CN101276417A - Method for filtering internet cartoon medium rubbish information based on content - Google Patents

Abstract

A method for filtering garbage information of Internet animation media based on contents in the fields of image processing and mode identification is disclosed, comprising the steps: 1, collecting at random a plurality of advertisement or non-advertisement animations from the Internet as training sample set, and extracting colour characteristics, texture characteristics, character information characteristics, animation length characteristics, animation geometric size characteristics and animation dynamic characteristics of all the animations; 2, inputting six characteristics of animations and type indexing into a support vector machine which obtains a model of the support vector machine describing the difference between advertisement and non-advertisement animation media; 3, according to the method for extracting the animation characteristics in the step 1, extracting six characteristics of animations to be tested and inputting into the model of the support vector machine so that the support vector machine is able to judge the animation belongs to either the advertisement animation or the non-advertisement animation. The invention can identify the advertisement animation or the non-advertisement animation as well as have high processing speed and identification accuracy.

Description

Content-based method for filtering internet cartoon medium rubbish information

Technical field

What the present invention relates to is the method for a kind of Flame Image Process and mode identification technology, particularly a kind of content-based method for filtering internet cartoon medium rubbish information.

Background technology

Along with Internet fast development, the rapid of broadband popularized, the reduction of carrying cost, and content of multimedia is also more and more on the net.The multimedia messages that comprises various types of data such as text, image, audio frequency, video etc., rapidly expansion becomes the main flow of information gradually, and people's the life and the development of society have been produced significant effects.Simultaneously, the development of video compression technology makes video quality when volume is more and more littler more and more higher on the contrary, and therefore, most of website all begins animation is incorporated among the webpage.Simultaneously, because popularizing of equipment such as Digital Video also emerges in large numbers like the mushrooms after rain for the website of individual's propagation and exchange video.

The characteristics of these video informations are, its quantity of information is very big on the one hand, are difficult to accurately describe with a small amount of mark, and different observers or same observer may provide different descriptions to same width of cloth image under different condition.This makes that text marking can not practical requirement under many circumstances.Its structuring degree is lower on the other hand, is unfavorable for effectively managing.The related content that how to effectively utilize in the multimedia messages has become a urgent problem.What deserves to be mentioned is, enjoy the while easily that scientific and technological prosperity brings, many junk information of propagating with flash such as advertisement also occurred, bring inconvenience to the user people.Existing multiple browser, as Internet Explorer, Firefox etc. provide the module of advertisement filter or similar functions, but these softwares are to realize by the key word of analyzing its filename mostly, can't correctly realize classification and filtering function when filename lacks meaning.

Find through retrieval the prior art document, Y.Alp Aslandogan etc. are at " IEEETransactions on Knowledge and Data Engineering " vol.11, no.1, Jan.1999, (" IEEE journal knowledge and data engineering " in January, 1999, the 11st volume, no.1) " image and video retrieval technology and the system " that delivers on (" Techniques and Systems for Image and VideoRetrieval ") proposed the method for the understanding and the retrieval of image and video in this article.For image: at first extract to describe color of image, texture and the shape facility of picture material, and other relevant informations (as filename, note etc.) of non-image content, by mating for the characteristics of image in the image library and comparing, obtain understanding and retrieval then for image.For video: at first cut apart, carry out target detection and tracking then, finish understanding and retrieval at last for video content for video lens.Yet, at the cartoon medium on the internet, said method has following deficiency: first, all there are bigger difference in cartoon medium and traditional image and video in duration, shot change, COLOR COMPOSITION THROUGH DISTRIBUTION, and its feature extracting method also is not suitable for analysis and understanding for internet cartoon medium; The second, for this new application of filtering internet cartoon medium rubbish information, the video features that proposes in this method does not have higher advertisement/non-ad classification resolving ability.

Summary of the invention

The objective of the invention is to overcome the deficiencies in the prior art, a kind of content-based method for filtering internet cartoon medium rubbish information is proposed, according to the characteristics of cartoon medium and the singularity of junk information categorical filtering, it is understood and filter according to the content of animation.

