CN101441717B - Method and system for detecting eroticism video - Google Patents

Abstract

The invention provides a pornographic-video detection method, which comprises the steps of utilizing a complexion model and a non-complexion model to calculate the complexion probability value and non-complexion probability value of every pixel in video frames of a to-be-detected video, establishing template images for the video frames according to the complexion probability value and non-complexion probability value of every pixel in the video frames, extracting image features from the video frames according to the template image, forming an observation sequence by use of the image features in the continuous video frames and inputting the observation sequence into a pornographic-shot model so as to detect whether the to-be-detected video is a pornographic video. The invention also provides a detection system adapted to the pornographic-video detection method. The detection method adopts the timing characteristics of the adjacent video frames during detection, thereby ensuring stronger robustness and higher detection accuracy of the whole video detection method.

Description

A kind of pornographic video detecting method and detection system

Technical field

The present invention relates to the Video Detection field, particularly a kind of pornographic video detecting method and corresponding detecting system.

Background technology

Along with broad application of Internet, the user also may run into many flames when obtaining a large amount of useful informations, wherein especially serious with pornographic video.Search engine has worsened this phenomenon because of it can offer the user with the station address of being grown up in a large number.Therefore, detection and the filtration of carrying out pornographic information has great importance.

Mainly there are two kinds of filter types in the pornographic Video Detection in past.First kind is to carry out shutoff by the network address that detects the porn site.But the network address of porn site normally often changes, and this system needs great amount of manpower to collect the network address of porn site, so the mode of network address shutoff is not a kind of feasible method.Another kind is to detect by video corresponding text content.Although the text content analysis method of video correspondence is more accurate, this method has very strong dependence to text message, and pornographic video does not have corresponding text introduction usually, so this method is not reliable method.

Based on above analysis, detecting its content according to the visual signature of video own is more reliable and feasible settling mode.At present, existing a large amount of researchist is engaged in the research of this aspect.In the prior art, the method for utilizing the visual signature of video self to detect pornographic video mainly comprises following step:

Whether 1, extract a frame of video from a video, be that pornographic frame detects to this frame.Because each frame all is equivalent to a static images, therefore can adopt whether existing porny detection method is that pornographic frame detects to this frame.In testing process, adopt the method for setting up complexion model usually, the area of skin color in the picture, colour of skin area etc. is detected, thereby know whether frame of video is pornographic frame.

2, in the video that will detect, extract the frame of video of some, repeat above-mentioned pornographic frame testing process, obtain the relevant detection result.

3, account for the ratio of the frame of video of extraction according to detected pornographic frame, judge whether the frame of video that will detect is pornographic video.

In the method for the above-mentioned type, most principal work concentrates on the detection accuracy rate of how setting up desirable complexion model and how improving porny.A limitation of these class methods is, the feature that it is adopted when pornographic is discerned remains the static nature of image, and compares with still image, and video itself also has other exclusive characteristic, as video temporal characteristics etc., these characteristics are not applied in the identification of pornographic video.

In actual applications, existing pornographic video method exists following restriction to detect the factor of accuracy rate:

A), in the actual detection process, because the diversity of ethnic group, the factors such as variation of light make to be difficult to set up perfect complexion model to realize the accurate detection of colour of skin point;

B), the human body attitude diversity, extraneous factor such as block and make the accurate detection that is difficult to realize porny;

C), be difficult to set up general rule, realize the accurate detection of pornographic video by the detection of porny for different videos.

In view of above-mentioned defective and the limitation that existing pornographic video method had, need a kind of new pornographic video detecting method.

Summary of the invention

Therefore, thereby task of the present invention is to overcome existing pornographic video detecting method can't be used video temporal characteristics influence detection accuracy rate in pornographic Video Detection process defective, thereby a kind of pornographic video detecting method with higher detection efficiency is provided.

