CN104573692A - Vehicle license plate binarizing method based on fuzzy degradation model - Google Patents

Abstract

The invention discloses a vehicle license plate binarizing method based on a fuzzy degradation model. The vehicle license plate binarizing method based on the fuzzy degradation model comprises the following steps: primarily binarizing; refining a framework; extracting a character center framework; extracting character center color; obtaining a binarizing threshold of a vehicle license plate image; finally binarizing. The vehicle license plate binarizing method can be used for effectively separating vehicle license plate characters and facilitating subsequent vehicle license character identification; meanwhile, the fuzzy degradation model is used for estimating a degradation relation between a center pixel and edge pixel of the vehicle license plate character so as to calculate the accurate threshold for distinguishing vehicle license plate character and the vehicle license plate background; the problem that the threshold cannot be self-adaptive under different conditions can be effectively avoided.

Description

A kind of license plate binary method based on blur degradation model

Technical field

The present invention relates to image processing field, particularly a kind of license plate binary method based on blur degradation model.

Background technology

Car license recognition is the very successful application of image processing and pattern recognition in modern intelligent transportation.The image arrived by camera collection or video, license plate area locates out by application image treatment and analyses technology from image, license plate image is divided into characters on license plate image independent one by one again, the each independent character picture of last application model recognition technology identification, thus split obtains final car plate result together.

Binaryzation is key one step in Vehicle License Plate Recognition System in Image semantic classification step, and the effect of binaryzation directly has influence on the accuracy of License Plate Character Segmentation and the discrimination of character recognition.But changeable due to license plate image shooting environmental, our actual acquisition to license plate image in sharpness, brightness and contrast etc., all there is difference.This brings difficulty with regard to carrying out the pretreated method of license plate image to traditional employing fixed threshold binaryzation.

Existing license plate binary algorithm is all carry out binaryzation based on the method for threshold value to car plate mostly, be directed to the calculating of threshold value, some employing local thresholds, some employing global thresholds, correspondence creates local binarization algorithm and overall Binarization methods, but due to the singularity of license plate image, some uneven illuminations, over-exposed, stained car plate is on apply during two kinds of methods, and effect is all undesirable.

A thorny difficult problem is also had to be exactly adhesion problems between character in actual applications, in the image pattern that low pixel camera head collects, image resolution ratio is low, a lot of details in image is caused to be all to simulate generation by the interpolation between pixel, the present license plate image of such information slip will allow between characters on license plate and form link, further character well cannot be separated with character, difficulty is caused to character recognition below.

Summary of the invention

The object of the present invention is to provide a kind of license plate binary method based on blur degradation model, effectively characters on license plate can be separated, be convenient to follow-up Recognition of License Plate Characters.

For achieving the above object, the present invention is by the following technical solutions:

Based on a license plate binary method for blur degradation model, comprise the following steps:

S1, preliminary binaryzation, carry out preliminary binaryzation to former license plate image, obtain two-value license plate image;

S2, skeleton refinement, carry out connected region refinement to the two-value license plate image obtained in step S1, obtains car plate skeleton image;

S3, character center skeletal extraction, to each skeleton point in the car plate skeleton image obtained in step S2, judge whether it belongs to the point on characters on license plate, is if so, then retained, and if not, is then deleted, thus obtain characters on license plate skeleton image;

S4, character center color extraction, according to each skeleton point in the characters on license plate skeleton image obtained in step S3, former license plate image extracts the color value of corresponding point, thus obtains character center color value;

S5, acquisition license plate image binary-state threshold, according to the character center color value that step S4 obtains, set up blur degradation model, calculate license plate image binary-state threshold;

S6, final binaryzation, based on the license plate image binary-state threshold that step S5 obtains, carry out binaryzation to former license plate image.

Preferably, in step sl, described preliminary binaryzation adopts Ostu Binarization methods.

