CN102163286A

CN102163286A - Pornographic image evaluating method

Info

Publication number: CN102163286A
Application number: CN 201010113826
Authority: CN
Inventors: 胡卫明; 左海强; 吴偶
Original assignee: Institute of Automation of Chinese Academy of Science
Current assignee: Renmin Zhongke Beijing Intelligent Technology Co ltd
Priority date: 2010-02-24
Filing date: 2010-02-24
Publication date: 2011-08-24
Anticipated expiration: 2030-02-24
Also published as: CN102163286B

Abstract

The invention discloses a pornographic image evaluating method. The method comprises the following steps of: dividing the complexion of an image; extracting 31-dimensional characteristics, namely the integral characteristics of the image, the local part characteristics of a human body and human body shape characteristics, to form characteristic vectors; and training an evaluation model, evaluating the pornographic degree of an input image, and outputting an evaluation result. By the method, whether the input image is a pornographic image or not can be judged, the pornographic degree of the image can be evaluated, 0 to 100 percent of evaluation results can be output, and a larger numerical value means the higher pornographic degree of the image, so that pornographic images is convenient to filter in a classified mode.

Description

Pornographic image evaluation method

Technical Field

The invention relates to the technical field of computer application, in particular to a pornographic image evaluation method.

Background

Currently, special actions for jointly remedying obscency and customs information of internet and mobile phone media are widely carried out nationwide and even worldwide. Because of the poor self-control ability of teenagers, nearly 80% of people are tempted by network bad information in the case of adolescent crimes. The system is especially necessary for helping teenagers avoid the erosion of pornographic information, reducing the possibility that the teenagers walk on a criminal road, effectively evaluating and filtering the pornographic degree of the image and intercepting the harmful information before the teenagers acquire the harmful information technically.

In the prior art, simple features such as skin color proportion and the like are mostly adopted, and the overall features of the image, the local part features of the human body and the shape features of the human body trunk cannot be comprehensively considered, so that the recognition rate is low, most output results are pornographic or normal logic values, and the pornographic degree of the image cannot be measured.

Disclosure of Invention

Technical problem to be solved

In view of the above, the present invention is directed to solve the problems that the recognition rate is low and the pornographic degree of an image cannot be measured in the prior art, and therefore, the present invention provides a pornographic image evaluation method capable of evaluating the pornographic degree of an input image and outputting an evaluation result.

(II) technical scheme

In order to achieve the above object, the present invention provides a pornographic image evaluating method comprising:

step 1: segmenting the skin color of an input image to obtain a skin color area;

step 2: extracting 31-dimensional features including the overall features of an input image, the local part features of a human body and the shape features of the human body on a large number of artificially marked image samples to form feature vectors;

and step 3: training an evaluation model, evaluating the pornographic degree of the input image, and outputting a pornographic image evaluation result.

(III) advantageous effects

According to the technical scheme, the invention has the following advantages:

the invention can not only judge whether the input image is the pornographic image, but also evaluate the pornographic degree of the image, improves the recognition rate and measures the pornographic degree of the image.

1. The erotic image evaluation method provided by the invention can effectively evaluate the erotic degree of the image due to comprehensive consideration of the overall characteristics of the image, the characteristics of the local part of the human body and the shape characteristics of the human body.

2. According to the method for evaluating the pornographic image, provided by the invention, the evaluation result of the pornographic degree of the image is output, the image is simply judged whether the pornographic degree is the same or not, 0% -100% of the evaluation result is output, and the higher the numerical value is, the higher the pornographic degree of the image is, so that the pornographic image can be conveniently graded and filtered.

Drawings

Fig. 1 is a flowchart of a pornographic image evaluation method.

FIG. 2 shows the results of the detection of the local part and the trunk of the human body

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to specific embodiments and the accompanying drawings.

The execution environment of the invention is composed of three modules which are realized by a computer:

the skin color extraction module has the main function of carrying out skin color segmentation on an input image to obtain a skin color area.

And the feature extraction module is mainly used for extracting 31-dimensional features including image overall features, human body local part features and human body trunk shape features on the input image to form feature vectors.

And thirdly, a model training and evaluating module, wherein the module has the main function of obtaining characteristics on a large number of manually marked image samples, inputting the characteristics into an evaluating model, and training the evaluating model to obtain model parameters. And evaluating the input image by using the trained evaluation model, and outputting an evaluation result.

A detailed description of the various step details involved in the solution of the invention is given below in connection with fig. 1.