To achieve these goals, the invention provides a kind of pornographic video detecting method, may further comprise the steps:

Step 1), in the frame of video of video to be detected, utilize complexion model and non-complexion model, calculate the skin color probability value and the non-skin color probability value of each picture element of described frame of video;

Step 2), according to the skin color probability value and the non-skin color probability value of picture element in the frame of video, for described frame of video is set up the masterplate image;

Step 3), according to described masterplate image, from described frame of video, extract characteristics of image;

Step 4), the characteristics of image in the successive video frames is formed observation sequence, whether described observation sequence is input in the pornographic camera lens model to detect video to be detected is pornographic video.

In the technique scheme, in described step 1), described complexion model and non-complexion model utilize the image in the image data base to set up.

In the technique scheme, describedly set up complexion model and non-complexion model may further comprise the steps:

Step 1-1), utilize the image in the image data base to set up training set; Image labeling in the described training set has colour of skin point and non-colour of skin point;

Step 1-2), utilize step 1-1) in the training set sample, adopt mixed Gauss model training complexion model and non-complexion model.

In the technique scheme, described masterplate image comprises two-value masterplate image and greyscale masterplate image.

In the technique scheme, described two-value masterplate image adopts following steps to set up:

At first, utilize colour of skin point and non-colour of skin point in described complexion model and the described frame of video of non-complexion model mark, obtain the bianry image corresponding with described frame of video;

Then, described bianry image is done connected region cut apart, obtain described two-value masterplate image.

In the technique scheme, described greyscale masterplate image adopts following steps to set up:

At first, calculate the initial gray-scale value of each picture element in the described frame of video according to described skin color probability value and non-skin color probability value;

Then, the initial gray-scale value on each picture element in the whole video frame is done linear normalization handle, obtain the final gray-scale value of each picture element, thereby obtain described greyscale masterplate image.

In the technique scheme, described linear normalization is handled and is adopted following formula:

$P_{\mathrm{Scale}}=\frac{P-P_{\mathrm{Min}}}{P_{\mathrm{Max}}-P_{\mathrm{Min}}}$

Wherein, P _ScaleRepresent described final gray-scale value, P _MinAnd P _MaxMinimum gradation value and the maximum gradation value of representing picture element in the piece image respectively.

In the technique scheme, in described step 3), described characteristics of image comprises following a kind of feature at least:

The average skin color probability of entire image, average skin color probability in all area of skin color, the outer average skin color probability of all area of skin color, largest connected region area, the average skin color probability in largest connected zone, angle and area thereof that the major axis in fitted ellipse zone, minor axis, major axis are rotated counterclockwise around the longitudinal axis.

In the technique scheme, in described step 4), described pornographic camera lens model utilizes the characteristics of image that is extracted in the frame of video of training set to set up.

In the technique scheme, described pornographic camera lens model is a hidden Markov model.

In the technique scheme, when adopting described hidden Markov model to detect pornographic video, with the input value of described observation sequence as described hidden Markov model, described hidden Markov model obtains one according to this input value and is used to represent whether camera lens is the probable value of pornographic camera lens, described probable value and a preset threshold are compared, if described probable value is greater than described threshold value, video then to be detected is pornographic video, otherwise is non-pornographic video.

In the technique scheme, the size of described threshold value is between 0.5 to 0.7.

The present invention also provides a kind of pornographic video detection system, comprising:

Skin color probability value and non-skin color probability value computing module, this module is used for the frame of video at video to be detected, utilizes complexion model and non-complexion model, calculates the skin color probability value and the non-skin color probability value of each picture element of described frame of video;

The masterplate image is set up module, and this module is according to the skin color probability value and the non-skin color probability value of picture element in the frame of video, for described frame of video is set up the masterplate image;

The image characteristics extraction module, this module is extracted characteristics of image according to described masterplate image from described frame of video;

Pornographic Video Detection module, this module are used for the characteristics of image of successive video frames is formed observation sequence, and whether then described observation sequence is input in the pornographic camera lens model to detect video to be detected is pornographic video.

The invention has the advantages that:

1, pornographic video detecting method of the present invention has adopted the temporal characteristics of adjacent video frames in testing process, makes that the robustness of whole video detection method is stronger, and it is higher to detect accuracy rate.