Preferably, described step S2 comprises step by step following:

S21, for single pixel, the set defining its 4 pixels in upper and lower, left and right is the 4-neighborhood of this pixel, the set defining 4 pixels in 4, its upper and lower, left and right pixel and diagonal thereof is the 8-neighborhood of this pixel, for each pixel p in prospect, definition intersects counts as in the 4-neighborhood of a p, gray-scale value meets total number of the pixel of condition below

g(P _k)-g(P _k-1)＝1，

Wherein, P _krepresent a kth pixel, g (P _k) be the gray-scale value of a kth pixel, 0≤k≤7, and the subscript mould 8 of pixel p during computing;

S22, to each pixel on two-value license plate image, count according to the gray-scale value of 4-neighborhood, the gray-scale value of 8-neighborhood and intersection, it is retained or deletes, finally obtain car plate skeleton image.

Preferably, described step S3 is realized step by step by following:

S31, for each skeleton point in the car plate skeleton image obtained in step S2, judge that in its 8-neighborhood, gray-scale value is the pixel quantity of 1, if equal 1 or be more than or equal to 3, then it is retained, otherwise perform step S32-S34;

S32, position relationship according to this pixel and surrounding pixel point, judge the line style of type of these pixel place lines, the described line style of type comprises vertical curve, horizontal line and parallax;

S33, from this pixel, on two-value license plate image, the vertical direction along the line style of type is searched for both sides, until run into black pixel point, the step-length sum of two-sided search is as the stroke width of this former character in pixel place;

If the stroke width of this former character in pixel place of S34 is in the normal stroke width range of characters on license plate, then judges that this pixel belongs to the point on characters on license plate and retained, otherwise deleted;

All skeleton points in S35, traversal car plate skeleton image, final acquisition characters on license plate skeleton image.

Preferably, described step S4 also comprises and being kept in array C by the character center color value of acquisition.

Preferably, described step S5 comprises step by step following:

S51, the observed reading of the color value in array C as stochastic variable X, X to be expressed as:

X={X ₁, X ₂..., X _t, wherein, X _i∈ C;

S52, set up the mixture gaussian modelling of stochastic variable X:

$P(X,{\mathrm{\μ}}_1,{\mathrm{\μ}}_2,{\mathrm{\σ}}_1,{\mathrm{\σ}}_2)=\underset{i=1}{\overset{2}{\mathrm{\Σ}}}{\mathrm{\ω}}_{i}g(X,{\mathrm{\μ}}_i,{\mathrm{\σ}}_i),$

Wherein, g (X, μ _i, σ _i) be gauss of distribution function, ω _ifor weight;

S53, initialization μ ₁=0, μ ₂=0, σ ₁=0, σ ₂=0, according to the actual value of stochastic variable X to μ _i, σ _i, ω _iupgrade;

The Gauss model that S54, acquisition weight are maximum, and ask for parameter μ, the σ of this Gauss model;

S55, set up the Gaussian distribution model of character center color value:

$G(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}{e}^{-(x^2+y^2)/2{\mathrm{\σ}}^2},$

To arbitrfary point (x, y), if its color value is g, then the condition of its point met on characters on license plate is:

|g-μ|＝3σ。

Preferably, in step s 6, describedly two-value is carried out to former license plate image turn to, by following formula, binaryzation carried out to former license plate image:

$b(x,y)=\left\{\begin{array}{cc}1,& |f(x,y)-\mathrm{\μ}|\≤3\mathrm{\σ}\\ 0,& |f(x,y)-\mathrm{\μ}|\≥3\mathrm{\σ}\end{array}\right.$

Wherein, b (x, y) is final car plate bianry image, and f (x, y) is the gray level image of former car plate.

After adopting technique scheme, the present invention is compared with background technology, and tool has the following advantages:

Characters on license plate can effectively be separated by the present invention, is convenient to follow-up Recognition of License Plate Characters; Adopt blur degradation model to estimate the degeneration relation of characters on license plate center pixel and edge pixel simultaneously, thus calculate correct threshold value to distinguish characters on license plate and car plate background, effectively avoiding different situations lower threshold value cannot adaptive problem.

Accompanying drawing explanation

Fig. 1 is workflow schematic diagram of the present invention.

Fig. 2 shows the two-value license plate image result obtained through preliminary binary conversion treatment.

