CN106778742A - A kind of automobile logo detection method suppressed based on Gabor filter background texture - Google Patents

Abstract

The invention discloses a kind of automobile logo detection method suppressed based on Gabor filter background texture, comprise the steps of：The first step, line tilt correction pretreatment is entered to image；Second step, carries out car plate detection in image after the pre-treatment, obtain license plate area；3rd step, based on priori, according to car plate and the position relationship of logo, obtains the logo coarse positioning region comprising logo pattern after License Plate；4th step, Gabor filtering is carried out to logo coarse positioning region, suppresses radiator-grid texture around logo, highlights car mark region；5th step, carries out gaussian filtering and mathematical morphology closed operation；6th step, selectes threshold value to gray level image thresholding, and confines detection target area, realizes logo fine positioning.The automobile logo detection method detection time is short, and verification and measurement ratio is high.

Description

A kind of automobile logo detection method suppressed based on Gabor filter background texture

Technical field

The present invention relates to automobile logo detection method, more particularly to a kind of logo suppressed based on Gabor filter background texture Detection method.

Background technology

With the rapid growth of social economy, the automobile consumption demand of China is increasingly vigorous at present, and automobile number constantly increases It is long, such as escaping behavior after traffic accident is also brought, vehicle is stolen to wait traffic problems.In order to determine illegal and vehicles peccancy, at present Generally use and the car plate of automobile is identified.But deck in recent years, board, car plate abrasion and license plate shading etc. are existing The appearance of elephant so that determining automobile only by identification car plate becomes unreliable.Logo is to contain vehicle and manufacturer The key image of information, is the important evidence of separation vehicle and identification.If logo can exactly be positioned, it will effectively carry Separation vehicle high and the accuracy rate of identification.

But logo species is abundant, various shapes, the surface without stabilization, while the environment residing for it is texture Complicated automobile grills region.Other logo is easily influenceed by weather：In the case of night or rainy day illumination deficiency, logo is difficult to distinguish Recognize；Under intense light conditions, logo is easily reflective.These characteristics cause that vehicle-logo location has larger difficulty, very challenging property.It is existing Vehicle-logo location method is ripe not enough, and it is low to there is verification and measurement ratio, the problems such as easily interference by illumination variation, so logo Detection results are also Need further raising.

The content of the invention

Goal of the invention：The problem that the present invention exists for prior art, there is provided one kind is based on Gabor filter background texture The automobile logo detection method of suppression.

Technical scheme：The automobile logo detection method suppressed based on Gabor filter background texture of the present invention is included：

(1) to shooting the vehicle image with certain angle of inclination for obtaining, vehicle symmetry axis is carried out based on SIFT operators Detection and slant correction；

(2) cascade classifier is obtained using the machine learning algorithm training of Harr+AdaBoost, and uses the cascade sort Device positioning licence plate region from the vehicle image after correction；

(3) based on priori, according to car plate and the position relationship of logo, obtained comprising car in the license plate area of positioning The logo coarse positioning region of case of marking on a map；

(4) Gabor filtering is carried out to logo coarse positioning region, suppresses radiator-grid texture around logo, highlight car mark region；

(5) gaussian filtering and mathematical morphology closed operation are carried out to car mark region；

(6) threshold value is selected to the gray level image thresholding that obtains step (5), and confines detection target area, obtain essence The logo of positioning.

Beneficial effect：Compared with prior art, its remarkable advantage is the present invention：

1) accuracy of detection is high：The method has certain anti-interference to illumination variation, has under different illumination conditions Verification and measurement ratio higher, is had preferably (especially under high light) under the illumination condition such as sun light direct beam and night to logo Locating effect；

2) real-time is good：The inventive method night or illumination deficiency under the conditions of, under the conditions of sunshine blaze With detection speed faster is suffered under conditions of uniform illumination, for highway bayonet socket capture image data for, can To realize online real-time processing；

3) applicable object is wide：Traditional background texture restrainable algorithms are often first to carry out texture side to logo coarse positioning region To judgement, different vehicle-logo location algorithms is then used according to different grain direction, once there is logo line in conventional method The situation of discriminating direction mistake is managed, the localization method of mistake will be used, necessarily cause positioning to fail, and what the present invention was used The method suppressed based on Gabor filter background texture is had applicability higher and need not carry out grain direction differentiation, can It is level to be applied to logo background simultaneously, and the vehicle of vertical and reticular texture carries out logo detection, so the present invention is a kind of suitable For different situations and the more preferable algorithm of robustness.

