CN101655972A - Image mosaic blinding method based on wavelet domain - Google Patents

Abstract

The invention relates to an image mosaic blinding method based on a wavelet domain in the technical field of image processing. In the invention, an image wavelet high-frequency sub-band coefficient isfit by generalized Gaussian distribution, an estimated fitting generalized Gaussian distribution density function is used as a probability density function of the actual high-frequency sub-band coefficient, therefore, statistic characteristics are extracted by using the difference of the generalized distribution and the actual distribution of the high-frequency sub-band coefficient. When splicingis tampered, since the statistic characteristics are changed, the invention can judge the authenticity of an image by using the characteristics. The invention can solve the problems of low precisionof high-order statistic characteristics and long operation time, thereby greatly improving the efficiency and the practicability of the splicing tampering blind detection. The invention can be used for detecting an actual splice image, tests prove that the method has strong real-time, can obtain good effect, and lays a better basis for further developing the field.

Description

Image mosaic blinding method based on wavelet field

Technical field

What the present invention relates to is a kind of detection method of technical field of image processing, specifically is a kind of image mosaic blinding method based on wavelet field.

Background technology

Based on digital picture be easy to distort, particularly the splicing of " grafting flowers on a tree " is distorted very commonly, makes the authenticity of digital picture differentiate that problem detects in the administration of justice, there is crucial meaning in fields such as news report.Distort at this, have and initiatively differentiate and passive discriminating dual mode.Initiatively differentiate it mainly is in digital picture, to add digital watermarking, but present camera does not mostly possess this function, so practicality is less relatively.And passive discriminating, though i.e. potentiality that grow a lot of blind Detecting under no any prior imformation condition differentiate that difficulty is bigger, thereby development are slower.The research of present blind checking method mainly contains two kinds: based on the detection of distorting vestige with based on the detection of picture material.The vestige that the former mainly stays after distort, equal aspect,, camera source inconsistent as illumination is differentiated, obtained good effect, but these methods is subjected to the restriction such as conditions such as illumination condition, camera model.Whether the latter then is a method of utilizing pattern-recognition, extracts the feature to the image mosaic sensitivity, differentiate image with sorter and distort.This method scope of application is extensive, and it is few to be restricted, but choosing of feature is most important, and its accuracy is not high at present.

Find by prior art documents, Tian-Tsong Ng, people such as Shih-Fu Chang and Qibin Sun are at document " Blind detection of photomontage using higher orderstatistics[C] " (" utilizing the distorted image blind Detecting of high-order statistic ") (IEEE Proceedingof International Symposium on Circuits and System.Vancouver, Canada:IEEE Press, ~ 691) 2004.688 (IEEE Circuits and Systems proceeding), proposed two spectrum signatures, and judged image true-false by sorter based on the image statistics feature.The scope of application that this method is distorted splicing is wider, but feature extraction is consuming time longer, and it is good inadequately to detect accuracy rate, only has 71.5%.

Summary of the invention

The object of the invention is at the deficiencies in the prior art, and a kind of image mosaic blinding method based on wavelet field is provided.By natural image small echo high-frequency sub-band coefficient distributed model is analyzed, to image characteristics extraction, judge by sorter, realize the image mosaic blind Detecting.

The present invention is achieved by the following technical solutions:

At first use natural image high frequency wavelet sub-band coefficients match generalized Gaussian distribution preferably (General Gaussian Distribution, GGD);

After the natural image process was distorted, the statistical property of the high-frequency sub-band coefficient that its wavelet decomposition obtains changed, thereby can't coincide well with generalized Gaussian distribution, therefore, to each high-frequency sub-band, utilized these characteristics to carry out feature extraction;

After image wavelet is decomposed, carry out parameter estimation with each high-frequency sub-band coefficient, compare with the actual distribution of high-frequency sub-band coefficient, and utilize difference extraction feature between them;

By support vector machine SVM (Support Vector Machine) classification, image true-false is differentiated then.