Step 101: and segmenting the skin color of the input image. Any skin color segmentation algorithm may be employed, there are mainly two types of skin color segmentation algorithms, one being at the pixel level and the other being at the region level. The skin color segmentation algorithm based on the region can comprehensively consider the attributes of skin color, such as texture, and the like, so that a better segmentation effect can be generally obtained.

Step 102: and extracting 31-dimensional features including the integral image features, the local human body part features and the human body trunk shape features to form feature vectors.

Integral characteristics of image

The overall features of the image consist of 5 parts:

image aspect ratio;

entropy of the image;

the entropy of a grayscale image is defined as:

<math><mrow><mi>Entropy</mi><mo>=</mo><mo>-</mo><munderover><mi>Σ</mi><mrow><mi>i</mi><mo>=</mo><mn>0</mn></mrow><mn>255</mn></munderover><msub><mi>p</mi><mi>i</mi></msub><mo>·</mo><msub><mi>log</mi><mn>2</mn></msub><msub><mi>p</mi><mi>i</mi></msub><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>1</mn><mo>)</mo></mrow></mrow></math>

in the formula, p_iIs the percentage of the number of pixels with gray value i to the total number of pixels in the whole image.

The skin color proportion of the image;

skin color ratio:

the number of skin color connected regions;

geometric moment of skin tone region (this feature has 10 dimensions);

the (p + q) order geometric moment of the image I (x, y) is defined as:

wherein x represents the image horizontal coordinate, y represents the image vertical coordinate, and p and q are arbitrary non-negative integers.

We extract the first 3 rd order geometric moments of the skin color region:

geometric moment of 0 th order (p + q ═ 0): m₀₀；

Geometric moment of order 1 (p + q ═ 1): m₁₀，M₀₁；

Geometric moment of order 2 (p + q ═ 2): m₂₀，M₁₁，M₀₂；

Geometric moment of order 3 (p + q ═ 3): m₃₀，M₂₁，M₁₂，M₀₃；

These 10 geometric moments are used as the geometric moment feature of the skin color region, which is expressed as follows: m₀₀，M₁₀，M₀₁，M₂₀，M₁₁，M₀₂，M₃₀，M₂₁，M₁₂，M₀₃。

Second, local part characteristics of human body

Next 4 features are related to the detection of the local part of the human body (see fig. 2), and we use a cascade classifier based on Haar (Haar) features to perform the detection of the local part of the human body, wherein the 4 features are expressed as follows:

the number of detected faces;

the number of breasts detected;

the detected private part number;

the proportion of the human face skin color to the whole image skin color;

shape characteristics of human body

The last 7 features, which are extracted from the maximum skin tone connected region and its fitting ellipse (see fig. 2), mainly describe the shape of the human torso, are expressed as follows:

the ratio of the maximum skin color communication area to the total skin color area;

the distance from the center of the maximum skin color connected region to the center of the image;

geometric invariant of the maximum skin tone connected region (this feature has 7 dimensions);

the geometrically invariant moments of the image I (x, y) have the properties of rotation, translation, scale invariance, etc. In order to calculate the geometric invariant moment of the image, the central moment μ of the image needs to be introduced first_pq：

<math><mrow><msub><mi>μ</mi><mi>pq</mi></msub><mo>=</mo><munder><mi>Σ</mi><mi>x</mi></munder><munder><mi>Σ</mi><mi>y</mi></munder><msup><mrow><mo>(</mo><mi>x</mi><mo>-</mo><mover><mi>x</mi><mo>&OverBar;</mo></mover><mo>)</mo></mrow><mi>p</mi></msup><msup><mrow><mo>(</mo><mi>y</mi><mo>-</mo><mover><mi>y</mi><mo>&OverBar;</mo></mover><mo>)</mo></mrow><mi>q</mi></msup><mi>I</mi><mrow><mo>(</mo><mi>x</mi><mo>,</mo><mi>y</mi><mo>)</mo></mrow><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>4</mn><mo>)</mo></mrow></mrow></math>

Wherein p, q are any non-negative integer,

is the centroid of the image area, M₁₀，M₀₁The image geometric moments defined for equation (3).

