CN106778742B - Car logo detection method based on Gabor filter background texture suppression - Google Patents

Car logo detection method based on Gabor filter background texture suppression Download PDF

Info

Publication number
CN106778742B
CN106778742B CN201611126129.2A CN201611126129A CN106778742B CN 106778742 B CN106778742 B CN 106778742B CN 201611126129 A CN201611126129 A CN 201611126129A CN 106778742 B CN106778742 B CN 106778742B
Authority
CN
China
Prior art keywords
license plate
vehicle
image
logo
point
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201611126129.2A
Other languages
Chinese (zh)
Other versions
CN106778742A (en
Inventor
路小波
陈聪
孙权
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southeast University
Original Assignee
Southeast University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Southeast University filed Critical Southeast University
Priority to CN201611126129.2A priority Critical patent/CN106778742B/en
Publication of CN106778742A publication Critical patent/CN106778742A/en
Application granted granted Critical
Publication of CN106778742B publication Critical patent/CN106778742B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/24—Aligning, centring, orientation detection or correction of the image
    • G06V10/245—Aligning, centring, orientation detection or correction of the image by locating a pattern; Special marks for positioning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses a car logo detection method based on Gabor filter background texture suppression, which comprises the following steps: firstly, carrying out inclination correction preprocessing on an image; secondly, detecting a license plate in the preprocessed image to obtain a license plate area; thirdly, based on prior knowledge, obtaining a vehicle logo coarse positioning area containing vehicle logo patterns after positioning the license plate according to the position relation between the license plate and the vehicle logo; fourthly, Gabor filtering is carried out on the vehicle logo coarse positioning area, radiating net textures around the vehicle logo are inhibited, and the vehicle logo area is highlighted; fifthly, Gaussian filtering and mathematical morphology closing operation are carried out; and sixthly, selecting a threshold value to thresh the gray image, framing a detection target area, and realizing accurate positioning of the car logo. The vehicle logo detection method is short in detection time and high in detection rate.

