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

Station caption identification method and system

Technical Field

The invention relates to the field of station caption identification, in particular to a station caption identification method and a system.

Background

The station mark of a television station is an important mark for distinguishing the television station, the station mark comprises important information of the television station such as the station name, the program orientation and the like, and the automatic identification of the station mark by utilizing the computer image processing and identifying technology becomes a research hotspot in recent years, and the station mark can effectively perform program monitoring, video content analysis and retrieval, user watching habit analysis and the like in daily application of the television.

The existing station logo identification methods mainly comprise: 1. acquiring station captions by a multi-frame difference method, and identifying by template matching; 2. recognition based on color histograms or shapes, etc. However, most of the methods have the condition that similar colors and transparent station marks are difficult to identify, and meanwhile, due to the interference of background and noise, the identification rate is low.

Accordingly, the prior art is yet to be improved and developed.

Disclosure of Invention

In view of the above disadvantages of the prior art, an object of the present invention is to provide a station caption identification method and system, which aims to solve the problems of the prior art that similar colors and transparent station captions are not easily identified and the identification rate is low.

The technical scheme of the invention is as follows:

a station caption identification method comprises the following steps:

A. extracting a positive sample and a negative sample of the station caption to be detected from a plurality of original images, wherein the positive sample is a regional image containing the station caption to be detected, and the negative sample is a regional image not containing the station caption to be detected;

B. performing characteristic corner detection on the extracted positive sample to obtain a characteristic corner set of the positive sample, and calculating the frequency of the characteristic corner of each pixel point in the positive sample through the characteristic corner set to obtain a characteristic corner probability distribution matrix of the station logo to be detected;

C. calculating the similarity of each positive sample and each negative sample through the characteristic corner probability distribution matrix to obtain a similarity set containing the similarities of all the positive samples and the negative samples, and calculating the similarity threshold of the station logo to be detected through the similarity set;

D. and calculating the similarity to be detected of the image to be detected, including the station logo to be detected, by the characteristic angular point probability distribution matrix, judging whether the similarity to be detected is greater than a similarity threshold value, if so, judging that the station logo to be detected is contained in the image to be detected, and if not, judging that the station logo to be detected is not contained in the image to be detected.

In the station caption identifying method, in the step a, for the same station caption, the relative position of each positive sample in the corresponding original image is the same, and the number ratio of the positive samples to the negative samples is 1: 1.5-1: 3.5.

in the station caption identifying method, in the step B, the process of obtaining the feature corner set of the positive sample includes:

b1, calculating the directional derivatives of the positive samples, and respectively storing the directional derivatives as an array I_xAnd array I_y，I_xIs the directional derivative of the x direction, I_yThe directional derivative in the y-direction;

b2, calculating a local autocorrelation matrix M for each pixel point in the positive sample by using the gaussian template, wherein, <math> <mrow> <mi>M</mi> <mo>=</mo> <mi>G</mi> <mrow> <mo>(</mo> <mover> <mi>s</mi> <mo>~</mo> </mover> <mo>)</mo> </mrow> <mo>&CircleTimes;</mo> <mfenced open='[' close=']'> <mtable> <mtr> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>x</mi> </msub> </mtd> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>y</mi> </msub> </mtd> </mtr> </mtable> </mfenced> <mo>,</mo> </mrow> </math>

is a Gaussian template;

b3, calculating a corner point quantity matrix I of each pixel point by M, wherein I = det (M) -k-tr²(M) where det is the determinant of the matrix, tr is the trace of the matrix, and k is [0.04,0.06 ]]A constant within a range;

b4, judging whether any point in the angular point quantity matrix I simultaneously satisfies that the element value of the point is larger than a threshold value and is a local maximum value in the point field, and if so, judging that the point is the characteristic angular point of the positive sample.

In the station caption identifying method, in the step B, the process of obtaining the characteristic corner probability distribution matrix of the station caption to be detected specifically includes:

b5, calculating the times n (x, y) of occurrence of characteristic corner points at the position of each pixel point (x, y) in all the positive samples, when the proportion of n (x, y) to the total amount of the positive samples is smaller than a preset value, judging that the corresponding pixel point (x, y) is not a characteristic corner point, and returning the value of the times (x, y) to zero, otherwise, judging that the pixel point is a characteristic corner point, and keeping the value of the times n (x, y);

b6, carrying out normalization operation on the frequency P (x, y) of the characteristic corner point appearing at the position of each pixel point (x, y) to obtain a probability distribution matrix Mp (x, y) of the characteristic corner point of the station logo to be detected,

the frequency P (x, y) of the characteristic corner appearing at the position of each pixel point (x, y) is the number n (x) of times of the characteristic corner appearing at the current pixel point (x, y)Y) to the total number of positive samples.

The station logo identification method, wherein the step C specifically includes:

c1, presetting a minimum recognition accuracy, a maximum recognition error rate and a maximum missing recognition rate;

c2, detecting the characteristic corner of each positive sample and each negative sample, and when detecting that the characteristic corner exists at any pixel point (x, y) in the positive sample or the negative sample, then expressing the characteristic corner information expression S of the positive sample or the negative sample at any pixel point (x, y)_i(x, y) =1, otherwise S_i(x, y) =0 to obtain all positive sample, negative sample feature angular point set S = { S = { S =₀,S₁,S₂,......,S_i......S_NIn which S is_iIs a determinant of w × h, N = NumSamples + Numnegative-1, w and h are the widths and heights of the positive sample and the negative sample, NumSamples is the total amount of the samples, and NumNepositive is the total amount of the negative samples;

c3, obtaining the similarity epsilon of each positive sample and each negative sample through the characteristic corner probability distribution matrix_i，So as to obtain all the feature sets of positive and negative samples epsilon = { [ epsilon ]₀,ε₁,ε₂,......,ε_i......ε_N}，ε_iRepresenting the similarity of the ith sample in all the positive samples and the negative samples;

and C4, training the station caption to be detected according to the minimum recognition correct rate, the maximum recognition error rate and the maximum missing recognition rate to obtain the similarity threshold value of the station caption to be detected.