Brief description of the drawings

Fig. 1 is the schematic flow sheet of the automobile logo detection method suppressed based on Gabor filter background texture of the invention；

Fig. 2 is the schematic flow sheet of License Plate；

Fig. 3 is the filtering schematic flow sheet of Gabor filter.

Specific embodiment

As shown in figure 1, the automobile logo detection method suppressed based on Gabor filter background texture of the present embodiment is included：

(1) to shooting the vehicle image with certain angle of inclination for obtaining, vehicle symmetry axis is carried out based on SIFT operators Detection and slant correction.

The step is specifically included：

(1-1) obtains the point vector of each characteristic point of vehicle image, wherein, the point vector of feature points i isx_i,y_iThe point coordinates is represented,Represent direction, s_iDimensional information is represented, i=1 ..., n, n is characterized The sum of point；

(1-2) is by the direction of characteristic pointBy after normalization, obtaining corresponding Feature Descriptor k_i, i=1 ..., n, and By the Feature Descriptor default dimension of generation a little SIFT feature vector；

(1-3) is by directly modification Feature Descriptor k_iGeneration mirror image m_i, then generated by matching characteristic point and mirror image possible Symmetrical feature point to (p_i,p_j), i, j=1 ..., n, i ≠ j；

(1-4) calculates every a pair possible symmetrical feature points to (p_i,p_j) angle confidence level Φ_ij, yardstick confidence level S_ij With apart from confidence level D_ij, then it is calculated total confidence level M_i,j, i, j=1 ..., n, i ≠ j；Wherein：

Angle confidence level Φ_ijComputing formula be：Wherein,Respectively point p_i With point p_jDirection, θ_ijIt is point p_iTo point p_jDirection；

Yardstick confidence level S_ijBy quantifying p_iAnd p_jThe similarity s of mesoscale_iAnd s_jTo obtain： Wherein σ_sIt is scale factor, is Gaussian function envelope along point p_iTo point p_jDirection standard variance；

Apart from confidence level D_ijFor：σ_dIt is that, apart from border, d is symmetric points p_iTo point p_jDistance；

Total confidence level M_i,jFor：

(1-5) calculates every a pair of symmetrical features point to (p_i,p_j) symmetry axis r_ij：r_ij=x_c cosθ_ij+y_c sinθ_ij, its Middle x_c, y_cRespectively characteristic point is to (p_i,p_j) length in x-axis and y-axis direction, θ_ijIt is point p_iTo point p_jDirection；

(1-6) finds main symmetry axis using linear Hough transform, and every a pair of symmetrical features point is to (p_i,p_j) empty to Hough Interior point (r_ij,θ_ij) with weights M_i,jBallot, obtains main symmetry axis；

(1-7) is using the angle of main symmetry axis as inclination of vehicle angle, θ；

Vehicle image is entered line tilt correction by (1-8) according to inclination of vehicle angle, θ, and the image after correction is：

Width, height, are the wide and height of original image, and width', height' are the wide and height of correction chart picture.

(2) cascade classifier is obtained using the machine learning algorithm training of Harr+AdaBoost, and uses the cascade sort Device positioning licence plate region from the vehicle image after correction.

Step (2) specifically includes following steps：

The positive sample and negative sample of (2-1) collection predetermined number, set up positive sample storehouse and negative example base respectively, wherein, just Sample refers to the car plate cut out during the high definition vehicle shot from road gate shines, and negative sample is from the non-car plate area in vehicle photograph The random background sample cut out in domain, comprising various background environments；

(2-2) extracts the Harr features of all positive negative samples, and goes out several AdaBoost with these Harr features trainings The strong classifier of algorithm；

Wherein, the training method of the strong classifier of AdaBoost algorithms is：

(2-2-1) sets training set S={ (a₁,b₁),...,(a_m,b_m) include m sample, wherein a_i∈ A (i=1, 2 ..., m) represent training sample, A is training sample set, b_i∈ B are a_iCorresponding diagnostic criterium, and have B={ 1, -1 }, for I-th j-th Weak Classifier h of training sample_j(a_i) be expressed as

