CN101923652A - Pornographic picture identification method based on joint detection of skin colors and featured body parts - Google Patents

Abstract

The invention belongs to the technical field of image processing and provides a pornographyic picture identification method based on joint detection of skin colors and featured body parts, which comprises the following steps: (1) preparatory stage of sample training for acquiring needed parameters of the identification stage of an image to be detected, the preparatory stage of sample training comprises color skin sample training and featured body part sample training; (2) the stage of reading the image to be detected; (3) the identification stage of the image to be detected for judging whether the image to be detected is a pornographyic picture or not according to the parameters acquired at the preparatory stage of sample training: the skin color and the featured body parts are detected in sequence, and if the area of the skin color region accounts for more than a half of the whole image and at least one featured body part emerges, the picture is judged to be a pornographic picture. The method of the invention has the advantages of being capable of reducing false detection rate of pornographyic pictures, and having wide range of application and high accuracy rate.

Description

A kind of porny recognition methods based on the colour of skin and characteristic portion joint-detection

Technical field

The invention belongs to technical field of image processing, be specifically related to a kind of porny recognition methods based on the colour of skin and characteristic portion joint-detection.

Background technology

Statistical data according in January, 2010 China's juvenile criminology research association announcement: the juvenile deliquency case accounts for 70% of China's crime total amount, and has 70% to be network pornography or violence in the inducement of juvenile deliquency case.The survey report of CNNIC in January, 2010 announcement shows: by the end of the year 2009, China netizen's 32.9% is the teenager below 19 years old.For follow-on physical and mental health, the highly vigilant of that remains network pornography will be a long-term and important job.

In mass network information, discerning porny rapidly and accurately and it is used as the evidence of fighting crime needs providing powerful support for of image processing techniques.In the method for various identification porny, according to their identifying object, roughly it can be divided into two classes: Face Detection and characteristic portion detect.

In Face Detection, method is exactly directly to judge area of skin color according to the color gamut of the colour of skin the most intuitively.Specifically, be exactly in the shades of colour space such as RGB, YCbCr, HIS etc. times, from the skin-color training sample, obtain the distributed area of its color, select a suitable color space and less color gamut to determine the colour of skin then.People such as Vladimir VezhnevetsVassili have summed up two big complexion models: a class printenv refers to Bayesian model; Another kind of have a parameter, based on the Gaussian function model, also developed afterwards and gauss hybrid models and corresponding iterative method thereof.Describe in the research of colour of skin character at the employing gauss hybrid models, people's such as Weiming Hu work is more representative, its innovation part is to come modeling by the quantity of histogram and corresponding color gray level, obtains a series of experimental results about different Gaussian probability-density function numbers at last.People such as D.A.Forsyth and M.M.Fleck has proposed the porny detection method based on the colour of skin and texture, at first is to judge area of skin color, judges the position of four limbs then by area of skin color, judges by the four limbs position whether image is pornographic image.

Characteristic portion such as nipple are the important evidence of judging porny.Although there is the colour of skin area of a lot of pictures bigger, probably be art pattern or bikini picture.People such as Fuangkhon come the recognition feature position with the isomorphism neural network then by image is carried out pre-service.People such as Yue Wang then adopt the adaboost algorithm to come the recognition feature position, strengthen identification to characteristic portion by the color characteristic of characteristic portion then.

Only a kind of method as the identification porny in Face Detection or the characteristic portion detection, existing patent of invention also is like this in domestic and international academic research mentioned above.Such as " a kind of pornographic image detection method " of application number 200410042877.3, by with the ratio of human face region area and area of skin color area judgement foundation as porny; " the content-based network pornography image and the bad image detecting system " of application number 200510048577.0 adopts Face Detection and discerns porny based on the attitude detection of Face Detection.Though the realization effect of prior art depends on concrete test sample book collection, can accomplish to discern automatically to a great extent porny.

