CN106384077A - Low bit rate video based camera recognition method and device - Google Patents

Abstract

The invention discloses a low bit rate video based camera recognition method and device. The camera recognition method comprises the steps of building a noise statistical model of the video, performing filtering processing by using a filter so as to acquire an order statistic distribution function of the noise statistical model; performing principal component analysis on the video, building a column span mapping matrix of a feature vector in a null space; calculating the direction and the threshold of the column span mapping matrix in a hyperplane so as to determine an optimization problem; determining a camera recognition condition according to an input-output relation between a camera separation sensor and a camera and the optimization problem; judging whether the video to be recognized meets the camera recognition condition or not; if so, determining the video to be recognized to be photographed by the camera to be recognized; and if not, determining the video to be recognized to be not photographed by the camera to be recognized. By adopting the embodiment of the invention, a time varying problem of quantization noises and an instability problem of light response non-homogeneous noises can be solved, the calculation amount is reduced, and the camera recognition accuracy is improved.

Description

A kind of camera recognition methodss based on low-bit rate video and device

Technical field

The present invention relates to field of computer technology, more particularly, to a kind of camera recognition methodss based on low-bit rate video and Device.

Background technology

With the popularization of various digital implementations and digital product, in social life, everyone is owned by the mobile phone of oneself substantially Or camera, and these digital products are commonly used to take pictures or shoot video.Offender also commonly uses camera to do illegal thing, such as makes Shooting someone secret photo or indecency video with threat or violent meanses, then threatening someone with these photos, so that being subject to Evil person is used by offender.And when victim hold photo go report to the police when, offender can also recognize that these photos are not He shoot so that being legally difficult to enough evidences to determine the crime of offender.But if can determine these photos It is captured by the camera of offender it is possible to restrain offender by law.Therefore, camera identification is recorded a video in child porn and is collected evidence In play an important role, whether be used for enrolling one group of given image and video by the camera of the suspect of test Whether, suspect can be conducive to restrain by law as a kind of reliable evidence.

Current camera recognition technology mainly by calculate camera photoresponse heterogeneity noise correlation coefficient Lai Determine, if this correlation coefficient is high, regard as the former camera shooting.But because the deposits of faith such as present video are to make an uproar The lower storage of sound pattern statistics, these noises mainly include quantizing noise and photoresponse heterogeneity noise.And in mobile device When being used for shooting video, these videos are all to compress preservation with low-down bit rate so that quantizing noise and photoresponse are non- Uniformity noise suffers serious distorting.The quantizing noise of each two field picture of one video all different it is meant that whole mould Type changes constantly.And prior art is just being to rely on this noise model and existing algorithm is identified, recognition result is not Stable and inaccurate, and arithmetic speed is not quick, required arithmetic hardware require higher it is impossible to apply in common computer On.

Content of the invention

The embodiment of the present invention proposes a kind of camera recognition methodss based on low-bit rate video and device, solves quantizing noise Time-varying problem and photoresponse heterogeneity noise instability problem, and reduce amount of calculation and improve camera identification standard Really rate.

The embodiment of the present invention provides a kind of camera recognition methodss based on low-bit rate video, including：

Set up the noise statistical model of video, and described noise statistical model is filtered process using wave filter, obtain Obtain the order statistic distribution function of described noise statistical model；Described video is several videos that camera to be identified shoots；

Principal component analysiss are carried out to described video, to extract described order statistic distribution function in the odd-order static state moment Characteristic vector；

Build the column span mapping matrix in kernel for the described characteristic vector, so that described characteristic vector all can be mapped in institute State in the column vector of column span mapping matrix；

Calculate described column span mapping matrix in the direction of hyperplane and threshold value, to determine optimization problem；

Input/output relation according to camera separation perceptron and described camera to be identified and described optimization problem, determine phase Machine identifies condition；

Obtain video to be identified, and judge whether described video to be identified meets described camera identification condition；

If it is satisfied, then determining that described video to be identified is shot and obtained by described camera to be identified；Otherwise, it determines it is described Video to be identified is not to be obtained by described camera shooting to be identified.