The present invention is achieved through the following technical solutions, and comprises the steps:

Step 1, the animation that several classification indexes of random acquisition are advertisement and non-advertisement from the internet is as training sample set, and extract the behavioral characteristics that training sample is concentrated the color characteristic of all animations, textural characteristics, Word message feature, animation length characteristic, animation geometries characteristic, animation, specific as follows:

According to the color histogram of each picture frame in the animation and the color characteristic of three passage colors of RGB average extraction animation;

Gray-scale map to each picture frame in the animation carries out the rich conversion of gal, extracts the textural characteristics of animation;

Each picture frame in the animation is carried out discrete cosine transform, the outline map of computed image, and then detected image frame Chinese version piece, the Word message feature of extraction animation;

Calculate animation duration span, extract the animation length characteristic;

Calculate the length and the width of picture frame in the animation, extract the animation geometries characteristic;

Calculate the average of the color distortion of each interframe of animation, extract the behavioral characteristics of animation.

Step 2, six features of all animations that the training sample that step 1 is obtained is concentrated and the classification index of corresponding animation are input in the support vector machine (SVM), support vector machine obtains parameters such as weights, biasing to all sample training, and obtains to describe the SVM model of advertisement and non-advertisement cartoon medium difference.

Step 3, for an animation to be tested, extract six features of animation to be tested according to the method for extracting animation feature in the step 1, and in the SVM model with these six feature input steps, two training gained, the SVM model is judged animation and is belonged to the commercial paper animation or belong to non-commercial paper animation, is filtered by medium for the animation of commercial paper.

Described extraction color characteristic is specially:

1. extract the color histogram of each picture frame, and calculate the average color histogram feature of all images frame;

2. calculate the color average of three Color Channels of each picture frame, and calculate the color average of all images frame.

Described texture feature extraction is specially: be meant for, use the Gabor wavelet transformation to come texture feature extraction,

At first, each color image frames is converted to gray-scale map: Y=0.299R+0.587G+0.114B, wherein: Y represents gray-scale value, R, G, B represent the value of three passages respectively;

Then, gray-scale map is carried out Gabor (Jia Bai) filtering, the Gabor wave filter is at the average of the output of { 0 °, 45 °, 90 °, 135 ° } four direction and the variance textural characteristics as this picture frame;

At last, ask the textural characteristics that on average obtains whole animation for the textural characteristics of all images frame.

Described extraction Word message feature is specially:

1. each picture frame color image in the animation is carried out gray processing and handle, obtain gray-scale map;

2. gray-scale map is carried out the distribution plan that discrete cosine transform obtains the gray scale edge;

3. according to gray scale marginal distribution figure, do projection respectively to level and vertical direction;

4. according to the distribution of horizontal direction edge block, set width experience threshold value and scan, width greater than the row of experience threshold value as line of text;

5. according to the distribution of line of text vertical direction edge block, setting height experience threshold value scans, and height is defined as text block greater than the piece of experience threshold value;

6. text block is screened, delete oriented pseudo-text block;

7. for the image that comprises text block, its text feature is designated as 1, otherwise is designated as 0.

The behavioral characteristics of described extraction animation is specially:

1. calculate the absolute value sum of the color distinction between all adjacent image frames;

2. ask on average for all interframe color distinction in the animation, obtain describing the behavioral characteristics of animation multidate information.

Described support vector machine, its kernel function are RBF (radially base) function, and the variance of RBF function is 0.0001.

Compared with prior art, the present invention has following beneficial effect: the present invention according to cartoon medium information on the internet in aspects such as duration, shot change, COLOR COMPOSITION THROUGH DISTRIBUTION and the traditional video and the difference of image, main difference in conjunction with advertisement animation and non-advertisement animation, adopted a kind of new cartoon medium feature representation method, on performance, be better than traditional feature representation method at image and visual classification at garbage information filtering.Through a large amount of experiment tests, the present invention can filter the rubbish animation information accurately and efficiently, and average accuracy is higher than 90%.

Description of drawings

Fig. 1 is the workflow diagram of the inventive method.