Yet, under some application scenarios,, require porny that machine adjudicates for human such as criminal investigation, lawsuit etc., have no objection.Such as the swimsuit photo,,, then should not list the evidence of investigation or lawsuit in the law enforcement aspect if its characteristic portion does not expose such as nipple, genitals etc. although the area that skin exposes is bigger.Although existing porny automatic identification technology can be found a large amount of doubtful porny, from these doubtful porny, filter out on the stricti jurise porny of produce (shenglvehao)in court legally, still need a large amount of artificial interpretations.

Summary of the invention

Purpose of the present invention is exactly in order to overcome above-mentioned weak point of the prior art, and a kind of porny recognition methods based on the colour of skin and characteristic portion joint-detection is provided.This method can effectively reduce the false drop rate of porny, and is applied widely, the accuracy rate height.

A kind of porny recognition methods based on the colour of skin and characteristic portion joint-detection of the present invention may further comprise the steps successively:

(1) preparatory stage of sample training, is used to obtain the required parameter of cognitive phase of image to be checked; This preparatory stage launches around colour of skin sample training and characteristic portion (be nipple, in the following literal, " nipple " used with " characteristic portion ") sample training respectively;

(2) read image to be checked;

(3) cognitive phase of image to be checked, be used for the parameter that preparatory stage obtained according to sample training, whether to image to be checked is that porny is judged, at first carry out Face Detection, carrying out characteristic portion then detects, if the area of area of skin color accounts for more than 1/2 of entire image, and occurred 1 characteristic portion at least, then this picture is judged as porny.

In technique scheme, colour of skin sample training in the step (1), at first collect the colour of skin sample of sufficient amount, obtain colour of skin grey level histogram, approach colour of skin grey level histogram by gauss hybrid models then, adopt EM iterative algorithm (Entropy Maximization, entropy maximization) to obtain model parameter.

In technique scheme, the characteristic portion sample training in the step (1) is at first collected the characteristic portion sample of sufficient amount, trains the sorter of a cascade then by the Adaboost algorithm.

In technique scheme, describedly train the sorter cascade quantity of a cascade more than 20 by the Adaboost algorithm.

In technique scheme, Face Detection described in the step (3) is that first length and width with image to be checked waits 10 parts of branches respectively, form 100 cells, get of the representative of the central point of these cells as this cell, if judging this central point is colour of skin point, the residing cell of this central point is colour of skin lattice so; When judging whether this central point is colour of skin point, the gray-scale value of this central point is decomposed on three components of R, G in the rgb space, B, be updated in the gauss hybrid models of in the preparatory stage, setting up, obtain the probability that it becomes colour of skin point, if these three components are all greater than its corresponding threshold value, then adjudicating this central point is colour of skin point, and the cell at place is a colour of skin lattice; Each cell for dividing all repeats above judgement, thereby obtains area of skin color; If the area of this area of skin color accounts for more than 1/2 of entire image, then proceed characteristic portion and detect, otherwise this image is not judged as porny.

In technique scheme, characteristic portion described in the step (3) detects and at first is used in the cascade classifier of training out with the Adaboost algorithm in the preparatory stage, image to be checked is carried out initial survey, then the initial survey result is carried out examination one by one according to following 4 restrictive conditions, as long as there is a condition not meet, just judge that this initial survey result is not a characteristic portion:

(a) this initial survey result's area is no more than 1/100 of entire image;

(b) this initial survey result is not the people on the face;

(c) the nearest colour of skin block of this initial survey result is no more than a block length (10 pixels);

(d) this initial survey result's edge is rounded;

Initial survey result by above 4 restrictive condition examinations is called inspection result eventually; If the inspection result has 1 at least eventually, so whole image to be checked just is judged as porny.

Beneficial effect of the present invention is effectively to reduce the false drop rate of porny, avoid the normal picture on the internet to be blocked as much as possible, effective collection at aspects such as criminal investigation, lawsuits for the porny evidence provides technical support, effectively reduces the workload of artificial interpretation.

