifically includes the following operations:

2.1 Take P points around any element on the block DCT coefficient absolute value matrix, denoted as g _p , p={1...P}. These points are arranged equally at an angle of 2π/P around the central point. When the surrounding points are not on the grid points of the matrix, the value of the surrounding grid points is used for interpolation estimation, and the distance between the surrounding points and the central point is recorded as R.

In the present invention, the preferred values of the P points are 8 points of the upper, lower, left, right, and upper left, lower left, upper right, and lower right of the arbitrary element, and the distance R between the peripheral point and the central point is preferably 1.

2.2 Compare the gray value of the peripheral points with the gray value of the center point in turn: when the gray value of the peripheral points is greater than the gray value of the center point, the comparison result is recorded as 1, otherwise it is 0; then P The comparison results of the points are arranged from right to left to form a 0/1 comparison result sequence of length P, which is used as a binary integer and converted to a decimal integer.

2.3 Process each element of the block DCT coefficient absolute value matrix in step 2.2 to obtain the corresponding decimal integers, and form a local two-dimensional feature histogram with all the decimal integers.

Step 3: In step 1, take the side length values b of different small square blocks to obtain the absolute value matrix of block DCT coefficients representing various scales, and obtain several corresponding local two-dimensional feature histograms according to the operation of step 2 Figure; The features generated by each local two-dimensional feature histogram are concatenated to form a complete statistical feature, and then the SVM classifier is used for learning and classification.

Each of the local two-dimensional feature histograms uniquely corresponds to a block DCT transformation with different side length values b in step 1.

The horizontal scale of the histogram ranges from 0 to 2 ^P -1, and the value of the histogram corresponding to each horizontal scale is taken as a feature, and the feature dimension is 2 ^P .

The classifier used is the mainstream SVM implementation LibSVM, specifically Chang CC, Lin CJ. LIBSVM: library for support vector machines. http://www.csie.ntu.edu.tw/cjlin/libsvm, 2001.

Step three specifically includes the following operations:

3.1 Obtain the absolute value matrix of block DCT coefficients of different scales by taking different side length values b, and obtain a local two-dimensional feature histogram as described in step 2 for each block DCT absolute value matrix.

3.2 Extract 2 ^P -dimensional features from each local two-dimensional feature histogram, and concatenate these features together to form a complete statistical feature.

3.3 Extract features from all the remaining pictures in the data set to be processed according to the above-mentioned feature extraction method, and divide the pictures into two parts, the training set and the test set, first input the features and category data of the training set into the classifier and obtain the classification model; then The classification model and the features of the test set are input into the classifier again to obtain the category discrimination of the test set; finally, the classification accuracy is obtained according to the known category of the test set.

The picture number ratio of described training set and test set is preferably 5:1;

The number of mosaic pictures and natural pictures contained in the training set and the test set is preferably 1:1.

The invention relates to a mosaic image detection system based on local two-dimensional features, including: a preprocessing module, a local two-dimensional histogram construction module, a feature extraction module and a classifier module, wherein: the preprocessing module and the local two-dimensional histogram construction The modules are connected and receive the original image for block DCT transformation to obtain the block DCT coefficient absolute value matrix and output it to the local two-dimensional histogram construction module. The local two-dimensional histogram construction module is connected to the feature extraction module and the block DCT coefficient is absolute After the value matrix is subjected to LBP operation, the local two-dimensional histogram is obtained and output to the feature extraction module. The feature extraction module is connected to the classifier module and the local two-dimensional histogram information is subjected to feature extraction operation to obtain classification features and output to the classifier module. The classifier module receives the classification feature and performs a classification operation to obtain the classification judgment of the original picture.

technical effect

Compared with the prior art, the present invention effectively captures the defects caused by splicing pictures by adopting multi-scale sub-block DCT transformation. The parameters of the optimal local two-dimensional feature algorithm are determined through experiments, P=8, R=1. The invention has higher detection accuracy than the prior art.