Embodiment

Below in conjunction with accompanying drawing embodiments of the invention are elaborated: present embodiment is being to implement under the prerequisite with the technical solution of the present invention, provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

As shown in Figure 1, present embodiment comprises the steps:

Step 1,1200 animations of random acquisition are as training sample set (wherein: the index of 420 animation classifications is that advertisement and the index of 780 animation classifications are non-advertisement) from the internet, and extract the behavioral characteristics that training sample is concentrated the color characteristic of all animations, textural characteristics, Word message feature, animation length characteristic, animation geometries characteristic, animation, specific as follows:

1. extract the animation color characteristic, comprise the steps:

At first, extract the color histogram of each picture frame, and calculate the average color histogram feature of all images frame, be specially:

$\mathrm{HIST}=\frac{1}{N}\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{HIST}}_i}{\mathrm{sum}\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{HIST}}_i\right)}$

Wherein: the color histogram after HIST represents on average, HIST _iThe color histogram of representing i picture frame, N are represented frame number altogether.

Then, calculate the color average of three Color Channels, as an other color characteristic, and the single order color moment of calculating all images frame, be specially:

$\mathrm{\μ}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\μ}}_i$

Wherein: μ _iThe color first moment of representing i picture frame, μ are represented the average color first moment of whole animation.Because the physical dimension of each two field picture equates, so following formula can be exchanged into:

$\mathrm{\μ}=\frac{1}{\mathrm{NWH}}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{w=1}{\overset{W}{\mathrm{\Σ}}}\underset{h=1}{\overset{H}{\mathrm{\Σ}}}{I}_i(w,h)$

Wherein, the width of W presentation video, the height of H presentation video, I _i(w, h) expression i frame coordinate is (w, color of pixel value h).

2. extract the cartoon grain feature: for each picture frame in the animation, use the Gabor wavelet transformation to come texture feature extraction, be specially:

At first, coloured image is converted to gray-scale map, method is as follows:

Y＝0.299R+0.587G+0.114B

Wherein: Y represents gray-scale value, R, and G, B represent the value of three passages respectively;

Then, gray-scale map is carried out Gabor filtering, the Gabor wave filter is at the average of the output of { 0 °, 45 °, 90 °, 135 ° } four direction and the variance textural characteristics as this picture frame;

Described Gabor wave filter, its formula is as follows:

$g(x,y)=\frac{U_h}{U_l}\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}\exp [(-\frac{x^2+y^2}{2{\mathrm{\σ}}^2})+2\mathrm{\πj}{U}_{h}x]$

g _m(x，y)＝a ^-mg(x′，y′)

x′＝a ^-m(xcosθ+ysinθ)

y′＝a ^-m(-xsinθ+ycosθ)

Wherein: θ is the direction factor of Gabor feature, gets θ={ 0 °, 45 °, 90 °, 135 ° } four direction and carries out filtering; U _hAnd U _lFor paying close attention to the bound of frequency, get U here _h=0.4, U _l=0.05, the frequency domain variance $\mathrm{\σ}={\left(\frac{0.6}{U_l}\right)}^2;$ Dimension scale factor a is set to 1, and filtering window is 9.

At last, ask the textural characteristics that on average obtains whole animation for the textural characteristics of all images frame.

3. extract the Word message feature, comprise the steps:

The first step is carried out gray processing to each picture frame that comprises in the animation and is handled, and is converted to gray-scale map: Y=0.299R+0.587G+0.114B

Wherein: R, G and B represent the value of three Color Channels respectively, and Y is input gray level figure;

Second step, rim detection: the piece that gray level image is cut into 8*8 carries out discrete cosine (DCT) conversion, get the texture value of the sub-piece of part DCT ac coefficient computed image, then these data texturings are being kept frequency domain figure picture of combination generation under the image subblock ordering permanence condition, the frequency domain figure picture characterizes the marginal information of original image to a certain extent.For obtaining high-quality frequency domain figure picture, need the gray level image interpolation is amplified, experience shows: effect was better when amplification was 5-7 times by number, adopted 6 times of interpolation to amplify in the present embodiment.

Described dct transform, its formula is:

${\mathrm{AC}}_{s,t}=\frac{1}{8}{C}_s{C}_t\underset{m=0}{\overset{7}{\mathrm{\Σ}}}\underset{n=0}{\overset{7}{\mathrm{\Σ}}}G(m,n)\cos \left[\frac{\mathrm{\π}}{16}(2m+1)s\right]\cos \left[\frac{\mathrm{\π}}{16}(2n+1)t\right]$

Wherein: ((m is that coordinate is (m, n) gray values of pixel points in the 8*8 piece n) to G for s, the t) coordinate after the expression conversion; C _sAnd C _tBe normalized parameter, satisfy

$C_s,C_t=\left\{\begin{array}{cc}\frac{1}{\sqrt{2}}& s,t=0\\ 0& \mathrm{otherwise}\end{array}\right.(0\≤s,t\≤7)$

For coordinate is (i, image subblock j), AC _0,0(i j) is the low frequency part of image subblock, is called DC coefficient, and all the other coefficients are called ac coefficient.