Description of drawings

Fig. 1 is a kind of porny recognition methods process flow diagram based on the colour of skin and characteristic portion joint-detection of the present invention.

Embodiment

A kind of porny recognition methods based on the colour of skin and characteristic portion joint-detection of the present invention mainly is divided into the preparatory stage of sample training and the cognitive phase of image to be checked, respectively these two stages is made an explanation below in conjunction with accompanying drawing.

One, the preparatory stage of sample training

When the embodiment of the preparatory stage of describing sample training, used mathematic sign and implication thereof are as shown in table 1.This preparatory stage is by setting up gauss hybrid models and adopting two processes of Adaboost algorithm training classifier (abbreviating " training Adaboost sorter " as) to form.This two process can be separated simultaneously and carried out, and also can successively carry out.When setting up the gauss hybrid models of the colour of skin, we have gathered 660 samples.When training Adaboost sorter, we have gathered 402 positive samples (nipple) and 7500 negative samples (not containing nipple).

Table 1 mathematic sign and implication thereof

A sets up gauss hybrid models

If the grey value profile of colour of skin sample is X={x ₁, x ₂... x _n).Owing to the grey value profile of colour of skin sample is common a plurality of peak values appear, so need to adopt the weighted mean of a plurality of Gaussian density functions (Gaussian Mixture Model Probability Density Function, or title gauss hybrid models) to approach.If the number of Gaussian density function is M, the gray-scale value probability density function of colour of skin sample then, p (x) can be expressed as:

$p\left(x\right)=\underset{k=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}g(x;{\mathrm{\μ}}_k,{\mathrm{\Σ}}_k)---\left(1\right)$

Wherein, g (x; μ _k, ∑ _k) be k Gaussian probability-density function, μ _kBe its central point, ∑ _kBe its covariance matrix, α _kBe its weighting coefficient, satisfy:

$\underset{k=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\α}}_k=1---\left(2\right)$

Usually, ∑ _kCan be expressed as:

Then k Gaussian probability-density function can be expressed as:

$g(x;{\mathrm{\μ}}_k,{{\mathrm{\σ}}_k}^2)=\frac{1}{2\mathrm{\π}{{\mathrm{\σ}}_k}^2}\exp [-\frac{{(x-{\mathrm{\μ}}_k)}^T(x-{\mathrm{\μ}}_k)}{2{{\mathrm{\σ}}_k}^2}]---\left(4\right)$

Following formula is carried out differential to each parameter, can obtain:

${\▿}_{{\mathrm{\μ}}_k}g(x;{\mathrm{\μ}}_k,{{\mathrm{\σ}}_k}^2)=g(x;{\mathrm{\μ}}_k,{{\mathrm{\σ}}_k}^2)\left(\frac{x-{\mathrm{\μ}}_k}{{\mathrm{\σ}}_k^2}\right)---\left(5\right)$

${\▿}_{{\mathrm{\σ}}_k}g(x;{\mathrm{\μ}}_k,{{\mathrm{\σ}}_k}^2)=g(x;{\mathrm{\μ}}_k,{{\mathrm{\σ}}_k}^2)(\frac{{(x-{\mathrm{\μ}}_k)}^T(x-{\mathrm{\μ}}_k)}{{\mathrm{\σ}}_k^3}-\frac{2}{{\mathrm{\σ}}_k})---\left(6\right)$

By (3) formula, (1) formula can be expressed as:

$p\left(x\right)=\underset{k=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}g(x;{\mathrm{\μ}}_k,{\mathrm{\σ}}_k^2)---\left(7\right)$

Can suppose x _i(i=1,2 ..., be incident independently mutually between n), X={x then takes place ₁..., x _nProbability density be:

$p(X,\mathrm{\μ},\mathrm{\Σ})=\underset{i=1}{\overset{n}{\mathrm{\Π}}}p(x_i;\mathrm{\μ},\mathrm{\Σ})---\left(8\right)$

Because X is event, so we wish to find out μ _kWith

Value makes p (x) that is formula (8) maximum.Make θ=[α ₁, α ₂..., α _M, μ ₁, μ ₂... μ _M,

], also promptly obtain the maximal value of following formula, be designated as formula (9).

