CN104156692A - Automobile logo sample training and recognition method based on air-inlet grille positioning - Google Patents

Abstract

The invention discloses an automobile logo sample training and recognition method based on air-inlet grille positioning. The training process comprises the following steps: a, obtaining air-inlet grille samples of different automobiles and converting the samples to grayness images; b, classifying the air-inlet grille samples obtained in the step a according to automobile brands, and respectively performing normalization gauge processing on various sample sets to enable the air-inlet grille samples to be unified to the same size; c, conducting Gabor function conversion on the various samples obtained in the step b to perform size conversion and rotation transformation to obtain a Gabor wavelet expression. The automobile logo sample training and recognition method has the following benefits: 1, the automobile air-inlet grille, rather than the automobile logo, is adopted as the identification target, so that more effective image information is guaranteed for identification of automobile brands; 2, the Gabor wavelet conversion is utilized to keep texture information in each direction of the air-inlet grille; 3, the template library for classifying provided by the invention has an elastic structure, so that flexible expansion and update can be realized.

Description

Based on automobile mark sample training and the recognition methods of air-inlet grille location

Technical field

The invention belongs to technical field of computer vision, relate to a kind of automobile mark recognition methods, be specifically related to a kind ofly realize automobile mark sample training method and the car mark automatic identifying method based on this sample training based on automotive air intake grid location.

Background technology

Along with expanding economy, the increase of people's income, automobile has entered increasing family and individual life, automobile is due to its convenient maneuverability flexibly, progressively become the preferred traffic instrument of people's trip, the demand driving of people to automobile the development of automobile industry, automobile market also thus competition increasingly sharpen.

Mass consumption person, in the time choosing automobile, in considering the factors such as the security of automobile, performance configuration, price, also considers emphatically the outward appearance of automobile.The demand of consumer to automobile, impel each main flow car manufactures constantly to innovate in automobile production technique, simultaneously, progressively more pay attention to oneself Automobile Design culture, strengthen brand recognition separately by distinctive appearance design separately, big-and-middle-sized automobile brand manufacturer is all that automobile product is made unique " types of facial makeup in Beijing operas " separately.

For traffic control department, the universal road vehicles that causes of automobile is more and more, brings huge challenge to traffic monitoring.To identify industrialization degree of ripeness higher for car plate now, the detection of car plate and discrimination allow this technology extensively commercial, but in traffic monitoring, for vehicle brand more the discriminator technology of refinement be but far from reaching commercialization degree, can carry out effectively identification to vehicle brand and have great significance for vehicle monitoring for vehicle supervision department.For example vehicle is set out on a journey and is travelled, must vehicle corresponding one by one with car plate, inconsistent if the information such as vehicle brand and corresponding car plate are registered this vehicle, may exist lawless person illegally to apply mechanically the behavior of the violation road traffic laws and regulationses such as car plate.

Prior art one related to the present invention: the Liu Zhi of Sichuan Chuandazhisheng Software Co., Ltd virtue waits people's application for a patent for invention, " utilizing the method for digital image processing techniques identification type of vehicle ", on Dec 22 2006 applying date, publication number is CN100485710C.

For automobile brand, identification adopts after the headstock of location in this invention, from headstock, orient all parts of headstock, according to the each image of component gray scale of headstock and texture information, utilize the outstanding headstock all parts of multiple dimensioned local energy function and gray level skeleton function information to realize cutting apart of each parts, to orient car mark, in car mark region, carry out stencil matching, the maximum coupling of each class template is only sorted, by maximum matching value and set threshold, maximum matching area is carried out to yardstick invariant feature extraction, according to the feature of extracting, car mark is identified.In image recognition flow process, identify the brand of vehicle according to the global feature information of the car mark obtaining and headstock.Defect is: 1, car is marked on headstock part to occupy region less, is difficult to ensure card for the effective Detection accuracy of car target; 2, the car target of different automobile brands truly has larger discrimination, but because its shape size is different, is difficult to effective tissue templates, and near inactive area car mark is excessive with respect to car sample body regional percentage, for car target template matches is brought larger challenge.

Prior art two related to the present invention: the people's such as the Sun Ming of the Third Research Institute of Ministry of Public Security rosy clouds application for a patent for invention, " a kind of car identifies other device and method ", on November 12 2013 applying date, publication number is CN103559492A.

This invention car mark identification module comprises feature extraction and Classification and Identification module, and its characteristic extracting module is for extracting mark on a map the constant Gradient Features in region of picture of the car of orienting.The car navigating to is marked on a map and looked like the quartern in the horizontal direction, vertical direction halves, thus image is divided into eight regions, calculate Grad and the gradient direction of each pixel that car is marked on a map in picture, and gradient direction is referred in 8 direction unit, add up respectively the quantity that is grouped into the pixel of the unit of 8 directions in eight regions, produce the constant Gradient Features vector in region of a 8*8 dimension, described learning classification device is the learning classification device based on algorithm of support vector machine.Defect is: car mark is divided into 8 unit by this invention, causes every subregion pixel less, and gradient information is not remarkable.

Meanwhile, in the evolution of auto industry design, the car manufactures of main flow when brand automobile, is not only limited to brand car target separately and solidifies under design, and the characteristic of air-inlet grille for automobile also presents its intrinsic inheritance separately.Represent the air-inlet grille of each automobile brand individual character, become " family's types of facial makeup in Beijing operas " of some auto vendor, experience personage, as long as observe automobile face, need not scrutinize car mark and can tell the affiliated brand of this automobile.Under U.S. JEEP board, different cars are, Cherokee as large in Fig. 1-1, and Fig. 1-2 guide person, Fig. 1-3 are visitor freely, and its air-inlet grille all has " seven vertical cores " design of JEEP family uniqueness.Germany's BMW board automobile as Fig. 1-4 BMW 1 be that Fig. 1-5 BMW 3 is that Fig. 1-6 BMW X5 etc. all has the air-inlet grille design of BMW peculiar " two kidney " shape.U.S.'s Chevrolet board automobile is as Fig. 1-7 Cruz, and Fig. 1-8 step sharp treasured, and the Sail air-inlet grille that is of waiting for bus in Fig. 1-9 all has the distinctive peltate moulding of Chevrolet, and " golden collar's knot " of middle roomy horizontal stripe and Chevrolet designs.

Summary of the invention

The object of the invention is to be difficult to ensure the problems such as the large and processing procedure gradient information of card, template matches difficulty is not obvious in order to solve car mark recognition methods car mark recognition accuracy in existing and background technology, proposed a kind of automobile mark sample training and recognition methods based on air-inlet grille location.

Technical scheme of the present invention is: a kind of automobile mark sample training method based on air-inlet grille location, it is characterized in that,

Training process comprises the following steps:

A, obtain the air-inlet grille sample of different automobile types, and be converted to gray level image;

B, the air-inlet grille sample in step a is classified according to automobile brand, and respectively Different categories of samples collection is normalized to scale processing, making air-inlet grille sample unified is same scale size;

C, Different categories of samples in step b is carried out to Gabor functional transformation, it is carried out to yardstick and rotational transform, obtain Gabor small echo expression formula;

D, by the Gabor wavelet character vector cascade of different directions, form the Gabor proper vector after the cascade of whole air-inlet grille sample;

E, by after set of eigenvectors mark, use SVM instrument to train, obtain the svm classifier device of air-inlet grille for automobile.

Further, in above-mentioned steps a, different automobile types refers to the difference of automobile brand and/or style.

Further, consider that air-inlet grille all has symmetry, in step a, only obtain image that air-inlet grille is positioned at axis of symmetry either side as sample.

Further, above-mentioned sample is square area.

Further, in above-mentioned steps c, Gabor functional transformation detailed process is: all kinds of brand air-inlet grille samples in step b are carried out to Gabor wavelet transformation, and g (x, y) is female wave function,

$g(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp [-\frac{1}{2}(\frac{x^2}{{\mathrm{\σ}}_x^2}+\frac{y^2}{{\mathrm{\σ}}_y^2})+2\mathrm{\πjWx}]$

It is carried out to yardstick and rotational transform, obtains Gabor small echo expression formula:

g _m,n(x,y)＝a ^-mG(x',y')

Wherein:

(x',y')＝a ^-m(xcosθ+ysinθ,-xsinθ+ycosθ)

a＞1

m,n∈Z

θ＝nπ/M

M is direction number, a ^-mfor scale factor.

Further, in above-mentioned steps c, all kinds of brand air-inlet grille samples are carried out to 0 degree, 45 degree, 90 degree, 135 degree and carry out Gabor filtering.

Automobile mark recognition methods based on above-mentioned sample training method, is characterized in that, comprises step:

F, locate the headstock region of automobile in pending frame of video;

G, headstock regional compartmentalization is processed, to dwindle region to be identified retaining in air-inlet grille region;

H, use moving window scan described region to be identified, and described moving window yardstick is consistent with step b Normalized Scale sample after treatment in training method;

I, 0 degree, 45 degree, 90 degree and 135 degree Gabor filtering are carried out in moving window region;

J, filtered Gabor proper vector is carried out to cascade, form the Gabor proper vector after classification moving window to be identified cascade;

K, Gabor proper vector after using svm classifier device in training method to described cascade are carried out svm classifier;

L, return to svm classifier result, obtain the brand of object vehicle.

Further, first above-mentioned steps g intercepts half region as region to be identified using axis of symmetry as boundary according to the symmetry of headstock.

Beneficial effect of the present invention: 1, the region of air-inlet grille for automobile is much larger than automobile mark, utilize air-inlet grille for automobile but not car mark as identifying object, ensured more effective image information, be beneficial to automobile brand identification; 2, utilize Gabor wavelet transformation, in different directions to the filtering of air-inlet grille region, air-inlet grille provincial characteristics vector is formed by the cascade of different directions Gabor wavelet character, has retained the texture information in air-inlet grille all directions; 3, the present invention has elastic construction for the template base of classification, can expand flexibly and upgrade.

Brief description of the drawings

Fig. 1-1, Fig. 1-2 and Fig. 1-3 are respectively the air-inlet grille structural representation of large Cherokee, guide person and free passenger car type under jeep brand;

Fig. 1-4, Fig. 1-5 and Fig. 1-6 are respectively German BMW board automobile 1 and are, 3 are and the air-inlet grille structural representation of X5 vehicle;

Fig. 1-7, Fig. 1-8 and Fig. 1-9 are respectively U.S.'s Chevrolet board automobile Cruz, step the air-inlet grille structural representation of sharp treasured and Sail car system;

Fig. 2-1 is based on air-inlet grille identification automobile brand method training process flow diagram;

Fig. 2-2 are based on air-inlet grille identification automobile brand method discriminator process flow diagram;

Fig. 3-1, Fig. 3-2 and Fig. 3-3 are respectively Jeep air-inlet grille sample, BMW air-inlet grille sample and Chevrolet air-inlet grille sample;

Fig. 3-4, Fig. 3-5, Fig. 3-6 and Fig. 3-7 are respectively at filtering angle theta=0 °, and 45 °, 90 °, Jeep automotive air intake grid sample instance 135 ° time;

Fig. 3-8, Fig. 3-9, Fig. 3-10 and Fig. 3-11 are respectively in filtering angle θ=0 °, 45 °, 90 °, 135 °, time BMW air intake grid sample instance;

As Fig. 3-12, Fig. 3-13, Fig. 3-14, Fig. 3-15 are in filtering angle θ=0 °, 45 °, 90 °, 135 °, time BMW air intake grid sample instance.

Embodiment

Embodiments of the invention are that principle according to the present invention designs, and below in conjunction with accompanying drawing and specific embodiment, the invention will be further elaborated.

As shown in Fig. 2-1 and Fig. 2-2, based on the automobile mark sample training method of air-inlet grille location, training process comprises the following steps:

A, obtain the air-inlet grille sample of different automobile types, and be converted to gray level image.Be specifically A board a by different the subordinaties such as known automobile A brand, B brand, C brand cars _ietc. type car, B board b _ietc. type car, C board c _ipress automobile brand classification etc. the air-inlet grille of type car, form the sample set of same brand different automobile types.Consider that air-inlet grille all has symmetry, therefore air-inlet grille sample can intercept all, also can only intercept half.When air-inlet grille sample intercepts a half, intercept quadrate region, and be converted into gray-scale map.B, the air-inlet grille sample in step a is classified according to automobile brand, and respectively Different categories of samples collection is normalized to scale processing, making air-inlet grille sample unified is same scale size.C, Different categories of samples in step b is carried out to Gabor functional transformation, it is carried out to yardstick and rotational transform, obtain Gabor small echo expression formula; Gabor functional transformation detailed process is: all kinds of brand air-inlet grille samples in step b are carried out to Gabor wavelet transformation, and g (x, y) is female wave function,

$g(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp [-\frac{1}{2}(\frac{x^2}{{\mathrm{\σ}}_x^2}+\frac{y^2}{{\mathrm{\σ}}_y^2})+2\mathrm{\πjWx}]$

It is carried out to yardstick and rotational transform, obtains Gabor small echo expression formula:

g _m,n(x,y)＝a ^-mG(x',y')

Wherein:

(x',y')＝a ^-m(xcosθ+ysinθ,-xsinθ+ycosθ)

a＞1

m,n∈Z

θ＝nπ/M

M is direction number, a ^-mfor scale factor.

According in Gabor filter angles being θ while spending, can significantly extract air-inlet grille sample correspondence the edge of direction, other angles are not obvious this feature, extracts respectively the filtered proper vector of Gabor of sample different directions.In the present invention program, respectively all kinds of brand air-inlet grille samples in b are carried out to 0 degree, 45 degree, 90 degree, 135 degree and carry out Gabor filtering.

D, by the Gabor wavelet character vector cascade of different directions, form the Gabor proper vector after the cascade of whole air-inlet grille sample.E, by after the set of eigenvectors mark of different brands air-inlet grille for automobile sample in steps d, use SVM instrument to train, obtain the svm classifier device of air-inlet grille for automobile.

Automobile mark recognition methods based on above-mentioned sample training method, comprises step: f, locate the headstock region of automobile in pending frame of video.G, headstock regional compartmentalization is processed, to dwindle region to be identified retaining in air-inlet grille region.This step is preferably the symmetry of automobile head and the feature to the design of vehicle head assembly of making full use of, suitably dwindle headstock identified region, get and in step f, detect that headstock region carries out empirical division, in retaining air-inlet grille approximate region, dwindle region to be identified as far as possible.H, use moving window scan described region to be identified, and described moving window yardstick is consistent with step b Normalized Scale sample after treatment in training method.I, 0 degree, 45 degree, 90 degree and 135 degree Gabor filtering are carried out in moving window region.J, filtered Gabor proper vector is carried out to cascade, form the Gabor proper vector after classification moving window to be identified cascade.K, Gabor proper vector after using svm classifier device in training method to described cascade are carried out svm classifier.L, return to svm classifier result, obtain the brand of object vehicle.

Described embodiment is mainly convenient to understanding of the present invention, and concrete automobile brand can not identified to strict restriction effect.

Be below further preferred implementation:

A, in the present embodiment, be freely visitor, BMW X5, Chevrolet Cruz etc. of Jeep by different the subordinaties such as known automobile brand Jeep, BMW, Chevrolet cars, intercept air-inlet grille sample, consider that air-inlet grille all has symmetry, therefore air-inlet grille sample only intercepts half, air-inlet grille sample intercepts square area, and is converted into gray-scale map.If Fig. 3-1 is Jeep air-inlet grille sample, Fig. 3-2 are BMW air-inlet grille sample, and Fig. 3-3 are Chevrolet air-inlet grille sample.B, the air-inlet grille sample in step a is classified according to automobile brand, respectively the sample set of all kinds of brand air-inlet grilles is normalized to scale processing, air-inlet grille sample is unified is 50*50 pixel.C, all kinds of brand air-inlet grille samples in step b are carried out to Gabor wavelet transformation, in the present embodiment, according to following algorithm, filtering angle theta gets respectively 0 °, 45 °, 90 °, carries out Gabor filtering 135 ° time:

f0＝0.2；

r＝1；

g＝1；

x1＝x*cos(theta)+y*sin(theta)；

y1＝-x*sin(theta)+y*cos(theta)；

gabor_k＝f0^2/(pi*r*g)*exp(-(f0^2*x1^2/r^2+f0^2*y1^2/g^2))*exp(i*2*pi*f0*x1)；

As Fig. 3-4, Fig. 3-5, Fig. 3-6, Fig. 3-7 are at filtering angle theta=0 °, 45 °, 90 °, Jeep automotive air intake grid sample instance 135 ° time.As Fig. 3-8, Fig. 3-9, Fig. 3-10, Fig. 3-11 are in filtering angle θ=0 °, 45 °, 90 °, 135 °, time BMW air intake grid sample instance.As Fig. 3-12, Fig. 3-13, Fig. 3-14, Fig. 3-15 are in filtering angle θ=0 °, 45 °, 90 °, 135 °, time BMW air intake grid sample instance.

Sample instance after Gabor filtering is observed, have following rule: be 0 while spending in Gabor filter angles, can significantly extract the edge of air-inlet grille sample vertical direction, other angles are not obvious; Be 45 while spending in Gabor filter angles, can significantly extract the edge of the left tilted direction of air-inlet grille sample, other angles are not obvious; Be 90 while spending in Gabor filter angles, can significantly extract the edge of air-inlet grille sample level direction, other angles are not obvious; Be 135 while spending in Gabor filter angles, can significantly extract the edge of the right tilted direction of air-inlet grille sample, other angles are not obvious.

D, the Gabor wavelet character cascade that different brands automobile in step c is tieed up at 0 degree, 45 degree, 90 degree, the filtered 11*11 of 135 degree, become the air-inlet grille sample Gabor proper vector that 4*11*11 ties up.E, by after the set of eigenvectors mark of different brands air-inlet grille for automobile sample in steps d, the LIBSVM instrument that uses the Lin Zhiren of Taiwan Univ. professor team to realize is trained, and obtains the LIBSVM sorter of air-inlet grille for automobile.

Identification link:

F, in the present embodiment, pending frame of video is used HAAR vehicle checking method, orients the head of vehicle.G, in the present embodiment, utilizes the symmetry of automobile head, dwindles headstock identified region, gets and horizontal 0 to 3/5 part in left side, headstock region, longitudinal 1/5 to 4/5 region from top side in step f, detected.H, obtain in step g after air-inlet grille approximate region, in this region, use with the moving window of same 50*50 pixel dimension in step b and scan.I, 0 degree, 45 degree, 90 degree, 135 degree are carried out in same step c in the moving window region in step h carry out Gabor filtering.J, the filtered 11*11 dimensional feature vector of four different directions Gabor of step I is carried out to cascade, form the 4*11*11 dimension Gabor proper vector after this classification moving window to be identified cascade.The LIBSVM sorter obtaining in k, integrating step e, carries out svm classifier to the cascade Gabor proper vector obtaining in step j.I, return to classification results, in the present embodiment, the vehicle brand that in this frame of video, classification obtains is Chevrolet, consistent with exact brand name, identifies successfully.

Those of ordinary skill in the art will appreciate that, embodiment described here is in order to help reader understanding's principle of the present invention, should be understood to that protection scope of the present invention is not limited to such special statement and embodiment.Those of ordinary skill in the art can make various other various concrete distortion and combinations that do not depart from essence of the present invention according to these technology enlightenments disclosed by the invention, and these distortion and combination are still in protection scope of the present invention.

Claims (8)

1. the automobile mark sample training method based on air-inlet grille location, is characterized in that,

Training process comprises the following steps:

A, obtain the air-inlet grille sample of different automobile types, and be converted to gray level image;

B, the air-inlet grille sample in step a is classified according to automobile brand, and respectively Different categories of samples collection is normalized to scale processing, making air-inlet grille sample unified is same scale size;

C, Different categories of samples in step b is carried out to Gabor functional transformation, it is carried out to yardstick and rotational transform, obtain Gabor small echo expression formula;

D, by the Gabor wavelet character vector cascade of different directions, form the Gabor proper vector after the cascade of whole air-inlet grille sample;

E, by after set of eigenvectors mark, use SVM instrument to train, obtain the svm classifier device of air-inlet grille for automobile.

2. this training method of car standard specimen according to claim 1, is characterized in that, in step a, different automobile types refers to the difference of automobile brand and/or style.

3. this training method of car standard specimen according to claim 2, is characterized in that, only obtains image that air-inlet grille is positioned at axis of symmetry either side as sample in step a.

4. this training method of car standard specimen according to claim 3, is characterized in that, sample is square area.

5. according to this training method of car standard specimen described in any one claim of claim 1-4, it is characterized in that, in step c, Gabor functional transformation detailed process is: all kinds of brand air-inlet grille samples in step b are carried out to Gabor wavelet transformation, g (x, y) be female wave function

$g(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp [-\frac{1}{2}(\frac{x^2}{{\mathrm{\σ}}_x^2}+\frac{y^2}{{\mathrm{\σ}}_y^2})+2\mathrm{\πjWx}]$

It is carried out to yardstick and rotational transform, obtains Gabor small echo expression formula:

g _m,n(x,y)＝a ^-mG(x',y')

Wherein:

(x',y')＝a ^-m(xcosθ+ysinθ,-xsinθ+ycosθ)

a＞1

m,n∈Z

θ＝nπ/M

M is direction number, a ^-mfor scale factor.

6. this training method of car standard specimen according to claim 5, is characterized in that, in above-mentioned steps c, all kinds of brand air-inlet grille samples is carried out to 0 degree, 45 degree, 90 degree, 135 degree and carries out Gabor filtering.

7. the automobile mark recognition methods based on above-mentioned sample training method, is characterized in that, comprises step:

F, locate the headstock region of automobile in pending frame of video;

G, headstock regional compartmentalization is processed, to dwindle region to be identified retaining in air-inlet grille region;

H, use moving window scan described region to be identified, and described moving window yardstick is consistent with step b Normalized Scale sample after treatment in training method;

I, 0 degree, 45 degree, 90 degree and 135 degree Gabor filtering are carried out in moving window region;

J, filtered Gabor proper vector is carried out to cascade, form the Gabor proper vector after classification moving window to be identified cascade;

K, Gabor proper vector after using svm classifier device in training method to described cascade are carried out svm classifier;

L, return to svm classifier result, obtain the brand of object vehicle.

8. car mark according to claim 7 recognition methods, is characterized in that, first step g intercepts half region as region to be identified using axis of symmetry as boundary according to the symmetry of headstock.