CN103530598A - Station logo identification method and system - Google Patents

Abstract

The invention discloses a station logo identification method and a system. The method comprises steps that similarity of each positive sample and negative sample is calculated via a characteristic angle point probability distribution matrix, a similarity set including similarity of the positive samples and the negative samples is acquired, and a similarity threshold value of a station logo needed to be detected is calculated via the similarity set; similarity to be detected of the station logo which is needed to be detected and is included by an image to be detected is calculated via the characteristic angle point probability distribution matrix, and whether the image to be detected includes the station logo needed to be detected is judged according to size relation between similarity to be detected and the similarity threshold value. According to the station logo identification method and the system, precision of station logo identification is enhanced, identification time is shortened, and identification efficiency is enhanced so that effective technical support is provided for multimedia technology video automatic searching, recording, analyzing and retrieval.

Description

A kind of TV station symbol recognition method and system

Technical field

The present invention relates to TV station symbol recognition field, relate in particular to a kind of TV station symbol recognition method and system.

Background technology

The station symbol of TV station is the important symbol of distinguishing TV station, the important messages such as the platform name that station symbol has comprised TV station, program orientation, utilizing Computer Image Processing recognition technology automatically to identify station symbol becomes study hotspot in recent years, and it can effectively carry out program monitoring, video content analysis and retrieval in the daily use of televisor, user watches custom analysis etc.

The TV station symbol recognition method existing at present mainly contains: 1, utilize Multi Frame Difference method to obtain station symbol, by template matches, identify; 2, based on color histogram or shape, identify etc.But these methods all exist mostly to Similar color and transparent station symbol situation not easy to identify, due to the interference of background and noise, discrimination is lower simultaneously.

Therefore, prior art has yet to be improved and developed.

Summary of the invention

In view of above-mentioned the deficiencies in the prior art, the object of the present invention is to provide a kind of TV station symbol recognition method and system, be intended to solve existing TV station symbol recognition technology and exist Similar color and transparent station symbol is not easy to identify, discrimination is lower problem.

Technical scheme of the present invention is as follows:

A method, wherein, comprises step:

A, from several original images, extract the positive sample of required monitor station target and negative sample, described positive sample is for containing required monitor station target area image, and described negative sample is not for containing required monitor station target area image;

B, the positive sample extracting is carried out to feature Corner Detection, obtain the characteristic angle point set of positive sample, and by the set of described characteristic angle point, align in sample each pixel and occur that the frequency of feature angle point calculates, obtain required monitor station target feature angle point probability distribution matrix;

C, by described feature angle point probability distribution matrix, calculate the similarity of each positive sample, negative sample, obtain the similarity collection that comprises all positive samples, negative sample similarity, by described similarity collection, calculate required monitor station target similarity threshold;

D, by described feature angle point probability distribution matrix, calculate image to be detected and comprise required monitor station target similarity to be detected, judge whether similarity to be detected is greater than similarity threshold, when being, judge in image to be detected and contain required detection station symbol, when no, judge and in image to be detected, do not contain required detection station symbol.

Described TV station symbol recognition method, wherein, in described steps A, for identical station symbol, the relative position of each positive sample in corresponding original image is identical, and positive sample is 1:1.5～1:3.5 with the quantity ratio of negative sample.

Described TV station symbol recognition method, wherein, in described step B, the process of obtaining the characteristic angle point set of positive sample comprises:

B1, calculate the directional derivative of positive sample, save as respectively array I _xwith array I _y, I _xfor the directional derivative of x direction, I _ydirectional derivative for y direction;

B2, to utilize Gauss's template be that in positive sample, each pixel calculates local autocorrelation matrix M, wherein, $M=G\left(\tilde{s}\right)\⊗\left[\begin{array}{cc}{I^2}_x& I_x{I}_y\\ I_x{I}_y& {I^2}_y\end{array}\right],$

for Gauss's template;

B3, by M, calculate the angle point moment matrix I of each pixel, wherein, I=det (M)-ktr ²(M), wherein det is determinant of a matrix, and tr is matrix trace, and k is the constant in [0.04,0.06] scope;

Any point in B4, judgement angle point moment matrix I, the element value that whether simultaneously meets this point is greater than a threshold value, and is the local maximum in this field, when meeting, judges the feature angle point that this point is positive sample simultaneously.

Described TV station symbol recognition method, wherein, in described step B, the process of obtaining required monitor station target feature angle point probability distribution matrix specifically comprises:

B5, calculate each pixel (x in all positive samples, y) there is the frequency n (x, y) of feature angle point in position, as n (x, y) be less than predetermined value with the ratio of positive sample total, judge that corresponding pixel (x, y) is not feature angle point, and by this number of times (x, y) value makes zero, otherwise be judged to be feature angle point, and retain the value of this frequency n (x, y);

B6, to occurring on each pixel (x, y) position that the frequency P (x, y) of feature angle point is normalized operation and obtains required monitor station target feature angle point probability distribution matrix Mp (x, y),

the frequency P (x, y) that occurs feature angle point on each pixel (x, y) position is that the frequency n (x, y) of feature angle point and the ratio of positive sample total appear in current pixel point (x, y).