The station caption identification method, wherein the step C4 specifically includes:

c41, presetting an initial similarity threshold, reclassifying all positive and negative samples if epsilon_iIf the initial similarity is greater than the initial similarity threshold value, determining epsilon_iCorresponding sampleIf the sample is a positive sample, otherwise, the sample is judged to be a negative sample;

c42, counting the number of correct identifications Nr, the number of incorrect identifications Nw and the number of missed identifications Nm under the condition of reclassifying all the positive samples and the negative samples, and calculating the identification accuracy, the identification error rate and the recognition rate of missed identifications under the condition of the initial similarity threshold according to the number of correct identifications Nr, the number of incorrect identifications Nw and the number of missed identifications Nm;

c43, judging whether the conditions are met under the condition of the initial similarity threshold value: if the recognition accuracy is greater than the minimum recognition accuracy, the recognition error rate is less than the maximum recognition error rate, and the missing recognition rate is less than the maximum missing recognition rate, the step C45 is executed, otherwise, the step C44 is executed;

c44, updating the initial similarity threshold by a preset step size, and returning to the step C41 for reclassification;

and C45, outputting the similarity threshold of the currently trained station caption.

The station logo identification method, wherein the step D specifically includes:

d1, traversing the feature corner probability distribution matrix Mp and the similarity threshold T of all the trained station captions, and extracting a station caption area from the image to be detected according to the position information and the size information of the positive sample of the current station caption;

d2, detecting characteristic corner points of the station caption area, acquiring characteristic corner point information of the station caption area and acquiring a characteristic corner point information expression S1 of the station caption area;

d3, calculating the similarity epsilon of the current station caption contained in the station caption area,

MP_i(x, y) is the value of the characteristic corner probability distribution matrix of the current station caption at the pixel point (x, y), and S1(x, y) is the value of the characteristic corner information expression S1 of the station caption region at the pixel point (x, y);

d4, according to the similarity epsilon and the similarity threshold T of the current station caption_kComparing when epsilon is more than or equal to T_kAnd if not, judging that the image to be detected contains the current station caption, otherwise, judging that the image to be detected does not contain the current station caption.

A station caption identification system, comprising:

the sample extraction module is used for extracting a positive sample and a negative sample of the station caption to be detected from a plurality of original images, wherein the positive sample is a regional image containing the station caption to be detected, and the negative sample is a regional image not containing the station caption to be detected;

the characteristic corner probability distribution matrix acquisition module is used for detecting characteristic corners of the extracted positive sample, acquiring a characteristic corner set of the positive sample, calculating the frequency of the characteristic corners of each pixel point in the positive sample through the characteristic corner set, and acquiring a characteristic corner probability distribution matrix of the station logo to be detected;

the similarity threshold acquisition module is used for calculating the similarity of each positive sample and each negative sample through the characteristic corner probability distribution matrix to obtain a similarity set containing the similarities of all the positive samples and the negative samples, and calculating the similarity threshold of the station logo to be detected through the similarity set;

and the station logo detection module is used for calculating the similarity to be detected of the station logo to be detected including the required detection through the characteristic angular point probability distribution matrix, judging whether the similarity to be detected is greater than a similarity threshold value, judging that the station logo to be detected is included in the image to be detected if the similarity to be detected is greater than the similarity threshold value, and judging that the station logo to be detected is not included in the image to be detected if the similarity to be detected is not greater than the similarity threshold value.

The station caption identification system, wherein the characteristic corner probability distribution matrix acquisition module comprises:

a direction derivative calculation unit for calculating the direction derivatives of the positive samples, which are stored as an array I_xAnd array I_y，I_xIs the directional derivative of the x direction, I_yThe directional derivative in the y-direction;

a local autocorrelation matrix calculation unit for calculating a local autocorrelation matrix M for each pixel point in the positive sample using a Gaussian template, wherein, <math> <mrow> <mi>M</mi> <mo>=</mo> <mi>G</mi> <mrow> <mo>(</mo> <mover> <mi>s</mi> <mo>~</mo> </mover> <mo>)</mo> </mrow> <mo>&CircleTimes;</mo> <mfenced open='[' close=']'> <mtable> <mtr> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>x</mi> </msub> </mtd> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>y</mi> </msub> </mtd> </mtr> </mtable> </mfenced> <mo>,</mo> </mrow> </math> is a Gaussian template;

a corner matrix calculation unit for calculating a corner matrix I of each pixel point by M, wherein I = det (M) -k-tr²(M) where det is the determinant of the matrix, tr is the trace of the matrix, and k is [0.04,0.06 ]]A constant within a range;

and the characteristic angular point information acquisition unit is used for judging whether any point in the angular point quantity matrix I simultaneously satisfies the condition that the element value of the point is greater than a threshold value and is a local maximum value in the field of the point, and when the point simultaneously satisfies the element value, judging that the point is the characteristic angular point of the positive sample.

The station caption identification system, wherein the characteristic corner probability distribution matrix obtaining module further comprises:

the superposition statistical unit is used for calculating the times n (x, y) of occurrence of characteristic angular points at the position of each pixel point (x, y) in all the positive samples, when the proportion of the n (x, y) to the total amount of the positive samples is smaller than a preset value, the corresponding pixel point (x, y) is judged not to be the characteristic angular point, the value of the times (x, y) is returned to zero, otherwise, the pixel point is judged to be the characteristic angular point, and the value of the times n (x, y) is reserved;

a characteristic angular point probability distribution matrix obtaining unit, which is used for carrying out normalization operation on the frequency P (x, y) of the characteristic angular point appearing on the position of each pixel point (x, y) to obtain a characteristic angular point probability distribution matrix Mp (x, y) of the station logo to be detected,

the frequency of the characteristic corner appearing at the position of each pixel point (x, y) is the ratio of the number n (x, y) of the characteristic corners appearing at the current pixel point (x, y) to the total number of positive samples.