Wherein, F_j(a_i) represent j-th value of Haar features, δ in subwindow_jRepresent the threshold value of setting, p_jRepresent control not The amount in the direction of equal sign；

Weak Classifier is trained for strong classifier by (2-2-2) with AdaBoost algorithms, is comprised the following steps that：

Initialization sample weights：

Wherein, w₁I () represents i-th initial weight of sample in first round training, p represents the sum of positive sample in S, q The sum of negative sample in S is represented, it is total sample number to have p+q=m, m；

For t=1,2 ..., T (T is iterations) is circulated as follows：

1. weights normalization

Wherein, w_tI () represents i-th weights of sample, i=1,2 ..., m in t wheel training；

2. its Weak Classifier h is trained to feature j_j, calculate its weighted error ε_j, i.e.,

And select the minimum grader h of weighted error_{t min}The grader circulated as this；

3. sample weights are updated according to equation below：

Wherein when classifying correct, h_t(a_i)=b_i, e_i=0, during classification error, h_t(a_i)≠b_i, e_i=1；

(2-2-3) obtains final strong classifier, as follows

WhereinA is window to be checked, h_tA () represents the Weak Classifier obtained in t wheel training, H (a) Result for 1 represent receive, 0 represent refusal.

(2-3) will train some strong classifiers for obtaining that car plate detection cascade classifier is constructed in the form of cascading；

(2-4) using the car plate detection cascade classifier for training as shown in Fig. 2 detect the car plate area in vehicle image Domain；

It is foundation that (2-5) uses car plate ratio characteristic and car plate position feature, and the license plate area that will be detected is picked up by mistake Screening.

Step (2-5) is used to exclude after positioning the region of flase drop.

Wherein, car plate ratio characteristic：Car plate is shaped as rectangle, and car plate has the font of fixed number of words and fixed size, real Border width and height are respectively 44 centimetres and 14 centimetres, and the ratio of width to height in image is basic 3：1 or so, in algorithm can to width and Highly it is sized scope limit；

Car plate position feature：Because real road monitoring camera is shot and car by fixed position ground induction coil signal triggering The installation site of board is in the partial below of whole vehicle, therefore vertical direction position of the car plate in entire image is basicly stable (more Tend to occur at image the latter half), the center can be deviateed in the hope of average car plate center according to statistical information The possibility that more remote candidate region turns into car plate is lower.

(3) it is can be found that by observing each part topological structure of the headstock of vehicle：It is characterized in it that logo is easiest to differentiate Position, its maximum difference with other interference regions is that logo does not appear in other positions above except car plate, and is done The position for disturbing region is random, uncertain, therefore can be by car plate position coarse localization logo approximate range；By to big Amount vehicle pictures experiment, according to car plate and the position relationship of logo, obtains comprising logo pattern in the license plate area of positioning Logo coarse positioning region is：

Wherein, X₁、X₂、Y₁、Y₂The left and right of the logo approximate range that respectively coarse positioning is obtained and up-and-down boundary, X_left、 X_rightThe respectively right boundary of license plate area, Y_upIt is license plate area coboundary, for license plate area highly, N is optional to height The height coefficient selected, is typically taken as N=3.

(4) Gabor filtering is carried out to logo coarse positioning region, suppresses radiator-grid texture around logo, highlight car mark region.

As shown in figure 3, step (4) specifically includes following steps：

(4-1) define Gabor filter two-dimensional Gabor kernel function h (x, y) be

Wherein, u₀It is function centre frequency, σ_x、σ_yIt is scale factor, respectively Gaussian function envelope is along x, y-axis direction Standard deviation, exp (2 π ju₀R₁) it is oscillating function, real part is cosine function, and imaginary part is SIN function,It is the direction of setting；

By changing Gabor kernel functions directionThe texture information of different directions in image can be extracted,Span ForIn the range of thisValue can describe all of direction, i.e.,WithSame direction is described, together Sample can extract texture information in image on different scale by changing the yardstick of Gabor kernel functions, without choosing in the present invention The extraction of multiple dimensioned upper texture information is selected, but under the particular dimensions for working well, have chosen 6 different directions Gabor filter constitutes wave filter group, and direction is respectively π/6, π/4, π/3,2 π/3,3 π/4,5 π/6.