Description

Car logo detection method based on Gabor filter background texture suppression
Technical Field
The invention relates to a car logo detection method, in particular to a car logo detection method based on Gabor filter background texture suppression.
Background
With the rapid increase of social economy, the current Chinese automobile consumption demand is increasingly vigorous, the number of automobiles is continuously increased, and traffic problems such as hit-and-run, vehicle theft and the like are brought. In order to determine illegal and violation vehicles, it is currently common to identify the license plate of the vehicle. However, in recent years, the occurrence of phenomena such as license plate overtaking, license plate reversing, license plate abrasion and license plate shielding makes the determination of the automobile by only recognizing the license plate unreliable. The car logo is a key image containing information of car models and manufacturers, and is an important basis for car classification and recognition. If the car logo can be accurately positioned, the accuracy of car classification and recognition can be effectively improved.
However, the car logo is rich in variety, various in shape and free of stable external features, and is located in an environment of a car grille area with complex textures. In addition, the car logo is susceptible to weather: at night or under the condition of insufficient illumination in rainy days, the car logo is difficult to identify; under the condition of strong light, the car logo is very easy to reflect light. The characteristics make the car logo have great difficulty in positioning and very challenging. The existing car logo positioning method is not mature enough, and has the problems of low detection rate, easy interference caused by illumination change and the like, so the car logo detection effect is still to be further improved.
Disclosure of Invention
The purpose of the invention is as follows: aiming at the problems in the prior art, the invention provides a car logo detection method based on Gabor filter background texture suppression.
The technical scheme is as follows: the car logo detection method based on Gabor filter background texture suppression comprises the following steps:
(1) the method comprises the steps of carrying out vehicle symmetry axis detection and inclination correction on a vehicle image which is shot and acquired and has a certain inclination angle based on an SIFT operator;
(2) training by using a machine learning algorithm of Harr + AdaBoost to obtain a cascade classifier, and positioning a license plate region from the corrected vehicle image by using the cascade classifier;
(3) based on prior knowledge, obtaining a vehicle logo coarse positioning area containing vehicle logo patterns in the positioned license plate area according to the position relation between the license plate and the vehicle logo;
(4) gabor filtering is carried out on the vehicle logo coarse positioning area, radiating network textures around the vehicle logo are inhibited, and the vehicle logo area is highlighted;
(5) performing Gaussian filtering and mathematical morphology closing operation on the car logo area;
(6) and (5) selecting a threshold value pair to threshold the gray level image obtained in the step (5), and framing the detection target area to obtain the precisely positioned car logo.
Has the advantages that: compared with the prior art, the invention has the following remarkable advantages:
1) the detection precision is high: the method has certain anti-interference performance on illumination change, has higher detection rate under different illumination conditions, and has better positioning effect on the car logo under the illumination conditions such as direct sunlight and night (especially under strong light);
2) the real-time property is as follows: the method has higher detection speed at night or under the condition of insufficient illumination, under the condition of direct sunlight and under the condition of uniform illumination, and can realize online real-time processing on the image data captured at the entrance of the highway;
3) the application range is wide: the traditional background texture suppression algorithm is that the texture direction of a vehicle logo coarse positioning area is judged firstly, then different vehicle logo positioning algorithms are adopted according to different texture directions, once the vehicle logo texture direction is judged wrongly in the traditional method, the wrong positioning method is used, and positioning failure is inevitably caused.
Drawings