The station caption identifying system, wherein the similarity threshold obtaining module comprises:

the device comprises a presetting unit, a judging unit and a judging unit, wherein the presetting unit is used for presetting a minimum identification correct rate, a maximum identification error rate and a maximum missing identification rate;

a characteristic corner set obtaining unit, configured to perform characteristic corner detection on each of the positive sample and the negative sample, and when it is detected that a characteristic corner exists at any pixel point (x, y) in the positive sample or the negative sample, obtain a characteristic corner information expression S of the positive sample or the negative sample at any pixel point (x, y)_i(x, y) =1, otherwise S_i(x, y) =0 to obtain all positive sample, negative sample feature angular point set S = { S = { S =₀,S₁,S₂,......,S_i......S_NIn which S is_iIs wA determinant of x h, N = NumSamples + NumNegative-1, w, h are the width and height of the positive and negative samples, NumSamples is the total amount of samples, NumNegative is the total amount of negative samples;

a similarity obtaining unit for obtaining the similarity epsilon of each positive sample and each negative sample through the characteristic corner probability distribution matrix_i，

So as to obtain all the feature sets of positive and negative samples epsilon = { [ epsilon ]₀,ε₁,ε₂,......,ε_i......ε_N}，ε_iRepresenting the similarity of the ith sample in all the positive samples and the negative samples;

and the similarity threshold acquisition unit is used for training the station caption to be detected according to the minimum recognition correct rate, the maximum recognition error rate and the maximum missing recognition rate to obtain the similarity threshold of the station caption to be detected.

Has the advantages that: according to the method, a large number of positive samples and negative samples are obtained, the characteristic angular points of the samples are detected, and a characteristic angular point probability distribution matrix and a similarity threshold value are obtained.

Drawings

Fig. 1 is a flowchart illustrating a station caption identification method according to a preferred embodiment of the present invention.

Fig. 2 is a specific flowchart of a process of acquiring a feature corner set of a positive sample in the station logo identification method of the present invention.

Fig. 3 is a specific flowchart of a process for acquiring a feature corner probability distribution matrix of a station logo to be detected in the station logo identification method of the present invention.

Fig. 4-7 are positive sample images of the central three sets of station markers.

Fig. 8 is a table mark feature corner point probability distribution diagram of three sets of table marks in the center.

Fig. 9 to 12 are positive sample images of the gansuwei station caption.

Fig. 13 is a table mark feature corner probability distribution diagram of the table mark of the defend view in Gansu province.

Fig. 14 is a flowchart of a similarity threshold acquisition process in the station caption identifying method of the present invention.

Fig. 15 is a specific flowchart of a process of obtaining an optimal similarity threshold in the station caption identifying method according to the present invention.

Fig. 16 is a flowchart of station logo recognition performed on an image to be detected in the station logo recognition method according to the present invention.

Fig. 17 is a block diagram of a station logo recognition system according to a preferred embodiment of the present invention.

Fig. 18 is a specific structural block diagram of a feature corner probability distribution matrix acquisition module in the system shown in fig. 17.

Fig. 19 is a block diagram of a detailed structure of a similarity threshold acquisition module in the system shown in fig. 17.

Fig. 20 is a block diagram showing a specific structure of the similarity threshold acquisition unit in fig. 19.

Fig. 21 is a block diagram showing a detailed structure of a beacon detection module in the system shown in fig. 17.

Detailed Description

The invention provides a station caption identification method and a station caption identification system, which are further described in detail below in order to make the purpose, technical scheme and effect of the invention clearer and clearer. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

Referring to fig. 1, fig. 1 is a flowchart illustrating a station caption identification method according to a preferred embodiment of the present invention, which includes:

s101, extracting a positive sample and a negative sample of a station logo to be detected from a plurality of original images, wherein the positive sample is an area image containing the station logo to be detected, and the negative sample is an area image not containing the station logo to be detected;

s102, performing characteristic corner detection on the extracted positive sample to obtain a characteristic corner set of the positive sample, and calculating the frequency of the characteristic corner of each pixel point in the positive sample through the characteristic corner set to obtain a characteristic corner probability distribution matrix of the station logo to be detected;

s103, calculating the similarity of each positive sample and each negative sample through the characteristic corner probability distribution matrix to obtain a similarity set containing the similarities of all the positive samples and the negative samples, and calculating the similarity threshold of the station logo to be detected through the similarity set;

s104, calculating the similarity to be detected of the image to be detected, including the station logo to be detected, through the characteristic angular point probability distribution matrix, judging whether the similarity to be detected is larger than a similarity threshold value, if so, judging that the station logo to be detected is contained in the image to be detected, and if not, judging that the station logo to be detected is not contained in the image to be detected.

Specifically, in step S101, if the station caption needs to be identified, samples of different station captions need to be extracted from a large number of original images to train the samples. The sample is divided into a positive sample and a negative sample, the positive sample is an area image containing a station caption to be detected, the negative sample is an area image not containing the station caption to be detected, generally, the station captions of a television station are all located in the upper left corner or the upper right corner of the original image, for the same station caption, the relative position of each station caption in the original image is fixed, so when the positive sample is collected, the relative position of the same type of positive sample in the original image is kept consistent except that the minimum circumscribed rectangle of the station caption is the size of the positive sample, for the negative sample, the negative sample can be obtained from a non-station caption area in the original image, and the size of the negative sample is required to be the same as that of the corresponding positive sample.

In this embodiment, the number of positive samples and negative samples determines the quality of the sample training result, and the ratio of the number of positive samples to the number of negative samples is preferably kept between 1: 1.5-1: 3.5, the invention preferably has a 1:2 ratio of the number of positive samples to the number of negative samples, and the training result is more reliable in the ratio. More preferably, the number of positive samples is kept at about 2000 (or about 2000) to improve the reliability of the training results without generating excessive calculation amount.

In step S102, the positive sample includes not only the station logo but also a possibly complex background, so that accurate feature corner detection needs to be performed on the positive sample to obtain a feature corner set in the positive sample, and then a feature corner probability distribution matrix is obtained according to a frequency of the feature corner, specifically, as shown in fig. 2, the process of obtaining the feature corner set in the positive sample specifically includes the steps of:

s201, calculating the directional derivatives of the positive samples, and respectively storing the directional derivatives as an array I_xAnd array I_y，I_xIs the directional derivative of the x direction, I_yThe directional derivative in the y-direction; in the invention, a Harris algorithm (an angular point detection algorithm) is adopted to detect the characteristic angular point of the positive sample, firstly, a Prewitt operator (edge detection of a first-order differential operator) or a Sobel operator (one of Sobel operators in image processing) is used to calculate the direction derivatives (namely gradients) of the positive sample in the x direction and the y direction, and an array I is used_x、I_yIs expressed in terms of the form.