(4-2) initializes Gabor filter：6 Gabor filter { h of different directions are built respectively_l(x, y) | l= 1 ..., 6 }, so as to constitute wave filter group, wherein, directionIt is respectively set as π/6, π/4, π/3,2 π/3,3 π/4,5 π/6,6 The scale parameter and bandwidth parameter of the Gabor filter in direction set as follows：Yardstick σ_x、σ_yValue be set as 1, bandwidth Sigma It is set to 2 π, and Sigma=σ u₀, therefore centre frequency u₀It is 2 π；

The logo coarse positioning area image that (4-3) obtains step (3) respectively with 6 two-dimensional Gabor cores of different directions Function h (x, y) carries out convolution operation and modulus, obtains 6 filtering image G₁(x, y), G₂(x, y), G₃(x, y), G₄(x, y), G₅ (x, y), G₆(x, y), wave filter group final output image is G (x, y)=(G₁(x,y)+G₂(x,y)+G₃(x,y)+G₄(x,y)+G₅ (x,y)+G₆(x, y))/6, wherein, G_l(x, y)=∫ ∫ f (x₀,y₀)h_l(x-x₀,y-y₀)dx₀dy₀, in formula, f (x, y) is that logo is thick Positioning region image；

Each pixel pk of wave filter group final output image G (x, y) that (4-4) is obtained to step (4-3) calculates it Attenuation coefficient λ_pk, k=1 ..., num, num are the total number of pixel, and calculation procedure is as follows：

1. to being hung down with Sobel horizontal edges detection method and Sobel respectively in the coarse positioning region that is obtained after License Plate Straight edge detection method obtains the image I comprising horizontal texture_xAnd the image I comprising vertical texture_y；

2. the attenuation coefficient for calculating each pixel pk is：

(x_pk,y_pk) it is the coordinate of pixel pk, α and β is two constant coefficients, and α=0.5, β=2 are taken in the present invention； Width_pai is the width of the license plate area that positioning is obtained；From formula：λ_pkSize and I_x/I_yAnd x_pkIt is relevant：I_x/I_yMore Level off to 0 or infinitely great the situation of texture (correspondence horizontally or vertically), λ_pkIt is smaller, similarly, x_pkMore deviate at width_pai/2 (the common symmetry axis of logo and car plate), λ_pkIt is smaller,

Each pixel of ripple device group output image G (x, y) is all multiplied by its corresponding attenuation coefficient λ by (4-5)_pk。

(5) gaussian filtering and mathematical morphology closed operation are carried out to car mark region.

The step filters noise for the image that step (4) is obtained with two-dimensional discrete Gaussian filter function, and carries out ash Degree level mathematical morphology closed operation, to make detection target interruption up or be broken, eliminates the minuscule hole in target.

(6) threshold value is selected to the gray level image thresholding that obtains step (5), and confines detection target area, obtain essence The logo of positioning.

Step (6) specifically includes following steps：

(6-1) selects the image T (x that step (5) is obtained_pk,y_pk) grey scale pixel value average 1/2 as threshold valueI.e.

(6-2) carries out thresholding using threshold value, and the image after thresholding is

The image center regular inspection of (6-3) from after thresholding surveys target area, and this target area is car mark region.

Claims (8)

1. it is a kind of based on Gabor filter background texture suppress automobile logo detection method, it is characterised in that the method includes：

(1) to shooting the vehicle image with certain angle of inclination for obtaining, the symmetrical shaft detection of vehicle is carried out based on SIFT operators And slant correction；

(2) obtain cascade classifier using the machine learning algorithm training of Harr+AdaBoost, and using the cascade classifier from Positioning licence plate region in vehicle image after correction；

(3) based on priori, according to car plate and the position relationship of logo, obtained comprising logo figure in the license plate area of positioning The logo coarse positioning region of case；

(4) Gabor filtering is carried out to logo coarse positioning region, suppresses radiator-grid texture around logo, highlight car mark region；

(5) gaussian filtering and mathematical morphology closed operation are carried out to car mark region；

(6) threshold value is selected to the gray level image thresholding that obtains step (5), and confines detection target area, obtain fine positioning Logo.