Description of drawings

Fig. 1 is a schematic flow chart of the present invention.

Fig. 2 is a schematic diagram illustrating the parameters of the LBP algorithm in the embodiment.

Fig. 3 is a schematic diagram of the system structure of the present invention.

Fig. 4 is the schematic diagram of embodiment processing effect;

In the figure: (1) is the original image, (2) is the LBP histogram under the 8×8 block DCT transformation, (3) is the LBP histogram under the 16×16 block DCT transformation, (4) is the 32× LBP histogram under 32-block DCT transformation; since the value of 0 point is abnormally large, it will affect the display of the histogram, so the value of 0 point is omitted here.

Detailed ways

The embodiments of the present invention are described in detail below. This embodiment is implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operating procedures are provided, but the protection scope of the present invention is not limited to the following implementation example.

Example 1

As shown in Figure 1, this embodiment includes the following steps:

Step 1 is to transform the original image through three modes of block DCT, and take the absolute value in (3) in Figure 1 to obtain three block DCT coefficient absolute value matrices, wherein: block DCT can be used in multiple Blocking methods of different sizes, different block sizes can capture different pixel transformation features. At the same time, the number of block patterns determines the dimension of the final statistical features. Therefore, the mode selection of the block DCT needs to take into account both the detection accuracy and the detection complexity.

The three block modes in (2) in Figure 1 given in Figure 1, 4x4, 8x8 and 16x16 have been proved by experiments to ensure the detection accuracy and have a smaller feature dimension.

Table 1 shows the detection accuracy of LBP _8,1 in different block DCT modes. The accuracy value is the average value of the detection results of 20 random sample groups, and the value in brackets is the mean square error. The image library used is Columbia University's spliced image detection library, see Ng TT, Chang SF., A dataset of authentic and spliced image blocks (dataset of real and spliced image blocks), (ADVENT Technical Report, #203-2004 -3, Columbia University.)

Table 1

The second step is to describe the generated block DCT matrix using local two-dimensional features, specifically:

The number of pixels used for comparison around a pixel is denoted as P. These pixels are equally arranged around the central point at an angle of 2π/P. When the surrounding pixel points are not on the image grid point, the interpolation representation of the value of its surrounding grid point is used. Another variable is R, which represents the distance between the surrounding comparison pixels and the center point. Since the value represented by each bit of a binary number is 2 ^p , the formula for converting a binary sequence to a decimal number can be expressed as:

${\mathrm{LBP}}_{P,R}(x_c,{\mathrm{the\; y}}_c)=\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\mathrm{the\; s}{(g_p-g_c)2}^p,$ Among them: x _c and y _c represent the position of the central pixel, g _p and g _c represent the absolute value of the central pixel and surrounding pixels, P and R represent the two parameters of the algorithm, P represents the number of surrounding points, R Represents the distance from the surrounding points to the center point; the threshold function s is: $\mathrm{the\; s}\left(x\right)=\left\{\begin{array}{c}1,x>\mathrm{\&sigma;}\\ 0,x<\mathrm{\&sigma;}\end{array},\right.$ Where: σ is a parameter, which is defined as the threshold of the threshold function s(x).

Combinations of P and R usually have (8,1), (8,2) and (8,3). In addition to considering these parameter combinations individually, they can also be combined to form larger features, that is, the multi-scale analysis. The results in Table 1 are the results of model (8,1), and the results of the multi-scale analysis are further given in Table 2 below.

Table 2 shows the results of multi-scale analysis

(P,R) (8,1)+(8,2) (8,1)+(8,2)+(8,3) Accuracy 90.45% 90.48%

After comparative experiments, weighing the detection accuracy and feature dimension, it is determined that a single scale P=8, R=1 is the best LBP descriptor parameter. The s function, by taking different σ, between 0 and 2, with a step size of 0.1, observes that the detection accuracy first increases and then decreases. When σ=0.9, the detection accuracy is the highest. The parameters used in (4) in Fig. 1 are as above.