S202, calculating a local autocorrelation matrix M for each pixel point in the positive sample by utilizing a Gaussian template, wherein, <math> <mrow> <mi>M</mi> <mo>=</mo> <mi>G</mi> <mrow> <mo>(</mo> <mover> <mi>s</mi> <mo>~</mo> </mover> <mo>)</mo> </mrow> <mo>&CircleTimes;</mo> <mfenced open='[' close=']'> <mtable> <mtr> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>x</mi> </msub> </mtd> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>y</mi> </msub> </mtd> </mtr> </mtable> </mfenced> <mo>,</mo> </mrow> </math>

is a Gaussian template;

s203, calculating a corner point quantity matrix I of each pixel point through a local autocorrelation matrix M, wherein I = det (M) -k-tr²(M) where det is the determinant of the matrix, tr is the trace of the matrix, and k is [0.04,0.06 ]]A constant in the range, wherein the constant is default, and the element value of each pixel point in the angular point quantity matrix I corresponds to the interest value of the corresponding point of the positive sample;

s204, judging whether any point in the angular point quantity matrix I simultaneously satisfies that the element value of the point is larger than a threshold thresh (different according to different algorithms) and is a local maximum value in the field of the point, and judging the point as a characteristic angular point when the point simultaneously satisfies the threshold thresh, so that each characteristic angular point in the positive sample can be found out.

The feature corners obtained in the above steps are not necessarily all the feature corners of the station caption, some may be pseudo station caption feature corners generated by a complex background, and further processing is required to obtain a feature corner probability distribution matrix of the station caption to be detected in order to eliminate the pseudo station caption feature corners, as shown in fig. 3, the process of obtaining the feature corner probability distribution matrix specifically includes the steps of:

s301, calculating the number n (x, y) of times of occurrence of a characteristic corner at the position of each pixel point (x, y) in all positive samples, and when the proportion of n (x, y) to the total number of the positive samples is smaller than a preset value Th, for example, smaller than 0.5, it can be considered that the characteristic corner occurring at the current pixel point (x, y) is generated by a background and is not the characteristic corner of a station logo, so that the corresponding pixel point (x, y) is judged not to be the characteristic corner, and the value of the number n (x, y) is returned to zero, otherwise, the pixel point is judged to be the characteristic corner, and the value of the number n (x, y) is reserved;

is 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 a predetermined value and NumSamples is the total number of positive samples.

S302, normalizing the frequency P (x, y) of the characteristic angular point appearing at the position of each pixel point (x, y) to obtain a probability distribution matrix Mp (x, y) of the characteristic angular point of the station logo to be detected,

the frequency P (x, y) of the characteristic corner appearing at the position of each pixel point (x, y) is the ratio of the number n (x, y) of the characteristic corners appearing at the current pixel point (x, y) to the total number of positive samples.

As shown in fig. 4 to 8, fig. 4 to 7 show images of several positive samples of the three sets of central station markers, and fig. 8 shows schematic diagrams formed by feature corner point probability distribution matrices obtained after training the positive samples of fig. 4 to 7, wherein in fig. 8, the darker the color represents that the feature corner point can better describe the features of the three sets of central station markers. Similarly, fig. 9 to 12 show images of several gansuwei satellite station calibration samples, and fig. 13 shows schematic diagrams formed by feature corner probability distribution matrices obtained after training the positive samples of fig. 9 to 12.

After the feature corner probability distribution matrix of the station caption to be detected is obtained, the similarity between the positive sample and the negative sample needs to be calculated by combining the obtained feature corner probability distribution matrix to obtain a similarity threshold of the station caption, as shown in fig. 14, the process specifically includes:

s401, presetting a minimum recognition accuracy (minRR), a maximum recognition error rate (maxFR) and a maximum missing recognition rate (maxMR);

wherein, the accuracy rate = (positive sample is identified as positive sample number + negative sample is identified as negative sample number)/positive sample and negative sample total sum;

error rate = (negative samples are identified as number of positive samples)/total number of negative samples;

leak recognition rate = (positive samples are recognized as negative number of samples)/total number of positive samples;

the minimum recognition accuracy (minRR), the maximum recognition error rate (maxFR), and the maximum missed recognition rate (maxMR) are respectively the minimum accuracy, the maximum error rate, and the maximum missed recognition rate that can be accepted during the training process, and are set as follows: minRR =99%, maxFR =5%, maxMR =5%, to obtain the optimal similarity threshold.

S402, detecting characteristic angular points of each positive sample and each negative sample, and when detecting that the characteristic angular points exist in any pixel point (x, y) in the positive sample or the negative sample, then expressing the characteristic angular point information expression S of the positive sample or the negative sample (single sample, ith sample)_i(x, y) =1, otherwise S_i(x, y) =0 to obtain all positive sample, negative sample feature angular point set S = { S = { S =₀,S₁,S₂,......,S_i......S_NIn which S is_iIs a w × h determinant, N = NumSamples + NumNegative-1, w (wide) and h (high) are the sizes of the positive sample and the negative sample, NumSamples is the total amount of the samples, and NumNegative is the total amount of the negative sample;

s403, obtaining the similarity epsilon of each positive sample and each negative sample through the feature corner probability distribution matrix_i，

So as to obtain all the feature sets of positive and negative samples epsilon = { [ epsilon ]₀,ε₁,ε₂,......,ε_i......ε_N}，ε_iRepresenting the similarity of the ith sample in all the positive samples and the negative samples;

s404, training the station caption to be detected according to the minimum recognition correct rate, the maximum recognition error rate and the maximum missing recognition rate to obtain the similarity threshold value of the station caption to be detected.