2, pornographic video detecting method of the present invention has adopted the temporal characteristics of adjacent video frames in testing process, makes to have reduced dependence to the complexion model accuracy rate in testing process;

3, pornographic video detecting method of the present invention is in testing process, and the accuracy rate of detection had both depended on the information of single frame image, also depends on the relation of consecutive frame feature, so reduced the dependence of single frame image being handled and detected accuracy rate in testing process;

4, pornographic video detecting method of the present invention does not need complicated area of skin color partitioning algorithm in testing process.

Description of drawings

Below, describe embodiments of the invention in conjunction with the accompanying drawings in detail, wherein:

Fig. 1 is a pornographic video detecting method of the present invention process flow diagram in one embodiment.

Embodiment

Below in conjunction with the drawings and specific embodiments, pornographic video detecting method of the present invention is described.

Compare with pornographic video detecting method of the prior art, pornographic video detecting method of the present invention has utilized the temporal characteristics of video when carrying out pornographic Video Detection.Related in the present invention video temporal characteristics is meant video consecutive frame image feature variation characteristic in time, the characteristic certain along with timetable reveals as the variation meeting of features such as color, texture, and this specific character just can be considered to the video temporal characteristics.Pornographic video detecting method of the present invention is when realizing, at first in each frame of video of video to be detected or video segment, extract characteristics of image, then the characteristics of image that is extracted in the continuous video frames is handled, judged according to the result who handles whether the video that is detected is pornographic video.Whether in above-mentioned testing process, need not single frames is that pornographic frame is made judgement.

Below in conjunction with Fig. 1, the specific implementation step of a preferred forms of pornographic video detecting method of the present invention is described further:

Step 10, set up complexion model and non-complexion model, particularly, the process of setting up of described complexion model and non-complexion model can comprise following two steps:

Step 11, set up training set.Setting up training set is that complexion model is set up the common step in the process, the main task that will finish in this step is to set up a large amount of image data bases, and to image labeling colour of skin point in the image data base and non-colour of skin point, the image behind the mark can be used for the follow-up colour of skin point model and the training process of non-colour of skin point model.In general, mark the operation of colour of skin point and non-colour of skin point in this step all by artificial realization.

Step 12, set up complexion model and non-complexion model.

In setting up the process of complexion model, the color of each colour of skin picture element can be considered to an observation X, and described observation X represents colour of skin point rgb value, and it is an input training data of setting up model.Described observation X can be considered to by specific single Gaussian probability-density function set G={p _i(X; θ), i=1,2 ... } in the weighted mean of limited element combine.Probability density function p (the X of this observation X; φ) can be expressed as the following form of mixed Gauss model (GMM):

$\left\{\begin{array}{c}p(X;\mathrm{\φ})=\underset{i=1}{\overset{g}{\mathrm{\Σ}}}{\mathrm{\π}}_i\·p_i(X;\mathrm{\θ})=\underset{i=1}{\overset{g}{\mathrm{\Σ}}}{\mathrm{\π}}_i\·p_i\left(X\right|i;\mathrm{\θ})=\underset{i=1}{\overset{g}{\mathrm{\Σ}}}{\mathrm{\π}}_i\·\frac{1}{{\left(2\mathrm{\π}\right)}^{d/2}{\left|{\mathrm{\Σ}}_i\right|}^{1/2}}{\exp }^{-\frac{1}{2}{(X-\mathrm{\μ})}^1{\left(\mathrm{\Σ}\right)}^{-}(X-\mathrm{\μ})}\\ \underset{i=1}{\overset{g}{\mathrm{\Σ}}}{\mathrm{\π}}_i=1\end{array}\right.$

Wherein, π _iBe weight, θ is the parameter vector of being made up of average and covariance matrix.Unknown parameter π _i(Expectation Maximization, note by abridging and be: EM), utilize in the resulting training set of step 11, colour of skin point that is marked and non-colour of skin point sample calculate can to pass through expectation-maximization algorithm with θ.The realization of EM algorithm is ripe prior art, and those of ordinary skill in the art is according to list of references 1: " pattern classification, Richard 0.Duda, Peter E.Hart, David G.Stork, China Machine Press, 2005 " realize this algorithm.In the present embodiment, select for use the RGB component constitutive characteristic vector of picture element to be used for model training; Select for use 16 single Gauss models to constitute for each GMM model.

