CN102799633B - Advertisement video detection method - Google Patents

Abstract

The invention discloses an advertisement video detection method, which comprises the following steps of: extracting at least one advertisement video and at least one non-advertisement video, constructing a video training set, extracting at least one audio characteristic of each video, and performing support vector machine training on all the audio characteristics serving as input samples; inputting a video to be detected, and extracting at least one audio characteristic of the video to be detected; calculating the audio characteristic of the video to be detected by using a support vector machine which is trained, and outputting a calculation result; and judging a video detection result according to the calculation result, and outputting the detection result. By the advertisement video detection method, the shortcomings in the prior art can be effectively overcome; and the method and the basis are supplied for quickly detecting and filtering advertisement videos.

Description

A kind of advertisement video detection method

Technical field

The present invention relates to video to detect and video detection technology field, particularly relate to a kind of advertisement video detection method.

Background technology

Along with the development of information age, as main carriers and the communication media of business information, advertisement video in the information interaction of daily life in occupation of more and more important effect.For common video tour personnel, they are not interested to advertisement video.Therefore they by commercial detection system, can navigate to advertisement video part rapidly, thus can carry out the operation of being correlated with, and such as delete for their skimble-skamble advertisement video part in one section of video frequency program, thus retain content of TV program.Such one side can save the time that they watch video, also can save the space of stored video data on the other hand.But in the face of so immense video data, the problem how detecting and filter relevant advertisement video section fast needs solution badly.Therefore there has been proposed advertisement video detection system, utilize it can quick position advertisement video section automatically.Through research in recent years, advertisement video detects and simply develops into the detection based on lens features now based on black/silent (black/noiseless) frame from initial, only considers from initial the detection that the feature detection of visual aspects develops into Voice & Video feature of today and combines.

For the feature difference between advertisement video and general programs video, there has been proposed a lot of purposes of commercial detection algorithms.According to detection algorithm based on feature different, the detection algorithm of present practical advertisement video system can be divided into following two classes substantially:

(1) based on the method (Logo-based methods) of mark

The method is advertisement video detection method the earliest.The method mainly utilizes the station symbol of TV station to detect.The station symbol of oneself can show when playing ordinary tv programme by TV station, and generally will conceal station symbol during broadcast advertisement, and this makes us whether can exist to distinguish advertisement video and ordinary tv programme video by monitor station target.Station symbol is generally divided into three kinds: static station symbol, translucent station symbol and dynamic station symbol.For the station symbol of different shape, there is corresponding detection method, thus realize the detection to advertisement video.Research wherein for static station symbol is more deep, and applies very extensive.But for translucent station symbol and dynamic station symbol, detect more difficult, so also there is no the detection method of comparative maturity.In addition, station symbol also can show when broadcast advertisement video by a lot of radio station now, and the detection method at this moment based on mark will lose efficacy.

(2) rule-based method (Rule-based methods)

Rule-based method is utilize the information of advertisement camera lens to carry out to detect mostly, and it distinguishes advertisement and ordinary tv programme video by a stack features Sum fanction.Because advertisement video and ordinary tv programme video also exist obvious difference in some characteristic aspect, therefore utilize the feature that these distinction are stronger, just can realize the detection to advertisement video.Detection can be realized by the average edge rate of change A-ECR (Average of Edge Change Ratio) of extraction one section of frame of video and edge variation variance V-ECR (Variance of Change Ratio) and average frame difference A-FD (Average of Frame Difference) and frame variance V-FD (Variance of FrameDifference) in video.In audio frequency, also there is the difference in some obvious features in the audio content of advertisement video and the audio content of ordinary tv programme, the general coefficient of audio frequency multi-dimension frequency (Mel-frequency Cepstral Coefficient) and audio-frequency information entropy such as can be utilized to realize detecting the segmentation of advertisement video.But the feature in research sound intermediate frequency is in the past general all in auxiliary video feature detection, could realize splitting detection more accurately to advertisement section by both combinations.

Summary of the invention

The object of the invention is to be to propose a kind of advertisement video detection method, the workload of the staff in video detection and editor field can be reduced in a large number, improve the degree of accuracy of video detection speed and detection.