As shown in fig. 15, step S404 can be specifically detailed as the following steps:

s501, presetting an initial similarity threshold, reclassifying all positive samples and negative samples, and if the similarity epsilon of the ith sample is epsilon_iIf the initial similarity is greater than the initial similarity threshold value, determining epsilon_iThe corresponding sample is a positive sample, otherwise, the corresponding sample is judged as a negative sample;

s502, counting the number of correct identifications Nr, the number of wrong identifications Nw and the number of missed identifications Nm under the condition of reclassifying all positive samples and negative samples, and calculating the identification correct rate pr, the identification error rate pw and the missed identification rate pm under the condition of an initial similarity threshold according to the number of correct identifications Nr, the number of wrong identifications Nw and the number of missed identifications Nm; wherein, $\mathrm{pr}=\frac{\mathrm{Nr}}{N},\mathrm{pw}=\frac{\mathrm{Nw}}{\mathrm{NumNegative}},\mathrm{pm}=\frac{\mathrm{Nm}}{\mathrm{NumNegative}}.$

s503, judging whether the initial similarity threshold meets the identification condition; the conditions are as follows: if the recognition accuracy is greater than the minimum recognition accuracy (pr is greater than minRR), the recognition error rate is less than the maximum recognition error rate (pw is less than maxFR), and the missing recognition rate is less than the maximum missing recognition rate (pm is less than maxMR), the step S505 is executed, otherwise, the step S504 is executed;

s504, updating the initial similarity threshold by a predetermined step size, for example, updating the initial similarity threshold by step = 0.05: t = T + step, and returns to step S501 to reclassify;

and S505, judging that the current initial similarity threshold is an accurate similarity threshold, and outputting the similarity threshold of the current training station caption.

Training the station captions (such as three sets of central station captions, Gansu station captions, Shandong station captions and the like) to be detected respectively according to the process, and obtaining the similarity threshold value of each station caption to be detected.

After the obtained feature corner point probability distribution matrix and similarity threshold of all the trained station captions are obtained, it can be detected whether the image to be detected contains the trained station caption, as shown in fig. 16, the detection process includes:

s601, traversing the feature corner probability distribution matrix Mp and the similarity threshold T of all the trained station captions, and extracting a station caption area from the image to be detected according to the position information of the positive sample of the current station caption (the relative position of the positive sample of the current station caption in the original image) and the size information of the positive sample (the size information of the positive sample of the current station caption); for example, when traversing to the kth trained bench mark, the obtained feature corner probability distribution matrix and similarity threshold are Mp respectively_k、T_k。

S602, performing characteristic corner detection on the station caption region, acquiring characteristic corner information of the station caption region and acquiring a characteristic corner information expression S1 of the station caption region, wherein the characteristic corner information expression S1 is acquired according to the same method as the step S402, namely, when a characteristic corner exists in any pixel point (x, y) in the station caption region, the value S1(x, y) of the characteristic corner information expression S1 at the pixel point (x, y) is 1, otherwise, the value S1(x, y) is 0;

s603, calculating the similarity epsilon of the station caption area containing the station caption to be detected,

wherein, MP_i(x, y) is the value of the characteristic corner probability distribution matrix of the current station caption at the pixel point (x, y), and S1(x, y) is the value of the characteristic corner information expression S1 at the pixel point (x, y);

s604, according to the similarity epsilon and the similarity threshold T of the station logo to be detected_kComparing when epsilon is more than or equal to T_kJudging that the image to be detected contains the kth station caption, otherwise, judging that the image to be detected does not contain the kth station caption, acquiring position information and size information of a positive sample of the next trained station caption, acquiring a characteristic corner point probability distribution matrix Mp and a similarity threshold value T of the next trained station caption, and repeatedly judging to obtain the position information and the size information of the positive sampleAnd identifying the station caption contained in the image to be detected.

Based on the above method, the present invention further provides a station logo recognition system, as shown in fig. 17, which includes:

the sample extraction module 100 is configured to extract a positive sample and a negative sample of a platform logo to be detected from a plurality of original images, where the positive sample is an area image containing the platform logo to be detected, and the negative sample is an area image not containing the platform logo to be detected;

a characteristic corner probability distribution matrix obtaining module 200, configured to perform characteristic corner detection on the extracted positive sample, obtain a characteristic corner set of the positive sample, and calculate, through the characteristic corner set, a frequency of a characteristic corner appearing in each pixel point in the positive sample, so as to obtain a characteristic corner probability distribution matrix of the station logo to be detected;

a similarity threshold obtaining module 300, configured to calculate, through the feature corner probability distribution matrix, a similarity of each positive sample and each negative sample to obtain a similarity set including similarities of all positive samples and all negative samples, and calculate, through the similarity set, a similarity threshold of the station logo to be detected;

and the station logo detection module 400 is used for calculating the similarity to be detected of the to-be-detected image containing the station logo to be detected through the characteristic angular point probability distribution matrix, judging whether the similarity to be detected is greater than a similarity threshold value, judging that the to-be-detected image contains the station logo to be detected if the similarity to be detected is greater than the similarity threshold value, and judging that the to-be-detected image does not contain the station logo to be detected if the similarity to be detected is not greater than the similarity threshold.

Further, as shown in fig. 18, the characteristic corner probability distribution matrix obtaining module 200 includes:

a direction derivative calculating unit 210 for calculating the direction derivatives of the positive samples, which are stored as an array I_xAnd array I_y，I_xIs the directional derivative of the x direction, I_yThe directional derivative in the y-direction;

local autocorrelation matrix meterA calculating unit 220, configured to calculate a local autocorrelation matrix M for each pixel point in the positive sample using a gaussian template, wherein, <math> <mrow> <mi>M</mi> <mo>=</mo> <mi>G</mi> <mrow> <mo>(</mo> <mover> <mi>s</mi> <mo>~</mo> </mover> <mo>)</mo> </mrow> <mo>&CircleTimes;</mo> <mfenced open='[' close=']'> <mtable> <mtr> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>x</mi> </msub> </mtd> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>y</mi> </msub> </mtd> </mtr> </mtable> </mfenced> <mo>,</mo> </mrow> </math>

is a Gaussian template;

a corner matrix computing unit 230, configured to compute a corner matrix I for each pixel point by M, where I = det (M) -k · tr²(M) where det is the determinant of the matrix, tr is the trace of the matrix, and k is [0.04,0.06 ]]A constant within a range;

and the characteristic corner point information acquisition unit 240 is configured to determine whether any one point in the corner point quantity matrix I simultaneously satisfies that an element value of the point is greater than a threshold and is a local maximum in the field of the point, and when the element value of the point is simultaneously satisfied, determine that the point is a characteristic corner point of the positive sample.