Fig. 3 shows the result of car plate skeleton image.

Fig. 4 shows the result of characters on license plate skeleton image.

Fig. 5 a shows the situation that the line style of type is vertical curve; It is horizontal situation that Fig. 5 b shows the line style of type; Fig. 5 c shows the situation that the line style of type is parallax.

Fig. 6 shows the car plate bianry image result obtained through final binary conversion treatment.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

Embodiment

Refer to Fig. 1, the invention discloses a kind of license plate binary method based on blur degradation model, comprise the following steps:

S1, preliminary binaryzation

Adopt Da-Jin algorithm (Ostu Binarization methods) to carry out preliminary binaryzation to former license plate image, obtain two-value license plate image.In this step, the acquiring method of Da-Jin algorithm threshold value is as follows:

For license plate image f (x, y), the segmentation threshold of characters on license plate and car plate background is T, and the ratio that the pixel belonging to characters on license plate accounts for entire image is designated as ω ₀, the average gray of these points is designated as μ ₀; The ratio that car plate background pixel point accounts for entire image is ω ₁, its average gray is ω ₁.The overall average gray scale of image is designated as μ, and inter-class variance is designated as σ.Then have:

σ＝ω ₀ω ₁(μ ₀-μ ₁) ²

T is from 0 to 255 for traversal, calculates the value of T when σ obtains maximal value, is the threshold value of Da-Jin algorithm.

The two-value license plate image of the binary conversion treatment acquisition of this step, be only used for carrying out guestimate to characters on license plate position, and be not used in the extraction of characters on license plate, the result of two-value license plate image as shown in Figure 2.

S2, skeleton refinement

Connected region refinement is carried out to the two-value license plate image obtained in step S1, obtains car plate skeleton image (as shown in Figure 3).This step realizes especially by following steps:

S21, for single pixel, the set defining its 4 pixels in upper and lower, left and right is the 4-neighborhood of this pixel, the set defining 4 pixels in 4, its upper and lower, left and right pixel and diagonal thereof is the 8-neighborhood of this pixel, for each pixel p in prospect, definition intersects counts as in the 4-neighborhood of a p, gray-scale value meets the sum of counting of the pixel of condition below

g(P _k)-g(P _k-1)＝1，

Wherein, P _krepresent a kth pixel, g (P _k) be the gray-scale value of a kth pixel, 0≤k≤7, and the subscript mould 8 of pixel p during computing.

S22, to each pixel on two-value license plate image, count according to the gray-scale value of 4-neighborhood, the gray-scale value of 8-neighborhood and intersection, it is retained or deletes, finally obtain car plate skeleton image.Step S22 realizes especially by following methods:

1) pixel needing to retain is determined.Each point successively in traversing graph picture, if current investigation point is p (its gray-scale value is 1), the 4-neighborhood gray scale sum of definition p is ∑ ⁴, the gray-scale value sum of the 8-neighborhood of p is ∑ ⁸, the intersection of p is counted as N _g.

First, in order to ensure that the image after refinement can reflect the outshot of target in original image, its 8-neighborhood being investigated to current point p and only deposits the point that single one gray-scale value is 1, then retaining this pixel in principle.

Secondly, investigate the 4-neighborhood point of current point p, if the gray-scale value of 4-neighborhood point is 1, i.e. ∑ ⁴=4, then can conclude that it is that the interior point of object edge should retain.

Then, investigate the 8-neighborhood of p, retaining 4-neighborhood sum is 1, i.e. ∑ ⁴=4 and intersect points N _gthe point of > 1, to ensure the graph connectedness after refinement.

Finally, in order to retain those, to refine to width be the points being difficult to before 2 determine whether to delete, adopt following template a, the b provided to detect, if one of Output rusults of template a, b is true, then retain current point to guarantee when edge thinning is to fracture unlikely during two-wire.

x	x	x
				0	p(x,y)＝1	1	0
x	x	x

Template a

x	0	x
				x	p(x,y)＝1	x
x	1	x
					0

Template b

2) pixel of Delete superfluous.While ensureing graph connectedness, realize the One-point Connection between the pixel on skeleton, adopt following template c, the d provided to detect, if the result that template c, d export is true, then delete current pixel point.