The invention provides a kind of advertisement video detection method, comprise the following steps:

Steps A, extraction are no less than the advertisement video of and are no less than the non-advertisement video of, set up video training set, extract the audio frequency characteristics being no less than of each video, all described audio frequency characteristics are carried out support vector machine training as input amendment;

Step B, input video to be detected, extract the audio frequency characteristics being no less than of video to be detected;

The support vector machine that step C, the audio frequency characteristics inputting video to be detected are used to complete training carries out calculating and exporting result of calculation;

Step D, judge video testing result according to described result of calculation, described testing result is exported.

Further, in steps A and step B, extract the audio frequency characteristics being no less than of video, further comprising the steps of:

Use decoding tool to decode to video, obtain video image set and audio-frequency information, described audio-frequency information is preserved with array form;

Be be no less than the camera lens of by described video image set-partition, extract the audio frequency characteristics of each camera lens;

Further, the described camera lens becoming to be no less than by described video image set-partition, further comprising the steps of:

Extract the color histogram of all video images, calculate the similar value of the color histogram of two adjacent video images, as interframe similar value;

According to the interframe similar value of described video, default high threshold, default Low threshold and default gradual shot length lowest threshold, determine shot boundary sequence.

Further, the color histogram of all video images of described extraction, calculates interframe similar value, further comprising the steps of:

Carry out piecemeal to video image, be divided into M × N block, wherein M is columns, and N is line number;

Spatially color histogram is extracted at hsv color to each image block of image, be wherein 1 interval luminance quantization, amount of tones is turned to 16 intervals, color saturation is quantified as 8 intervals, each two field picture obtains M × N number of color histogram having 16 × 8 × 1=128 bin, and is normalized the color histogram obtained;

Adopt the similar value of the color histogram of formulae discovery two adjacent video images below:

$S(a,b)=\stackrel{M}{\mathrm{\Σ}}\stackrel{N}{\mathrm{\Σ}}\stackrel{128}{\mathrm{\Σ}}{w}_{\mathrm{pq}}\min (H_a(p,q,i),H_b(p,q,i)),$

Wherein the weight of the capable block of q is arranged, H for being in p _a(p, q, i) is the value of i-th bin of the color histogram of p × q block of a frame, a and b is the sequence number of video frame image, and min function asks for value less in two values.

Further, the described interframe similar value according to whole video, default high threshold, default Low threshold and default gradual shot length lowest threshold, determine shot boundary sequence, further comprising the steps of:

Input the interframe similar value sequence S={s of whole video sequence ₁, s ₂..., s _n, preset high threshold T _h, preset Low threshold T _l, preset gradual shot length lowest threshold T _gradual, wherein T _gradual=10;

According to each interframe similarity s _ijudge:

If s _i< T _land gradual change mark is not set, judge that camera lens there occurs sudden change, the boundary information of output mutation camera lens;

If s _i< T _hbut be provided with gradual change mark, then gradual shot length Length adds 1;

If s _i>=T _lbut s _i< T _hbut gradual change mark is not set, then gradual change mark is set, record present image frame position alternatively camera lens start border, and start to count gradual shot length Length;

If s _i>=T _hand be provided with gradual change mark, then check gradual shot length, if Length > is T _gradualthen judge to there occurs gradual change, export the boundary information of gradual shot, otherwise judge it is not gradual change, cancel gradual change mark, and gradual shot length Length again zero setting;

If s _i>=T _hand gradual change mark is not set, then judge shot change not to occur;

Output lens border sequence.

Further, the audio frequency characteristics of described each camera lens, specifically comprises: zero-crossing rate, short-time energy, DFT coefficient, spectral centroid and Mel frequency cepstral coefficient.

Further, in steps A, all described audio frequency characteristics are carried out support vector machine training as input amendment, further comprising the steps of:

The audio frequency characteristics of advertisement video camera lens is designated 1, and the audio frequency characteristics of non-advertisement video camera lens is designated-1;

By nonlinear transformation, the input space is transformed to a higher dimensional space, in described higher dimensional space, ask for optimum linearity classifying face;

Formula is below adopted to obtain described optimum linearity classifying face:

$S\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{w}_i\mathrm{\&phi;}\left(x_i\right)\mathrm{\&phi;}\left(X\right)+b,$