Further, the described noise statistical model setting up video, and using wave filter, described noise statistical model is entered Row Filtering Processing, obtains the order statistic distribution function of described noise statistical model, specially：

Obtain the P captured by camera C to be identified_cIndividual video, and defineFor P_cIn individual videoFrame The real-valued overall noise model of the M × N of video, usesRepresent m row n-th rowElement；

By wavelet filter and Wiener filter to describedIt is filtered processing, obtain：

$X_{l_{p_c},p_c,c}(m,n)=\frac{W_{l_{p_c},p_c,c}(m,n){\hat{I}}_{l_{p_c},p_c,c}(m,n)}{{\hat{I}}_{l_{p_c},p_c,c}^2(m,n)};$

Wherein, describedFor described P_cIn individual videoTwo field picture；For the letter being filtered out Breath；

Then described order statistic distribution functionFor：

$g_{l_{p_c},p_c,c}\left(k\right)=\frac{1}{MN}\underset{b=0}{\overset{B-1}{\Σ}}{\left({\tilde{X}}_{l_{p_c},p_c,c}\right(b\left)\right)}^k{f}_{l_{p_c},p_c,c}\left(b\right);$

Wherein,Respective table indicating value for b-th grid；Probability close Degree distribution function.

Further, described principal component analysiss are carried out to described video, existed with extracting described order statistic distribution function The characteristic vector in odd-order static state moment, specially：

Described order statistic distribution function odd-order static state the moment vector beThen：

$g_{l_{p_c},p_c,c}={\left[\begin{array}{ccc}{g}_{l_{p_c},p_c,c}\left(3\right)& g_{l_{p_c},p_c,c}\left(5\right)& g_{l_{p_c},p_c,c}\left(7\right)\end{array}\right]}^T;$

$R_{p_c,c}=\frac{1}{L_{p_c}^2}\underset{{\stackrel{\‾}{l}}_{p_c}=0}{\overset{L_{p_c}-1}{\Σ}}\underset{{\tilde{l}}_{p_c}=0}{\overset{L_{p_c}-1}{\Σ}}(g_{{\stackrel{\‾}{l}}_{p_c},p_c,c}-g_{{\tilde{l}}_{p_c},p_c,c}){(g_{{\stackrel{\‾}{l}}_{p_c},p_c,c}-g_{{\tilde{l}}_{p_c},p_c,c})}^T;$

$y_{l_{p_c},p_c,c}=g_{l_{p_c},p_c,c}^T{v}_{p_c,c,0};$

Wherein,WithForK-th eigenvalue and corresponding characteristic vector；

Described order statistic distribution function odd-order static state the moment characteristic vector be：

Further, described build described characteristic vector kernel column span mapping matrix so that described feature to Amount all can be mapped in the column vector of described column span mapping matrix, specially：

Separation matrix in note class and the separation matrix between class are respectively Γ₁And Γ₂, then have：

${\mathrm{\Γ}}_1=\frac{1}{C}\underset{c=0}{\overset{C-1}{\Σ}}\frac{1}{{P_c}^2}\underset{{\stackrel{\‾}{p}}_c=0}{\overset{P_c-1}{\Σ}}\underset{{\tilde{p}}_c=0}{\overset{P_c-1}{\Σ}}(Y_{{\tilde{p}}_c,c}-Y_{{\stackrel{\‾}{p}}_c,c}){(Y_{{\tilde{p}}_c,c}-Y_{{\stackrel{\‾}{p}}_c,c})}^T;$

${\mathrm{\Γ}}_2=\frac{1}{C(C-1)}\underset{c_1=0}{\overset{C-1}{\Σ}}\underset{c_1\≠c_2}{\overset{C-1}{\underset{c_2=0}{\Σ}}}\frac{1}{P_{c_1}{P}_{c_2}}\underset{{\tilde{p}}_{c_1}=0}{\overset{P_{c_1}-1}{\Σ}}\underset{{\stackrel{\‾}{p}}_{c_2}=0}{\overset{P_{c_2}-1}{\Σ}}(Y_{{\tilde{p}}_{c_1},c}-Y_{{\stackrel{\‾}{p}}_{c_2},c}){(Y_{{\tilde{p}}_{c_1},c}-Y_{{\stackrel{\‾}{p}}_{c_2},c})}^T;$

Note α is mapping hyperplane, then the object function of minimum inter- object distance and maximum between class distance is defined as：And

Then described column span mapping matrix is：

Wherein,Characteristic vector for classification

Further, the described column span mapping matrix of described calculating is in the direction of hyperplane and threshold value, to determine that optimization asks Topic, specially：

Define the direction of linear hyperplane and threshold value is respectivelyWithThen have：