0	1	x
			1	p(x,y)＝1	1

x

0

0

Template c

x	1	0
			0	p(x,y)＝1	1
0	0	x

Template d

3) unnecessary branch line is removed.In order to branch line unnecessary in the skeleton image after removal of images refinement, obtain smooth image framework, to through step 1), 2) refined image of acquisition after process, again investigate the intersection points N of every bit on its skeleton _gif, N _gbe not more than 1, then deleted.The car plate skeleton image of final acquisition as shown in Figure 3.

S3, character center skeletal extraction

To each skeleton point in the car plate skeleton image obtained in step S2, judge whether it belongs to the point on characters on license plate, is if so, then retained, and if not, is then deleted, thus obtain characters on license plate skeleton image (as shown in Figure 4).This step realizes especially by following steps:

S31, for each skeleton point in the car plate skeleton image obtained in step S2, judge that in its 8-neighborhood, gray-scale value is the pixel quantity of 1, if equal 1 or be more than or equal to 3, then it is retained, otherwise perform step S32-S34.

S32, position relationship according to this pixel and surrounding pixel point, judge the line style of type of these pixel place lines, the described line style of type comprises vertical curve, horizontal line and parallax.The line style of type be the situation of vertical curve as shown in Figure 5 a, the line style of type be horizontal situation as shown in Figure 5 b, the line style of type be the situation of parallax as shown in Figure 5 c.

S33, from this pixel, on two-value license plate image, the vertical direction along the line style of type is searched for both sides, until run into black pixel point, now, the step-length of two-sided search is designated as S respectively _land S _r, the stroke width of this former character in pixel place is designated as W, W=S _l+ S _r.

If S34 W is in the normal stroke width range of characters on license plate, then judges that this pixel belongs to the point on characters on license plate and retained, otherwise deleted.That is, W demand fulfillment:

w ₁≤W≤w ₂，

Wherein, w ₁=pH/20, w ₂=pH/8, pH are the height of license plate image.

All skeleton points in S35, traversal car plate skeleton image, car plate skeleton image only can retain the point on the center framework of characters on license plate, final acquisition characters on license plate skeleton image.

S4, character center color extraction

According to each skeleton point in the characters on license plate skeleton image obtained in step S3, former license plate image extracts the color value of corresponding point, thus obtain character center color value.

If former license plate image is I (x, y), the characters on license plate skeleton image obtained in step 3 is T (x, y).Traversing graph as T (x, y), if T (x _i, y _i)=1, then by I (x _i, y _i) gray-scale value be kept in an array C.After having traveled through, then save the gray-scale value of the pixel at most of characters on license plate central axis place in array C.

S5, acquisition license plate image binary-state threshold

According to the character center color value that step S4 obtains, set up blur degradation model, calculate license plate image binary-state threshold.This step realizes especially by following steps:

S51, using the color value in array C as stochastic variable X, then the observed reading of X is expressed as:

X={X ₁, X ₂..., X _t, wherein, X _i∈ C.

S52, set up the mixture gaussian modelling of stochastic variable X:

$P(X,{\mathrm{\μ}}_1,{\mathrm{\μ}}_2,{\mathrm{\σ}}_1,{\mathrm{\σ}}_2)=\underset{i=1}{\overset{2}{\mathrm{\Σ}}}{\mathrm{\ω}}_{i}g(X,{\mathrm{\μ}}_i,{\mathrm{\σ}}_i),$

Wherein, g (X, μ _i, σ _i) be gauss of distribution function, ω _ifor weight.

S53, initialization μ ₁=0, μ ₂=0, σ ₁=0, σ ₂=0, according to the actual value of stochastic variable X to μ _i, σ _i, ω _iupgrade.