Wherein, φ () is a nonlinear mapping function, X={x _i∈ R ^d} _i=1...nand y _i={-1,1} represents corresponding training set and corresponding class mark respectively, and b is side-play amount, and any one corresponds to a nonzero value w _isample x _ifor support vector, kernel function adopts gaussian kernel function:

$K(x,y)=\mathrm{\&phi;}\left(x\right)\&CenterDot;\mathrm{\&phi;}\left(y\right)=\exp \left(\frac{-{\left|\right|x-y\left|\right|}^2}{{\mathrm{\&delta;}}^2}\right).$

Further, step C is further comprising the steps of:

Input tape detects the audio feature vector x of video, brings formula into export result of calculation:

$L\left(x\right)=\mathrm{\&theta;}\left(S\left(x\right)\right)=\left\{\begin{array}{ccc}+1& \mathrm{if}& S\left(x\right)>0\\ -1& \mathrm{if}& S\left(x\right)<0\end{array}\right.,$

Wherein, θ () is an indicator function.During L (x)=+ 1, video lens to be detected is judged as advertisement camera lens, otherwise is judged as non-advertisement video camera lens.

Further, step D is further comprising the steps of:

If be continuous print more than the advertisement camera lens of default high level, merged into an advertisement video section;

If the advertisement camera lens being less than default high level is continuous print, but the distance of its nearest advertisement video section is less than predetermined low level, be then incorporated to described nearest advertisement video section;

If the camera lens number between adjacent advertisement video section is lower than predetermined low level, then described adjacent advertisement video section is merged into an advertisement video section;

Repeat above step, until do not have new advertisement video section to occur, if the advertisement video section in video to be detected is greater than preset value, then described video to be detected is judged as advertisement video, output detections result.

Technique effect of the present invention is:

For the staff in video detection and editing field provides a kind of advertisement video detection method, the workload of the staff in video detection and editor field can be reduced in a large number, improve the degree of accuracy of video detection speed and detection.

Other features and advantages of the present invention will be set forth in the following description, and, partly become apparent from instructions, or understand by implementing the present invention.Object of the present invention and other advantages realize by structure specifically noted in write instructions, claims and accompanying drawing and obtain.

Below by drawings and Examples, technical scheme of the present invention is described in further detail.

Accompanying drawing explanation

Accompanying drawing is used to provide a further understanding of the present invention, and forms a part for instructions, together with embodiments of the present invention for explaining the present invention, is not construed as limiting the invention.In the accompanying drawings:

Fig. 1 is the process flow diagram of advertisement video detection method in the specific embodiment of the invention.

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described, should be appreciated that preferred embodiment described herein is only for instruction and explanation of the present invention, is not intended to limit the present invention.

Fig. 1 is the process flow diagram of advertisement video detection method in the specific embodiment of the invention.As shown in Figure 1, the flow process of advertisement video detection method, specifically comprises the following steps:

Step 101, the extraction advertisement video of 1000 and the non-advertisement video of 1000, set up video training set.

Step 102, use decoding tool are decoded to video, obtain video image set and audio-frequency information, are preserved by described audio-frequency information with array form;

Step 103, carry out piecemeal to all video images, be divided into M × N block, wherein M is columns, and N is line number;

Spatially color histogram is extracted at hsv color to each image block of image, be wherein 1 interval luminance quantization, amount of tones is turned to 16 intervals, color saturation is quantified as 8 intervals, each two field picture obtains M × N number of color histogram having 16 × 8 × 1=128 bin, and is normalized the color histogram obtained;

In order to reduce the complexity of compute histograms, carried out interlacing every column scan to image, image size becomes 1/4 of original image, reduces computation complexity;

Adopt the similar value of the color histogram of formulae discovery two adjacent video images below:

$S(a,b)=\stackrel{M}{\mathrm{\&Sigma;}}\stackrel{N}{\mathrm{\&Sigma;}}\stackrel{128}{\mathrm{\&Sigma;}}{w}_{\mathrm{pq}}\min (H_a(p,q,i),H_b(p,q,i)),$

Wherein the weight of the capable block of q is arranged, H for being in p _a(p, q, i) is the value of i-th bin of the color histogram of p × q block of a frame, a and b is the sequence number of video frame image, and min function asks for value less in two values.