With

If there isWithMeet following two formula, then simultaneouslyWithLinear separability；

With

Then described optimization problem is：

$\begin{array}{cc}& \underset{(w_{c_1,c_2},{\mathrm{\δ}}_{c_1,c_2})}{\max }{\mathrm{\ϵ}}_{c_1,c_2},\\ subjectto& \begin{array}{c}{w}_{c_1,c_2}^T{A}_{p_{c_1},c_1}+{\mathrm{\δ}}_{c_1,c_2}\≥{\mathrm{\ϵ}}_{c_1,c_2}for{p}_{c_1}=0,\mathrm{...},P_{c_1}-1,\\ w_{c_1,c_2}^T{A}_{p_{c_2},c_2}+{\mathrm{\δ}}_{c_1,c_2}\≤-{\mathrm{\ϵ}}_{c_1,c_2}for{p}_{c_2}=0,\mathrm{...},P_{c_2}-1,\end{array}\end{array}$

$\begin{array}{cc}and& {\mathrm{\ϵ}}_{c_1,c_2}\≥\stackrel{\‾}{\mathrm{\ϵ}}\end{array};$

Wherein,It is the vigorousness specification controlling linear programming problem, and

Further, the described input/output relation according to camera separation perceptron and described camera to be identified and described excellent Change problem, determines camera identification condition, specially：

Define c_iThe separation perceptron of individual camera and c_jThe input/output relation of individual camera is And c_j≠c_i；

Then described camera identification condition is：

${\mathrm{\ψ}}_{c_i,c_j}\left(A_{p_{c_i},c_i}\right)=Q(w_{c_i,c_j}^T{A}_{p_{c_i},c_i}+{\mathrm{\δ}}_{c_i,c_j})=1;$

With

Correspondingly, the embodiment of the present invention also provides a kind of camera identifying device based on low-bit rate video, including：

Filtering module, for setting up the noise statistical model of video, and is entered to described noise statistical model using wave filter Row Filtering Processing, obtains the order statistic distribution function of described noise statistical model；Described video is that camera to be identified shoots Several videos；

Principal component analysiss module, for carrying out principal component analysiss to described video, to extract described order statistic distribution Function is in the characteristic vector in odd-order static state moment；

Mapping matrix builds module, for building the column span mapping matrix in kernel for the described characteristic vector, so that institute State characteristic vector all can be mapped in the column vector of described column span mapping matrix；

Optimization problem determining module, for calculating described column span mapping matrix in the direction of hyperplane and threshold value, with true Determine optimization problem；

Identification condition determining module, for the input/output relation according to camera separation perceptron and described camera to be identified With described optimization problem, determine camera identification condition；

Judge module, for obtaining video to be identified, and judges whether described video to be identified meets described camera identification Condition；

In described judge module, recognition result determining module, for determining that described video to be identified meets described camera identification During condition, determine that described video to be identified is shot and obtained by described camera to be identified；And for true in described judge module When fixed described video to be identified is unsatisfactory for described camera identification condition, determine that described video to be identified is not by described phase to be identified Machine shoots and obtains.

Implement the embodiment of the present invention, have the advantages that：

A kind of camera recognition methodss based on low-bit rate video provided in an embodiment of the present invention and device, first to video Noise statisticses model is processed, and video is filtered process with wave filter, then confirms which photograph frame to video Processed；Again by the principal component analysiss to video, to reduce the quantity of video, and to be minimized using linear discriminant analysiss It is spaced and maximizes interval in the class of characteristic vector in class, with construction feature vector in the row moment of span mapping matrix of kernel, count Calculate it in the direction of hyperplane and threshold value, have confirmed that optimization problem；Separate perceptron and phase finally according to optimization problem with camera The input/output relation of machine, determines camera identification condition, when camera to be identified and video to be identified meet this identification condition, really This video fixed is shot and obtained by this camera.Pass through compared to prior art to calculate the phase of the photoresponse heterogeneity noise of camera Close coefficient to carry out camera identification, the present invention can solve the problems, such as the time-varying of quantizing noise and the shakiness of photoresponse heterogeneity noise Qualitative question, and reduce amount of calculation and improve accuracy rate.

Brief description

Fig. 1 is that a kind of flow process of embodiment of the camera recognition methodss based on low-bit rate video that the present invention provides is illustrated Figure；

Fig. 2 is a kind of structural representation of embodiment of the camera identifying device based on low-bit rate video that the present invention provides Figure.