2. the automobile logo detection method suppressed based on Gabor filter background texture according to claim 1, its feature exists In：Step (1) is specifically included：

(1-1) obtains the point vector of each characteristic point of vehicle image, wherein, the point vector of feature points i isx_i,y_iThe point coordinates is represented,Represent direction, s_iDimensional information is represented, i=1 ..., n, n is characterized The sum of point；

(1-2) is by the direction of characteristic pointBy after normalization, obtaining corresponding Feature Descriptor k_i, i=1 ..., n, and by institute The SIFT feature vector of the default dimension of Feature Descriptor generation a little；

(1-3) is by directly modification Feature Descriptor k_iGeneration mirror image m_i, then it is possible right by matching characteristic point and mirror image generation Claim characteristic point to (p_i,p_j), i, j=1 ..., n, i ≠ j；

(1-4) calculates every a pair possible symmetrical feature points to (p_i,p_j) angle confidence level Φ_ij, yardstick confidence level S_ijWith away from From confidence level D_ij, then it is calculated total confidence level M_i,j, i, j=1 ..., n, i ≠ j；Wherein：

Angle confidence level Φ_ijComputing formula be：Wherein,Respectively point p_iAnd point p_jDirection, θ_ijIt is point p_iTo point p_jDirection；

Yardstick confidence level S_ijBy quantifying p_iAnd p_jThe similarity s of mesoscale_iAnd s_jTo obtain：Its Middle σ_sIt is scale factor, is Gaussian function envelope along point p_iTo point p_jDirection standard variance；

Apart from confidence level D_ijFor：σ_dIt is that, apart from border, d is symmetric points p_iTo point p_jDistance；

Total confidence level M_i,jFor：

(1-5) calculates every a pair of symmetrical features point to (p_i,p_j) symmetry axis r_ij：r_ij=x_ccosθ_ij+y_csinθ_ij, wherein x_c, y_c Respectively characteristic point is to (p_i,p_j) length in x-axis and y-axis direction, θ_ijIt is point p_iTo point p_jDirection；

(1-6) finds main symmetry axis using linear Hough transform, and every a pair of symmetrical features point is to (p_i,p_j) in hough space Point (r_ij,θ_ij) with weights M_i,jBallot, obtains main symmetry axis；

(1-7) is using the angle of main symmetry axis as inclination of vehicle angle, θ；

Vehicle image is entered line tilt correction by (1-8) according to inclination of vehicle angle, θ, and the image after correction is：

$\left\{\begin{array}{c}widt{h}^{\′}=width.cos\mathrm{\θ}+height.sin\mathrm{\θ}\\ heigh{t}^{\′}=width.sin\mathrm{\θ}+height.cos\mathrm{\θ}\end{array}\right.$

Width, height, are the wide and height of original image, and width', height' are the wide and height of correction chart picture.

3. the automobile logo detection method suppressed based on Gabor filter background texture according to claim 1, its feature exists In：Step (2) is specifically included：

The positive sample and negative sample of (2-1) collection predetermined number, set up positive sample storehouse and negative example base respectively, wherein, positive sample Refer to the car plate cut out during the high definition vehicle shot from road gate shines, negative sample is from the non-license plate area in vehicle photograph The random background sample cut out, comprising various background environments；

(2-2) extracts the Harr features of all positive negative samples, and goes out several AdaBoost algorithms with these Harr features trainings Strong classifier；

(2-3) will train some strong classifiers for obtaining that car plate detection cascade classifier is constructed in the form of cascading；

(2-4) detects the license plate area in vehicle image using the car plate detection cascade classifier for training；

It is foundation that (2-5) uses car plate ratio characteristic and car plate position feature, and the license plate area that will be detected is picked up screening by mistake.