Described TV station symbol recognition method, wherein, described step C specifically comprises:

C1, set in advance a minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks;

C2, each positive sample and negative sample are carried out to feature Corner Detection, in positive sample or negative sample being detected at arbitrary pixel (x, while y) there is feature angle point, described positive sample or negative sample are at the characteristic angle dot information expression formula S of arbitrary pixel (x, y) _i(x, y)=1, otherwise S _i(x, y)=0, to obtain all positive samples, negative sample characteristic angle point set S={S ₀, S ₁, S ₂..., S _i... S _n, S wherein _ifor the determinant of w * h, N=NumSamples+NumNegative-1, w, h are the wide and high of positive sample and negative sample, and NumSamples is sample total, and NumNegative is negative sample total amount;

C3, by feature angle point probability distribution matrix, obtain the similarity ε of each positive sample and negative sample _i, thereby obtain all positive samples and negative sample feature set ε={ ε ₀, ε ₁, ε ₂..., ε _i... ε _n, ε _ithe similarity that represents i sample in all positive samples and negative sample;

C4, according to described minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks, the station symbol of required detection is trained, obtain the similarity threshold of the station symbol of required detection.

Described TV station symbol recognition method, wherein, described step C4 specifically comprises:

C41, a default initial similarity threshold, reclassify all positive samples and negative sample, if ε _ibe greater than prima facies like bottom valve value, judge ε _icorresponding sample is positive sample, otherwise is judged to be negative sample;

C42, statistics are reclassifying in situation correct identification number Nr, wrong identification number Nw and are leaking identification number Nm to all positive samples, negative sample, according to described correct identification number Nr, wrong identification number Nw and leak identification number Nm and calculate recognition correct rate, identification error rate under initial similarity threshold condition, leak discrimination;

Whether C43, judgement, under initial similarity threshold condition, satisfy condition: recognition correct rate is greater than minimum recognition correct rate, and identification error rate is less than maximum identification error rate, and leakage discrimination is less than the maximum discrimination that leaks, when meeting, proceed to step C45, otherwise proceed to step C44;

C44, with the step-length of being scheduled to, initial similarity threshold is upgraded, and return to step C41 and reclassify;

C45, export current institute training station target similarity threshold.

Described TV station symbol recognition method, wherein, described step D specifically comprises:

D1, travel through feature angle point probability distribution matrix Mp and the similarity threshold T of the station symbol of all training, and from image to be detected, extract station symbol region according to the positive sample position information of current station symbol and positive sample size information;

D2, feature Corner Detection is carried out in station symbol region, the characteristic angle dot information expression formula S1 that obtains the characteristic angle dot information in station symbol region and obtain station symbol region;

The similarity ε of D3, the current station symbol of calculating station symbol district inclusion,

mP _i(x, y) is the value that the feature angle point probability distribution matrix of current station symbol is located at pixel (x, y), and S1 (x, y) is the value that the characteristic angle dot information expression formula S1 in station symbol region locates at pixel (x, y);

D4, according to the similarity threshold T of described similarity ε and current station symbol _kcompare, as ε>=T _ktime, judge in this image to be detected and comprise current station symbol, otherwise be judged to be not containing current station symbol.

A system, wherein, comprising:

Sample extraction module, for extract the positive sample of required monitor station target and negative sample from several original images, described positive sample is for containing required monitor station target area image, and described negative sample is not for containing required monitor station target area image;

Feature angle point probability distribution matrix acquisition module, for the positive sample extracting is carried out to feature Corner Detection, obtain the characteristic angle point set of positive sample, and by the set of described characteristic angle point, align in sample each pixel and occur that the frequency of feature angle point calculates, obtain required monitor station target feature angle point probability distribution matrix;

Similarity threshold acquisition module, be used for by described feature angle point probability distribution matrix, calculate the similarity of each positive sample, negative sample, obtain the similarity collection that comprises all positive samples, negative sample similarity, by described similarity collection, calculate required monitor station target similarity threshold;

Station symbol detection module, for calculating image to be detected by described feature angle point probability distribution matrix, comprise required monitor station target similarity to be detected, judge whether similarity to be detected is greater than similarity threshold, when being, judge in image to be detected and contain required detection station symbol, when no, judge and in image to be detected, do not contain required detection station symbol.

Described TV station symbol recognition system, wherein, described feature angle point probability distribution matrix acquisition module comprises:

Directional derivative computing unit, for calculating the directional derivative of positive sample, saves as respectively array I _xwith array I _y, I _xfor the directional derivative of x direction, I _ydirectional derivative for y direction;

Local autocorrelation matrix calculation modules, is that each pixel of positive sample calculates local autocorrelation matrix M for utilizing Gauss's template, wherein, $M=G\left(\tilde{s}\right)\⊗\left[\begin{array}{cc}{I^2}_x& I_x{I}_y\\ I_x{I}_y& {I^2}_y\end{array}\right],$ for Gauss's template;

Angle point moment matrix computing unit, for calculate the angle point moment matrix I of each pixel by M, wherein, I=det (M)-ktr ²(M), wherein det is determinant of a matrix, and tr is matrix trace, and k is the constant in [0.04,0.06] scope;

Characteristic angle dot information acquiring unit, for judging any point of angle point moment matrix I, the element value that whether simultaneously meets this point is greater than a threshold value, and is the local maximum in this field, when meeting, judges the feature angle point that this point is positive sample simultaneously.

Described TV station symbol recognition system, wherein, described feature angle point probability distribution matrix acquisition module also comprises:

, for calculating each pixel (x, y) position of all positive samples, there is the frequency n (x of feature angle point in stack statistic unit, y), when n (x, y) is less than predetermined value with the ratio of positive sample total, judge corresponding pixel (x, y) not feature angle point, and the value of this number of times (x, y) is made zero, otherwise be judged to be feature angle point, and retain the value of this frequency n (x, y);

Feature angle point probability distribution matrix acquiring unit, for obtaining required monitor station target feature angle point probability distribution matrix Mp (x, y) to occurring on each pixel (x, y) position that the frequency P (x, y) of feature angle point is normalized to operate,

the frequency that occurs feature angle point on each pixel (x, y) position is that the frequency n (x, y) of feature angle point and the ratio of positive sample total appear in current pixel point (x, y).