Specific embodiment

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete Site preparation description is it is clear that described embodiment is only a part of embodiment of the present invention, rather than whole embodiments.It is based on Embodiment in the present invention, it is every other that those of ordinary skill in the art are obtained under the premise of not making creative work Embodiment, broadly falls into the scope of protection of the invention.

Referring to Fig. 1, it is a kind of schematic flow sheet of embodiment of the camera recognition methodss of low bit rate that the present invention provides. 101 to step that the method comprising the steps of 108.Each step is specific as follows：

Step 101：Set up the noise statistical model of video, and noise statistical model is filtered locate using wave filter Reason, obtains the order statistic distribution function of noise statistical model；Video is several videos that camera to be identified shoots.

In this example, it is assumed that had C camera and captured by the C camera P out_cIndividual video, obtains P captured by camera C to be identified_cIndividual video.In P_cExtract in individual videoFrame video, and defineForThe real-valued overall noise model of the M × N of the video of frame, usesRepresent m row n-th rowUnit Element.

By wavelet filter and Wiener filter to describedIt is filtered processing, specially：Vacation lets f be one Individual wavelet low-pass filter, andRepresent the P that the C camera shoots_cThe of individual videoTwo field picture.DefinitionFor the output of wavelet filter,For the information being filtered out, then haveWithObtain after Wiener filter again

Then willBe converted to a vector, and this vector representation isReuse mN+nIt is expressed asThen have：

${\stackrel{\‾}{X}}_{l_{p_c},p_c,c}(mN+n)=X_{l_{p_c},p_c,c}(m,n);$

Then order statistic distribution functionFor：

$g_{l_{p_c},p_c,c}\left(k\right)=\frac{1}{MN}\underset{b=0}{\overset{B-1}{\Σ}}{\left({\tilde{X}}_{l_{p_c},p_c,c}\right(b\left)\right)}^k{f}_{l_{p_c},p_c,c}\left(b\right);$

Wherein,Respective table indicating value for b-th grid；ForProbability close Degree distribution function,ForTotal amount of element.

AssumeDifference approximation to function in laplacian distribution, then have WhereinIt is the control parameter of distribution function spread speed.BecauseIt is an even function and x^2k-1 It is an odd function, so having

Due toThe static moment to the photoresponse heterogeneity noise of video is highstrung, institute Camera identification is processed with a kind of present invention proposition odd-order static state moment with global noise model.

Step 102：Principal component analysiss are carried out to video, with sequence of extraction statistics distribution function in the odd-order static state moment Characteristic vector.

In the present embodiment, due in actual camera identification process, needing to processIndividual data, These data volumes are very big, need the computer of very strong computing capability just to enable, therefore the present invention adopts principal component analysiss side Method, to reduce data volume, is identified with the camera realizing common computer.

In the present embodiment, step 102 is specially：Note order statistic distribution function is in the vector in odd-order static state moment ForThen have：

$g_{l_{p_c},p_c,c}={\left[\begin{array}{ccc}{g}_{l_{p_c},p_c,c}\left(3\right)& g_{l_{p_c},p_c,c}\left(5\right)& g_{l_{p_c},p_c,c}\left(7\right)\end{array}\right]}^T;$

$R_{p_c,c}=\frac{1}{L_{p_c}^2}\underset{{\stackrel{\‾}{l}}_{p_c}=0}{\overset{L_{p_c}-1}{\Σ}}\underset{{\tilde{l}}_{p_c}=0}{\overset{L_{p_c}-1}{\Σ}}(g_{{\stackrel{\‾}{l}}_{p_c},p_c,c}-g_{{\tilde{l}}_{p_c},p_c,c}){(g_{{\stackrel{\‾}{l}}_{p_c},p_c,c}-g_{{\tilde{l}}_{p_c},p_c,c})}^T;$

$y_{l_{p_c},p_c,c}=g_{l_{p_c},p_c,c}^T{v}_{p_c,c,0};$

Wherein,WithForK-th eigenvalue and corresponding characteristic vector；

Described order statistic distribution function odd-order static state the moment characteristic vector be：

Step 103：The vectorial column span mapping matrix in kernel of construction feature, so that characteristic vector all can be mapped in row In the column vector of span mapping matrix.