The superposition counting unit 250 is used for calculating the times n (x, y) of occurrence of characteristic angular points at the position of each pixel point (x, y) in all the positive samples, when the proportion of the n (x, y) to the total amount of the positive samples is smaller than a preset value, the corresponding pixel point (x, y) is judged not to be the characteristic angular point, the value of the times (x, y) is returned to zero, otherwise, the pixel point is judged to be the characteristic angular point, and the value of the times n (x, y) is reserved;

a characteristic corner probability distribution matrix obtaining unit 260, configured to perform normalization operation on the frequency of the characteristic corner appearing at the position of each pixel point (x, y) to obtain a characteristic corner probability distribution matrix Mp (x, y) of the station logo to be detected,

the frequency of the characteristic corner appearing at the position of each pixel point (x, y) is the ratio of the number n (x, y) of the characteristic corners appearing at the current pixel point (x, y) to the total number of positive samples.

Further, as shown in fig. 19, the similarity threshold obtaining module 300 includes:

a presetting unit 310, configured to preset a minimum recognition accuracy, a maximum recognition error rate, and a maximum missed recognition rate;

a characteristic corner set obtaining unit 320, configured to perform characteristic corner detection on each of the positive sample and the negative sample, and when it is detected that a characteristic corner exists at any pixel point (x, y) in the positive sample or the negative sample, obtain a characteristic corner information expression S of the positive sample or the negative sample at any pixel point (x, y)_i(x, y) =1, otherwise S_i(x, y) =0 to obtain all positive sample, negative sample feature angular point set S = { S = { S =₀,S₁,S₂,......,S_i......S_NIn which S is_iIs a determinant of w × h, N = NumSamples + Numnegative-1, w and h are the widths and heights of the positive sample and the negative sample, NumSamples is the total amount of the samples, and NumNepositive is the total amount of the negative samples;

a similarity obtaining unit 330 for passing the feature corner pointsThe probability distribution matrix obtains the similarity epsilon of each positive sample and each negative sample_i，

So as to obtain all the feature sets of positive and negative samples epsilon = { [ epsilon ]₀,ε₁,ε₂,......,ε_i......ε_N}，ε_iRepresenting the similarity of the ith sample in all the positive samples and the negative samples;

and the similarity threshold obtaining unit 340 is configured to train the station caption to be detected according to the minimum recognition accuracy, the maximum recognition error rate, and the maximum missed recognition rate, so as to obtain a similarity threshold of the station caption to be detected.

Further, as shown in fig. 20, the similarity threshold acquisition unit 340 includes:

an initial similarity threshold presetting subunit 341, configured to preset an initial similarity threshold, and reclassify all positive samples and negative samples, if ∈ is detected_iIf the initial similarity is greater than the initial similarity threshold value, determining epsilon_iThe corresponding sample is a positive sample, otherwise, the corresponding sample is judged as a negative sample;

the statistics subunit 342 is configured to count the correct identification number Nr, the incorrect identification number Nw, and the missing identification number Nm under the condition of reclassifying all the positive samples and the negative samples, and calculate, according to the correct identification number Nr, the incorrect identification number Nw, and the missing identification number Nm, an identification correct rate, an identification error rate, and a missing identification rate under the initial similarity threshold condition;

a determining subunit 343, configured to determine whether the condition is met under the initial similarity threshold condition: when the recognition accuracy is higher than the minimum recognition accuracy, the recognition error rate is lower than the maximum recognition error rate, and the missing recognition rate is lower than the maximum missing recognition rate, the updating subunit is switched to if the recognition accuracy is higher than the minimum recognition accuracy, and the outputting subunit is switched to if the recognition error rate is lower than the maximum missing recognition rate;

an updating subunit 344, configured to update the initial similarity threshold by a predetermined step size, and reclassify;

the output subunit 345 is configured to output the similarity threshold of the currently trained station caption.

Further, as shown in fig. 21, the station caption detecting module 400 specifically includes:

a traversing unit 410, configured to traverse the feature corner probability distribution matrix Mp and the similarity threshold T of all the trained station captions, and extract a station caption area from the image to be detected according to the positive sample position information and the positive sample size information of the current station caption;

the station caption area detection unit 420 is configured to perform characteristic corner detection on the station caption area, obtain characteristic corner information of the station caption area, and obtain a characteristic corner information expression S1 of the station caption area;

a station caption region similarity calculation unit 430, configured to calculate a similarity epsilon that the station caption region includes the current station caption,

MP_i(x, y) is the value of the characteristic corner probability distribution matrix of the current station caption at the pixel point (x, y), and S1(x, y) is the value of the characteristic corner information expression S1 at the pixel point (x, y);

an identifying unit 440, configured to identify a similarity threshold T between the current station caption and the similarity epsilon_kComparing when epsilon is more than or equal to T_kAnd if not, judging that the image to be detected contains the current station caption, otherwise, judging that the image to be detected does not contain the current station caption.

In summary, according to the invention, a large number of positive samples and negative samples are obtained, and the characteristic corner detection is performed on the samples to obtain the characteristic corner probability distribution matrix and the similarity threshold, through the training process, the station caption contained in the image to be detected can be accurately and quickly identified in a complex background, so that the accuracy of station caption identification is improved, the identification time is shortened, and the identification efficiency is improved, thereby providing effective technical support for automatic video search, recording, analysis and retrieval in the multimedia technology.

It is to be understood that the invention is not limited to the examples described above, but that modifications and variations may be effected thereto by those of ordinary skill in the art in light of the foregoing description, and that all such modifications and variations are intended to be within the scope of the invention as defined by the appended claims.