$J\left(\mathrm{\θ}\right)=\mathrm{In}\left[\underset{i=1}{\overset{n}{\mathrm{\Π}}}p\left(x_i\right)\right]=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Inp}\left(x_i\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{In}\underset{k=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}g(x_i;{\mathrm{\μ}}_k,{{\mathrm{\σ}}_k}^2)---\left(9\right)$

So to parameter μ _kAnd σ _kDifferentiate:

${\▿}_{{\mathrm{\μ}}_k}J\left(\mathrm{\θ}\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_k\left(x_i\right)\left(\frac{x_i-{\mathrm{\μ}}_k}{{\mathrm{\σ}}_k^2}\right)---\left(10\right)$

${\▿}_{{\mathrm{\σ}}_k}J\left(\mathrm{\θ}\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_k\left(x_i\right)[\frac{{(x_i-{\mathrm{\μ}}_k)}^T(x_i-{\mathrm{\μ}}_k)}{{\mathrm{\σ}}_k^2}-\frac{2}{{\mathrm{\σ}}_k}]---\left(11\right)$

Wherein, ${\mathrm{\β}}_k\left(x\right)=\frac{{\mathrm{\α}}_{k}g(x;{\mathrm{\μ}}_k,{\mathrm{\σ}}_k^2)}{\underset{j=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\α}}_{j}g(x;{\mathrm{\μ}}_j,{\mathrm{\σ}}_j^2)}.$

Make that formula (10) and formula (11) are zero, can get:

${\mathrm{\μ}}_k=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i{\mathrm{\β}}_k\left(x_i\right)}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_k\left(x_i\right)}---\left(12\right)$

${\mathrm{\σ}}_k^2=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_k\left(x_i\right){(x_i-{\mathrm{\μ}}_k)}^T(x_i-{\mathrm{\μ}}_k)}{2\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_k\left(x_i\right)}---\left(13\right)$

In (9) formula to parameter alpha _kWhen differentiate gets, because of needs are considered (2) formula, thus introduce Lagrangian factor lambda, and definition fresh target function is:

$J_{\mathrm{new}}=J\left(\mathrm{\θ}\right)+\mathrm{\λ}(1-\underset{k=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\α}}_k)$

Order : K=1,2 ..., M.Simultaneously, consider (2) formula, can get:

λ=n, so:

${\mathrm{\α}}_k=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_k\left(x_i\right)---\left(14\right)$

Based on (12), (13) and (14),, find the solution parameter vector θ=[α now by the method for iteration ₁, α ₂..., α _M, μ ₁, μ ₂... μ _M,

].

1, sets an initial parameters value

Can make

And use the K-means mode to calculate the central point of clustering, with this as

Value.

2, use

Calculate β _k(x _i), i=1 ... n.

3, β _k(x _i) substitution (12) tries to achieve new μ _k, be designated as

4,

μ in the substitution (13) _k, try to achieve new σ _k, be designated as

5, β _k(x _i) substitution (14) tries to achieve new α _k, be designated as

6, order

If

Less than a certain minimum tolerance value, then stop.Otherwise order

And rebound step 2.