In the present embodiment, step 103 is specially：Separation matrix in note class and the separation matrix between class are respectively Γ₁ And Γ₂, then have：

${\mathrm{\Γ}}_1=\frac{1}{C}\underset{c=0}{\overset{C-1}{\Σ}}\frac{1}{{P_c}^2}\underset{{\stackrel{\‾}{p}}_c=0}{\overset{P_c-1}{\Σ}}\underset{{\tilde{p}}_c=0}{\overset{P_c-1}{\Σ}}(Y_{{\tilde{p}}_c,c}-Y_{{\stackrel{\‾}{p}}_c,c}){(Y_{{\tilde{p}}_c,c}-Y_{{\stackrel{\‾}{p}}_c,c})}^T;$

${\mathrm{\Γ}}_2=\frac{1}{C(C-1)}\underset{c_1=0}{\overset{C-1}{\Σ}}\underset{c_1\≠c_2}{\overset{C-1}{\underset{c_2=0}{\Σ}}}\frac{1}{P_{c_1}{P}_{c_2}}\underset{{\tilde{p}}_{c_1}=0}{\overset{P_{c_1}-1}{\Σ}}\underset{{\stackrel{\‾}{p}}_{c_2}=0}{\overset{P_{c_2}-1}{\Σ}}(Y_{{\tilde{p}}_{c_1},c}-Y_{{\stackrel{\‾}{p}}_{c_2},c}){(Y_{{\tilde{p}}_{c_1},c}-Y_{{\stackrel{\‾}{p}}_{c_2},c})}^T;$

Note α is mapping hyperplane, then the object function of minimum inter- object distance and maximum between class distance is defined as：And

If the way according to prior art is：Add constraint and add α^TΓ₂α=1, the then optimization of Prescribed Properties is asked Topic can be converted into the optimization problem of unconfined condition：It solves and isHaveExisting is by intrinsic Method looking forAnd α^*.And α^*It is respectivelyEigenvalue and characteristic vector.But due to Γ₂Eigenvalue connect It is bordering on 0, andIt is also unstable, so being solved using prior art is not a good method.The therefore present invention Propose by looking forKernel column span solving this optimization problem.

The column span mapping matrix of the present invention is：

Wherein,Span beExtract the sample point of 30 in an experiment.So havingWherein k=0 ..., 29.IfThen have One new vectorCan be drawn by equation below：This vector can conduct Last characteristic vector is used for classifying.

Step 104：Calculate column span mapping matrix in the direction of hyperplane and threshold value, to determine optimization problem.

In the present embodiment, step 104 is specially：Define the direction of linear hyperplane and threshold value is respectivelyWithThen have：

With

DefinitionIf there isWithMeet following two public affairs simultaneously Formula, thenWithLinear separability；

With

With

With

If： WithBe arbitrarily to initial value.Then have：

${\stackrel{\‾}{w}}_{c_1,c_2}(k+1)={\stackrel{\‾}{w}}_{c_1,c_2}\left(k\right)+\frac{t_{c_1,c_2}\left(k\right)-Q\left({\stackrel{\‾}{w}}_{c_1,c_2}^T\right(k\left){\stackrel{\‾}{A}}_{c_1,c_2}\right(k\left)\right)}{2}{\stackrel{\‾}{A}}_{c_1,c_2}\left(k\right)$

Then described optimization problem is：

$\begin{array}{cc}& \underset{(w_{c_1,c_2},{\mathrm{\δ}}_{c_1,c_2})}{\max }{\mathrm{\ϵ}}_{c_1,c_2},\\ subjectto& \begin{array}{c}{w}_{c_1,c_2}^T{A}_{p_{c_1},c_1}+{\mathrm{\δ}}_{c_1,c_2}\≥{\mathrm{\ϵ}}_{c_1,c_2}for{p}_{c_1}=0,\mathrm{...},P_{c_1}-1,\\ w_{c_1,c_2}^T{A}_{p_{c_2},c_2}+{\mathrm{\δ}}_{c_1,c_2}\≤-{\mathrm{\ϵ}}_{c_1,c_2}for{p}_{c_2}=0,\mathrm{...},P_{c_2}-1,\end{array}\end{array}$

$\begin{array}{cc}and& {\mathrm{\ϵ}}_{c_1,c_2}\≥\stackrel{\‾}{\mathrm{\ϵ}}\end{array};$

Wherein,It is the vigorousness specification controlling linear programming problem, and

IfWithFor every a pair (c_i,c_j) It is all linear separability, that is, claim this data to be linear separability, wherein c two-by-two_j≠c_i,.