FIG. 1 is a schematic flow chart of a car logo detection method based on Gabor filter background texture suppression according to the present invention;
FIG. 2 is a schematic flow chart of license plate location;
fig. 3 is a schematic diagram of a filtering flow of the Gabor filter.
Detailed Description
As shown in fig. 1, the car logo detection method based on Gabor filter background texture suppression of the present embodiment includes:
(1) and carrying out vehicle symmetry axis detection and inclination correction on the shot vehicle image with a certain inclination angle based on an SIFT operator.
The method specifically comprises the following steps:
(1-1) acquiring a point vector of each feature point of the vehicle image, wherein the point vector of the feature point i is defined as
Figure BDA0001175233340000021
xi,yiThe coordinates of the point are represented by,
Figure BDA0001175233340000025
indicates the direction, siRepresenting scale information, i is 1, …, n, n is the total number of the feature points;
(1-2) mixingDirection of characteristic point
Figure BDA0001175233340000022
After normalization, the corresponding feature descriptor k is obtainediI is 1, …, n, and generating SIFT feature point vectors with preset dimensions from feature descriptors of all points;
(1-3) by directly modifying the feature descriptor kiGenerating mirror image miAnd generating possible symmetrical feature point pairs (p) by matching the feature points and the mirror imagesi,pj)，i，j＝1,…,n，i≠j；
(1-4) calculating each pair of possible symmetrical feature point pairs (p)i,pj) Angle confidence of (phi)ijScale confidence SijAnd distance confidence DijThen, the total confidence coefficient M is obtained by calculationi,jI, j ≠ 1, …, n, i ≠ j; wherein:
angle confidence phiijThe calculation formula of (2) is as follows:
Figure BDA0001175233340000023
wherein the content of the first and second substances,
Figure BDA0001175233340000024
are respectively a point piAnd point pjDirection of (a), thetaijIs a point piTo point pjThe direction of (a);
scale confidence SijBy quantizing piAnd pjMesoscale similarity siAnd sjTo obtain:
Figure BDA0001175233340000031
wherein sigmasIs a scale factor, is a point p along the envelope of the Gaussian functioniTo point pjThe standard deviation of the direction of (a);
distance confidence DijComprises the following steps:
Figure BDA0001175233340000032
σdis a distance boundary, d is a symmetry point piTo point pjThe distance of (d);
total confidence Mi,jComprises the following steps:
Figure BDA0001175233340000033
(1-5) calculating each pair of symmetrical characteristic point pairs (p)i,pj) Axis of symmetry rij：rij＝xccosθij+ycsinθijWherein x isc，ycAre respectively a pair of characteristic points (p)i,pj) Length in x-and y-directions, θijIs a point piTo point pjThe direction of (a);
(1-6) finding a main symmetry axis by using a linear Hough transform, wherein each pair of symmetric characteristic point pairs (p)i,pj) For points (r) in Hough spaceij,θij) By weight Mi,jVoting to obtain a main symmetry axis;
(1-7) taking the angle of the main symmetry axis as the vehicle inclination angle theta;
(1-8) performing inclination correction on the vehicle image according to the vehicle inclination angle theta, wherein the corrected image is as follows:
Figure BDA0001175233340000034
width and height which are the width and height of the original image, and width' and height which are the width and height of the corrected image.
(2) And training by using a machine learning algorithm of Harr + AdaBoost to obtain a cascade classifier, and positioning the license plate region from the corrected vehicle image by using the cascade classifier.
The step (2) specifically comprises the following steps:
(2-1) collecting a preset number of positive samples and negative samples, and respectively establishing a positive sample library and a negative sample library, wherein the positive samples refer to license plates cut out from high-definition vehicle photos shot at a road gate, and the negative samples refer to background samples randomly cut out from non-license plate areas in the vehicle photos and contain various background environments;
(2-2) extracting Harr characteristics of all positive and negative samples, and training a plurality of strong classifiers of an AdaBoost algorithm by using the Harr characteristics;
the training method of the strong classifier of the AdaBoost algorithm comprises the following steps:
(2-2-1) setting a training set S { (a)1,b1),...,(am,bm) Contains m samples, where aiE.a (i ═ 1, 2.. times.m) denotes training samples, a is the set of training samples, biE B is aiCorresponding discrimination indicator, and B ═ 1, -1, for the jth weak classifier h of the ith training samplej(ai) Is shown as