4. the automobile logo detection method suppressed based on Gabor filter background texture according to claim 3, its feature exists In：The training method of the strong classifier of AdaBoost algorithms is described in step (2-2)：

(2-2-1) sets training set S={ (a₁,b₁),...,(a_m,b_m) include m sample, wherein a_i∈ A (i=1,2 ..., m) Training sample is represented, A is training sample set, b_i∈ B are a_iCorresponding diagnostic criterium, and have B={ 1, -1 }, for i-th training J-th Weak Classifier h of sample_j(a_i) be expressed as

$h_j\left(a_i\right)=\left\{\begin{array}{cc}1,& p_j{F}_j\left(a_i\right)<p_j{\mathrm{\&delta;}}_j\\ 0,& otherwise\end{array}\right.$

Wherein, F_j(a_i) represent j-th value of Haar features, δ in subwindow_jRepresent the threshold value of setting, p_jRepresent the control sign of inequality Direction amount；

Weak Classifier is trained for strong classifier by (2-2-2) with AdaBoost algorithms, is comprised the following steps that：

Initialization sample weights：

$w_1\left(i\right)=\left\{\begin{array}{cc}\frac{1}{2p},& b_i=1\\ \frac{1}{2q},& b_i=-1\end{array}\right.$

Wherein, w₁I () represents i-th initial weight of sample in first round training, p represents the sum of positive sample in S, and q represents S The sum of middle negative sample, it is total sample number to have p+q=m, m；

For t=1,2 ..., T (T is iterations) is circulated as follows：

1. weights normalization

$w_t\left(i\right)\&LeftArrow;\frac{w_t\left(i\right)}{\underset{j=1}{\overset{m}{\&Sigma;}}{w}_t\left(j\right)}$

Wherein, w_tI () represents i-th weights of sample, i=1,2 ..., m in t wheel training；

2. its Weak Classifier h is trained to feature j_j, calculate its weighted error ε_j, i.e.,

${\mathrm{\&epsiv;}}_j=\underset{i=1}{\overset{m}{\&Sigma;}}{w}_t\left(i\right)|h_j\left(a_i\right)-b_i|$

And select the minimum grader h of weighted error_tminThe grader circulated as this；

3. sample weights are updated according to equation below：

$w_{t+1}\left(i\right)=w_t\left(i\right){\left(\frac{{\mathrm{\&epsiv;}}_t}{1-{\mathrm{\&epsiv;}}_t}\right)}^{1-e_i}=\left\{\begin{array}{cc}{w}_t\left(i\right)\frac{{\mathrm{\&epsiv;}}_t}{1-{\mathrm{\&epsiv;}}_t},& h_t\left(a_i\right)=b_i\\ w_t\left(i\right),& h_t\left(a_i\right)\&NotEqual;b_i\end{array}\right.$

Wherein when classifying correct, h_t(a_i)=b_i, e_i=0, during classification error, h_t(a_i)≠b_i, e_i=1；

(2-2-3) obtains final strong classifier, as follows

$H\left(a\right)=sign\left(\underset{t=1}{\overset{T}{\&Sigma;}}{\mathrm{\&alpha;}}_t{h}_t\left(a\right)\right)=\left\{\begin{array}{cc}1,& \underset{t=1}{\overset{T}{\&Sigma;}}{\mathrm{\&alpha;}}_t{h}_t\left(a\right)\&GreaterEqual;\frac{1}{2}\underset{t=1}{\overset{T}{\&Sigma;}}{\mathrm{\&alpha;}}_t\\ -1,& otherwise\end{array}\right.$

WhereinA is window to be checked, h_tA () represents the Weak Classifier obtained in t wheel training, the knot of H (a) Fruit represents receiving for 1, and 0 represents refusal.

5. the automobile logo detection method suppressed based on Gabor filter background texture according to claim 1, its feature exists In：Step (3) specifically includes following steps：

Based on priori, according to car plate and the position relationship of logo, obtained comprising logo pattern in the license plate area of positioning Logo coarse positioning region, wherein, logo coarse positioning region is：

$\left\{\begin{array}{c}{X}_1=X_{left}\\ X_2=X_{right}\\ Y_1=Y_{up}-N*height\\ Y_2=Y_{up}\end{array}\right.$

Wherein, X₁、X₂、Y₁、Y₂The left and right of the logo approximate range that respectively coarse positioning is obtained and up-and-down boundary, X_left、X_rightPoint Not Wei license plate area right boundary, Y_upIt is license plate area coboundary, for license plate area highly, N is selectable height to height Degree coefficient.