B training Adaboost sorter

If two tuple (c _i, y _i) expression samples pictures c _iAnd positive and negative attribute y _i, wherein, i=1,2 ..., n.If sample c _iBe positive sample, then y _iBe designated as 1; Otherwise y _iBe designated as 0.If positive sample has L _qIt is individual,

The method of training Adaboost sorter is as follows with the false code of Matlab style:

For t＝1:1:T

1. normalized weight w _t(i):

$w_t\left(i\right)=\frac{w_t\left(i\right)}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_t\left(i\right)}$

2. for each feature, be Weak Classifier h of its training specially _j(detailed under the training method), the error of its classification is:

${\mathrm{\ϵ}}_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_j\left(i\right)|h_j\left(c_i\right)-y_i|$

3. select one to make ε _jMinimum sorter h _j, be designated as h _t

4. renewal weight:

$w_{t+1}\left(i\right)=w_t\left(i\right){\mathrm{\γ}}_t^{1-e_i}$

Wherein, work as c _iCorrectly divided time-like, e _i=0, otherwise, e _i=1; And

End

At last, the strong classifier h (c) of generation is:

$h\left(c\right)=\left\{\begin{array}{c}1,\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_t{h}_t\left(c\right)\&GreaterEqual;\frac{1}{2}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_t\\ 0,\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_t{h}_t\left(c\right)<\frac{1}{2}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_t\end{array}\right.$

Wherein, ${\mathrm{\&delta;}}_t=\log \frac{1}{{\mathrm{\&gamma;}}_t}.$

Weak Classifier h _j(c) can be expressed as:

Wherein, D _jThe direction of the expression sign of inequality can be according to the actual conditions adjustment.S _jBe eigenwert f _jThreshold value.The training method of Weak Classifier is: with the eigenwert f of sample _jArrange by ascending order, again with threshold value S _jBe located to allow and be classified the wrong minimum place of sample size.

Two, the cognitive phase of image to be checked

In Face Detection, earlier image division to be checked being become the length of side is the grid of 10 pixels, forms 100 cells.In order to improve the speed of Face Detection, get of the representative of the central point of these cells as this cell, that is to say, be colour of skin point if judge this central point, the residing cell of this central point is colour of skin lattice so.When judging whether this central point is colour of skin point, the gray-scale value of this central point is decomposed on three components of R, G in the rgb space, B, be updated in the gauss hybrid models of in the preparatory stage, setting up, take all factors into consideration modeling time and modeling accuracy, the parameter of the gauss hybrid models that the present invention set up is as shown in table 2, obtains the probability that it becomes colour of skin point, if these three components are all greater than its corresponding threshold value, then adjudicating this central point is colour of skin point, and the cell at place is a colour of skin lattice.

The parameter list of table 2 gauss hybrid models (M=5)

Each cell for dividing all repeats above judgement, thereby obtains area of skin color, the area that obtains this area of skin color accounts for the ratio S of entire image area, detect if S, then proceeds characteristic portion greater than 1/2, otherwise this image is not judged as porny.

In the detection of characteristic portion, at first be used in one 25 grades the cascade classifier of training out with the Adaboost algorithm in the preparatory stage, image to be checked is carried out initial survey.Then the initial survey result is carried out examination one by one according to following 4 restrictive conditions,, just judges that this initial survey result is not a characteristic portion as long as there is a condition not meet:

1, this initial survey result's area is no more than 1/100 of entire image;

2, this initial survey result is not the people on the face;

When judging whether the initial survey result satisfies this restrictive condition, at first, adopt people's face detection algorithm to detect people's face and whether exist, if there is no people's face, then the initial survey result certainly not on the face the people, promptly the initial survey result meets this restrictive condition.If detected people's face, whether the overlay area of then then judging this initial survey result exists superposed part with the overlay area of people's face.If overlap, think that then the initial survey result is positioned at people's face, otherwise this initial survey result be not the people on the face.

3, the nearest colour of skin block of this initial survey result is no more than a block length (10 pixels);

The consideration that this restrictive condition is set is to get rid of the initial survey result far apart from area of skin color.If these initial surveies result not on human body skin, has been not a characteristic portion just naturally.

4, this initial survey result's edge is rounded.