Step 105：Input/output relation according to camera separation perceptron and camera to be identified and optimization problem, determine phase Machine identifies condition.

In the present embodiment, step 105 is specially：Define c_iThe separation perceptron of individual camera and c_jIndividual camera defeated Entering output relation isAnd c_j≠c_i；

Then described camera identification condition is：

${\mathrm{\ψ}}_{c_i,c_j}\left(A_{p_{c_i},c_i}\right)=Q(w_{c_i,c_j}^T{A}_{p_{c_i},c_i}+{\mathrm{\δ}}_{c_i,c_j})=1;$

With

Step 106：Obtain video to be identified, and judge whether video to be identified meets camera identification condition；If it is satisfied, Then execution step 107, otherwise, execution step 108.

Step 107：Determine that video to be identified is shot and obtained by camera to be identified.

Step 108：Determine that video to be identified is not to be shot by camera to be identified and obtains.

In order to better illustrate the beneficial effect of technical solution of the present invention, can be found in the recognition result of table 1 below -4.Experiment The recognition result accuracy rate that result can be seen that method proposed by the invention is 100%, traditional correlation coefficient method accurate Rate is 30%, and the accuracy rate of traditional Nearest Neighbor Method method is 30%, and the accuracy rate of support vector machine method is 100%.But It is because support vector machine method amount of calculation required in the training process is far longer than the method that we are proposed, so supporting Operation time needed for vector machine is also long, be not as rapid as method proposed by the present invention.So method proposed by the present invention obtains The result going out much is better than the result that traditional correlation coefficient method and traditional Nearest Neighbor Method method draw, and it is proposed that The required in the training process amount of calculation of method be also far smaller than traditional support vector machine method.

The recognition result of table 1 method proposed by the invention

The recognition result of the traditional correlation coefficient method of table 2

The recognition result of the traditional Nearest Neighbor Method method of table 3

The recognition result of table 4 support vector machine method

Correspondingly, referring to Fig. 2, Fig. 2 is the one kind of the camera identifying device based on low-bit rate video that the present invention provides The structural representation of embodiment.This camera identifying device includes：Filtering module 201, principal component analysiss module 202, mapping matrix Build module 203, optimization problem determining module 204, identification condition determining module 205, judge module 206 and recognition result to determine Module 207.

Wherein, filtering module 201 is used for setting up the noise statistical model of video, and counts mould using wave filter to this noise Type is filtered processing, and obtains the order statistic distribution function of noise statistical model；If video is camera to be identified shooting Dry video.

Principal component analysiss module 202 is used for carrying out principal component analysiss to video, is existed with sequence of extraction statistics distribution function The characteristic vector in odd-order static state moment.

Mapping matrix builds module 203 and is used for the column span mapping matrix in kernel for the construction feature vector, so that feature Vector all can be mapped in the column vector of column span mapping matrix.

Optimization problem determining module 204 is used for calculating column span mapping matrix in the direction of hyperplane and threshold value, to determine Optimization problem.

Identification condition determining module 205 be used for input/output relation according to camera separation perceptron and camera to be identified and Optimization problem, determines camera identification condition.

Judge module 206 is used for obtaining video to be identified, and judges whether video to be identified meets camera identification condition.

Recognition result determining module 207 is used for when judge module 206 determines that video to be identified meets camera identification condition, Determine that video to be identified is shot and obtained by camera to be identified；And for determining that video to be identified is unsatisfactory for phase in judge module During machine identification condition, determine that video to be identified is not to be shot and obtained by camera to be identified.

Therefore, a kind of camera recognition methodss based on low-bit rate video provided in an embodiment of the present invention and device, First the noise statisticses model of video is processed, and video is filtered process with wave filter, then confirm to video Which photograph frame is processed；Again by the principal component analysiss to video, to reduce the quantity of video, and divided using linear discriminant Analysis, to minimize interval and the interior interval of class maximizing characteristic vector in class, is reflected in the row moment of span of kernel with construction feature vector Penetrate matrix, calculate it in the direction of hyperplane and threshold value, have confirmed that optimization problem；Separate sense finally according to optimization problem with camera Know the input/output relation of device and camera, determine camera identification condition, when camera to be identified and video to be identified meet this identification During condition, determine that this video is shot and obtained by this camera.The photoresponse passing through to calculate camera compared to prior art is non-homogeneous Carrying out camera identification, the present invention can solve the problems, such as time-varying and the photoresponse heterogeneity of quantizing noise to the correlation coefficient of property noise The instability problem of noise, and reduce amount of calculation and improve accuracy rate.