Claims (11)

1. A station caption identification method is characterized by comprising the following steps:

A. extracting a positive sample and a negative sample of the station caption to be detected from a plurality of original images, wherein the positive sample is a regional image containing the station caption to be detected, and the negative sample is a regional image not containing the station caption to be detected;

B. performing characteristic corner detection on the extracted positive sample to obtain a characteristic corner set of the positive sample, and calculating the frequency of the characteristic corner of each pixel point in the positive sample through the characteristic corner set to obtain a characteristic corner probability distribution matrix of the station logo to be detected;

C. calculating the similarity of each positive sample and each negative sample through the characteristic corner probability distribution matrix to obtain a similarity set containing the similarities of all the positive samples and the negative samples, and calculating the similarity threshold of the station logo to be detected through the similarity set;

D. and calculating the similarity to be detected of the image to be detected, including the station logo to be detected, by the characteristic angular point probability distribution matrix, judging whether the similarity to be detected is greater than a similarity threshold value, if so, judging that the station logo to be detected is contained in the image to be detected, and if not, judging that the station logo to be detected is not contained in the image to be detected.

2. The station caption identifying method according to claim 1, wherein in the step a, for the same station caption, the relative position of each positive sample in the corresponding original image is the same, and the number ratio of the positive samples to the negative samples is 1: 1.5-1: 3.5.

3. the station logo recognition method according to claim 1, wherein in the step B, the process of acquiring the feature corner point set of the positive sample comprises:

b1, calculating the directional derivatives of the positive samples, and respectively storing the directional derivatives as an array I_xAnd array I_y，I_xIs the directional derivative of the x direction, I_yThe directional derivative in the y-direction;

b2, calculating a local autocorrelation matrix M for each pixel point in the positive sample by using the gaussian template, wherein, <math> <mrow> <mi>M</mi> <mo>=</mo> <mi>G</mi> <mrow> <mo>(</mo> <mover> <mi>s</mi> <mo>~</mo> </mover> <mo>)</mo> </mrow> <mo>&CircleTimes;</mo> <mfenced open='[' close=']'> <mtable> <mtr> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>x</mi> </msub> </mtd> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>y</mi> </msub> </mtd> </mtr> </mtable> </mfenced> <mo>,</mo> </mrow> </math>

is a Gaussian template;

b3, calculating a corner point quantity matrix I of each pixel point by M, wherein I = det (M) -k-tr²(M) where det is the determinant of the matrix, tr is the trace of the matrix, and k is [0.04,0.06 ]]A constant within a range;

b4, judging whether any point in the angular point quantity matrix I simultaneously satisfies that the element value of the point is larger than a threshold value and is a local maximum value in the point field, and if so, judging that the point is the characteristic angular point of the positive sample.

4. The station caption identifying method according to claim 1, wherein in the step B, the process of obtaining the characteristic corner point probability distribution matrix of the station caption to be detected specifically comprises:

b5, calculating the times n (x, y) of occurrence of characteristic corner points at the position of each pixel point (x, y) in all the positive samples, when the proportion of n (x, y) to the total amount of the positive samples is smaller than a preset value, judging that the corresponding pixel point (x, y) is not a characteristic corner point, and returning the value of the times (x, y) to zero, otherwise, judging that the pixel point is a characteristic corner point, and keeping the value of the times n (x, y);

b6, for each pixel point (x, y)Normalizing the frequency P (x, y) of the characteristic corner points at the positions to obtain a probability distribution matrix Mp (x, y) of the characteristic corner points of the station logo to be detected,

the frequency P (x, y) of the characteristic corner appearing at the position of each pixel point (x, y) is the ratio of the number n (x, y) of the characteristic corners appearing at the current pixel point (x, y) to the total number of positive samples.

5. The station caption identification method according to claim 1, wherein the step C specifically includes:

c1, presetting a minimum recognition accuracy, a maximum recognition error rate and a maximum missing recognition rate;

c2, detecting the characteristic corner of each positive sample and each negative sample, and when detecting that the characteristic corner exists at any pixel point (x, y) in the positive sample or the negative sample, then expressing the characteristic corner information expression S of the positive sample or the negative sample at any pixel point (x, y)_i(x, y) =1, otherwise S_i(x, y) =0 to obtain all positive sample, negative sample feature angular point set S = { S = { S =₀,S₁,S₂,......,S_i......S_NIn which S is_iIs a determinant of w × h, N = NumSamples + Numnegative-1, w and h are the widths and heights of the positive sample and the negative sample, NumSamples is the total amount of the samples, and NumNepositive is the total amount of the negative samples;

c3, obtaining the similarity epsilon of each positive sample and each negative sample through the characteristic corner probability distribution matrix_i，

So as to obtain all the feature sets of positive and negative samples epsilon = { [ epsilon ]₀,ε₁,ε₂,......,ε_i......ε_N}，ε_iRepresenting the similarity of the ith sample in all the positive samples and the negative samples;

and C4, training the station caption to be detected according to the minimum recognition correct rate, the maximum recognition error rate and the maximum missing recognition rate to obtain the similarity threshold value of the station caption to be detected.

6. The station caption identification method according to claim 5, wherein the step C4 specifically comprises:

c41, presetting an initial similarity threshold, reclassifying all positive and negative samples if epsilon_iIf the initial similarity is greater than the initial similarity threshold value, determining epsilon_iThe corresponding sample is a positive sample, otherwise, the corresponding sample is judged as a negative sample;

c42, counting the number of correct identifications Nr, the number of incorrect identifications Nw and the number of missed identifications Nm under the condition of reclassifying all the positive samples and the negative samples, and calculating the identification accuracy, the identification error rate and the recognition rate of missed identifications under the condition of the initial similarity threshold according to the number of correct identifications Nr, the number of incorrect identifications Nw and the number of missed identifications Nm;

c43, judging whether the conditions are met under the condition of the initial similarity threshold value: if the recognition accuracy is greater than the minimum recognition accuracy, the recognition error rate is less than the maximum recognition error rate, and the missing recognition rate is less than the maximum missing recognition rate, the step C45 is executed, otherwise, the step C44 is executed;

c44, updating the initial similarity threshold by a preset step size, and returning to the step C41 for reclassification;

and C45, outputting the similarity threshold of the currently trained station caption.