As mentioned before, characteristic portion and nipple in the patent of the present invention are used with, consider that there is a circle mammary areola on every side in nipple, and the shape of this mammary areola are near circular.Therefore can detect mammary areola by the method for hough change detection circle.If, be not characteristic portion naturally then so this initial survey result's edge is not circular.

Initial survey result by above 4 restrictive condition examinations is called inspection result eventually, puts into inspection set eventually, and the element of set adds up to K, and initial value is zero.If at this moment K is more than or equal to 1, entire image just is judged as porny so.In sum, when (S+K)/2＞1, entire image just is judged as porny.

Claims (6)

1. porny recognition methods based on the colour of skin and characteristic portion joint-detection is characterized in that this method may further comprise the steps successively:

(1) preparatory stage of sample training, is used to obtain the required parameter of cognitive phase of image to be checked; The preparatory stage of this sample training launches around colour of skin sample training and characteristic portion sample training respectively;

(2) read image to be checked;

(3) cognitive phase of image to be checked, be used for the parameter that preparatory stage obtained according to sample training, whether to image to be checked is that porny is judged, at first carry out Face Detection, carrying out characteristic portion then detects, if the area of area of skin color accounts for more than 1/2 of entire image, and occurred 1 characteristic portion at least, then this picture is judged as porny.

2. the porny recognition methods based on the colour of skin and characteristic portion joint-detection according to claim 1, it is characterized in that: the colour of skin sample training in the step (1), at first collect the colour of skin sample of sufficient amount, obtain colour of skin grey level histogram, approach colour of skin grey level histogram by gauss hybrid models then, adopt the EM iterative algorithm to obtain model parameter.

3. the porny recognition methods based on the colour of skin and characteristic portion joint-detection according to claim 1, it is characterized in that: the characteristic portion sample training in the step (1), at first collect the characteristic portion sample of sufficient amount, train the sorter of a cascade then by the Adaboost algorithm.

4. the porny recognition methods based on the colour of skin and characteristic portion joint-detection according to claim 3 is characterized in that: describedly train the sorter cascade quantity of a cascade more than 20 by the Adaboost algorithm.

5. the porny recognition methods based on the colour of skin and characteristic portion joint-detection according to claim 1, it is characterized in that: the Face Detection described in the step (3) is that first length and width with image to be checked waits 10 parts of branches respectively, form 100 cells, get of the representative of the central point of these cells as this cell, if judging this central point is colour of skin point, the residing cell of this central point is colour of skin lattice so; When judging whether this central point is colour of skin point, the gray-scale value of this central point is decomposed on three components of R, G in the rgb space, B, be updated in the gauss hybrid models of in the preparatory stage, setting up, obtain the probability that it becomes colour of skin point, if these three components are all greater than its corresponding threshold value, then adjudicating this central point is colour of skin point, and the cell at place is a colour of skin lattice; Each cell for dividing all repeats above judgement, thereby obtains area of skin color; If the area of this area of skin color accounts for more than 1/2 of entire image, then proceed characteristic portion and detect, otherwise this image is not judged as porny.

6. the porny recognition methods based on the colour of skin and characteristic portion joint-detection according to claim 1, it is characterized in that: the characteristic portion described in the step (3) detects and at first is used in the cascade classifier of training out with the Adaboost algorithm in the preparatory stage, image to be checked is carried out initial survey, then the initial survey result is carried out examination one by one according to following 4 restrictive conditions, as long as there is a condition not meet, just judge that this initial survey result is not a characteristic portion:

(a) this initial survey result's area is no more than 1/100 of entire image;

(b) this initial survey result is not the people on the face;

(c) the nearest colour of skin block of this initial survey result is no more than i.e. 10 pixels of a block length;

(d) this initial survey result's edge is rounded;

Initial survey result by above 4 restrictive condition examinations is called inspection result eventually; If the inspection result has 1 at least eventually, so whole image to be checked just is judged as porny.

Images