The Gauss model that S54, acquisition weight are maximum, and ask for parameter μ, the σ of this Gauss model.Step S54 realizes especially by following methods:

1) weights omega _kcarry out according to following renewal expression formula in the renewal of t:

ω _k,t＝(1-α)ω _k,t-1+αM _k,t

Wherein, α is the learning rate of this newer, and 1/ α represents is a time constant, is used for characterizing the speed upgraded.M _k,ta two-valued variable, as this model and X _iduring coupling, M _k,t=1, otherwise M _k,t=0.

2) if Gaussian distribution model g is (X, μ _k, σ _k) and X _icoupling, average μ _k, variances sigma _k ²renewal show according to following renewal expression and carry out:

μ _k,t＝(1-ρ)μ _k,t-1+ρX _t

${\mathrm{\σ}}_{k,t}^2=(1-\mathrm{\ρ}){\mathrm{\σ}}_{k,t-1}^2+\mathrm{\ρ}{(X_t-{\mathrm{\μ}}_{k,t})}^T(X_t-{\mathrm{\μ}}_{k,t})$

ρ＝αg(X _t|μ _k,σ _k)

Wherein α is a constant, g (X _t| μ _k, σ _k) be the density function of Gaussian distribution.By the ρ calculated be exactly average and variance upgrade time learning rate.When model does not have and X _iduring coupling, average μ _k, variance keep original value constant.

Because in array C, more point is the point at characters on license plate center, and noise spot occupies the minority, so after mixed Gauss model is added up, the Gaussian distribution that weight is large, then for describing the statistical model of characters on license plate center pixel distribution.

Through step above, we can in the hope of being used for the parameter μ of that model describing character center pixel and σ.Wherein μ and σ is the parameter of that Gauss model that in mixture model, weight is the highest.

S55, according to image blurring principle, image from the close-by examples to those far off time, the decline of picture quality can by and convolution simulation can only be carried out by a gauss low frequency filter and original image.Thus, the Gaussian distribution model of character center color value is set up:

$G(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}{e}^{-(x^2+y^2)/2{\mathrm{\σ}}^2},$

To arbitrfary point (x, y), if its color value is g, then the condition of its point met on characters on license plate is:

|g-μ|＝3σ。

S6, final binaryzation

Based on the license plate image binary-state threshold that step S5 obtains, carry out binaryzation to former license plate image, the final car plate bianry image obtained as shown in Figure 6.In this step, by following formula, binaryzation is carried out to former license plate image:

$b(x,y)=\left\{\begin{array}{cc}1,& |f(x,y)-\mathrm{\μ}|\≤3\mathrm{\σ}\\ 0,& |f(x,y)-\mathrm{\μ}|\≥3\mathrm{\σ}\end{array}\right.$

Wherein, b (x, y) is final car plate bianry image, and f (x, y) is the gray level image of former car plate.

The above; be only the present invention's preferably embodiment, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; the change that can expect easily or replacement, all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claim.

Claims (7)

1., based on a license plate binary method for blur degradation model, it is characterized in that, comprise the following steps:

S1, preliminary binaryzation, carry out preliminary binaryzation to former license plate image, obtain two-value license plate image;

S2, skeleton refinement, carry out connected region refinement to the two-value license plate image obtained in step S1, obtains car plate skeleton image;

S3, character center skeletal extraction, to each skeleton point in the car plate skeleton image obtained in step S2, judge whether it belongs to the point on characters on license plate, is if so, then retained, and if not, is then deleted, thus obtain characters on license plate skeleton image;

S4, character center color extraction, according to each skeleton point in the characters on license plate skeleton image obtained in step S3, former license plate image extracts the color value of corresponding point, thus obtains character center color value;

S5, acquisition license plate image binary-state threshold, according to the character center color value that step S4 obtains, set up blur degradation model, calculate license plate image binary-state threshold;

S6, final binaryzation, based on the license plate image binary-state threshold that step S5 obtains, carry out binaryzation to former license plate image.

2. a kind of license plate binary method based on blur degradation model as claimed in claim 1, is characterized in that, in step sl, described preliminary binaryzation adopts Ostu Binarization methods.