Described image characteristics extraction comprises the steps:

1. after image carries out wavelet transform (DWT),, suppose coefficient obedience generalized Gaussian distribution to each high-frequency sub-band, the big or small high-frequency sub-band coefficient w for M*N of utilization (i, j), find the solution following equation:

${\mathrm{\δ}}_{\mathrm{\θ}}^2=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{w}^2(i,j)$

$\frac{\mathrm{\Γ}(1/v)\mathrm{\Γ}(5/v)}{{\mathrm{\Γ}}^2(3/v)}=\frac{1}{{\mathrm{\δ}}_{\mathrm{\θ}}^4}\frac{1}{\mathrm{NN}}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{w}^4(i,j)$

Wherein Γ () is the gamma function.Can obtain the form parameter v and the variance parameter δ of generalized Gaussian distribution probability density function _θAnd the parameter value substitution following formula that estimates can be obtained the generalized Gaussian distribution probability density function: $\mathrm{\π}\left(\mathrm{\θ}\right)=\frac{\mathrm{v\η}\left(v\right)}{2\mathrm{\Γ}(1/v)}\frac{1}{{\mathrm{\δ}}_{\mathrm{\θ}}}\exp \{-{\left[\mathrm{\η}\left(v\right)\right|\mathrm{\θ}|/{\mathrm{\δ}}_{\mathrm{\θ}}]}^v\},$ Wherein $\mathrm{\η}\left(v\right)=\sqrt{\mathrm{\Γ}(3/v)/\mathrm{\Γ}(1/v)}.$ Because the parameter v of natural image is generally between 0.5 to 0.8, with parameter v as a feature.

2. with the coefficient minimum value of high-frequency sub-band coefficient to be divided between the maximal value K interval, i.e. burst length L=(Max-Min)/K; Design factor drops on each interval interior Probability p (θ _i) (i=1,2...K), then the actual probabilities density fonction is approximately in the value of each interval center: g (θ _i)=p (θ _i)/L;

3. in each interval center, poor between the high-frequency sub-band coefficient distribution density function of realistic border image and its anticipation function value: delta (θ _i)=| g (θ _i)-π (θ _i) |, with its mean value $\stackrel{\‾}{\mathrm{delta}\left({\mathrm{\θ}}_i\right)}=\frac{1}{K}\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\mathrm{delta}\left({\mathrm{\θ}}_i\right),$ Delta (θ in all are interval _i) * L and $\mathrm{sum}=\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\mathrm{delta}\left({\mathrm{\θ}}_i\right)*L$ As two other feature;

Described svm classifier process comprises the steps:

1. use the image construction training set (comprising true picture and stitching image) of some known class, each width of cloth image to training set extracts above-mentioned wavelet field feature, and with different zone bits indicate classification under its image (as zone bit be 1 expression its be true picture ,-1 expression stitching image);

2. svm classifier device kernel function is selected radial basis function for use, and the training set feature is sent into support vector machine SVM, utilizes cross validation, obtains to make cross validation accuracy rate the highest optimized parameter C and g, and trains support vector machine SVM with it;

3. to test pattern, extract above-mentioned wavelet field feature, and judge its affiliated classification with the support vector machine SVM that trained.

The splicing that the present invention is directed to image is distorted, and utilizes the generalized Gaussian distribution feature extraction feature of natural image wavelet sub-band coefficient, and classifies with sorter, has solved image mosaic blind Detecting problem well.For the image that will differentiate,, can provide its result of determination through sorter by it being extracted statistical nature based on generalized Gaussian distribution.

The present invention has the high and high characteristics of accuracy rate of efficient.Adopt method of the present invention, can reduce the complexity of algorithm greatly, can improve image characteristics extraction speed significantly.And when handling with sorter, classification speed is influenced hardly, so efficient improves a lot, and accuracy rate has also nearly improved 3% than the classification results of two spectrum signatures.

Description of drawings

Fig. 1 is a process flow diagram of the present invention;

Fig. 2 extracts process flow diagram for Generalized Gaussian statistical nature of the present invention.