The above is the preferred embodiment of the present invention it is noted that for those skilled in the art For, under the premise without departing from the principles of the invention, some improvements and modifications can also be made, these improvements and modifications are also considered as Protection scope of the present invention.

Claims (7)

1. a kind of camera recognition methodss based on low-bit rate video are it is characterised in that include：

Set up the noise statistical model of video, and described noise statistical model is filtered process using wave filter, obtain institute State the order statistic distribution function of noise statistical model；Described video is several videos that camera to be identified shoots；

Principal component analysiss are carried out to described video, to extract the spy in the odd-order static state moment for the described order statistic distribution function Levy vector；

Build the column span mapping matrix in kernel for the described characteristic vector, so that described characteristic vector all can be mapped in described row In the column vector of span mapping matrix；

Calculate described column span mapping matrix in the direction of hyperplane and threshold value, to determine optimization problem；

Input/output relation according to camera separation perceptron and described camera to be identified and described optimization problem, determine that camera is known Other condition；

Obtain video to be identified, and judge whether described video to be identified meets described camera identification condition；

If it is satisfied, then determining that described video to be identified is shot and obtained by described camera to be identified；Otherwise, it determines described wait to know Other video is not to be obtained by described camera shooting to be identified.

2. the camera recognition methodss based on low-bit rate video according to claim 1 are it is characterised in that described foundation regards The noise statistical model of frequency, and described noise statistical model is filtered process using wave filter, obtain described noise statistics The order statistic distribution function of model, specially：

Obtain the P captured by camera C to be identified_cIndividual video, and defineFor P_cIn individual videoThe video of frame M × N real-valued overall noise model, useRepresent m row n-th rowElement；

By wavelet filter and Wiener filter to describedIt is filtered processing, obtain：

$X_{l_{p_c},p_c,c}(m,n)=\frac{W_{l_{p_c},p_c,c}(m,n){\hat{I}}_{l_{p_c},p_c,c}(m,n)}{{\hat{I}}_{l_{p_c},p_c,c}^2(m,n)};$

Wherein, describedFor described P_cIn individual videoTwo field picture；For the information being filtered out；

Then described order statistic distribution functionFor：

$g_{l_{p_c},p_c,c}\left(k\right)=\frac{1}{MN}\underset{b=0}{\overset{B-1}{\Σ}}{\left({\tilde{X}}_{l_{p_c},p_c,c}\right(b\left)\right)}^k{f}_{l_{p_c},p_c,c}\left(b\right);$

Wherein,Respective table indicating value for b-th grid；ForProbability density divide Cloth function.

3. the camera recognition methodss based on low-bit rate video according to claim 2 it is characterised in that described to described Video carries out principal component analysiss, to extract the characteristic vector in the odd-order static state moment for the described order statistic distribution function, tool Body is：

Described order statistic distribution function odd-order static state the moment vector beThen：

$g_{l_{p_c},p_c,c}={\left[\begin{array}{ccc}{g}_{l_{p_c},p_c,c}\left(3\right)& g_{l_{p_c},p_c,c}\left(5\right)& g_{l_{p_c},p_c,c}\left(7\right)\end{array}\right]}^T;$

$R_{p_c,c}=\frac{1}{L_{p_c}^2}\underset{{\stackrel{\‾}{l}}_{p_c}=0}{\overset{L_{p_c}-1}{\Σ}}\underset{{\tilde{l}}_{p_c}=0}{\overset{L_{p_c}-1}{\Σ}}(g_{{\stackrel{\‾}{l}}_{p_c},p_c,c}-g_{{\tilde{l}}_{p_c},p_c,c}){(g_{{\stackrel{\‾}{l}}_{p_c},p_c,c}-g_{{\tilde{l}}_{p_c},p_c,c})}^T;$

$y_{l_{p_c},p_c,c}=g_{l_{p_c},p_c,c}^T{v}_{p_c,c,0};$

Wherein,WithForK-th eigenvalue and corresponding characteristic vector；

Described order statistic distribution function odd-order static state the moment characteristic vector be：