Described TV station symbol recognition system, wherein, described similarity threshold acquisition module comprises:

Set in advance unit, for setting in advance a minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks;

Characteristic angle point set acquiring unit, for each positive sample and negative sample are carried out to feature Corner Detection, in positive sample or negative sample being detected at arbitrary pixel (x, while y) there is feature angle point, described positive sample or negative sample are at the characteristic angle dot information expression formula S of arbitrary pixel (x, y) _i(x, y)=1, otherwise S _i(x, y)=0, to obtain all positive samples, negative sample characteristic angle point set S={S ₀, S ₁, S ₂..., S _i... S _n, S wherein _ifor the determinant of w * h, N=NumSamples+NumNegative-1, w, h are the wide and high of positive sample and negative sample, and NumSamples is sample total, and NumNegative is negative sample total amount;

Similarity acquiring unit, for obtaining the similarity ε of each positive sample and negative sample by feature angle point probability distribution matrix _i,

thereby obtain all positive samples and negative sample feature set ε={ ε ₀, ε ₁, ε ₂..., ε _i... ε _n, ε _ithe similarity that represents i sample in all positive samples and negative sample;

Similarity threshold acquiring unit, for according to described minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks, the station symbol of required detection being trained, obtains the similarity threshold of the station symbol of required detection.

Beneficial effect: the present invention is by obtaining a large amount of positive sample, negative sample, and sample is carried out to feature Corner Detection, obtain feature angle point probability distribution matrix and similarity threshold, by above-mentioned training process, the present invention can identify accurately and fast the station symbol comprising in image to be detected in complicated background, improved the accuracy of TV station symbol recognition, shortened the time of identification, improved recognition efficiency, thereby for the video automatic search of multimedia technology, include, analyze and retrieval provides effective technical support.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of a kind of TV station symbol recognition method of the present invention preferred embodiment.

Fig. 2 is the process particular flow sheet that obtains the characteristic angle point set of positive sample in TV station symbol recognition method of the present invention.

Fig. 3 is the process particular flow sheet that obtains required monitor station target feature angle point probability distribution matrix in TV station symbol recognition method of the present invention.

Fig. 4 to Fig. 7 is the positive sample image of central authorities' three cover station symbols.

Fig. 8 is the station symbol feature angle point probability distribution graph of central authorities' three cover station symbols.

Fig. 9 to Figure 12 is the positive sample image of Gansu satellite TV station symbol.

Figure 13 is the station symbol feature angle point probability distribution graph of Gansu satellite TV station symbol.

Figure 14 is the process flow diagram of similarity threshold acquisition process in TV station symbol recognition method of the present invention.

Figure 15 obtains the particular flow sheet of Best similarity degree threshold process in TV station symbol recognition method of the present invention.

Figure 16 treats the process flow diagram that detected image is carried out TV station symbol recognition in TV station symbol recognition method of the present invention.

Figure 17 is the structured flowchart of TV station symbol recognition system of the present invention preferred embodiment.

Figure 18 is the concrete structure block diagram of feature angle point probability distribution matrix acquisition module in system shown in Figure 17.

Figure 19 is the concrete structure block diagram of similarity threshold acquisition module in system shown in Figure 17.

Figure 20 is the concrete structure block diagram of similarity threshold acquiring unit in Figure 19.

Figure 21 is the concrete structure block diagram of station symbol detection module in system shown in Figure 17.

Embodiment

The invention provides a kind of TV station symbol recognition method and system, for making object of the present invention, technical scheme and effect clearer, clear and definite, below the present invention is described in more detail.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not intended to limit the present invention.

Refer to Fig. 1, Fig. 1 is the process flow diagram of a kind of TV station symbol recognition method of the present invention preferred embodiment, and as shown in the figure, it comprises:

S101, from several original images, extract the positive sample of required monitor station target and negative sample, described positive sample is for containing required monitor station target area image, and described negative sample is not for containing required monitor station target area image;

S102, the positive sample extracting is carried out to feature Corner Detection, obtain the characteristic angle point set of positive sample, and by the set of described characteristic angle point, align in sample each pixel and occur that the frequency of feature angle point calculates, obtain required monitor station target feature angle point probability distribution matrix;

S103, by described feature angle point probability distribution matrix, calculate the similarity of each positive sample, negative sample, obtain the similarity collection that comprises all positive samples, negative sample similarity, by described similarity collection, calculate required monitor station target similarity threshold;

S104, by described feature angle point probability distribution matrix, calculate image to be detected and comprise required monitor station target similarity to be detected, judge whether similarity to be detected is greater than similarity threshold, when being, judge in image to be detected and contain required detection station symbol, when no, judge and in image to be detected, do not contain required detection station symbol.

Specifically, in step S101, if need identify station symbol, first need to from a large amount of original images, extract the sample of different station symbols, with to sample training.Sample is divided into again positive sample and negative sample, positive sample refers to and contains required monitor station target area image, negative sample refers to not containing required monitor station target area image, in general, the station symbol of TV station is all positioned at the upper left corner or the region, the upper right corner of original image, for identical station symbol, each station symbol is fixed at the relative position of original image, so when aligning sample and gather, except guaranteeing that the minimum boundary rectangle of station symbol is positive sample size, also should make the relative position of similar positive sample in original image be consistent, for negative sample, can in the non-station symbol region from original image, obtain, the size of negative sample need be with corresponding positive sample measure-alike.

In the present embodiment, the quantity of positive sample and negative sample has determined the quality of sample training result, and the quantity of positive sample and negative sample remains between 1:1.5～1:3.5 than preferably, and the preferably positive sample of the present invention is 1:2 with the quantity ratio of negative sample, under this ratio, training result is more reliable.Preferred, the quantity of positive sample remains on 2000(or 2000 left and right) be advisable, to improve the reliability of training result, and be unlikely to produce too much calculated amount.