7. The station caption identification method according to claim 1, wherein the step D specifically includes:

d1, traversing the feature corner probability distribution matrix Mp and the similarity threshold T of all the trained station captions, and extracting a station caption area from the image to be detected according to the position information and the size information of the positive sample of the current station caption;

d2, detecting characteristic corner points of the station caption area, acquiring characteristic corner point information of the station caption area and acquiring a characteristic corner point information expression S1 of the station caption area;

d3, calculating the similarity epsilon of the current station caption contained in the station caption area,

MP_i(x, y) is the value of the characteristic corner probability distribution matrix of the current station caption at the pixel point (x, y), and S1(x, y) is the value of the characteristic corner information expression S1 of the station caption region at the pixel point (x, y);

d4, according to the similarity epsilon and the similarity threshold T of the current station caption_kComparing when epsilon is more than or equal to T_kAnd if not, judging that the image to be detected contains the current station caption, otherwise, judging that the image to be detected does not contain the current station caption.

8. A station caption identification system, comprising:

the sample extraction module is used for extracting a positive sample and a negative sample of the station caption to be detected from a plurality of original images, wherein the positive sample is a regional image containing the station caption to be detected, and the negative sample is a regional image not containing the station caption to be detected;

the characteristic corner probability distribution matrix acquisition module is used for detecting characteristic corners of the extracted positive sample, acquiring a characteristic corner set of the positive sample, calculating the frequency of the characteristic corners of each pixel point in the positive sample through the characteristic corner set, and acquiring a characteristic corner probability distribution matrix of the station logo to be detected;

the similarity threshold acquisition module is used for calculating the similarity of each positive sample and each negative sample through the characteristic corner probability distribution matrix to obtain a similarity set containing the similarities of all the positive samples and the negative samples, and calculating the similarity threshold of the station logo to be detected through the similarity set;

and the station logo detection module is used for calculating the similarity to be detected of the station logo to be detected including the required detection through the characteristic angular point probability distribution matrix, judging whether the similarity to be detected is greater than a similarity threshold value, judging that the station logo to be detected is included in the image to be detected if the similarity to be detected is greater than the similarity threshold value, and judging that the station logo to be detected is not included in the image to be detected if the similarity to be detected is not greater than the similarity threshold value.

9. The station logo recognition system according to claim 8, wherein the characteristic corner point probability distribution matrix obtaining module comprises:

a direction derivative calculation unit for calculating the direction derivatives of the positive samples, which are stored as an array I_xAnd array I_y，I_xIs the directional derivative of the x direction, I_yThe directional derivative in the y-direction;

a local autocorrelation matrix calculation unit for calculating a local autocorrelation matrix M for each pixel point in the positive sample using a Gaussian template, wherein, <math> <mrow> <mi>M</mi> <mo>=</mo> <mi>G</mi> <mrow> <mo>(</mo> <mover> <mi>s</mi> <mo>~</mo> </mover> <mo>)</mo> </mrow> <mo>&CircleTimes;</mo> <mfenced open='[' close=']'> <mtable> <mtr> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>x</mi> </msub> </mtd> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>I</mi> <mi>x</mi> </msub> <msub> <mi>I</mi> <mi>y</mi> </msub> </mtd> <mtd> <msub> <msup> <mi>I</mi> <mn>2</mn> </msup> <mi>y</mi> </msub> </mtd> </mtr> </mtable> </mfenced> <mo>,</mo> </mrow> </math>

is a Gaussian template;

a corner matrix calculation unit for calculating a corner matrix I of each pixel point by M, wherein I = det (M) -k-tr²(M) where det is the determinant of the matrix, tr is the trace of the matrix, and k is [0.04,0.06 ]]A constant within a range;

and the characteristic angular point information acquisition unit is used for judging whether any point in the angular point quantity matrix I simultaneously satisfies the condition that the element value of the point is greater than a threshold value and is a local maximum value in the field of the point, and when the point simultaneously satisfies the element value, judging that the point is the characteristic angular point of the positive sample.

10. The station logo recognition system according to claim 9, wherein the characteristic corner point probability distribution matrix obtaining module further comprises:

the superposition statistical unit is used for calculating the times n (x, y) of occurrence of characteristic angular points at the position of each pixel point (x, y) in all the positive samples, when the proportion of the n (x, y) to the total amount of the positive samples is smaller than a preset value, the corresponding pixel point (x, y) is judged not to be the characteristic angular point, the value of the times (x, y) is returned to zero, otherwise, the pixel point is judged to be the characteristic angular point, and the value of the times n (x, y) is reserved;

a characteristic angular point probability distribution matrix obtaining unit, which is used for carrying out normalization operation on the frequency P (x, y) of the characteristic angular point appearing on the position of each pixel point (x, y) to obtain a characteristic angular point probability distribution matrix Mp (x, y) of the station logo to be detected,

the frequency of the characteristic corner appearing at the position of each pixel point (x, y) is the ratio of the number n (x, y) of the characteristic corners appearing at the current pixel point (x, y) to the total number of positive samples.

11. The station logo recognition system as claimed in claim 8, wherein the similarity threshold acquisition module comprises:

the device comprises a presetting unit, a judging unit and a judging unit, wherein the presetting unit is used for presetting a minimum identification correct rate, a maximum identification error rate and a maximum missing identification rate;

a characteristic corner set acquisition unit for detecting the characteristic corners of each of the positive and negative samples, when the characteristic of any pixel point (x, y) in the positive or negative sample is detectedWhen the corner is present, the characteristic corner information expression S of the positive sample or the negative sample at any pixel point (x, y)_i(x, y) =1, otherwise S_i(x, y) =0 to obtain all positive sample, negative sample feature angular point set S = { S = { S =₀,S₁,S₂,......,S_i......S_NIn which S is_iIs a determinant of w × h, N = NumSamples + Numnegative-1, w and h are the widths and heights of the positive sample and the negative sample, NumSamples is the total amount of the samples, and NumNepositive is the total amount of the negative samples;

a similarity obtaining unit for obtaining the similarity epsilon of each positive sample and each negative sample through the characteristic corner probability distribution matrix_i，

So as to obtain all the feature sets of positive and negative samples epsilon = { [ epsilon ]₀,ε₁,ε₂,......,ε_i......ε_N}，ε_iRepresenting the similarity of the ith sample in all the positive samples and the negative samples;

and the similarity threshold acquisition unit is used for training the station caption to be detected according to the minimum recognition correct rate, the maximum recognition error rate and the maximum missing recognition rate to obtain the similarity threshold of the station caption to be detected.

Images