By aforementioned calculation, after obtaining observing the probability density function of X, also just set up corresponding complexion model.

Process and the above-mentioned process of setting up complexion model of setting up non-complexion model are similar, all adopted aforesaid mixed Gauss model (GMM), both differences are, set up in the process of complexion model, the input data are the rgb value of colour of skin point, and set up in the process of non-complexion model, the input data are the rgb value of non-colour of skin point.All the other operation stepss have had relevant description in above stated specification, no longer do repeat specification at this.

The operation of setting up complexion model and non-complexion model of this step is not the necessary step of pornographic Video Detection, if having existing complexion model and non-complexion model in pornographic Video Detection process, then this step can be omitted.

Step 20, in a frame of video, the complexion model and the non-complexion model that obtain by step 10, each picture element in the frame of video is calculated skin color probability value and non-skin color probability value, according to the comparison between skin color probability value and the non-skin color probability value, know whether some picture elements are colour of skin point.In this step, the probability that picture element of skin color probability value representation is a colour of skin point, similarly, the probability that picture element of non-skin color probability value representation is not a colour of skin point.The rgb value of colour of skin picture element is input in the GMM model of complexion model, and resulting output result is exactly the skin color probability value.Same, the rgb value of non-colour of skin picture element is input in the GMM model of non-complexion model, resulting output result just is non-skin color probability value.Use P _SkinExpression skin color probability value is used _PNon-SkinRepresent non-skin color probability value, if satisfy P _Skin＞P _Non-Skin, then this picture element is a colour of skin point, otherwise this picture element is non-colour of skin point.

Step 30, set up the masterplate image for frame of video.In the present embodiment, the masterplate image of setting up for frame of video comprises two-value masterplate image and greyscale masterplate image, and their constructive process there are differences, and is described step by step below.

Step 31, set up two-value masterplate image.In previous step, utilized complexion model and non-complexion model that whether the picture element of frame of video is judged for colour of skin point, therefore, in this step, colour of skin point is made marks with " 1 ", non-colour of skin point is made marks with " 0 ", thereby can obtain the bianry image corresponding with original image, this bianry image is done connected region again cut apart, just obtain the two-value masterplate image Mask of original image correspondence ₁The realization of this step all is a prior art, and wherein the associative operation of cutting apart about connected region can be referring to list of references 2 " computer vision, Jia Yun gets, Beijing: Science Press, 2000 ".

Step 32, set up the greyscale masterplate image.In this step, at first calculate the initial gray-scale value P of each picture element in the frame of video, resulting skin color probability value is relevant with non-skin color probability value in the calculating of initial gray-scale value and the abovementioned steps, and its computing formula is:

$P=\frac{P_{\mathrm{Skin}}}{P_{\mathrm{Skin}}+P_{\mathrm{Non}-\mathrm{Skin}}};$

Then the initial gray-scale value on each picture element in the whole video frame is done linear normalization and handle, obtain the final gray-scale value P of each picture element _Scale, after all picture elements all carry out the linear normalization processing in the frame of video, also just obtain the greyscale masterplate image M ask corresponding with original image ₂In the said process, the formula that linear normalization is handled is as follows:

$P_{\mathrm{Scale}}=\frac{P-P_{\mathrm{Min}}}{P_{\mathrm{Max}}-P_{\mathrm{Min}}}$

Wherein, P _MinAnd P _MaxMinimum gradation value and the maximum gradation value of representing picture element in the piece image respectively.