Figure BDA0001175233340000041
Wherein, Fj(ai) Representing the value of the jth Haar feature in the subwindow, δjIndicating a set threshold value, pjA quantity representing a direction of controlling the unequal sign;
(2-2-2) training the weak classifier into a strong classifier by using an AdaBoost algorithm, and specifically comprising the following steps:
initializing sample weights:
Figure BDA0001175233340000042
wherein, w1(i) Representing the initial weight of the ith sample in the first round of training, p representing the total number of positive samples in S, q representing the total number of negative samples in S, wherein p + q is m, and m is the total number of samples;
for T ═ 1,2, …, T (T is the number of iterations), the following loop is performed:
① weight normalization
Figure BDA0001175233340000043
Wherein, wt(i) Representing the weight of the ith sample in the t round of training, wherein i is 1, 2.
② training weak classifier h of feature jjCalculating its weighted error εjI.e. by
Figure BDA0001175233340000044
And selects the classifier h with the lowest weighted errort minAs a classifier for this cycle;
③ the sample weights are updated according to the following formula:
Figure BDA0001175233340000045
wherein when the classification is correct, ht(ai)＝bi，eiWhen the classification is wrong, h is 0t(ai)≠bi，ei＝1；
(2-2-3) obtaining the final strong classifier as follows
Figure BDA0001175233340000051
Wherein
Figure BDA0001175233340000052
a is the window to be inspected, ht(a) The weak classifiers obtained in the t-th round of training are shown, and the result of H (a) is 1, which indicates acceptance, and 0, which indicates rejection.
(2-3) constructing a license plate detection cascade classifier by a plurality of strong classifiers obtained by training in a cascade mode;
(2-4) as shown in FIG. 2, detecting a license plate region in the vehicle image by using a trained license plate detection cascade classifier;
and (2-5) carrying out error picking and screening on the detected license plate region by adopting the license plate proportion characteristic and the license plate position characteristic as the basis.
And (2-5) eliminating the false detection area after positioning.
Wherein, the license plate proportion is characterized in that: the license plate is rectangular, the license plate has fixed word number and fixed size fonts, the actual width and height are 44 cm and 14 cm respectively, and the aspect ratio in the image is basically 3: 1 or so, the algorithm can set size range limits for the width and the height;
the position characteristics of the license plate: because the actual road monitoring camera is triggered and shot by the fixed position ground sensing coil signal, and the installation position of the license plate is below the whole vehicle, the vertical position of the license plate in the whole image is basically stable (more tends to appear at the lower half part of the image), the average license plate center position can be obtained according to statistical information, and the probability that the candidate area which is farther away from the center position becomes the license plate is lower.
(3) The topological structures of all parts of the vehicle head can be observed to find that: the car logo is characterized by being located at the most easily distinguishable position, the biggest difference between the car logo and other interference areas is that the car logo cannot be located at other positions except above a license plate, and the positions of the interference areas are random and uncertain, so that the rough range of the car logo can be roughly located according to the position of the license plate; through a large number of vehicle picture tests, according to the position relation of the license plate and the vehicle logo, a vehicle logo coarse positioning area containing vehicle logo patterns is obtained in the positioned license plate area as follows:
Figure BDA0001175233340000053
wherein, X1、X2、Y1、Y2Left and right and upper and lower boundaries, X, of rough range of vehicle logo obtained by rough positioningleft、XrightAre the left and right boundaries of the license plate region, YupThe upper boundary of the license plate region is height, and N is an optional height coefficient, which is generally taken as N-3.
(4) And carrying out Gabor filtering on the coarse positioning area of the car logo, inhibiting the texture of the heat dissipation net around the car logo and highlighting the car logo area.
As shown in fig. 3, the step (4) specifically includes the following steps:
(4-1) defining the two-dimensional Gabor kernel function h (x, y) of the Gabor filter as
Figure BDA0001175233340000061
Wherein u is0As a function of center frequency, σx、σyIs a scale factor, which is the standard deviation of the envelope of the Gaussian function along the directions of the x axis and the y axis, exp (2 pi ju)0R1) Is an oscillation function, the real part is a cosine function, the imaginary part is a sine function,
Figure BDA00011752333400000611
is a set direction;
by changing the Gabor kernel function direction
Figure BDA0001175233340000064
Texture information in different directions in the image can be extracted,
Figure BDA0001175233340000069
has a value range of
Figure BDA0001175233340000066
Within this range
Figure BDA0001175233340000065
The values may describe all directions, i.e.
Figure BDA0001175233340000067
And
Figure BDA0001175233340000068
the invention does not select the extraction of the texture information on multiple scales, but selects 6 Gabor filters in different directions to form a filter bank under a specific scale with good effect, and the directions are pi/6, pi/4, pi/3, 2 pi/3, 3 pi/4 and 5 pi/6 respectively.
(4-2) initializing Gabor filter: gabor filters { h) of 6 different directions are respectively constructedl(x, y) | l ═ 1,. ·,6}, thereby constituting a filter bank in which directions are set forth
Figure BDA00011752333400000610
The dimension parameters and bandwidth parameters of the Gabor filter set to be pi/6, pi/4, pi/3, 2 pi/3, 3 pi/4, 5 pi/6 and 6 directions respectively are set as follows: dimension σx、σyAre all set to 1, the bandwidth Sigma is set to 2 pi, and Sigma ═ Sigma u0So the center frequency u0Is 2 pi;
(4-3) respectively carrying out convolution operation on the vehicle logo coarse positioning area image obtained in the step (3) and two-dimensional Gabor kernel functions h (x, y) in 6 different directions and carrying out modulus extraction to obtain 6 filtering images G1(x,y)，G2(x,y)，G3(x,y)，G4(x,y)，G5(x,y)，G6(x, y), the final output image of the filter bank is G (x, y) ═ G1(x,y)+G2(x,y)+G3(x,y)+G4(x,y)+G5(x,y)+G6(x, y))/6, wherein Gl(x,y)＝∫∫f(x0,y0)hl(x-x0,y-y0)dx0dy0In the formula, f (x, y) is an image of the vehicle logo coarse positioning area;
(4-4) calculating the attenuation coefficient lambda of each pixel point pk of the final output image G (x, y) of the filter bank obtained in the step (4-3)pkAnd k is 1., num, num is the total number of the pixel points, and the calculation steps are as follows:
① obtaining an image I containing horizontal textures by respectively using a Sobel horizontal edge detection method and a Sobel vertical edge detection method in a coarse positioning area obtained after license plate positioningxAnd an image I comprising vertical texturey；
② calculating the attenuation coefficient of each pixel pk as:
Figure BDA0001175233340000071
(xpk,ypk) The coordinates of a pixel point pk are α and β are two constant coefficients, α is 0.5, β is 2, width _ pai is the width of a license plate region obtained by positioning, and the formula shows that lambda ispkSize and I ofx/IyAnd xpkThe following steps are involved: i isx/IyThe closer to 0 or infinity (corresponding to the case of horizontal or vertical texture), λpkThe smaller, for the same reason, xpkThe more deviated from width _ pai/2 (the common symmetry axis of the car logo and the license plate), the more lambdapkThe smaller the size of the tube is,
(4-5) multiplying each pixel point of the wave filter group output image G (x, y) by the corresponding attenuation coefficient lambdapk。
(5) And performing Gaussian filtering and mathematical morphology closing operation on the car logo area.
And (4) filtering noise of the image obtained in the step (4) by using a two-dimensional discrete Gaussian filter function, and performing gray-scale morphological close operation to close the discontinuity or fracture of the detected target and eliminate fine holes in the target.
(6) And (5) selecting a threshold value pair to threshold the gray level image obtained in the step (5), and framing the detection target area to obtain the precisely positioned car logo.
The step (6) specifically comprises the following steps:
(6-1) selecting the image T (x) obtained in the step (5)pk,ypk) 1/2 of the mean value of the pixel gray values as a threshold value
Figure BDA0001175233340000074
Namely, it is
Figure BDA0001175233340000072
(6-2) thresholding by using a threshold value, wherein the thresholded image is
Figure BDA0001175233340000073
And (6-3) framing and detecting a target area from the thresholded image, wherein the target area is a car logo area.

