CN103207998A - License plate character segmentation method based on support vector machine - Google Patents

Abstract

The invention discloses a motor vehicle license plate character segmentation method based on a support vector machine, and belongs to the field of computer image processing. The method comprises the following steps of: obtaining all possible segmentation results (only one result is accurate) according to a vertical projection integral curve of a known segmentation result license plate image, calculating the character width variance V of each segmentation result, the projection value sum sumvalue at each character segmentation point, the proportion sum T of each character pixel in a segmented image and the segmentation trust value S to obtain feature vectors R (V, sumvalue, T, S), inputting the feature vectors to the support vector machine with the aid of a tag for distinguishing accurate segmentation, and training the feature vectors to obtain a classification model; and performing image acquisition, positioning, binarization and normalization on a license plate to be identified to obtain support vector machine feature vectors R (V, sumvalue, T, S) of all possible segmentation results according to the vertical projection integral curve likewise, and inputting the feature vectors to the classification model of the support vector machine to obtain a segmentation result corresponding to the final classification result. The method has the advantages of high segmentation accuracy, high efficiency and low requirement for imaging quality.

Description

A kind of registration number character dividing method based on support vector machine

Technical field

The invention belongs to the computer image processing technology field, relate to mode identification technology, specifically refer to a kind of automotive number plate Character segmentation method based on support vector machine.

Background technology

Automotive number plate identification is called for short car plate identification, plays a part crucially in the intelligent transportation system the inside, no matter is at toll on the road and bridge, cell management, parking lot management, and still in traffic surveillance and control system, car plate identification all is also of paramount importance part of basis the most.Usually, the car plate recognizer can be divided into three parts: car plate location, Character segmentation and character recognition.On preceding two-part basis, how to carry out accurate character recognition, just become the final major issue that influences system recognition rate.

Current, characters on license plate is cut apart and is mainly contained following several method:

1, based on the registration number character dividing method of Template Location.

This method by a character duration Character mother plate and a character pitch width between partiting template, allow Character mother plate and interval masterplate among license plate area, slide then, ask for the ratio of the interior pixel value sum of Character mother plate and the pixel value sum in the partiting template.When ratio is obtained minimal value, the position of current character template is judged to be the cut-point position of character.This method must at first be known this priori of character duration, and generally we can't obtain the size of character duration.

2, based on the registration number character dividing method of car plate vertical projection integrated curve.

This method at first obtains the vertical integral projection curve of license plate area, according to the Wave crest and wave trough structure of character projection characters on license plate is cut apart then.This method can obtain effect preferably usually, but when there was adhesion in characters on license plate, cutting apart quality can descend to some extent.

3, based on the registration number character dividing method of connected component labeling.

This method is carried out binaryzation to license plate image earlier, removes then and disturbs connected domain, simultaneously the situation that characters on license plate ruptures is merged processing, adopts the method for zone marker to obtain the regional extent of each character at last, thereby finally realizes cutting apart of character.This method has all proposed very high requirement to the noise of car plate and the integrality of character.

4, based on the registration number character dividing method of color information:

This method is carried out the analysis of color uncontinuity according to the difference of license plate area background color and character color, realizes cutting apart of character.This method has proposed requirement to the chromatic information of car plate, and much bigger with respect to the computational processing of gray level image based on the processing of coloured image.

These methods have all obtained good segmentation effect at specific car plate, should have his own strong points.Wherein second method is the most stable, and is also not high relatively to the requirement of car plate quality.The scholar who has has also proposed registration number character dividing method that second method and first method are combined.The present invention uses support vector machine, utilizes its good classification capacity, and directly the car plate that is come out in the location is cut apart, and has obtained good effect.

Summary of the invention

The invention provides a kind of registration number character dividing method based on support vector machine, this method adopts to be cut apart the automotive number plate character based on the support vector machine of Gaussian radial basis function, compare the scheme with other same domains, it is of the present invention that to cut apart accuracy rate higher, efficient is higher, and image quality is required lower characteristics.

Technical solution of the present invention is as follows:

A kind of automotive number plate Character segmentation method based on support vector machine as shown in Figure 1, may further comprise the steps:

Step 1: choose the license plate image that N opens the correct segmentation result of the known characters on license plate of identical size (N is more than or equal to 500), make up the support vector machine disaggregated model.Specifically comprise the steps:

Step 1-1: choose the license plate image that N opens the correct segmentation result of the known characters on license plate of identical size (N is more than or equal to 500), according to the car plate vertical projection integrated curve of every described license plate image, calculate every all possible characters on license plate segmentation result of described license plate image.Having only a kind of among these results is correct segmentation result, other all be wrong.