3. a kind of license plate binary method based on blur degradation model as claimed in claim 1 or 2, it is characterized in that, described step S2 comprises step by step following:

S21, for single pixel, the set defining its 4 pixels in upper and lower, left and right is the 4-neighborhood of this pixel, the set defining 4 pixels in 4, its upper and lower, left and right pixel and diagonal thereof is the 8-neighborhood of this pixel, for each pixel p in prospect, definition intersects counts as in the 4-neighborhood of a p, gray-scale value meets total number of the pixel of condition below

g(P _k)-g(P _k-1)＝1，

Wherein, P _krepresent a kth pixel, g (P _k) be the gray-scale value of a kth pixel, 0≤k≤7, and the subscript mould 8 of pixel p during computing;

S22, to each pixel on two-value license plate image, count according to the gray-scale value of 4-neighborhood, the gray-scale value of 8-neighborhood and intersection, it is retained or deletes, finally obtain car plate skeleton image.

4. a kind of license plate binary method based on blur degradation model as claimed in claim 3, is characterized in that, described step S3 is realized step by step by following:

S31, for each skeleton point in the car plate skeleton image obtained in step S2, judge that in its 8-neighborhood, gray-scale value is the pixel quantity of 1, if equal 1 or be more than or equal to 3, then it is retained, otherwise perform step S32-S34;

S32, position relationship according to this pixel and surrounding pixel point, judge the line style of type of these pixel place lines, the described line style of type comprises vertical curve, horizontal line and parallax;

S33, from this pixel, on two-value license plate image, the vertical direction along the line style of type is searched for both sides, until run into black pixel point, the step-length sum of two-sided search is as the stroke width of this former character in pixel place;

If the stroke width of this former character in pixel place of S34 is in the normal stroke width range of characters on license plate, then judges that this pixel belongs to the point on characters on license plate and retained, otherwise deleted;

All skeleton points in S35, traversal car plate skeleton image, final acquisition characters on license plate skeleton image.

5. a kind of license plate binary method based on blur degradation model as claimed in claim 4, is characterized in that, described step S4 also comprises and being kept in array C by the character center color value of acquisition.

6. a kind of license plate binary method based on blur degradation model as claimed in claim 5, it is characterized in that, described step S5 comprises step by step following:

S51, the observed reading of the color value in array C as stochastic variable X, X to be expressed as:

X={X ₁, X ₂..., X _t, wherein, X _i∈ C;

S52, set up the mixture gaussian modelling of stochastic variable X:

$P(X,{\mathrm{\μ}}_1,{\mathrm{\μ}}_2,{\mathrm{\σ}}_1,{\mathrm{\σ}}_2)=\underset{i=1}{\overset{2}{\mathrm{\Σ}}}{\mathrm{\ω}}_{i}g(X,{\mathrm{\μ}}_i,{\mathrm{\σ}}_i),$

Wherein, g (X, μ _i, σ _i) be gauss of distribution function, ω _ifor weight;

S53, initialization μ ₁=0, μ ₂=0, σ ₁=0, σ ₂=0, according to the actual value of stochastic variable X to μ _i, σ _i, ω _iupgrade;

The Gauss model that S54, acquisition weight are maximum, and ask for parameter μ, the σ of this Gauss model;

S55, set up the Gaussian distribution model of character center color value:

$G(x,y)=\frac{1}{{2\mathrm{\π\σ}}^2}{e}^{-(x^2+y^2)/{2\mathrm{\σ}}^2},$

To arbitrfary point (x, y), if its color value is g, then the condition of its point met on characters on license plate is:

|g-μ|＝3σ。

7. a kind of license plate binary method based on blur degradation model as claimed in claim 6, is characterized in that, in step s 6, describedly carries out two-value to former license plate image and turns to and carry out binaryzation by following formula to former license plate image:

$b(x,y)=\left\{\begin{array}{cc}1,& |f(x,y)-\mathrm{\μ}|\≤3\mathrm{\σ}\\ 0,& |f(x,y)-\mathrm{\μ}|\≥3\mathrm{\σ}\end{array}\right.$

Wherein, b (x, y) is final car plate bianry image, and f (x, y) is the gray level image of former car plate.