4. the camera recognition methodss based on low-bit rate video according to claim 3 are it is characterised in that described structure institute State the column span mapping matrix in kernel for the characteristic vector, so that described characteristic vector all can be mapped in described column span mapping square In the column vector of battle array, specially：

Separation matrix in note class and the separation matrix between class are respectively Γ₁And Γ₂, then have：

${\mathrm{\Γ}}_1=\frac{1}{C}\underset{c=0}{\overset{C-1}{\Σ}}\frac{1}{{P_c}^2}\underset{{\stackrel{\‾}{p}}_c=0}{\overset{P_c-1}{\Σ}}\underset{{\tilde{p}}_c=0}{\overset{P_c-1}{\Σ}}(Y_{{\tilde{p}}_c,c}-Y_{{\stackrel{\‾}{p}}_c,c}){(Y_{{\tilde{p}}_c,c}-Y_{{\stackrel{\‾}{p}}_c,c})}^T;$

${\mathrm{\Γ}}_2=\frac{1}{C(C-1)}\underset{c_1=0}{\overset{C-1}{\Σ}}\underset{c_1\≠c_2}{\overset{C-1}{\underset{c_2=0}{\mathrm{\Σ}}}}\frac{1}{P_{c_1}{P}_{c_2}}\underset{{\tilde{p}}_{c_1}=0}{\overset{P_{c_1}-1}{\Σ}}\underset{{\stackrel{\‾}{p}}_{c_2}=0}{\overset{P_{c_2}-1}{\Σ}}(Y_{{\tilde{p}}_{c_1},c_1}-Y_{{\stackrel{\‾}{p}}_{c_2},c_2}){(Y_{{\tilde{p}}_{c_1},c_1}-Y_{{\stackrel{\‾}{p}}_{c_2},c_2})}^T;$

Note α is mapping hyperplane, then the object function of minimum inter- object distance and maximum between class distance is defined as：And J (z α)=J (α)_,

Then described column span mapping matrix is：

Wherein,Characteristic vector for classification

5. the camera recognition methodss based on low-bit rate video according to claim 4 are it is characterised in that described calculating institute State column span mapping matrix in the direction of hyperplane and threshold value, to determine optimization problem, specially：

Define the direction of linear hyperplane and threshold value is respectivelyWithThen have：

With

If there isWithMeet following two formula, then simultaneouslyWith Linear separability；

With

Then described optimization problem is：

Wherein,It is the vigorousness specification controlling linear programming problem, and

6. the camera recognition methodss based on low-bit rate video according to claim 5 it is characterised in that described according to phase The input/output relation of machine separation perceptron and described camera to be identified and described optimization problem, determine camera identification condition, tool Body is：

Define c_iThe separation perceptron of individual camera and c_jThe input/output relation of individual camera is And c_j≠c_i；

Then described camera identification condition is：

${\mathrm{\ψ}}_{c_i,c_j}\left(A_{p_{c_i},c_i}\right)=Q(w_{c_i,c_j}^T{A}_{p_{c_i},c_i}+{\mathrm{\δ}}_{c_i,c_j})=1;$

With

7. a kind of camera identifying device based on low-bit rate video is it is characterised in that include：

Filtering module, for setting up the noise statistical model of video, and is filtered to described noise statistical model using wave filter Ripple is processed, and obtains the order statistic distribution function of described noise statistical model；If described video is camera to be identified shooting Dry video；

Principal component analysiss module, for carrying out principal component analysiss to described video, to extract described order statistic distribution function Characteristic vector in the odd-order static state moment；

Mapping matrix builds module, for building the column span mapping matrix in kernel for the described characteristic vector, so that described spy Levy vector all can be mapped in the column vector of described column span mapping matrix；

Optimization problem determining module, for calculating described column span mapping matrix in the direction of hyperplane and threshold value, excellent to determine Change problem；

Identification condition determining module, for the input/output relation according to camera separation perceptron and described camera to be identified and institute State optimization problem, determine camera identification condition；

Judge module, for obtaining video to be identified, and judges whether described video to be identified meets described camera identification condition；

In described judge module, recognition result determining module, for determining that described video to be identified meets described camera identification condition When, determine that described video to be identified is shot and obtained by described camera to be identified；And for determining institute in described judge module When stating video to be identified and being unsatisfactory for described camera identification condition, determine that described video to be identified is not to be clapped by described camera to be identified Take the photograph and obtain.