Step 3: Use the absolute value matrix of multi-scale block DCT coefficients obtained in step 1, and the statistical features obtained in step 2 to form the feature vector of machine learning, use LibSVM for learning and testing, and finally obtain the category judgment of the original image.

In LibSVM, it is necessary to determine the kernel function used, and the Gaussian RBF kernel function is selected after comparison. The Gaussian RBF kernel function has two variables: the penalty variable C and the Gaussian kernel width γ.

Use grid search to determine the best combination of C and γ, where: the search range of C is 2 ^{-1,1,3,5} , and the search range of γ is 2 ^{{-5,-3,-1, 1}} .

The above parameter pairs of C and γ constitute a grid. Perform cross-validation on the training set to find the best parameter pair, and then use this parameter pair for final testing. The above results in Table 1 and Table 2 are obtained by the above method.

As shown in Figure 3, in order to realize the system of above-mentioned method, this system comprises: preprocessing module, local two-dimensional histogram building module, feature extraction module and classifier module, wherein: preprocessing module and local two-dimensional histogram building module After connecting and receiving the original image for block DCT transformation, the block DCT coefficient absolute value matrix is obtained and output to the local two-dimensional histogram construction module, the local two-dimensional histogram construction module is connected with the feature extraction module and the block DCT coefficient absolute value After the matrix is subjected to LBP operation, the local two-dimensional histogram is obtained and output to the feature extraction module. The feature extraction module is connected to the classifier module and the local two-dimensional histogram information is subjected to feature extraction operation to obtain classification features and output to the classifier module. The detector module receives the classification features and performs classification operations to obtain the classification judgment of the original picture.

Table 3 shows the comparison of the detection accuracy between the present invention and two existing mainstream detection methods on the mosaic image detection library of Columbia University. Shi's Markov Algorithm, Shi, Yun Q., Chunhua Chen, and WenChen."A natural image model approach to splicing detection." (A natural image model for detecting spliced images) Proceedings of the 9th workshop on Multimedia&security.ACM, 2007, and Ng's bispectrum algorithm, Ng, Tian-Tsong, Shih-Fu Chang, and Qibin Sun."Blind detection of photomontage using higher order statistics." (Blind detection of stitched pictures based on high-order statistics) Circuits and Systems,2004.ISCAS'04.Proceedings of the2004International Symposium on.Vol.5.IEEE,2004.

table 3

this invention Markov Bispectrum

Accuracy 89.9% 86.6% 72.3%

The above table shows that the algorithm of the present invention has better detection accuracy than the existing technology.

Claims (10)

1. the stitching image detection method based on local two dimensional character is characterized in that, may further comprise the steps:

Step 1: treat arbitrary image that deal with data concentrates and carry out multiple dimensioned block DCT transform, the piecemeal DCT matrix of coefficients that obtains, and piecemeal DCT coefficient all taken absolute value, obtain piecemeal DCT coefficient absolute value matrix;

Step 2: adopt local two dimensional character to come token image to splice the variation of the statistical nature that causes, piecemeal DCT coefficient absolute value matrix is converted into local two dimensional character histogram;

Step 3: the length of side value of getting different square tiles in the step 1 is the corresponding absolute value matrix that obtains representing the piecemeal DCT coefficient of various yardsticks then, and obtains the local two dimensional character histogram of several correspondences according to the operation of step 2; The feature that each local two dimensional character histogram is generated is together in series and consists of a complete statistical nature, then utilizes the svm classifier device to learn and classify.

2. method according to claim 1 is characterized in that, described block DCT transform refers to: select a kind of side length b, pending image is divided into the square tiles that the length of side is the formed objects of b, then carry out dct transform in each square tiles.