Claims (7)

1. A car logo detection method based on Gabor filter background texture suppression is characterized by comprising the following steps:
(1) the method comprises the steps of carrying out vehicle symmetry axis detection and inclination correction on a vehicle image which is shot and acquired and has a certain inclination angle based on an SIFT operator;
(2) training by using a machine learning algorithm of Harr + AdaBoost to obtain a cascade classifier, and positioning a license plate region from the corrected vehicle image by using the cascade classifier;
(3) based on prior knowledge, obtaining a vehicle logo coarse positioning area containing vehicle logo patterns in the positioned license plate area according to the position relation between the license plate and the vehicle logo;
(4) gabor filtering is carried out on the vehicle logo coarse positioning area, radiating network textures around the vehicle logo are inhibited, and the vehicle logo area is highlighted; the method specifically comprises the following steps:
(4-1) defining the two-dimensional Gabor kernel function h (x, y) of the Gabor filter as
Figure FDA0002227550510000011
Wherein u is0As a function of center frequency, σx、σyIs a scale factor, which is the standard deviation of the envelope of the Gaussian function along the directions of the x axis and the y axis, exp (2 pi ju)0R1) Is an oscillation function, the real part is a cosine function, the imaginary part is a sine function,
Figure FDA0002227550510000012
Figure FDA0002227550510000013
is a set direction;
(4-2) initializing Gabor filter: gabor filters { h) of 6 different directions are respectively constructedl(x, y) | l ═ 1,. ·,6}, thereby constituting a filter bank in which directions are set forth
Figure FDA0002227550510000014
The dimension parameters and bandwidth parameters of the Gabor filter set to be pi/6, pi/4, pi/3, 2 pi/3, 3 pi/4, 5 pi/6 and 6 directions respectively are set as follows: dimension σx、σyAre all set to 1, the bandwidth Sigma is set to 2 pi, and Sigma ═ Sigma u0Center frequency ofu0Is 2 pi;
(4-3) respectively carrying out convolution operation on the vehicle logo coarse positioning area image obtained in the step (3) and two-dimensional Gabor kernel functions h (x, y) in 6 different directions and carrying out modulus extraction to obtain 6 filtering images G1(x,y)，G2(x,y)，G3(x,y)，G4(x,y)，G5(x,y)，G6(x, y), the final output image of the filter bank is G (x, y) ═ G1(x,y)+G2(x,y)+G3(x,y)+G4(x,y)+G5(x,y)+G6(x, y))/6, wherein Gl(x,y)＝∫∫f(x0,y0)hl(x-x0,y-y0)dx0dy0In the formula, f (x, y) is an image of the vehicle logo coarse positioning area;
(4-4) calculating the attenuation coefficient lambda of each pixel point pk of the final output image G (x, y) of the filter bank obtained in the step (4-3)pkAnd k is 1., num, num is the total number of the pixel points, and the calculation steps are as follows:
① obtaining an image I containing horizontal textures by respectively using a Sobel horizontal edge detection method and a Sobel vertical edge detection method in a coarse positioning area obtained after license plate positioningxAnd an image I comprising vertical texturey；
② calculating the attenuation coefficient of each pixel pk as:
Figure FDA0002227550510000021
(xpk,ypk) α and β are two constant coefficients for the coordinates of the pixel point pk, and width _ pai is the width of the license plate region obtained by positioning;
(4-5) multiplying each pixel point of the wave filter group output image G (x, y) by the corresponding attenuation coefficient lambdapk；
(5) Performing Gaussian filtering and mathematical morphology closing operation on the car logo area;
(6) and (5) selecting a threshold value pair to threshold the gray level image obtained in the step (5), and framing the detection target area to obtain the precisely positioned car logo.
2. The method for detecting the car logo based on the Gabor filter background texture suppression according to claim 1, wherein the method comprises the following steps: the step (1) specifically comprises the following steps:
(1-1) acquiring a point vector of each feature point of the vehicle image, wherein the point vector of the feature point i is defined as
Figure FDA0002227550510000022
xi,yiThe coordinates of the point are represented by,
Figure FDA0002227550510000023
indicates the direction, siRepresenting scale information, i is 1, …, n, n is the total number of the feature points;
(1-2) Direction of feature points
Figure FDA0002227550510000024
After normalization, the corresponding feature descriptor k is obtainediI is 1, …, n, and generating SIFT feature point vectors with preset dimensions from feature descriptors of all points;
(1-3) by directly modifying the feature descriptor kiGenerating mirror image miAnd generating possible symmetrical feature point pairs (p) by matching the feature points and the mirror imagesi,pj)，i，j＝1,…,n，i≠j；
(1-4) calculating each pair of possible symmetrical feature point pairs (p)i,pj) Angle confidence of (phi)ijScale confidence SijAnd distance confidence DijThen, the total confidence coefficient M is obtained by calculationi,jI, j ≠ 1, …, n, i ≠ j; wherein:
angle confidence phiijThe calculation formula of (2) is as follows:
Figure FDA0002227550510000025
wherein the content of the first and second substances,
Figure FDA0002227550510000026
are respectively a point piAnd point pjDirection of (a), thetaijIs a point piTo point pjThe direction of (a);
scale confidence SijBy quantizing piAnd pjMesoscale similarity siAnd sjTo obtain:
Figure FDA0002227550510000027
wherein sigmasIs a scale factor, is a point p along the envelope of the Gaussian functioniTo point pjThe standard deviation of the direction of (a);
distance confidence DijComprises the following steps:
Figure FDA0002227550510000031
σdis a distance boundary, d is a symmetry point piTo point pjThe distance of (d);
total confidence Mi,jComprises the following steps:
Figure FDA0002227550510000032
(1-5) calculating each pair of symmetrical characteristic point pairs (p)i,pj) Axis of symmetry rij：rij＝xccosθij+ycsinθijWherein x isc，ycAre respectively a pair of characteristic points (p)i,pj) Length in x-and y-directions, θijIs a point piTo point pjThe direction of (a);
(1-6) finding a main symmetry axis by using a linear Hough transform, wherein each pair of symmetric characteristic point pairs (p)i,pj) For points (r) in Hough spaceij,θij) By weight Mi,jVoting to obtain a main symmetry axis;