In step S102, in positive sample, not only comprised station symbol, also the background that has comprised the more complicated that may occur, so need to align sample carries out feature Corner Detection accurately and obtains the characteristic angle point set in positive sample, then according to the frequency acquisition feature angle point probability distribution matrix that occurs feature angle point, specifically, as shown in Figure 2, the process of obtaining the characteristic angle point set in positive sample specifically comprises step:

S201, calculate the directional derivative of positive sample, save as respectively array I _xwith array I _y, I _xfor the directional derivative of x direction, I _ydirectional derivative for y direction; In the present invention, adopt Harris algorithm (a kind of Corner Detection Algorithm) to detect the feature angle point of positive sample, first need to use Prewitt operator (rim detection of first order differential operator) or Sobel operator (Sobel Operator, one of operator in image processing) calculate the directional derivative (being gradient) of positive sample x direction and y direction, and with array I _x, I _yform represent.

S202, to utilize Gauss's template be that in positive sample, each pixel calculates local autocorrelation matrix M, wherein, $M=G\left(\tilde{s}\right)\⊗\left[\begin{array}{cc}{I^2}_x& I_x{I}_y\\ I_x{I}_y& {I^2}_y\end{array}\right],$

for Gauss's template;

S203, by local autocorrelation matrix M, calculate the angle point moment matrix I of each pixel, wherein, I=det (M)-ktr ²(M), wherein det is determinant of a matrix, and tr is matrix trace, and k is the constant in [0.04,0.06] scope, and this constant is given tacit consent to, and in angle point moment matrix I, the element value of each pixel is corresponding to the interest value of positive sample respective point;

Any point in S204, judgement angle point moment matrix I, the element value that whether simultaneously meets this point is greater than a threshold value thresh(according to different algorithms, different), and be the local maximum in this field, when meeting simultaneously, judge that this point is as feature angle point, can find out like this each the feature angle point in positive sample.

The feature angle point that above-mentioned steps is obtained might not be all the feature angle point of station symbol, some may be the pseudo-station symbol feature angle point producing due to complex background, in order to eliminate pseudo-station symbol feature angle point, obtain required monitor station target feature angle point probability distribution matrix, also need to be further processed, as shown in Figure 3, the process of obtaining feature angle point probability distribution matrix specifically comprises step:

S301, calculate each pixel (x in all positive samples, y) there is the frequency n (x of feature angle point in position, y), as n (x, y) be less than predetermined value Th with the ratio of positive sample total, for example be less than 0.5, can think that the feature angle point that current pixel point (x, y) occurs is produced by background, rather than the feature angle point of station symbol, so judge that corresponding pixel (x, y) is not feature angle point, and by this frequency n (x, y) value makes zero, otherwise be judged to be feature angle point, and retain the value of this frequency n (x, y);

Be formulated as follows:

$n(x,y)=\left\{\begin{array}{cc}0& n(x,y)/\mathrm{NumSamplse}<\mathrm{Th}\\ n(x,y)& \mathrm{else}\end{array}\right.,$ Th is predetermined value, and NumSamples is positive sample total.

S302, to occurring on each pixel (x, y) position that the frequency P (x, y) of feature angle point is normalized operation and obtains required monitor station target feature angle point probability distribution matrix Mp (x, y),

the frequency P (x, y) that occurs feature angle point on each pixel (x, y) position is that the frequency n (x, y) of feature angle point and the ratio of positive sample total appear in current pixel point (x, y).

As shown in Fig. 4 to Fig. 8, Fig. 4 to Fig. 7 shows the image of several positive samples of central authorities' three cover station symbols, Fig. 8 shows the formed schematic diagram of feature angle point probability distribution matrix to obtaining after the positive sample training of Fig. 4 to Fig. 7, wherein, in Fig. 8, color is darker represents that this feature angle point more can describe the features of central authorities' three cover station symbols preferably.Equally, Fig. 9 to Figure 12 shows the image of the positive sample of several Gansu satellite TV's station symbol, and Figure 13 shows the formed schematic diagram of feature angle point probability distribution matrix to obtaining after the positive sample training of Fig. 9 to Figure 12.

Obtained after required monitor station target feature angle point probability distribution matrix, need the similarity that aligns sample and negative sample in conjunction with the feature angle point probability distribution matrix obtaining to calculate, to obtain the similarity threshold of station symbol, as shown in figure 14, this process specifically comprises:

S401, set in advance a minimum recognition correct rate (minRR), maximum identification error rate (maxFR) and the maximum discrimination (maxMR) that leaks;

Wherein, accuracy=(positive specimen discerning become positive number of samples+negative sample to be identified as negative sample number)/positive sample and negative sample total amount and;

Error rate=(negative sample is identified as positive number of samples)/negative sample total amount;

Leak discrimination=(positive specimen discerning becomes negative sample number)/positive sample total;

And minimum recognition correct rate (minRR), maximum identification error rate (maxFR) and the maximum discrimination (maxMR) that leaks are respectively the minimum accuracy that can accept in training process, maximum error rate and maximum leakage discrimination, for example arrange as follows: minRR=99%, maxFR=5%, maxMR=5%, to obtain best similarity threshold.