Step 40, current video frame is extracted characteristics of image.In the present embodiment, the characteristics of image that is extracted comprises the average skin color probability of entire image, average skin color probability in all area of skin color, the outer average skin color probability of all area of skin color, largest connected region area, the angle that the average skin color probability in largest connected zone, the major axis in fitted ellipse zone, minor axis, major axis are rotated counterclockwise around the longitudinal axis and area thereof etc.Those of ordinary skill in the art should understand, and in other embodiment, can select wherein several from above-mentioned characteristics of image, also can select other characteristics of image for use.

Leaching process to each characteristics of image describes below:

Step 41, utilize resulting greyscale masterplate image in the previous step, extract the average skin color probability P of view picture video frame images _Image, its computing formula is as follows:

$P_{\mathrm{Image}}=\frac{1}{M\×N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{Gray}(i,j)$

Wherein, Gray (i, j) presentation video Mask ₂(i, gray-scale value j), M and N are represented the wide and high of greyscale masterplate image respectively in middle position.

Step 42, utilize resulting masterplate image in the previous step, calculate average skin color probability in all area of skin color.When calculating, as long as computed image Mask ₁Middle gray scale is the some correspondence image Mask of " 1 " ₂The mean value of gray scale of picture element.

Step 43, utilize resulting masterplate image in the previous step, calculate the outer average skin color probability of all area of skin color.When calculating, as long as computed image Mask ₁Middle gray scale is the some correspondence image Mask of " 0 " ₂The mean value of gray scale of picture element.

Step 44, calculate largest connected region area.When calculating, computed image Mask ₁The maximum area relative pixel of picture element that each connected region that middle gray scale is " 1 " comprises is counted.

The average skin color probability in step 45, the largest connected zone of calculating.When calculating, the largest connected zone that obtains in the calculation procedure 44 is at image M ask ₂The average gray value of corresponding picture element.

Angle and area thereof that major axis, minor axis, the major axis in step 46, calculating fitted ellipse zone is rotated counterclockwise around the longitudinal axis.In this step, described fitted ellipse is a figure that is used for characterizing largest connected region shape, and this oval shape, size can be used for reflecting shape, the size in largest connected zone.Before correlation computations is carried out in the fitted ellipse zone, at first calculate the center in largest connected zone coordinate (r, c), its computing formula is as follows:

Wherein, y _sRepresent each picture element gray-scale value, r _sAnd c _sRepresent horizontal ordinate and the ordinate of picture element S respectively, R represents largest connected zone, A represent all picture element gray-scale values in the piece image and.

Calculate the second moment in largest connected zone then, its computing formula is as follows:

The capable square of second order: ${\mathrm{\μ}}_{\mathrm{rr}}=\frac{1}{A}\underset{s\∈R}{\mathrm{\Σ}}{(r_s-r)}^2{y}_s$

Second order row square: ${\mathrm{\μ}}_{\mathrm{cc}}=\frac{1}{A}\underset{s\∈R}{\mathrm{\Σ}}{(c_s-c)}^2{y}_s$

The second order mixed moment: ${\mathrm{\μ}}_{\mathrm{rc}}=\frac{1}{A}\underset{s\∈R}{\mathrm{\Σ}}(r_s-r)(c_s-c)y_s;$

Then, use the largest connected zone of ellipse fitting, and calculate the major axis l of this elliptic region _Major, minor axis l _MinorWith major axis around the angle θ that the longitudinal axis is rotated counterclockwise, have following four kinds of situations:

A, work as μ _Rc=0, and μ _Rr＞μ _CcThe time:

B, work as μ _Rc=0, and μ _Rr≤ μ _CcThe time:

C. work as μ _Rc≠ 0, and μ _Rr≤ μ _CcThe time:

$\left\{\begin{array}{c}{l}_{\mathrm{major}}=\sqrt{8(({\mathrm{\μ}}_{\mathrm{rr}}+{\mathrm{\μ}}_{\mathrm{cc}})+\sqrt{{({\mathrm{\μ}}_{\mathrm{rr}}-{\mathrm{\μ}}_{\mathrm{cc}})}^2+4{{\mathrm{\μ}}_{\mathrm{rc}}}^2})}\\ l_{\min \mathrm{or}}=\sqrt{8(({\mathrm{\μ}}_{\mathrm{rr}}+{\mathrm{\μ}}_{\mathrm{cc}})-\sqrt{{({\mathrm{\μ}}_{\mathrm{rr}}-{\mathrm{\μ}}_{\mathrm{cc}})}^2+4{{\mathrm{\μ}}_{\mathrm{rc}}}^2})}\\ \mathrm{\θ}=\mathrm{arcan}\frac{-2{\mathrm{\μ}}_{\mathrm{rc}}}{({\mathrm{\μ}}_{\mathrm{rr}}-{\mathrm{\μ}}_{\mathrm{cc}})+\sqrt{{({\mathrm{\μ}}_{\mathrm{cc}}-{\mathrm{\μ}}_{\mathrm{rr}})}^2+4{{\mathrm{\μ}}_{\mathrm{rc}}}^2}}\end{array}\right.$

D. work as μ _Rc≠ 0, and μ _Rr＞μ _CcThe time:

$\left\{\begin{array}{c}{l}_{\mathrm{major}}=\sqrt{8(({\mathrm{\μ}}_{\mathrm{rr}}+{\mathrm{\μ}}_{\mathrm{cc}})+\sqrt{{({\mathrm{\μ}}_{\mathrm{rr}}-{\mathrm{\μ}}_{\mathrm{cc}})}^2+4{{\mathrm{\μ}}_{\mathrm{rc}}}^2})}\\ l_{\min \mathrm{or}}=\sqrt{8(({\mathrm{\μ}}_{\mathrm{rr}}+{\mathrm{\μ}}_{\mathrm{cc}})-\sqrt{{({\mathrm{\μ}}_{\mathrm{rr}}-{\mathrm{\μ}}_{\mathrm{cc}})}^2+4{{\mathrm{\μ}}_{\mathrm{rc}}}^2})}\\ \mathrm{\θ}=\arctan \frac{({\mathrm{\μ}}_{\mathrm{rr}}-{\mathrm{\μ}}_{\mathrm{cc}})+{\sqrt{{({\mathrm{\μ}}_{\mathrm{cc}}-{\mathrm{\μ}}_{\mathrm{rr}})}^2+4{{\mathrm{\μ}}_{\mathrm{rc}}}^2}}_{\mathrm{rc}}}{-2\mathrm{\μ}}\end{array}\right.$

At last, calculate the area A of fitted ellipse _e:

$A_e=4\mathrm{\π}\sqrt{{\mathrm{\μ}}_{\mathrm{rr}}{\mathrm{\μ}}_{\mathrm{cc}}-{{\mathrm{\μ}}_{\mathrm{rc}}}^2}$

Step 50, set up pornographic camera lens model, and utilize this model to realize the detection of pornographic video.In pornographic camera lens, not only single-frame images comprises colour of skin information, and successive frame comprises relevant colour of skin information, therefore in order to characterize this space characteristics and temporal characteristics better, utilize hidden Markov model (HMM) in the statistical model to set up pornographic camera lens model among the present invention and be used for the detection of pornographic video.This step comprises model training stage and test phase, and specifying of two stages is as follows:

Step 51, model training stage.In this stage, at first to set up training set.Setting up training set is exactly manually to set up pornographic video segment data set, successive frame to each video segment extracts following feature: the average skin color probability of entire image, average skin color probability in all area of skin color, the outer average skin color probability of all area of skin color, largest connected region area, the average skin color probability in largest connected zone, the major axis in fitted ellipse zone, minor axis, angle and area thereof that major axis is rotated counterclockwise around the longitudinal axis, with the successive frame behind the said extracted characteristics of image as training sample.The concrete operations that image extracts have had specific description in aforesaid explanation, no longer do duplicate explanation herein.