Embodiment

Below in conjunction with accompanying drawing embodiments of the invention are elaborated: implement under the widely used in the world Columbia University of the present embodiment picture library; provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

Columbia University's picture library comprises the picture that the true picture of 933 width of cloth and 912 width of cloth are distorted through splicing, can obtain at http://www.ee.columbia.edu/ln/dvmm/downloads/AuthSplicedDataSet/dlform.html.Based on the analysis of image wavelet high-frequency sub-band coefficients statistics feature, picture be spliced distort after, its statistical nature changes, and after feature extraction, classifies through the svm classifier device, its authenticity can form a prompt judgement out.Detailed process is:

The first step: picked at random is true, distort each 770 width of cloth of picture forms training set, all the other pictures are formed test set: the true picture of 933 width of cloth is numbered 1 to 933, utilize random function, produce one group of random number between the 1-933, the numbering of picture according to random number rearranged, get preceding 770 width of cloth as the training set part, residue is as the test set part; In like manner, carry out similar operation on the pictures distorting, can set up training set and test set;

Second step: image in the training set is carried out extracting based on the statistical nature of wavelet field:

(1) to each width of cloth picture in the training set, carries out 2 grades of wavelet transforms with the Daubechies2 small echo.To the matrix of coefficients of each high-frequency sub-band, on the one hand, to being divided into K=256 interval between the maximal value, promptly burst length L=(Max-Min)/K obtains its actual probability density function at the value g of each interval center (θ with the coefficient minimum value _i); On the other hand, utilize parameter estimation method to obtain generalized Gaussian distribution variance parameter δ _θWith form parameter v, the value of parameter v as a feature, and is calculated its probability density function π (θ);

(2) in each interval center, the poor delta (θ between the high-frequency sub-band coefficient distribution density function of realistic border image and its anticipation function value _i), with its mean value $\stackrel{\‾}{\mathrm{delta}\left({\mathrm{\θ}}_i\right)}=\frac{1}{K}\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\mathrm{delta}\left({\mathrm{\θ}}_i\right),$ Delta (θ in all are interval _i) * L and $\mathrm{sum}=\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\mathrm{delta}\left({\mathrm{\θ}}_i\right)*L$ As two other feature;

The 3rd step: training svm classifier device:, indicate the affiliated classification of its image (it is true picture 1 expression ,-1 expression stitching image) with different zone bits to the training set characteristics of image that obtains.The svm classifier device can be selected libsvm for use, and kernel function is selected radial basis function for use.Characteristics of image in the training set utilizes cross validation, obtains to make cross validation accuracy rate the highest optimized parameter C and g, and with these parameters sorter is trained;

The 4th step: utilize the svm classifier device of training to carry out the judgement of the affiliated classification of image: similar with feature extraction mode in the training set, the feature of each width of cloth image by trained listening group, can obtain result of determination in the extraction test set.With classification under result of determination and the test pattern reality relatively, statistical decision result's accuracy.

In this case study on implementation, the accuracy that finally obtains is 74.25%, is 1.9 seconds to the feature extraction of every width of cloth picture and type identification averaging time.And the accuracy of conventional two spectrum statistical natures are 71.48%, and picture feature is extracted and type identification averaging time is 91.6 seconds.As can be seen, than conventional method, present embodiment all is being significantly improved aspect accuracy and the efficient.

Present embodiment comes match image wavelet high-frequency sub-band coefficient with generalized Gaussian distribution, and with the generalized Gaussian distribution density function of the match probability density function as actual high-frequency sub-band coefficient, utilizes the difference of the two to extract statistical nature.This feature extracting method still belongs to initiative.When the generation splicing is distorted, because the change of statistical nature can utilize these features that the true and false of image is judged.Experiment shows that present embodiment not only is better than classic method on the accuracy of differentiating, and efficient has had very big change.