3. method according to claim 2 is characterized in that, the side length b of described square tiles is 8 multiple; As the big or small aliquant b of pending image, then add 0 in the rightmost side and the lower side of pending image, until it meets the integral multiple of b; Resulting piecemeal DCT matrix of coefficients is the matrix of a M * N behind the process block DCT transform, and pending picture size is that the two is equirotal to M ' * N ' usually; When added 0 row and 0 row in the pending image rightmost side and lower side after, resulting matrix of coefficients can be greater than pending image.

4. method according to claim 1 is characterized in that, described step 2 comprises following operation:

2.1 on piecemeal DCT coefficient absolute value matrix arbitrary element around get P point, be denoted as g _p, p={1 ... P}; These points are minute arrangements such as 2 π/P angle around central point; When peripheral point not on matrix grid point, utilize its peripheral net point numerical value to carry out interpolation estimation, the distance of peripheral point and central point is designated as R;

2.2 the gray-scale value of peripheral point is compared with the gray-scale value of central point successively: when the gray-scale value of the peripheral point gray-scale value greater than central point, then comparative result is denoted as 1, otherwise is 0; Then the comparative result of P point is arranged according to right-to-left, consisted of 0/1 comparison result sequence that length is P, this comparison result sequence also is used for being converted to decimal integer as bigit;

2.3 each element on the piecemeal DCT coefficient absolute value matrix is carried out the processing of step 2.2, obtain corresponding decimal integer, all decimal integers are consisted of a local two dimensional character histogram.

5. method according to claim 4, it is characterized in that, described on piecemeal DCT coefficient absolute value matrix arbitrary element around get P and put and to refer to: in the upper and lower, left and right of described arbitrary element and upper left, lower-left, upper right, bottom right totally 8 points, the distance R of peripheral point and central point gets 1.

6. method according to claim 4 is characterized in that, described histogrammic horizontal scale value is 0 to 2 ^P-1, histogrammic value corresponding to each horizontal scale as a feature, characteristic dimension 2 ^P

7. method according to claim 1 is characterized in that, described step 3 comprises following operation:

3.1 obtain the piecemeal DCT coefficient absolute value matrix of different scale by getting different length of side value b, each piecemeal DCT absolute value matrix adopted step 2 is described to obtain local two dimensional character histogram;

3.2 extract 2 from each local two dimensional character histogram ^PDimensional feature, and these features are serially connected form a complete statistical nature;

3.3 all the other all pictures of pending data centralization are extracted feature according to above-mentioned feature extracting method, simultaneously picture is divided into training set and test set two parts, first feature and the categorical data of training set are inputted sorter and obtained disaggregated model; Again the feature of disaggregated model and test set is inputted sorter again, the classification that obtains test set is differentiated; Last classification according to known test set obtains classify accuracy.

8. method according to claim 7 is characterized in that, the number of pictures ratio of described training set and test set is 5:1.

9. method according to claim 7 is characterized in that, the splicing picture that comprises in described training set and the test set and the number of natural picture are 1:1.

10. detection system that is used for realizing the described method of above-mentioned arbitrary claim, it is characterized in that, comprise: pretreatment module, local two-dimensional histogram makes up module, characteristic extracting module and classifier modules, wherein: pretreatment module links to each other with local two-dimensional histogram structure module and receives original image and carries out obtaining piecemeal DCT coefficient absolute value matrix and export local two-dimensional histogram to making up module behind the block DCT transform, local two-dimensional histogram structure module links to each other with characteristic extracting module and piecemeal DCT coefficient absolute value matrix is carried out obtaining local two-dimensional histogram and outputing to characteristic extracting module after the LBP computing, characteristic extracting module is connected with classifier modules and local two-dimensional histogram information is carried out obtaining characteristic of division after the feature extraction computing and outputs to classifier modules, and classifier modules receives characteristic of division and carries out obtaining the classification of original image is judged after the sort operation.

Images