Step 1-2: the feature of every kind of characters on license plate segmentation result of calculation procedure 1-1 gained specifically comprises:

A) the variance V of character duration,

$V=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(l_i-\stackrel{\‾}{l})}^2---\left(1\right)$

Wherein:

l _i=candidateSeg(i+1)-candidateSeg(i)（2）

$\stackrel{\‾}{l}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{l}_i---\left(3\right)$

(1) formula is to (3) formula, i=1, and 2 ..., n, n are the characters on license plate number, the width of i character of candidateSeg (i) expression, l _iWhat represent is the poor of character duration,

Represent average character duration.

B) each projection value sum sumvalue of Character segmentation point place.In ideal conditions, the interval trough value between character and the character is very little, and the projection value sum at its each Character segmentation point place is minimum.So, the correctness that the projection value sum at each Character segmentation point place is cut apart with judgement as a kind of feature.

C) each character pixels proportion sum T in the character picture that is partitioned into.In ideal conditions, in the binary image of car plate, each character pixels proportion in the character picture that splits is maximum, so, the correctness that the ratio sum T that each character pixels is shared is cut apart with judgement as a kind of feature.

D) cut apart degree of belief value S, S is the correctness that the Character segmentation scheme the is carried out tolerance from the priori of car plate shape.For every kind of segmentation result, according to the head and the tail position of Character segmentation, calculate the ratio of width to height r of car plate; According to the split position of each character, calculate the width w of each character again _kMean breadth meanv with character.If r, w _kSatisfy following condition with meanv:

r _min<r<r _max

w _kmeanv<1.35

Then obtain S=255, otherwise the confidence level that obtains current scheme is S=0.R wherein _MinAnd r _MaxBe respectively minimum the ratio of width to height and maximum the ratio of width to height of known existing car plate.

Step 1-3: with variance V, each projection value sum sumvalue of Character segmentation point place of step 1-2 gained characteristic character width, each character pixels proportion sum T and cut apart degree of belief value S and form a support vector machine proper vector R(V in the character picture that is partitioned into, sumvalue, T, S); Each all can obtain a support vector machine proper vector R(V according to car plate vertical projection integrated curve gained segmentation result, sumvalue, T S), is correct because have only a segmentation result, give each proper vector R(V, sumvalue, T, S) label, correct segmentation result label is made as " 1 ", and all the other segmentation result labels are made as " 0 "; Like this, all are had the support vector machine proper vector R(V of label, sumvalue, T, S) the input support vector machine is trained, and finally obtains a support vector machine disaggregated model.

Step 2: gather license plate image to be identified, after car plate location, binary conversion treatment, be normalized to the license plate image license plate image of a size with the correct segmentation result of known characters on license plate described in the step 1, be normalization license plate image to be split.

Step 3: the vertical projection integrated curve according to the normalization license plate image to be split of step 2 gained obtains all possible characters on license plate segmentation result.

Step 4: with reference to the support vector machine proper vector R(V of described calculation procedure 3 all segmentation results of gained of step 1-2, sumvalue, T, S), with all support vector machine proper vector R(V, sumvalue, T, S) input step 1 final gained support vector machine disaggregated model carries out Classification and Identification, the support vector machine disaggregated model finally is identified as that support vector machine proper vector R(V of " 1 ", sumvalue, T, S) corresponding segmentation result is as final characters on license plate segmentation result.

Innovation part of the present invention is:

1, the neck that is applied in identification that support vector machine is main is incited somebody to action, and the present invention is applied to its method cutting apart of characters on license plate cleverly, carries out cutting apart of characters on license plate from the angle of Classification and Identification.Compare with the method for the Character segmentation of existing character drop shadow curve, this method can be good at handling those car plates poor quality's picture, situation for the adhesion of the character that exists in the car plate has good robustness, the efficient of correctly cutting apart is compared higher with it, its antimierophonic ability is stronger.

2, adopt Gaussian radial basis function to reduce the quantity of support vector to reach the purpose that reduces the sort operation calculated amount.The method that the present invention adopts can reduce the calculated amount of algorithm in certain degree than existing registration number character dividing method based on color information, improves the efficient of algorithm.Under the suitable situation of accuracy rate, this method can be good at solving the problem of efficiency of algorithm.

Description of drawings

Fig. 1 schematic flow sheet of the present invention.

Embodiment

Range of application of the present invention is automotive number plate image automatic recognition system, mainly solves the segmentation problem of automotive number plate character.

Here provide and adopt method of the present invention to carry out the example process that characters on license plate is cut apart:

1. write the Character segmentation algoritic module according to method of the present invention.