Claims (4)

1, a kind of image mosaic blinding method based on wavelet field is characterized in that:

At first use the high frequency wavelet sub-band coefficients match generalized Gaussian distribution of natural image;

After the natural image process was distorted, the statistical property of the high-frequency sub-band coefficient that its wavelet decomposition obtains changed, thereby can't coincide well with generalized Gaussian distribution, therefore, to each high-frequency sub-band, utilized these characteristics to carry out feature extraction;

After image wavelet is decomposed, carry out parameter estimation with each high-frequency sub-band coefficient, compare with the actual distribution of high-frequency sub-band coefficient, and utilize difference extraction feature between them;

By the support vector machine svm classifier, image true-false is differentiated then.

2, the image mosaic blinding method based on wavelet field according to claim 1 is characterized in that, described image characteristics extraction comprises the steps:

1. after image carries out wavelet transform,, suppose coefficient obedience generalized Gaussian distribution to each high-frequency sub-band, the big or small high-frequency sub-band coefficient w for M*N of utilization (i, j), find the solution following equation:

$\begin{array}{c}{\mathrm{\δ}}_{\mathrm{\θ}}^2=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{w}^2(i,j)\\ \frac{\mathrm{\Γ}(1/v)\mathrm{\Γ}(5/v)}{{\mathrm{\Γ}}^2(3/v)}=\frac{1}{{\mathrm{\δ}}_{\mathrm{\θ}}^4}\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{w}^4(i,j)\end{array}$

Wherein: Γ () is the gamma function, can obtain the form parameter v and the variance parameter δ of generalized Gaussian distribution probability density function _θAnd the parameter value substitution following formula that estimates can be obtained the generalized Gaussian distribution probability density function: $\mathrm{\π}\left(\mathrm{\θ}\right)=\frac{\mathrm{v\η}\left(v\right)}{2\mathrm{\Γ}(1/v)}\frac{1}{{\mathrm{\δ}}_{\mathrm{\θ}}}\exp \{-{\left[\mathrm{\η}\left(v\right)\right|\mathrm{\θ}|/{\mathrm{\δ}}_{\mathrm{\θ}}]}^v\}$

Wherein $\mathrm{\η}\left(v\right)=\sqrt{\mathrm{\Γ}(3/v)/\mathrm{\Γ}(1/v)};$ Parameter v is as a feature;

2. with the coefficient minimum value of high-frequency sub-band coefficient to be divided between the maximal value K interval, i.e. burst length L=(Max-Min)/K; Design factor drops on each interval interior Probability p (θ _i) (i=1,2...K), then the actual probabilities density fonction is approximately in the value of each interval center: g (θ _i)=p (θ _i)/L;

3. in each interval center, poor between the high-frequency sub-band coefficient distribution density function of realistic border image and its anticipation function value: delta (θ _i)=| g (θ _i)-π (θ _i) |, with its mean value $\stackrel{\‾}{\mathrm{delta}\left({\mathrm{\θ}}_i\right)}=\frac{1}{K}\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\mathrm{delta}\left({\mathrm{\θ}}_i\right),$ Delta (θ in all are interval _i) * L and $\mathrm{sum}=\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\mathrm{delta}\left({\mathrm{\θ}}_i\right)*L$ As two other feature.

3, the image mosaic blinding method based on wavelet field according to claim 2 is characterized in that, described support vector machine svm classifier process comprises the steps:

1. with comprising the image construction training set of true picture and stitching image known class, each width of cloth image of training set is extracted above-mentioned wavelet field feature, and indicate classification under its image with zone bit;

2. svm classifier device kernel function is selected radial basis function for use, and the training set feature is sent into support vector machine SVM, utilizes cross validation, obtains to make cross validation accuracy rate the highest optimized parameter C and g, and trains support vector machine SVM with it;

3. to test pattern, extract above-mentioned wavelet field feature, and judge its affiliated classification with the support vector machine SVM that trained.

4, the image mosaic blinding method based on wavelet field according to claim 3 is characterized in that, described under classification, be meant zone bit be 1 expression its for true picture ,-1 expression stitching image.

Images