In the 3rd step,, calculate each image subblock texture value T according to above-mentioned DCT coefficient _AC(i j), and then adopts and based on the piece texture filter in zone texture image is carried out filtering, obtains outline map.

Described sub-piece texture value T _AC(its formula is for i, j) computing method:

$T_{\mathrm{AC}}(i,j)=\underset{m=1}{\overset{3}{\mathrm{\Σ}}}\left|{\mathrm{AC}}_{0,m}(i,j)\right|+\underset{n=1}{\overset{3}{\mathrm{\Σ}}}\left|{\mathrm{AC}}_{n,0}(i,j)\right|+{\mathrm{AC}}_{l,1}(i,j)$

Wherein: AC is the frequency spectrum parameter behind the dct transform described in second step.

Described piece texture filter filtering method based on the zone, its formula is:

$E(i,j)=\left\{\begin{array}{cc}1& T_{\mathrm{AC}}(i,j)>\max (\mathrm{\α}\×\stackrel{\‾}{T_{\mathrm{AC}}},{\mathrm{MT}}_{\mathrm{Th}})\\ 0& \mathrm{otherwise}\end{array}\right.$

$\stackrel{\‾}{T_{\mathrm{AC}}}=\frac{1}{M\×N}\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}{T}_{\mathrm{AC}}(i,j)$

Wherein: M, N represent the number of subimage block on level and the vertical direction, MT respectively _ThIt is minimum texture value requirement.Get MT in the present embodiment _Th=230 and α=1.5.(after will exporting E and adjusting to original size, its isolated pixel point of filtering obtains original edge figure to output E for i, the outline map after j) representative is amplified.

In the 4th step, adopt based on projecting method and carry out the text block location: at first according to the distribution of horizontal direction edge block, select an experience threshold value (get in the present embodiment picture width 1/3) to scan, the row that meets the certain altitude requirement as line of text; Secondly according to the distribution of line of text vertical direction edge block, select an experience threshold value (get in the present embodiment picture height 1/10) to scan equally, determine the position of text block.At last, text block is screened, delete oriented pseudo-character block according to the rate of filling up, aspect rate and character boundary.

A large amount of experiments show, compare word area detection (being whether to comprise character area in the detected image), and the accuracy rate of character area location, identification is much lower.Therefore, whether the employed character features of present embodiment only detects character area and exists, and only needs a Boolean type variable can represent in 1 representative image character area is arranged, and does not have character area in 0 representative image.

4. extract the animation duration features: the frame number that the totalframes of the duration=animation of animation/per second is play.

The difference on the duration at advertisement animation and non-advertisement animation, present embodiment also with the information of animation duration as one of foundation of classification and identification.

5. extract the geometries characteristic of animation: geometries characteristic comprises: the picture frame width and the height that 1) with the pixel are unit; 2) length breadth ratio of picture frame (width of the length/image of image).

6. extract the animation behavioral characteristics

The difference on camera lens switches in view of advertisement animation and non-advertisement animation, color distortion is described the behavioral characteristics of animation between the average frame of introducing animation, and the mean value of frame difference is $\mathrm{mean}\left(\frac{1}{N-1}\underset{i=2}{\overset{N}{\mathrm{\Σ}}}\right|F_i-F_{i-1}\left|\right)$

Wherein: N is the frame number of an animation, F _i, F _I-1Represent the vector that the rgb color value of all pixels in the image of i frame, i-1 frame is constituted respectively.

Step 2, six features of all animations that the training sample that step 1 is obtained is concentrated and the classification index of corresponding animation are input in the support vector machine (SVM), support vector machine obtains parameters such as weights, biasing to all sample training, and obtains to describe the SVM model of advertisement/non-advertisement cartoon medium difference.

Described support vector machine, the kernel function of SVM wherein are the RBF function, and the variance of RBF function is 0.0001.