After setting up training set, set up hidden Markov model (HMM) again.Hidden Markov model is the algorithm of a maturation, and it is equivalent to a black box, has the account form of oneself.After data that the user will comprise certain feature were input to this black box, this black box can be set up corresponding model according to the feature of data.When the data of being imported are when including the continuous data frame of characteristics of image, to have implied temporal characteristics between these adjacent datas, after the continuous data frame is input to hidden Markov model, also comprised temporal characteristics among the resulting output result.HMM is a dual random process, and it is by two ingredients: a. Markov chain, and this chain is used to the transfer of the state of describing, and describes with transition probability; B. general stochastic process, the relation between this process prescription state and observation sequence is described with the probability of observed value.

HMM can represent with a five-tuple:

λ＝(N，M，π，A，B)，

Wherein, N is a state number; π is the initial state probability, π={ π ₁, π ₂... π _N; A is state-transition matrix A={ α _Ij}=p (Q _t=j|Q _T-1=i); B is state output probability B={b _j(o _t)=p (O _t=o _t| Q _t=j), be the probability distribution of observed events corresponding states, just export the probability of each observed events for a state; M is the observed event number of each state correspondence.

HMM can be used for continuous state transformed and is described, by training sample, and promptly given observation sequence O={O ₁, O ₂..., O _TObtain this model parameter λ by the EM algorithm, make:

${\mathrm{\λ}}^{*}=\arg \underset{\mathrm{\λ}}{\max }P\left(O\right|\mathrm{\λ})$

Like this, this HMM model can be used to describe the state transitions relation of the observation sequence of given classification.The detailed introduction of relevant HMM, visible list of references 3 " pattern classification, Richard O.Duda, Peter E.Hart, David G.Stork, China Machine Press, 2005 ".

The above-mentioned model training stage be not in the pornographic Video Detection process must the experience stage, if existing hidden Markov model has been arranged in pornographic Video Detection, the then above-mentioned stage can save.

Step 52, test phase.In aforesaid step, behind the composition of the feature described in the successive frame extraction step 40 of video or video segment wherein observation sequence, with of the input of this observation sequence as HMM, one of HMM output is used to represent whether camera lens is the probable value P of pornographic camera lens, if P is greater than preset threshold TH, then this camera lens is judged as pornographic camera lens, otherwise is non-pornographic camera lens.In the present embodiment, the span of TH is chosen for 0.6 in the present embodiment between 0.5-0.7.

Also provide in the present embodiment and the corresponding pornographic video detection system of pornographic video detecting method, comprised in this system:

Skin color probability value and non-skin color probability value computing module, this module is used for the frame of video at video to be detected, utilizes complexion model and non-complexion model, calculates the skin color probability value and the non-skin color probability value of each picture element of described frame of video;

The masterplate image is set up module, and this module is according to the skin color probability value and the non-skin color probability value of picture element in the frame of video, for described frame of video is set up the masterplate image;

The image characteristics extraction module, this module is extracted characteristics of image according to described masterplate image from described frame of video;

Pornographic Video Detection module, this module are used for the characteristics of image of successive video frames is formed observation sequence, and whether then described observation sequence is input in the pornographic camera lens model to detect video to be detected is pornographic video.

The concrete workflow of above-mentioned each module has had detailed explanation in the correlation step of pornographic video detecting method, therefore, no longer do the description of repeatability herein.

It should be noted last that above embodiment is only unrestricted in order to technical scheme of the present invention to be described.Although the present invention is had been described in detail with reference to embodiment, those of ordinary skill in the art is to be understood that, technical scheme of the present invention is made amendment or is equal to replacement, do not break away from the spirit and scope of technical solution of the present invention, it all should be encompassed in the middle of the claim scope of the present invention.

Claims (13)

1. pornographic video detecting method may further comprise the steps:

Step 1), in the frame of video of video to be detected, utilize complexion model and non-complexion model, calculate the skin color probability value and the non-skin color probability value of each picture element of described frame of video;

Step 2), according to the skin color probability value and the non-skin color probability value of picture element in the frame of video, for described frame of video is set up the masterplate image;

Step 3), according to described masterplate image, from described frame of video, extract characteristics of image;

Step 4), the characteristics of image in the successive video frames is formed observation sequence, whether described observation sequence is input in the pornographic camera lens model to detect video to be detected is pornographic video.