6. the automobile logo detection method suppressed based on Gabor filter background texture according to claim 1, its feature exists In：Step (4) specifically includes following steps：

(4-1) define Gabor filter two-dimensional Gabor kernel function h (x, y) be

$h(x,y)=\frac{1}{2{\mathrm{\&pi;\&sigma;}}_x{\mathrm{\&sigma;}}_y}\exp \&lsqb;-\frac{1}{2}(\frac{{R_1}^2}{{{\mathrm{\&sigma;}}_x}^2}+\frac{{R_2}^2}{{{\mathrm{\&sigma;}}_y}^2})\&rsqb;.\exp \left(2{\mathrm{\&pi;ju}}_0{R}_1\right)$

Wherein, u₀It is function centre frequency, σ_x、σ_yIt is scale factor, respectively Gaussian function envelope is along x, y-axis direction Standard deviation, exp (2 π ju₀R₁) it is oscillating function, real part is cosine function, and imaginary part is SIN function, It is the direction of setting；

(4-2) initializes Gabor filter：6 Gabor filter { h of different directions are built respectively_l(x, y) | l=1 ..., 6 }, so as to constitute wave filter group, wherein, directionIt is respectively set as π/6, π/4, π/3,2 π/3,3 π/4,5 π/6,6 directions The scale parameter and bandwidth parameter of Gabor filter set as follows：Yardstick σ_x、σ_yValue be set as 1, bandwidth Sigma is set to 2 π, and Sigma=σ u₀, therefore centre frequency u₀It is 2 π；

The logo coarse positioning area image that (4-3) obtains step (3) respectively with 6 two-dimensional Gabor kernel function h of different directions (x, y) carries out convolution operation and modulus, obtains 6 filtering image G₁(x, y), G₂(x, y), G₃(x, y), G₄(x, y), G₅(x, y), G₆(x, y), wave filter group final output image is G (x, y)=(G₁(x,y)+G₂(x,y)+G₃(x,y)+G₄(x,y)+G₅(x,y)+ G₆(x, y))/6, wherein, G_l(x, y)=∫ ∫ f (x₀,y₀)h_l(x-x₀,y-y₀)dx₀dy₀, in formula, f (x, y) is logo coarse positioning area Area image；

Each pixel pk of wave filter group final output image G (x, y) that (4-4) is obtained to step (4-3) calculates its decay Coefficient lambda_pk, k=1 ..., num, num are the total number of pixel, and calculation procedure is as follows：

1. to using Sobel horizontal edges detection method and Sobel vertical edges in the coarse positioning region that is obtained after License Plate respectively Edge detection method obtains the image I comprising horizontal texture_xAnd the image I comprising vertical texture_y；

2. the attenuation coefficient for calculating each pixel pk is：

${\mathrm{\&lambda;}}_{pk}=e^{-\mathrm{\&alpha;}\left|ln\frac{I_x}{I_y}\right|}.e^{-\mathrm{\&beta;}\left|ln\frac{x_{pk}}{width\_pai/2}\right|}$

(x_pk,y_pk) it is the coordinate of pixel pk, α and β is two constant coefficients, and width_pai is the license plate area that positioning is obtained Width；

Each pixel of ripple device group output image G (x, y) is all multiplied by its corresponding attenuation coefficient λ by (4-5)_pk。

7. the automobile logo detection method suppressed based on Gabor filter background texture according to claim 1, its feature exists In：Step (5) specifically includes following steps：

Noise is filtered with two-dimensional discrete Gaussian filter function for the image that step (4) is obtained, and carries out number of greyscale levels shape State closed operation, to make detection target interruption up or be broken, eliminates the minuscule hole in target.

8. the automobile logo detection method suppressed based on Gabor filter background texture according to claim 1, its feature exists In：Step (6) specifically includes following steps：

(6-1) selects the image T (x that step (5) is obtained_pk,y_pk) grey scale pixel value average 1/2 as threshold valueI.e.

(6-2) carries out thresholding using threshold value, and the image after thresholding is

$T^{\&prime;}(x_{pk},y_{pk})=\left\{\begin{array}{c}1,T(x_{pk},y_{pk})>\frac{1}{2}\stackrel{\&OverBar;}{t}\\ 0,T(x_{pk},y_{pk})\&le;\frac{1}{2}\stackrel{\&OverBar;}{t}\end{array}\right.$

The image center regular inspection of (6-3) from after thresholding surveys target area, and this target area is car mark region.