The 0-3 order central moments of the image can be expressed as:

μ₀₀＝M₀₀，

μ₀₁＝0，

μ₁₀＝0，

μ₁₁＝M₁₁-xM₀₁＝M₁₁-yM₁₀，

μ₂₀＝M₂₀-xM₁₀，

μ₀₂＝M₀₂-yM₀₁，

μ₂₁＝M₂₁-2xM₁₁-yM₂₀+2x²M₀₁， (5)

μ₁₂＝M₁₂-2yM₁₁-xM₀₂+2y²M₁₀，

μ₃₀＝M₃₀-3xM₂₀+2x2M₁₀，

μ₀₃＝M₀₃-3yM₀₂+2y2M₀₁.

the central moments of the images have only the property of being shift invariant. For p + q ≧ 2, the moment η with scale invariant properties can be constructed using the following formula_pq：

On the basis, the following geometric invariant moment is extracted as a feature from the maximum skin color connected region of the image, and the feature h is used as a feature₁To characteristic h₇As shown below:

h₁＝η₂₀+η₀₂，

h₂＝(η₂₀-η₀₂)²+(2η₁₁)²，

h₃＝(η₃₀-3η₁₂)²+(3η₂₁-η₀₃)²，

h₄＝(η₃₀+η₁₂)²+(η₂₁+η₀₃)²，

h₅＝(η₃₀-3η₁₂)(η₃₀+η₁₂)[(η₃₀+η₁₂)²-3(η₂₁+η₀₃)²]+

(3η₂₁-η₀₂)[3(η₃₀+η₁₂)²-(η₂₁+η₀₃)²]，

h₆＝(η₂₀-3η₁₂)[(η₃₀+η₁₂)²-3(η₂₁+η₀₃)²]+4η₁₁(η₃₀+η₁₂)(η₂₁+η₀₃)

h₇＝(3η₂₁-3η₀₃)(η₃₀+η₁₂)[(η₃₀+η₁₂)²-3(η₂₁+η₀₃)²]+ (7)

(η₃₀-3η₁₂)(η₂₁+η₀₃)[3(η₃₀+η₁₂)²-(η₂₁+η₀₃)²].

roundness of the maximum skin color connected region;

we calculate the circularity R of the maximum skin tone connected region using the following formula:

R = \frac{{| | F_{1} | |}^{2}}{{| | F_{- 1} | |}^{2}} - - - (8)

wherein,

n is the number of pixel points on the outline of the maximum skin color communication area, k is an integer of 0_k，y_k) Is the coordinate of the point on the outline of the maximum skin tone communication area, and j is the imaginary unit (j)²＝-1)，F₁、F_-1Respectively shows the wheels for calculating the maximum skin color connected region in the anticlockwise direction and the clockwise directionThe first order component of the fourier transform of the profile. When R is 0, the maximum skin color connected region is a circular region; when R is 1, the maximum skin color connected region is a straight line; when R is between 0 and 1, it indicates that the region of maximum skin tone cell is an ellipse.

Irregularity of the maximum skin tone connected region;

the irregularity IR of the maximum skin color connected region represents the degree of deviation of the region outline from a circle, and the irregularity IR of the maximum skin color connected region is calculated by adopting the following formula:

wherein σ²The variance of all points on the contour of the maximum skin tone connected region is represented and is given by:

fitting the ratio of the major axis and the minor axis of an ellipse to the maximum skin color connected region;

fitting the inclination angle of the main shaft of the ellipse in the maximum skin color communication area;

step 103: and training an evaluation model, evaluating the pornographic degree of the input image and outputting an evaluation result.

The evaluation model selects a random forest classifier of a regression mode, the random forest classifier is composed of a series of decision trees, and in the regression mode, the output of a random forest algorithm is the average value of the output results of all the decision trees.

Firstly, a training image library is marked manually, wherein a normal image is marked with a value of 0.0, and a pornographic image is marked with a value of 1.0. Extracting features from the marked image library samples, inputting the features into an evaluation model, training the evaluation model to obtain model parameters, and automatically learning the distribution characteristics of the mood images and the normal images on the extracted 31-dimensional feature space by training the evaluation model. And evaluating the input image by using the trained evaluation model, judging the input image independently by each decision tree in the random forest classifier, and outputting a decision result of 0 or 1. The final output result is the average value of the decision results, namely the evaluation result between 0% and 100%, and represents the pornographic degree of the image, and the larger the numerical value is, the higher the pornographic degree of the image is, so that the hierarchical filtering of the pornographic image is facilitated.

The above description is only an embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can understand the changes or substitutions within the technical scope of the present invention, and therefore, the scope of the present invention should be subject to the claims.

Claims

1. A pornographic image evaluation method, comprising:

2. The pornographic image evaluation method according to claim 1, wherein the evaluation model adopts a random forest algorithm and a regression mode of the random forest algorithm.

3. The pornographic image evaluation method according to claim 1, wherein the evaluation result is an evaluation result that outputs 0% to 100%, and a larger numerical value indicates a higher pornographic degree of the image, facilitating hierarchical filtering of the pornographic image.