2. pornographic video detecting method according to claim 1 is characterized in that, in described step 1), described complexion model and non-complexion model utilize the image in the image data base to set up.

3. pornographic video detecting method according to claim 2 is characterized in that, describedly sets up complexion model and non-complexion model may further comprise the steps:

Step 1-1), utilize the image in the image data base to set up training set; Image labeling in the described training set has colour of skin point and non-colour of skin point;

Step 1-2), utilize step 1-1) in the training set sample, adopt mixed Gauss model training complexion model and non-complexion model.

4. pornographic video detecting method according to claim 1 is characterized in that, described masterplate image comprises two-value masterplate image and greyscale masterplate image.

5. pornographic video detecting method according to claim 4 is characterized in that, described two-value masterplate image adopts following steps to set up:

At first, utilize colour of skin point and non-colour of skin point in described complexion model and the described frame of video of non-complexion model mark, obtain the bianry image corresponding with described frame of video;

Then, described bianry image is done connected region cut apart, obtain described two-value masterplate image.

6. pornographic video detecting method according to claim 4 is characterized in that, described greyscale masterplate image adopts following steps to set up:

At first, calculate the initial gray-scale value of each picture element in the described frame of video according to described skin color probability value and non-skin color probability value;

Then, the initial gray-scale value on each picture element in the whole video frame is done linear normalization handle, obtain the final gray-scale value of each picture element, thereby obtain described greyscale masterplate image.

7. pornographic video detecting method according to claim 6 is characterized in that, described linear normalization is handled and adopted following formula:

$P_{\mathrm{Scale}}=\frac{P-P_{\mathrm{Min}}}{P_{\mathrm{Max}}-P_{\mathrm{Min}}}$

Wherein, P _ScaleRepresent described final gray-scale value, P _MinAnd P _MaxMinimum gradation value and the maximum gradation value of representing picture element in the piece image respectively; P represents the initial gray-scale value of each picture element in the whole video frame.

8. pornographic video detecting method according to claim 1 is characterized in that, in described step 3), described characteristics of image comprises following a kind of feature at least:

The average skin color probability of entire image, average skin color probability in all area of skin color, the outer average skin color probability of all area of skin color, largest connected region area, the average skin color probability in largest connected zone, angle and area thereof that the major axis in fitted ellipse zone, minor axis, major axis are rotated counterclockwise around the longitudinal axis.

9. pornographic video detecting method according to claim 3 is characterized in that, in described step 4), described pornographic camera lens model utilizes the characteristics of image that is extracted in the frame of video of training set to set up.

10. pornographic video detecting method according to claim 9 is characterized in that, described pornographic camera lens model is a hidden Markov model.

11. pornographic video detecting method according to claim 10, it is characterized in that, when adopting described hidden Markov model to detect pornographic video, with the input value of described observation sequence as described hidden Markov model, described hidden Markov model obtains one according to this input value and is used to represent whether camera lens is the probable value of pornographic camera lens, described probable value and a preset threshold are compared, if described probable value is greater than described threshold value, video then to be detected is pornographic video, otherwise is non-pornographic video.

12. pornographic video detecting method according to claim 11 is characterized in that the size of described threshold value is between 0.5 to 0.7.

13. a pornographic video detection system comprises:

Skin color probability value and non-skin color probability value computing module, this module is used for the frame of video at video to be detected, utilizes complexion model and non-complexion model, calculates the skin color probability value and the non-skin color probability value of each picture element of described frame of video;

The masterplate image is set up module, and this module is according to the skin color probability value and the non-skin color probability value of picture element in the frame of video, for described frame of video is set up the masterplate image;

The image characteristics extraction module, this module is extracted characteristics of image according to described masterplate image from described frame of video;

Pornographic Video Detection module, this module are used for the characteristics of image of successive video frames is formed observation sequence, and whether then described observation sequence is input in the pornographic camera lens model to detect video to be detected is pornographic video.

Images