(1-7) taking the angle of the main symmetry axis as the vehicle inclination angle theta;
(1-8) performing inclination correction on the vehicle image according to the vehicle inclination angle theta, wherein the corrected image is as follows:
Figure FDA0002227550510000033
width and height which are the width and height of the original image, and width' and height which are the width and height of the corrected image.
3. The method for detecting the car logo based on the Gabor filter background texture suppression according to claim 1, wherein the method comprises the following steps: the step (2) specifically comprises the following steps:
(2-1) collecting a preset number of positive samples and negative samples, and respectively establishing a positive sample library and a negative sample library, wherein the positive samples refer to license plates cut out from high-definition vehicle photos shot at a road gate, and the negative samples refer to background samples randomly cut out from non-license plate areas in the vehicle photos and contain various background environments;
(2-2) extracting Harr characteristics of all positive and negative samples, and training a plurality of strong classifiers of an AdaBoost algorithm by using the Harr characteristics;
(2-3) constructing a license plate detection cascade classifier by a plurality of strong classifiers obtained by training in a cascade mode;
(2-4) detecting a license plate region in the vehicle image by adopting a trained license plate detection cascade classifier;
and (2-5) false detection and screening are carried out on the detected license plate area according to the license plate proportion characteristic and the license plate position characteristic.
4. The method for detecting the car logo based on the Gabor filter background texture suppression according to claim 3, wherein the method comprises the following steps: the training method of the strong classifier of the AdaBoost algorithm in the step (2-2) comprises the following steps:
(2-2-1) setting a training set S { (a)1,b1),...,(am,bm) Contains m samples, where aiE.a represents a training sample, i is 1,2iE B is aiCorresponding discrimination indicator, and B ═ 1, -1, for the jth weak classifier h of the ith training samplej(ai) Is shown as
Figure FDA0002227550510000041
Wherein, Fj(ai) Representing the value of the jth Haar feature in the subwindow, δjIndicating a set threshold value, pjA quantity representing a direction of controlling the unequal sign;
(2-2-2) training the weak classifier into a strong classifier by using an AdaBoost algorithm, and specifically comprising the following steps:
initializing sample weights:
Figure FDA0002227550510000042
wherein, w1(i) Representing the initial weight of the ith sample in the first round of training, p representing the total number of positive samples in S, q representing the total number of negative samples in S, wherein p + q is m, and m is the total number of samples;
for T1, 2, T being the number of iterations, the following loop is performed:
① weight normalization
Figure FDA0002227550510000043
Wherein, wt(i) Representing the weight of the ith sample in the t round of training, wherein i is 1, 2.
② training weak classifier h of feature jjCalculating its weighted error εjI.e. by
Figure FDA0002227550510000044
And selects the classifier h with the lowest weighted errortminAs a classifier for this cycle;
③ the sample weights are updated according to the following formula:
Figure FDA0002227550510000045
wherein when the classification is correct, ht(ai)＝bi，ei＝0，When the classification is wrong, ht(ai)≠bi，ei＝1；
(2-2-3) obtaining the final strong classifier as follows
Figure FDA0002227550510000051
Wherein
Figure FDA0002227550510000052
a is the window to be inspected, ht(a) The weak classifiers obtained in the t-th round of training are shown, and the result of H (a) is 1, which indicates acceptance, and 0, which indicates rejection.
5. The method for detecting the car logo based on the Gabor filter background texture suppression according to claim 1, wherein the method comprises the following steps: the step (3) specifically comprises the following steps:
based on prior knowledge, according to the position relation of the license plate and the vehicle logo, a vehicle logo coarse positioning area containing a vehicle logo pattern is obtained in the positioned license plate area, wherein the vehicle logo coarse positioning area is as follows:
Figure FDA0002227550510000053
wherein, X1、X2、Y1、Y2Left and right and upper and lower boundaries, X, of rough range of vehicle logo obtained by rough positioningleft、XrightAre the left and right boundaries of the license plate region, YupThe upper boundary of the license plate area, height is the height of the license plate area, and N is a selectable height coefficient.
6. The method for detecting the car logo based on the Gabor filter background texture suppression according to claim 1, wherein the method comprises the following steps: the step (5) specifically comprises the following steps:
and (4) filtering noise of the image obtained in the step (4) by using a two-dimensional discrete Gaussian filter function, and performing gray-scale morphological closed operation to close the discontinuity or fracture of the detected target and eliminate fine holes in the target.
7. The method for detecting the car logo based on the Gabor filter background texture suppression according to claim 1, wherein the method comprises the following steps: the step (6) specifically comprises the following steps:
(6-1) selecting the image T (x) obtained in the step (5)pk,ypk) 1/2 of the mean value of the pixel gray values as a threshold value
Figure FDA0002227550510000054
Namely, it is
Figure FDA0002227550510000055
(6-2) thresholding by using a threshold value, wherein the thresholded image is
Figure FDA0002227550510000061
And (6-3) framing and detecting a target area from the thresholded image, wherein the target area is a car logo area.
CN201611126129.2A 2016-12-09 2016-12-09 Car logo detection method based on Gabor filter background texture suppression Active CN106778742B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201611126129.2A CN106778742B (en) 2016-12-09 2016-12-09 Car logo detection method based on Gabor filter background texture suppression