2. collect and satisfy the characters on license plate picture that the inventive method requires, be used for the training to character recognition algorithm, preserve training result.

3. set up monitor camera device at the vehicular traffic road cross, the collection vehicle photo.

4. orient car plate with algorithm of locating license plate of vehicle.

5. with successfully locating the car plate picture of coming out, the Character segmentation module of importing the present invention's design into is cut apart, and obtains the segmentation result of character picture.

In sum, the present invention utilizes the good classification ability of support vector machine, with the set of eigenvectors of characters on license plate image as data to be sorted.The characters on license plate image of binaryzation has reduced the influence that the character imaging is subjected to illumination.Adopt the support vector machine of Gaussian radial basis function also to reduce the quantity of training back support vector, and then reduced the operand in the assorting process.According to the character position method that segmentation is cut apart to characters on license plate, reduced quantity to be sorted, reduce operand from another angle.Consider the characteristic of support vector machine, carry out the method for cross validation when having proposed training, thereby find better parameter, obtain better classifying quality.

Claims (2)

1. automotive number plate Character segmentation method based on support vector machine may further comprise the steps:

Step 1: choose the license plate image that N opens the correct segmentation result of the known characters on license plate of identical size, make up the support vector machine disaggregated model; Specifically comprise the steps:

Step 1-1: choose the license plate image that N opens the correct segmentation result of the known characters on license plate of identical size, according to the car plate vertical projection integrated curve of every described license plate image, calculate every all possible characters on license plate segmentation result of described license plate image; Having only a kind of among these results is correct segmentation result, other all be wrong;

Step 1-2: the feature of every kind of characters on license plate segmentation result of calculation procedure 1-1 gained specifically comprises:

A) the variance V of character duration,

$V=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(l_i-\stackrel{\&OverBar;}{l})}^2---\left(2\right)$

Wherein:

l _i=candidateSeg(i+1)-candidateSeg(i) （2）

$\stackrel{\&OverBar;}{l}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{l}_i---\left(3\right)$

(1) formula is to (3) formula, i=1, and 2 ..., n, n are the characters on license plate number, the width of i character of candidateSeg (i) expression, l _iWhat represent is the poor of character duration,

Represent average character duration;

B) each projection value sum sumvalue of Character segmentation point place;

C) each character pixels proportion sum T in the character picture that is partitioned into;

D) cut apart degree of belief value S, S is the correctness that the Character segmentation scheme the is carried out tolerance from the priori of car plate shape; For every kind of segmentation result, according to the head and the tail position of Character segmentation, calculate the ratio of width to height r of car plate; According to the split position of each character, calculate the width w of each character again _kMean breadth meanv with character.If r, w _kSatisfy following condition with meanv:

r _min<r<r _max

w _kmeanv<1.35

Then obtain S=255, otherwise the confidence level that obtains current scheme is S=0; R wherein _MinAnd r _MaxBe respectively minimum the ratio of width to height and maximum the ratio of width to height of known existing car plate;

Step 1-3: with variance V, each projection value sum sumvalue of Character segmentation point place of step 1-2 gained characteristic character width, each character pixels proportion sum T and cut apart degree of belief value S and form a support vector machine proper vector R(V in the character picture that is partitioned into, sumvalue, T, S); Each all can obtain a support vector machine proper vector R(V according to car plate vertical projection integrated curve gained segmentation result, sumvalue, T S), is correct because have only a segmentation result, give each proper vector R(V, sumvalue, T, S) label, correct segmentation result label is made as " 1 ", and all the other segmentation result labels are made as " 0 "; Like this, all are had the support vector machine proper vector R(V of label, sumvalue, T, S) the input support vector machine is trained, and finally obtains a support vector machine disaggregated model;

Step 2: gather license plate image to be identified, after car plate location, binary conversion treatment, be normalized to the license plate image license plate image of a size with the correct segmentation result of known characters on license plate described in the step 1, be normalization license plate image to be split;

Step 3: the vertical projection integrated curve according to the normalization license plate image to be split of step 2 gained obtains all possible characters on license plate segmentation result;

Step 4: with reference to the support vector machine proper vector R(V of described calculation procedure 3 all segmentation results of gained of step 1-2, sumvalue, T, S), with all support vector machine proper vector R(V, sumvalue, T, S) input step 1 final gained support vector machine disaggregated model carries out Classification and Identification, the support vector machine disaggregated model finally is identified as that support vector machine proper vector R(V of " 1 ", sumvalue, T, S) corresponding segmentation result is as final characters on license plate segmentation result.

2. the automotive number plate Character segmentation method based on support vector machine according to claim 1 is characterized in that described N is more than or equal to 500.

Images