S402, each positive sample and negative sample are carried out to feature Corner Detection, in positive sample or negative sample being detected at arbitrary pixel (x, while y) there is feature angle point, described positive sample or negative sample (single sample, i sample) characteristic angle dot information expression formula S _i(x, y)=1, otherwise S _i(x, y)=0, to obtain all positive samples, negative sample characteristic angle point set S={S ₀, S ₁, S ₂..., S _i... S _n, S wherein _ifor the determinant of w * h, N=NumSamples+NumNegative-1, w(is wide), h(is high) be the size of positive sample and negative sample, NumSamples is sample total, NumNegative is negative sample total amount;

S403, by feature angle point probability distribution matrix, obtain the similarity ε of each positive sample and negative sample _i,

thereby obtain all positive samples and negative sample feature set ε={ ε ₀, ε ₁, ε ₂..., ε _i... ε _n, ε _ithe similarity that represents i sample in all positive samples and negative sample;

S404, according to described minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks, the station symbol of required detection is trained, obtain the similarity threshold of the station symbol of required detection.

As shown in figure 15, step S404 can specifically be refined as following steps:

S501, a default initial similarity threshold, reclassify all positive samples and negative sample, if the similarity ε of i sample _ibe greater than prima facies like bottom valve value, judge ε _icorresponding sample is positive sample, otherwise is judged to be negative sample;

S502, statistics are reclassifying in situation correct identification number Nr, wrong identification number Nw and are leaking identification number Nm to all positive samples, negative sample, according to described correct identification number Nr, wrong identification number Nw and leak identification number Nm and calculate recognition correct rate pr, identification error rate pw under initial similarity threshold condition, leak discrimination pm; Wherein, $\mathrm{pr}=\frac{\mathrm{Nr}}{N},\mathrm{pw}=\frac{\mathrm{Nw}}{\mathrm{NumNegative}},\mathrm{pm}=\frac{\mathrm{Nm}}{\mathrm{NumNegative}}.$

Whether S503, judgement, under initial similarity threshold condition, meet the condition of identification; This condition is: recognition correct rate is greater than minimum recognition correct rate (pr > minRR), identification error rate is less than maximum identification error rate (pw < maxFR), and leakage discrimination is less than the maximum discrimination (pm < maxMR) that leaks, when meeting, proceed to step S505, otherwise proceed to step S504;

S504, with the step-length of being scheduled to, initial similarity threshold is upgraded, for example, with step-length step=0.05, initial similarity threshold is upgraded: T=T+step, and return to step S501 and reclassify;

S505, judge that current initial similarity threshold, as similarity threshold accurately, exports current institute training station target similarity threshold.

According to said process, required detection station symbol (such as central San Tao, Gansu satellite TV, Shandong satellite TV etc.) is trained respectively, can obtain each required monitor station target similarity threshold.

After all feature angle point probability distribution matrixes and similarity threshold of station symbols after training that obtain, can treat in detected image and whether containing the station symbol of training to some extent, detect, as shown in figure 16, its testing process comprises:

S601, travel through feature angle point probability distribution matrix Mp and the similarity threshold T of the station symbol of all training, and from image to be detected, extract station symbol region according to the positive sample position information of current station symbol (the positive sample of current station symbol is at the relative position of original image) and positive sample size information (the positive size information of current station symbol); For example, while traveling through to k station symbol of training, the feature angle point probability distribution matrix and the similarity threshold that get are respectively Mp _k, T _k.

S602, feature Corner Detection is carried out in station symbol region, the characteristic angle dot information expression formula S1 that obtains the characteristic angle dot information in station symbol region and obtain station symbol region, this characteristic angle dot information expression formula S1 obtains according to the same method of step S402, that is, and and as the arbitrary pixel (x in station symbol region, while y) there is feature angle point, described characteristic angle dot information expression formula S1 is 1 at the value S1 of described pixel (x, y) (x, y), otherwise S1 (x, y) is 0;

S603, the required monitor station target of calculating station symbol district inclusion similarity ε,

wherein, MP _i(x, y) is the value that the feature angle point probability distribution matrix of current station symbol is located at pixel (x, y), and S1 (x, y) is the value that characteristic angle dot information expression formula S1 locates at pixel (x, y);

S604, according to described similarity ε and required monitor station target similarity threshold T _kcompare, as ε>=T _ktime, judge in this image to be detected and comprise k station symbol, otherwise be judged to be not containing k station symbol, and obtain positive sample position information and the positive sample size information of the station symbol of next training, obtain feature angle point probability distribution matrix Mp and the similarity threshold T of the station symbol of next training, and repeat judgement, to identify the station symbol being comprised in image to be detected.

Based on said method, the present invention also provides a kind of TV station symbol recognition system, and as shown in figure 17, it comprises:

Sample extraction module 100, for extract the positive sample of required monitor station target and negative sample from several original images, described positive sample is for containing required monitor station target area image, and described negative sample is not for containing required monitor station target area image;

Feature angle point probability distribution matrix acquisition module 200, for the positive sample extracting is carried out to feature Corner Detection, obtain the characteristic angle point set of positive sample, and by the set of described characteristic angle point, align in sample each pixel and occur that the frequency of feature angle point calculates, obtain required monitor station target feature angle point probability distribution matrix;

Similarity threshold acquisition module 300, be used for by described feature angle point probability distribution matrix, calculate the similarity of each positive sample, negative sample, obtain the similarity collection that comprises all positive samples, negative sample similarity, by described similarity collection, calculate required monitor station target similarity threshold;

Station symbol detection module 400, for calculating image to be detected by described feature angle point probability distribution matrix, comprise required monitor station target similarity to be detected, judge whether similarity to be detected is greater than similarity threshold, when being, judge in image to be detected and contain required detection station symbol, when no, judge and in image to be detected, do not contain required detection station symbol.

Further, as shown in figure 18, described feature angle point probability distribution matrix acquisition module 200 comprises:

Directional derivative computing unit 210, for calculating the directional derivative of positive sample, saves as respectively array I _xwith array I _y, I _xfor the directional derivative of x direction, I _ydirectional derivative for y direction;

Local autocorrelation matrix calculation modules 220, is that each pixel of positive sample calculates local autocorrelation matrix M for utilizing Gauss's template, wherein, $M=G\left(\tilde{s}\right)\&CircleTimes;\left[\begin{array}{cc}{I^2}_x& I_x{I}_y\\ I_x{I}_y& {I^2}_y\end{array}\right],$

for Gauss's template;

Angle point moment matrix computing unit 230, for calculate the angle point moment matrix I of each pixel by M, wherein, I=det (M)-ktr ²(M), wherein det is determinant of a matrix, and tr is matrix trace, and k is the constant in [0.04,0.06] scope;

Characteristic angle dot information acquiring unit 240, for judging any point of angle point moment matrix I, the element value that whether simultaneously meets this point is greater than a threshold value, and is the local maximum in this field, when meeting, judge the feature angle point that this point is positive sample simultaneously.

, for calculating each pixel (x, y) position of all positive samples, there is the frequency n (x of feature angle point in stack statistic unit 250, y), when n (x, y) is less than predetermined value with the ratio of positive sample total, judge corresponding pixel (x, y) not feature angle point, and the value of this number of times (x, y) is made zero, otherwise be judged to be feature angle point, and retain the value of this frequency n (x, y);

Feature angle point probability distribution matrix acquiring unit 260, for obtaining required monitor station target feature angle point probability distribution matrix Mp (x, y) to occurring on each pixel (x, y) position that the frequency of feature angle point is normalized to operate,

the frequency that occurs feature angle point on each pixel (x, y) position is that the frequency n (x, y) of feature angle point and the ratio of positive sample total appear in current pixel point (x, y).

Further, as shown in figure 19, described similarity threshold acquisition module 300 comprises:

Set in advance unit 310, for setting in advance a minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks;

Characteristic angle point set acquiring unit 320, for each positive sample and negative sample are carried out to feature Corner Detection, in positive sample or negative sample being detected at arbitrary pixel (x, while y) there is feature angle point, described positive sample or negative sample are at the characteristic angle dot information expression formula S of arbitrary pixel (x, y) _i(x, y)=1, otherwise S _i(x, y)=0, to obtain all positive samples, negative sample characteristic angle point set S={S ₀, S ₁, S ₂..., S _i... S _n, S wherein _ifor the determinant of w * h, N=NumSamples+NumNegative-1, w, h are the wide and high of positive sample and negative sample, and NumSamples is sample total, and NumNegative is negative sample total amount;

Similarity acquiring unit 330, for obtaining the similarity ε of each positive sample and negative sample by feature angle point probability distribution matrix _i,

thereby obtain all positive samples and negative sample feature set ε={ ε ₀, ε ₁, ε ₂..., ε _i... ε _n, ε _ithe similarity that represents i sample in all positive samples and negative sample;

Similarity threshold acquiring unit 340, for according to described minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks, the station symbol of required detection being trained, obtains the similarity threshold of the station symbol of required detection.

Further, as shown in figure 20, described similarity threshold acquiring unit 340 comprises:

The default subelement 341 of initial similarity threshold, for a default initial similarity threshold, reclassifies all positive samples and negative sample, if ε _ibe greater than prima facies like bottom valve value, judge ε _icorresponding sample is positive sample, otherwise is judged to be negative sample;

Statistics subelement 342, for adding up all positive samples, negative sample being reclassified to correct number Nr, wrong identification number Nw and the leakage identification number Nm of identifying in situation, according to described correct identification number Nr, wrong identification number Nw and leakage identification number Nm, calculate recognition correct rate, identification error rate, the leakage discrimination under initial similarity threshold condition;

Judgment sub-unit 343, for judging under initial similarity threshold condition, whether satisfy condition: recognition correct rate is greater than minimum recognition correct rate, identification error rate is less than maximum identification error rate, and leakage discrimination is less than the maximum discrimination that leaks, when meeting, proceed to renewal subelement, otherwise proceed to output subelement;

Upgrade subelement 344, for the step-length to be scheduled to, initial similarity threshold is upgraded, and reclassify;

Output subelement 345, for exporting current institute training station target similarity threshold.

Further, as shown in figure 21, described station symbol detection module 400 specifically comprises:

Traversal unit 410 for traveling through feature angle point probability distribution matrix Mp and the similarity threshold T of the station symbol of all training, and extracts station symbol region according to the positive sample position information of current station symbol and positive sample size information from image to be detected;

Station symbol region detecting unit 420, for feature Corner Detection is carried out in station symbol region, the characteristic angle dot information expression formula S1 that obtains the characteristic angle dot information in station symbol region and obtain station symbol region;

Station symbol Regional Similarity computing unit 430, for calculating the similarity ε of the current station symbol of station symbol district inclusion,

mP _i(x, y) is the value that the feature angle point probability distribution matrix of current station symbol is located at pixel (x, y), and S1 (x, y) is the value that characteristic angle dot information expression formula S1 locates at pixel (x, y);

Recognition unit 440, for according to the similarity threshold T of described similarity ε and current station symbol _kcompare, as ε>=T _ktime, judge in this image to be detected and comprise current station symbol, otherwise be judged to be not containing current station symbol.

In sum, the present invention is by obtaining a large amount of positive sample, negative sample, and sample is carried out to feature Corner Detection, obtain feature angle point probability distribution matrix and similarity threshold, by above-mentioned training process, the present invention can identify accurately and fast the station symbol comprising in image to be detected in complicated background, improved the accuracy of TV station symbol recognition, shortened the time of identification, improved recognition efficiency, thereby for the video automatic search of multimedia technology, include, analyze and retrieval provides effective technical support.

Should be understood that, application of the present invention is not limited to above-mentioned giving an example, and for those of ordinary skills, can be improved according to the above description or convert, and all these improvement and conversion all should belong to the protection domain of claims of the present invention.

Claims (11)

1. a TV station symbol recognition method, is characterized in that, comprises step:

A, from several original images, extract the positive sample of required monitor station target and negative sample, described positive sample is for containing required monitor station target area image, and described negative sample is not for containing required monitor station target area image;

B, the positive sample extracting is carried out to feature Corner Detection, obtain the characteristic angle point set of positive sample, and by the set of described characteristic angle point, align in sample each pixel and occur that the frequency of feature angle point calculates, obtain required monitor station target feature angle point probability distribution matrix;

C, by described feature angle point probability distribution matrix, calculate the similarity of each positive sample, negative sample, obtain the similarity collection that comprises all positive samples, negative sample similarity, by described similarity collection, calculate required monitor station target similarity threshold;

D, by described feature angle point probability distribution matrix, calculate image to be detected and comprise required monitor station target similarity to be detected, judge whether similarity to be detected is greater than similarity threshold, when being, judge in image to be detected and contain required detection station symbol, when no, judge and in image to be detected, do not contain required detection station symbol.

2. TV station symbol recognition method according to claim 1, is characterized in that, in described steps A, for identical station symbol, the relative position of each positive sample in corresponding original image is identical, positive sample: compare for 1:1.5～1:3.5 with the quantity of negative sample.

3. TV station symbol recognition method according to claim 1, is characterized in that, in described step B, the process of obtaining the characteristic angle point set of positive sample comprises:

B1, calculate the directional derivative of positive sample, save as respectively array I _xwith array I _y, I _xfor the directional derivative of x direction, I _ydirectional derivative for y direction;

B2, to utilize Gauss's template be that in positive sample, each pixel calculates local autocorrelation matrix M, wherein, $M=G\left(\tilde{s}\right)\&CircleTimes;\left[\begin{array}{cc}{I^2}_x& I_x{I}_y\\ I_x{I}_y& {I^2}_y\end{array}\right],$

for Gauss's template;

B3, by M, calculate the angle point moment matrix I of each pixel, wherein, I=det (M)-ktr ²(M), wherein det is determinant of a matrix, and tr is matrix trace, and k is the constant in [0.04,0.06] scope;

Any point in B4, judgement angle point moment matrix I, the element value that whether simultaneously meets this point is greater than a threshold value, and is the local maximum in this field, when meeting, judges the feature angle point that this point is positive sample simultaneously.

4. TV station symbol recognition method according to claim 1, is characterized in that, in described step B, the process of obtaining required monitor station target feature angle point probability distribution matrix specifically comprises:

B5, calculate each pixel (x in all positive samples, y) there is the frequency n (x, y) of feature angle point in position, as n (x, y) be less than predetermined value with the ratio of positive sample total, judge that corresponding pixel (x, y) is not feature angle point, and by this number of times (x, y) value makes zero, otherwise be judged to be feature angle point, and retain the value of this frequency n (x, y);

B6, to occurring on each pixel (x, y) position that the frequency P (x, y) of feature angle point is normalized operation and obtains required monitor station target feature angle point probability distribution matrix Mp (x, y),

the frequency P (x, y) that occurs feature angle point on each pixel (x, y) position is that the frequency n (x, y) of feature angle point and the ratio of positive sample total appear in current pixel point (x, y).

5. TV station symbol recognition method according to claim 1, is characterized in that, described step C specifically comprises:

C1, set in advance a minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks;

C2, each positive sample and negative sample are carried out to feature Corner Detection, in positive sample or negative sample being detected at arbitrary pixel (x, while y) there is feature angle point, described positive sample or negative sample are at the characteristic angle dot information expression formula S of arbitrary pixel (x, y) _i(x, y)=1, otherwise S _i(x, y)=0, to obtain all positive samples, negative sample characteristic angle point set S={S ₀, S ₁, S ₂..., S _i... S _n, S wherein _ifor the determinant of w * h, N=NumSamples+NumNegative-1, w, h are the wide and high of positive sample and negative sample, and NumSamples is sample total, and NumNegative is negative sample total amount;

C3, by feature angle point probability distribution matrix, obtain the similarity ε of each positive sample and negative sample _i,

thereby obtain all positive samples and negative sample feature set ε={ ε ₀, ε ₁, ε ₂..., ε _i... ε _n, ε _ithe similarity that represents i sample in all positive samples and negative sample;

C4, according to described minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks, the station symbol of required detection is trained, obtain the similarity threshold of the station symbol of required detection.

6. TV station symbol recognition method according to claim 5, is characterized in that, described step C4 specifically comprises:

C41, a default initial similarity threshold, reclassify all positive samples and negative sample, if ε _ibe greater than prima facies like bottom valve value, judge ε _icorresponding sample is positive sample, otherwise is judged to be negative sample;

C42, statistics are reclassifying in situation correct identification number Nr, wrong identification number Nw and are leaking identification number Nm to all positive samples, negative sample, according to described correct identification number Nr, wrong identification number Nw and leak identification number Nm and calculate recognition correct rate, identification error rate under initial similarity threshold condition, leak discrimination;

Whether C43, judgement, under initial similarity threshold condition, satisfy condition: recognition correct rate is greater than minimum recognition correct rate, and identification error rate is less than maximum identification error rate, and leakage discrimination is less than the maximum discrimination that leaks, when meeting, proceed to step C45, otherwise proceed to step C44;

C44, with the step-length of being scheduled to, initial similarity threshold is upgraded, and return to step C41 and reclassify;

C45, export current institute training station target similarity threshold.

7. TV station symbol recognition method according to claim 1, is characterized in that, described step D specifically comprises:

D1, travel through feature angle point probability distribution matrix Mp and the similarity threshold T of the station symbol of all training, and from image to be detected, extract station symbol region according to the positive sample position information of current station symbol and positive sample size information;

D2, feature Corner Detection is carried out in station symbol region, the characteristic angle dot information expression formula S1 that obtains the characteristic angle dot information in station symbol region and obtain station symbol region;

The similarity ε of D3, the current station symbol of calculating station symbol district inclusion,

mP _i(x, y) is the value that the feature angle point probability distribution matrix of current station symbol is located at pixel (x, y), and S1 (x, y) is the value that the characteristic angle dot information expression formula S1 in station symbol region locates at pixel (x, y);

D4, according to the similarity threshold T of described similarity ε and current station symbol _kcompare, as ε>=T _ktime, judge in this image to be detected and comprise current station symbol, otherwise be judged to be not containing current station symbol.

8. a TV station symbol recognition system, is characterized in that, comprising:

Sample extraction module, for extract the positive sample of required monitor station target and negative sample from several original images, described positive sample is for containing required monitor station target area image, and described negative sample is not for containing required monitor station target area image;

Feature angle point probability distribution matrix acquisition module, for the positive sample extracting is carried out to feature Corner Detection, obtain the characteristic angle point set of positive sample, and by the set of described characteristic angle point, align in sample each pixel and occur that the frequency of feature angle point calculates, obtain required monitor station target feature angle point probability distribution matrix;

Similarity threshold acquisition module, be used for by described feature angle point probability distribution matrix, calculate the similarity of each positive sample, negative sample, obtain the similarity collection that comprises all positive samples, negative sample similarity, by described similarity collection, calculate required monitor station target similarity threshold;

Station symbol detection module, for calculating image to be detected by described feature angle point probability distribution matrix, comprise required monitor station target similarity to be detected, judge whether similarity to be detected is greater than similarity threshold, when being, judge in image to be detected and contain required detection station symbol, when no, judge and in image to be detected, do not contain required detection station symbol.

9. TV station symbol recognition system according to claim 8, is characterized in that, described feature angle point probability distribution matrix acquisition module comprises:

Directional derivative computing unit, for calculating the directional derivative of positive sample, saves as respectively array I _xwith array I _y, I _xfor the directional derivative of x direction, I _ydirectional derivative for y direction;

Local autocorrelation matrix calculation modules, is that each pixel of positive sample calculates local autocorrelation matrix M for utilizing Gauss's template, wherein, $M=G\left(\tilde{s}\right)\&CircleTimes;\left[\begin{array}{cc}{I^2}_x& I_x{I}_y\\ I_x{I}_y& {I^2}_y\end{array}\right],$

for Gauss's template;

Angle point moment matrix computing unit, for calculate the angle point moment matrix I of each pixel by M, wherein, I=det (M)-ktr ²(M), wherein det is determinant of a matrix, and tr is matrix trace, and k is the constant in [0.04,0.06] scope;

Characteristic angle dot information acquiring unit, for judging any point of angle point moment matrix I, the element value that whether simultaneously meets this point is greater than a threshold value, and is the local maximum in this field, when meeting, judges the feature angle point that this point is positive sample simultaneously.

10. TV station symbol recognition system according to claim 9, is characterized in that, described feature angle point probability distribution matrix acquisition module also comprises:

, for calculating each pixel (x, y) position of all positive samples, there is the frequency n (x of feature angle point in stack statistic unit, y), when n (x, y) is less than predetermined value with the ratio of positive sample total, judge corresponding pixel (x, y) not feature angle point, and the value of this number of times (x, y) is made zero, otherwise be judged to be feature angle point, and retain the value of this frequency n (x, y);

Feature angle point probability distribution matrix acquiring unit, for obtaining required monitor station target feature angle point probability distribution matrix Mp (x, y) to occurring on each pixel (x, y) position that the frequency P (x, y) of feature angle point is normalized to operate,

the frequency that occurs feature angle point on each pixel (x, y) position is that the frequency n (x, y) of feature angle point and the ratio of positive sample total appear in current pixel point (x, y).

11. TV station symbol recognition systems according to claim 8, is characterized in that, described similarity threshold acquisition module comprises:

Set in advance unit, for setting in advance a minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks;

Characteristic angle point set acquiring unit, for each positive sample and negative sample are carried out to feature Corner Detection, in positive sample or negative sample being detected at arbitrary pixel (x, while y) there is feature angle point, described positive sample or negative sample are at the characteristic angle dot information expression formula S of arbitrary pixel (x, y) _i(x, y)=1, otherwise S _i(x, y)=0, to obtain all positive samples, negative sample characteristic angle point set S={S ₀, S ₁, S ₂..., S _i... S _n, S wherein _ifor the determinant of w * h, N=NumSamples+NumNegative-1, w, h are the wide and high of positive sample and negative sample, and NumSamples is sample total, and NumNegative is negative sample total amount;

Similarity acquiring unit, for obtaining the similarity ε of each positive sample and negative sample by feature angle point probability distribution matrix _i,

thereby obtain all positive samples and negative sample feature set ε={ ε ₀, ε ₁, ε ₂..., ε _i... ε _n, ε _ithe similarity that represents i sample in all positive samples and negative sample;

Similarity threshold acquiring unit, for according to described minimum recognition correct rate, maximum identification error rate and the maximum discrimination that leaks, the station symbol of required detection being trained, obtains the similarity threshold of the station symbol of required detection.

Images