Input the interframe similar value sequence S={s of whole video sequence ₁, s ₂..., s _n, preset high threshold T _h, preset Low threshold T _l, preset gradual shot length lowest threshold T _gradual, wherein T _gradual=10;

According to each interframe similarity s _ijudge:

If s _i< T _land gradual change mark is not set, judge that camera lens there occurs sudden change, the boundary information of output mutation camera lens;

If s _i< T _hbut be provided with gradual change mark, then gradual shot length Length adds 1;

If s _i>=T _lbut s _i< T _hbut gradual change mark is not set, then gradual change mark is set, record present image frame position alternatively camera lens start border, and start to count gradual shot length Length;

If s _i>=T _hand be provided with gradual change mark, then check gradual shot length, if Length > is T _gradualthen judge to there occurs gradual change, export the boundary information of gradual shot, otherwise judge it is not gradual change, cancel gradual change mark, and gradual shot length Length again zero setting;

If s _i>=T _hand gradual change mark is not set, then judge shot change not to occur;

Output lens border sequence, completes the segmentation by camera lens of video.

Step 104, extract the audio frequency characteristics of each camera lens, specifically comprise: zero-crossing rate, short-time energy, DFT coefficient, spectral centroid and Mel frequency cepstral coefficient.

Step 105, all described audio frequency characteristics are carried out support vector machine training as input amendment, the audio frequency characteristics of advertisement video camera lens is designated 1, and the audio frequency characteristics of non-advertisement video camera lens is designated-1;

By nonlinear transformation, the input space is transformed to a higher dimensional space, in described higher dimensional space, ask for optimum linearity classifying face;

Formula is below adopted to obtain described optimum linearity classifying face:

$S\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{w}_i\mathrm{\&phi;}\left(x_i\right)\mathrm{\&phi;}\left(X\right)+b,$

Wherein, φ () is a nonlinear mapping function, X={x _i∈ R ^d} _i=1...nand y _i={-1,1} represents corresponding training set and corresponding class mark respectively, and b is side-play amount, and any one corresponds to a nonzero value w _isample x _ifor support vector, kernel function adopts gaussian kernel function:

$K(x,y)=\mathrm{\&phi;}\left(x\right)\&CenterDot;\mathrm{\&phi;}\left(y\right)=\exp \left(\frac{-{\left|\right|x-y\left|\right|}^2}{{\mathrm{\&delta;}}^2}\right).$

Step 106, input video to be detected, use method described in step 102 to step 104 to extract the audio frequency characteristics being no less than of video to be detected.

Step 107, input tape detect the audio feature vector x of video, bring formula into

Wherein, θ () is an indicator function.During L (x)=+ 1, video lens to be detected is judged as advertisement camera lens, otherwise is judged as non-advertisement video camera lens.

If the advertisement camera lens of step 108 5 is continuous print, merged into an advertisement video section;

If the advertisement camera lens being less than 5 is continuous print, but the distance of its nearest advertisement video section is less than 3 camera lenses, be then incorporated to described nearest advertisement video section;

If the camera lens number between adjacent advertisement video section is less than 3, then described adjacent advertisement video section is merged into an advertisement video section;

Repeat above step, until do not have new advertisement video section to occur, if the advertisement video section in video to be detected is greater than preset value, then described video to be detected is judged as advertisement video, output detections result.

Last it is noted that the foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, although with reference to previous embodiment to invention has been detailed description, for a person skilled in the art, it still can be modified to the technical scheme described in foregoing embodiments, or carries out equivalent replacement to wherein portion of techniques feature.Within the spirit and principles in the present invention all, any amendment done, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (2)

1. an advertisement video detection method, is characterized in that, comprises the following steps:

A, extraction are no less than the advertisement video of and are no less than the non-advertisement video of, set up video training set, extract the audio frequency characteristics being no less than of each video, the audio frequency characteristics of advertisement video camera lens is designated 1, the audio frequency characteristics of non-advertisement video camera lens is designated-1, by nonlinear transformation, the input space is transformed to a higher dimensional space, in described higher dimensional space, ask for optimum linearity classifying face, adopt formula below obtain described optimum linearity classifying face, wherein, φ () is a nonlinear mapping function, X={x _i∈ R ^d} _i=1...nand y _i={-1,1} represents corresponding training set and corresponding class mark respectively, and b is side-play amount, and any one corresponds to a nonzero value w _isample x _ifor support vector, kernel function adopts gaussian kernel function:

$K(x,y)=\mathrm{\&phi;}\left(x\right)\&CenterDot;\mathrm{\&phi;}\left(y\right)=\exp \left(\frac{-{\left|\right|x-y\left|\right|}^2}{{\mathrm{\&delta;}}^2}\right);$

B, input video to be detected, extract the audio frequency characteristics being no less than of video to be detected;

C, input tape detect the audio feature vector x of video, bring formula into $S\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{w}_i\mathrm{\&phi;}\left(x_i\right)\mathrm{\&phi;}\left(X\right)+b,$ Export result of calculation:

$L\left(x\right)=\mathrm{\&theta;}\left(S\left(x\right)\right)=\left\{\begin{array}{ccc}+1& \mathrm{if}& S\left(x\right)>0\\ -1& \mathrm{if}& S\left(x\right)<0\end{array}\right.,$

Wherein, θ () is an indicator function, during L (x)=+ 1, video lens to be detected is judged as advertisement camera lens, otherwise is judged as non-advertisement video camera lens;

If D is continuous print more than the advertisement camera lens of default high level, merged into an advertisement video section; If the advertisement camera lens being less than default high level is continuous print, but the distance of its nearest advertisement video section is less than predetermined low level, be then incorporated to described nearest advertisement video section; If the camera lens number between adjacent advertisement video section is lower than predetermined low level, then described adjacent advertisement video section is merged into an advertisement video section; Repeat above step, until do not have new advertisement video section to occur, if the advertisement video section in video to be detected is greater than preset value, then described video to be detected is judged as advertisement video, output detections result;

In wherein said steps A and step B, extract the audio frequency characteristics being no less than of video, further comprising the steps:

Use decoding tool to decode to video, obtain video image set and audio-frequency information, described audio-frequency information is preserved with array form;

Carry out piecemeal to video image, be divided into M × N block, wherein M is columns, and N is line number;

Spatially color histogram is extracted at hsv color to each image block of image, be wherein 1 interval luminance quantization, amount of tones is turned to 16 intervals, color saturation is quantified as 8 intervals, each two field picture obtains M × N number of color histogram having 16 × 8 × 1=128 bin, and is normalized the color histogram obtained;

Adopt the similar value of the color histogram of formulae discovery two adjacent video images below, the interframe similar value as video:

$S(a,b)=\underset{p=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{q=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{128}{\mathrm{\&Sigma;}}}{w}_{\mathrm{pq}}\min (H_a(p,q,i),H_b(p,q,i)),$

Wherein w _pqthe weight of the capable block of q is arranged, H for being in p _a(p, q, i) is the value of i-th bin of the color histogram of p × q block of a frame, a and b is the sequence number of video frame image, and min function asks for value less in two values,

Input the interframe similar value sequence S={s of whole video sequence ₁, s ₂..., s _n, preset high threshold T _h, preset Low threshold T _l, preset gradual shot length lowest threshold T _gradual, wherein T _gradual=10,

According to each interframe similarity s _ijudge:

If s _i< T _land gradual change mark is not set, judge that camera lens there occurs sudden change, the boundary information of output mutation camera lens,

If s _i< T _hbut be provided with gradual change mark, then gradual shot length Length adds 1,

If s _i>=T _lbut s _i< T _hbut gradual change mark is not set, then gradual change mark is set, record present image frame position alternatively camera lens start border, and start to count gradual shot length Length,

If s _i>=T _hand be provided with gradual change mark, then check gradual shot length, if Length > is T _gradualthen judge to there occurs gradual change, export the boundary information of gradual shot, otherwise judge it is not gradual change, cancel gradual change mark, and gradual shot length Length again zero setting,

If s _i>=T _hand gradual change mark is not set, then judge shot change not to occur,

Output lens border sequence;

Extract the audio frequency characteristics of each camera lens.

2. a kind of advertisement video detection method according to claim 1, it is characterized in that, the audio frequency characteristics of described each camera lens, specifically comprises: zero-crossing rate, short-time energy, DFT coefficient, spectral centroid and Mel frequency cepstral coefficient.