Step 3, for an animation to be tested, extract six features of animation to be tested according to the method for extracting animation feature in the step 1, and in the SVM model with these six feature input steps, two training gained, obtain the test set filter result by the SVM model, promptly judge animation and belong to the commercial paper animation or belong to non-commercial paper animation, filter by medium for the animation of commercial paper.

The present embodiment method is by testing from the animation on the internet surplus 1000, and present embodiment has higher correct recognition rata to flash, and the average error rate is lower than 10%.

Claims (7)

1, a kind of content-based method for filtering internet cartoon medium rubbish information is characterized in that, comprises the steps:

Step 1, the animation that several classification indexes of random acquisition are advertisement and non-advertisement from the internet is as training sample set, and extract color characteristic, textural characteristics, Word message feature, length characteristic, geometries characteristic, the behavioral characteristics that training sample is concentrated all animations, specific as follows:

According to the color histogram of each picture frame in the animation and the color characteristic of three passage colors of RGB average extraction animation,

Gray-scale map to each picture frame in the animation carries out the rich conversion of gal, extracts the textural characteristics of animation,

Each picture frame in the animation is carried out discrete cosine transform, the outline map of computed image, and then detected image frame Chinese version piece, the Word message feature of extraction animation,

Calculate animation duration span, extract the length characteristic of animation,

Calculate the length and the width of picture frame in the animation, extract the geometries characteristic of animation,

Calculate the average of the color distortion of each interframe of animation, extract the behavioral characteristics of animation;

Step 2, six features of all animations that the training sample that step 1 is obtained is concentrated and the classification index of corresponding animation are input in the support vector machine, support vector machine obtains weights, offset parameter to all sample training, and obtains to describe the supporting vector machine model of advertisement and non-advertisement cartoon medium difference;

Step 3, for an animation to be tested, extract six features of animation to be tested according to the method for extracting animation feature in the step 1, and in the supporting vector machine model with these six feature input steps, two training gained, supporting vector machine model is judged animation and is belonged to the commercial paper animation or belong to non-commercial paper animation, is filtered by medium for the animation of commercial paper.

2, content-based method for filtering internet cartoon medium rubbish information according to claim 1 is characterized in that, described extraction color characteristic is specially:

1. extract the color histogram of each picture frame, and calculate the average color histogram feature of all images frame;

2. calculate the color average of three Color Channels of each picture frame, and calculate the color average of all images frame.

3, content-based method for filtering internet cartoon medium rubbish information according to claim 1 is characterized in that, described texture feature extraction is specially:

At first, each color image frames is converted to gray-scale map: Y=0.299R+0.587G+0.114B, wherein: Y represents gray-scale value, R, G, B represent the value of three passages respectively;

Then, gray-scale map is carried out Gabor filtering, the Gabor wave filter is at the average of the output of { 0 °, 45 °, 90 °, 135 ° } four direction and the variance textural characteristics as this picture frame;

At last, ask the textural characteristics that on average obtains whole animation for the textural characteristics of all images frame.

4, content-based method for filtering internet cartoon medium rubbish information according to claim 1 is characterized in that, described extraction Word message feature is specially:

1. each picture frame color image in the animation is carried out gray processing and handle, obtain gray-scale map;

2. gray-scale map is carried out the distribution plan that discrete cosine transform obtains the gray scale edge;

3. according to gray scale marginal distribution figure, do projection respectively to level and vertical direction;

4. according to the distribution of horizontal direction edge block, set width experience threshold value and scan, width greater than the row of experience threshold value as line of text;

5. according to the distribution of line of text vertical direction edge block, setting height experience threshold value scans, and height is defined as text block greater than the piece of experience threshold value;

6. text block is screened, delete oriented pseudo-text block;

7. for the image that comprises text block, its text feature is designated as 1, otherwise is designated as 0.

5, content-based method for filtering internet cartoon medium rubbish information according to claim 1 is characterized in that, the behavioral characteristics of described extraction animation is specially:

1. calculate the absolute value sum of the color distinction between all adjacent image frames;

2. ask on average for all interframe color distinction in the animation, obtain describing the behavioral characteristics of animation multidate information.

6, content-based method for filtering internet cartoon medium rubbish information according to claim 1 is characterized in that, described support vector machine, its kernel function are radial basis function.

7, content-based method for filtering internet cartoon medium rubbish information according to claim 6 is characterized in that, described radial basis function, its variance are 0.0001.

Images