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201611126129.2A CN106778742B (en) 2016-12-09 2016-12-09 Car logo detection method based on Gabor filter background texture suppression

Publications (2)

Publication Number Publication Date
CN106778742A CN106778742A (en) 2017-05-31
CN106778742B true CN106778742B (en) 2020-03-31

Family

ID=58877560

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201611126129.2A Active CN106778742B (en) 2016-12-09 2016-12-09 Car logo detection method based on Gabor filter background texture suppression

Country Status (1)

Country Link
CN (1) CN106778742B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108961137B (en) * 2018-07-12 2023-05-23 中山大学 Image steganalysis method and system based on convolutional neural network
CN111915897B (en) * 2019-05-10 2021-11-16 北京万集科技股份有限公司 Method and device for determining position of identification area, storage medium and electronic device
CN110867083B (en) * 2019-11-20 2021-06-01 浙江宇视科技有限公司 Vehicle monitoring method, device, server and machine-readable storage medium
CN111723335B (en) * 2020-05-21 2023-03-24 河海大学 Target symmetry axis detection method based on concentric circumference filter
CN113873442B (en) * 2021-09-08 2023-08-04 宁波大榭招商国际码头有限公司 Positioning method for external collection card
CN114220101A (en) * 2021-11-25 2022-03-22 慧之安信息技术股份有限公司 Method and system for reducing interference of natural light change on license plate marker extraction

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101520841A (en) * 2009-03-10 2009-09-02 北京航空航天大学 Real-time and anti-interference method for positioning license plate in high-definition TV video
CN101937508A (en) * 2010-09-30 2011-01-05 湖南大学 License plate localization and identification method based on high-definition image
CN102419820A (en) * 2011-08-18 2012-04-18 电子科技大学 Method for rapidly detecting car logo in videos and images
CN104156692A (en) * 2014-07-07 2014-11-19 叶茂 Automobile logo sample training and recognition method based on air-inlet grille positioning
CN105740886A (en) * 2016-01-25 2016-07-06 宁波熵联信息技术有限公司 Machine learning based vehicle logo identification method

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7840059B2 (en) * 2006-09-21 2010-11-23 Microsoft Corporation Object recognition using textons and shape filters

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101520841A (en) * 2009-03-10 2009-09-02 北京航空航天大学 Real-time and anti-interference method for positioning license plate in high-definition TV video
CN101937508A (en) * 2010-09-30 2011-01-05 湖南大学 License plate localization and identification method based on high-definition image
CN102419820A (en) * 2011-08-18 2012-04-18 电子科技大学 Method for rapidly detecting car logo in videos and images
CN104156692A (en) * 2014-07-07 2014-11-19 叶茂 Automobile logo sample training and recognition method based on air-inlet grille positioning
CN105740886A (en) * 2016-01-25 2016-07-06 宁波熵联信息技术有限公司 Machine learning based vehicle logo identification method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
Detecting Symmetry and Symmetric Constellations of Features;Gareth Loy et al.;《ECCV"06 Proceedings of the 9th European conference on Computer Vision-Volume Part II》;20060531;正文第3.1-3.2节 *

Also Published As

Publication number Publication date
CN106778742A (en) 2017-05-31

Similar Documents

Publication Publication Date Title
CN106778742B (en) Car logo detection method based on Gabor filter background texture suppression
CN109740478B (en) Vehicle detection and identification method, device, computer equipment and readable storage medium
CN108319964B (en) Fire image recognition method based on mixed features and manifold learning
CN108510467B (en) SAR image target identification method based on depth deformable convolution neural network
WO2020173022A1 (en) Vehicle violation identifying method, server and storage medium
JP4723582B2 (en) Traffic sign detection method
Chen et al. Vehicle type categorization: A comparison of classification schemes
CN101859382B (en) License plate detection and identification method based on maximum stable extremal region
CN106650731B (en) Robust license plate and vehicle logo recognition method
CN106682586A (en) Method for real-time lane line detection based on vision under complex lighting conditions
CN105740886B (en) A kind of automobile logo identification method based on machine learning
CN106384345B (en) A kind of image detection and flow statistical method based on RCNN
CN107194393B (en) Method and device for detecting temporary license plate
Agrawal et al. Target detection in SAR images using SIFT
CN107784263B (en) Planar rotation face detection method based on improved accelerated robust features
Liu et al. Multi-type road marking recognition using adaboost detection and extreme learning machine classification
CN109858438A (en) A kind of method for detecting lane lines based on models fitting
CN106407951A (en) Monocular vision-based nighttime front vehicle detection method
Xiao et al. Traffic sign detection based on histograms of oriented gradients and boolean convolutional neural networks
CN111368742A (en) Double-yellow traffic marking reconstruction identification method and system based on video analysis
CN104966064A (en) Pedestrian ahead distance measurement method based on visual sense
CN110516547B (en) Fake-licensed vehicle detection method based on weighted non-negative matrix factorization
CN111832463A (en) Deep learning-based traffic sign detection method
CN108647679B (en) Car logo identification method based on car window coarse positioning
CN115512310A (en) Vehicle type recognition method and system based on face features under video monitoring

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant