CN102521587B - License plate location method - Google Patents

Abstract

The invention discloses a license plate location method, which comprises the following steps of: (1) preprocessing an original vehicle image I (x,y) to obtain an enhanced vehicle image I1 (x,y); (2) correcting the color values of points in the vehicle image I1 (x,y) to obtain a corrected vehicle image I2 (x,y); (3) scanning the corrected vehicle image I2 (x,y), and extracting a blue-and-yellow-concentrated region as a license plate candidate region; and (4) checking a license plate region, i.e. judging the size of the current region, accounting the current region as the license plate region if the region is in a set range, and finishing license plate location. With the adoption of the license plate location method disclosed by the invention, the license plates in vehicle images acquired under general conditions can be located, and the license plates in vehicle images photographed under complicated conditions, such as dim light, rain, snow, smoke and the like, can also be located, so that the license plate location method has important practical value in license plate recognition.

Description

A kind of license plate locating method

Technical field

The invention belongs to computer picture recognition field, be specifically related to a kind of license plate locating method.

Background technology

Car plate identification (license plate recognition, LPR) technology is intelligent transportation system (intelligent transport system, ITS) important means of management, its task is automatically to locate and identify vehicle license plate by gathering, analyze, process image.LPR system can be widely used in public security bayonet, parking lot management, road management violating the regulations, highway inspection, monitor the important events such as black board vehicle.Car plate location in Vehicle License Plate Recognition System is the basis of License Plate Segmentation, character recognition, is also the committed step that improves car plate discrimination.Many scholars are in exploration and the research carried out aspect car plate location.At present, the method for car plate location is mainly divided into based on two large classes gray level image and based on coloured image.

Based on the license plate locating method of gray level image, the coloured image before this equipment being collected converts gray level image to, abandons the colouring information of image, then consider the information such as the gray scale, texture, edge of license plate area, realizes the location of car plate.This localization method is simple, quick, particularly for clear picture, and the uncomplicated license plate image of background, accuracy rate is higher, but is directed to background complexity, the weak license plate image of contrast, and this localization method positioning licence plate is just more difficult.

Given this, scholars forward research emphasis on the license plate locating method based on coloured image that does not abandon color of image information to.This method is the coloured image that directly use equipment collects, realize car plate location, take full advantage of the image information collecting, improve the accuracy rate of car plate location, still, due to the difference of image capture device with affected by extraneous factor, make the picture quality collecting be difficult to precisive and prediction, and that license plate locating method based on coloured image is affected by picture quality is larger, particularly for picture contrast not by force, partially dark situation, be difficult to accurate positioning licence plate.

According to Retinex theory, the color of object is to be determined the reflection potential of long wave, medium wave and shortwave is common by object, and the reflection potential of object in certain wave band is the intrinsic attribute of object itself, there is no dependence with light source.

Summary of the invention

For above problem, the invention provides a kind of license plate locating method based on coloured image, by considering the impact of light power on vehicle image and the demand of coloured image license plate locating method, adjust in vehicle image and irradiate component and reflecting component ratio, realize the identification location to car plate.

Realize the concrete numerical procedure that object of the present invention adopts as follows:

Step 1, original vehicle image I (x, y) is carried out to pre-service, the image I 1 (x, y) after being enhanced;

In step 2, correction image I1 (x, y), the color value of each point, obtains revising rear vehicle image I 2 (x, y);

Step 3, scan image I2 (x, y), extract blueness, yellow concentrated region as vehicle candidate region.

Step 4, verification vehicle region.Judge current region size, if region is within the scope of setting, thinks license plate area, otherwise cast out.

The present invention has considered the strong and weak impact of vehicle image of light and the demand of coloured image vehicle positioning method, adjust and in vehicle image, irradiated component and reflecting component ratio, image overall contrast and local contrast are improved, strengthen vehicle image shade details, overcome tradition and can not accurately not locate based on coloured image license plate locating method that illumination is poor, the not problem of strong image of contrast, expand the usable range based on coloured image positioning licence plate, promoted the accuracy of car plate location.

Accompanying drawing explanation

Fig. 1 is overall flow figure of the present invention;

Fig. 2 is image pretreatment process figure;

Fig. 3 is example vehicle image;

Fig. 4 adjusts the vehicle image irradiating after component and reflecting component;

Fig. 5 is the vehicle image of revising after color;

Fig. 6 is the car plate of the inventive method location.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in further detail.

Step 1, image pre-service, as shown in Figure 2, result as shown in Figure 4 for flow process.

(1) vehicle image (as shown in Figure 3) function table is shown as to following form:

I(x，y)＝R(x，y)L(x，y)

Wherein (x, y) is the position of image mid point, and I represents original vehicle image, and R represents reflected light component, represents light irradiation component, and light irradiation component L describes the brightness of surrounding environment, irrelevant with scenery; And reflected light component to be R refer to scenery reflection potential is irrelevant with illumination, the detailed information that it has comprised scenery.

(2) taken the logarithm in function both sides, so by light irradiation component and reflected light component is expressed as and form:

LogI＝Log(R.L)＝LogR+LogL

(3) by Gauss's template of n different scale, original image I is done to convolution, each convolution is equivalent to original image to do low-pass filtering one time, can obtain the image D after n width low-pass filtering ₁, D ₂, D ₃d _n,

expression yardstick is σ _igaussian filter function, i ∈ [1, n] (preferred n=3 in the present embodiment, σ ₁, σ ₂, σ ₃preferably be respectively 10,50,240):

$F{(x,y)}_{{\mathrm{\σ}}_i}=\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_i}\exp (-\frac{(x^2+y^2)}{2{{\mathrm{\σ}}_i}^2}),(r^2=x^2+y^2)$

$D_i(x,y)=I(x,y)*F{(x,y)}_{{\mathrm{\σ}}_i}$

(4), in log-domain, with the image subtraction after original image and every width low-pass filtering, just can obtain the image of n width high frequency enhancement:

R _i(x，y)＝LogI(x，y)-LogD _i(x，y)

(5) to the n width high frequency enhancement image weighted sum obtaining in (4):

$I1(x,y)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ω}}_i{R}_i(x,y)$ $\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ω}}_i=1$

ω _ibe the weight of i width high frequency enhancement image, I1 (x, y) is the vehicle image (ω in this example after enhancing ₁=ω ₂=ω ₃=1/3).

In step 2, correction image I1 (x, y), the color value of each point obtains I2 (x, y), and revised result as shown in Figure 5.

(1) image is transformed into HSV space from rgb space, in hsv color model in each component h, s, v and RGB color model the corresponding relation of r, g, b component suc as formula shown in (1)-(3).

Luminance component v is

v＝k _max/255

(1)

Saturation degree component s is

$s=\left\{\begin{array}{cc}\mathrm{\δ}/k_{\max }& k_{\max }>0\\ 0& k_{\max }=0\end{array}\right.$

(2)

Chromatic component h is

$h=\left\{\begin{array}{cc}0& s=0\\ (g-b)/\mathrm{\δ}& s\≠0,r=k_{\max }\\ 2+(b-r)/\mathrm{\δ}& s\≠0,g=k_{\max }\\ 4+(r-g)/\mathrm{\δ}& s\≠0,b=k_{\max }\end{array}\right.$

h＝h×60；h＝h+360(h＜0)

(3)

Wherein, and h ∈ [0,360), s ∈ [0,1], v ∈ [0,1], k _max=max (r, g, b), k _min=min (r, g, b), δ=k _max-k _min.

(2) choose black, white, blue, yellow 4 kinds of colors as benchmark color, calculate R (x according to formula (4), y) distance of each pixel color value and 4 kinds of benchmark colors in, is with the benchmark color of this pixel distance minimum the color value that this point is new.Table 1 is the rgb value and corresponding HSV value of 4 kinds of benchmark colors.

Distance between two kinds of colors is:

d ₁＝(v ₁-v ₂) ²+(s ₁×cos(h ₁)-s ₂×cos(h ₂)) ²+(s ₁×sin(h ₁)-s ₂×sin(h ₂)) ²

Wherein, (h ₁, s ₁, v ₁) and (h ₂, s ₂, v ₂) be respectively and treat the front HSV value of correction of my good adjusting point and the HSV value as benchmark color, δ is scale-up factor, δ ∈ (0,1).

The rgb value of table 14 kind of benchmark color and corresponding HSV value

Step 3, scan revised image I 2 (x, y) (as shown in Figure 5), extract blueness, yellow concentrated region as vehicle candidate region.

(1) to image I 2 (x, y) line by line scan, the position of point that in every a line is vehicle background color (common car plate background color, for blue, yellow, is selected blue as car plate background color in this example) is recorded in array S, with the point of a line, be recorded in same a line of S.

(2) analyze in S and be in the point in every a line, if the distance of two points is less than s_min, think that these two points are continuous, 2 are linked to be a line segment, if be greater than s_min, discontinuous, the first and last position of line segment is recorded in array L_cur, judge after points all in same a line, the line segment being recorded in L_cur is screened, if line segment length is greater than l_min's, the position of line segment is recorded in array L, in L with having many line segments in a line, empty L_cur, until analyzed row all in S, in the present embodiment, preferably s_min gets 40, l_min gets 20.

(3) with the line segment initialization rectangle in array L, each line segment is initialized as a rectangle, horizontal two line numbers that limit is line segment place of rectangle, rectangular longitudinal to two limits be the row number at two some places of line segment, initialized rectangle is stored in array R.

(4) merge neighbouring rectangle in array R, if differing, longitudinal bar limit of two rectangles is less than 3 pixels, the upper widthwise edge of the lower widthwise edge below rectangle of top rectangle is separated by and is less than 2 pixels, merge, rectangle longitudinal edge invariant position after merging, the widthwise edge of top is the upper widthwise edge of former top rectangle, and lower widthwise edge is the lower widthwise edge of below rectangle, and two original rectangles are removed.

Step 4, verification vehicle region.Judge current region size, if region is within the scope of setting, thinks license plate area, otherwise cast out.Rectangle shown in R storage is judged, if rectangle length is greater than w_min, width is greater than h_min, thinks license plate area, and finally in original image, relevant position intercepts out car plate.W_min=40, h_min=20, intercepts result as shown in Figure 6.

Claims (1)

1. a license plate locating method, specifically comprises the steps:

(1) original vehicle image I (x, y) is carried out to pre-service, the vehicle image I1 (x, y) after being enhanced;

(2) color value of each point in correction vehicle image I1 (x, y), obtains revised vehicle image I2 (x, y);

(3) scan revised vehicle image I2 (x, y), extract blue and yellow concentrated region as license plate candidate area;

(4) verification license plate area, judges current license plate candidate area size, if current license plate candidate area is within the scope of setting, thinks license plate area, completes car plate location;

Wherein, in described step (1), carry out pretreated detailed process and be:

(1.1) license plate image function I (x, y) is expressed as to following form:

I(x,y)＝R(x,y)L(x,y)

Wherein (x, y) is the position coordinates of image mid point, and I (x, y) represents original vehicle image, and R (x, y) represents reflected light component, and L (x, y) represents light irradiation component;

(1.2) taken the logarithm in the expression formula both sides of license plate image function I (x, y), so by light irradiation component and reflected light component is expressed as and form:

LogI=Log(R.L)=LogR+LogL

(1.3) respectively original image I (x, y) is done to convolution by Gauss's template of n different scale, obtain the image Di (x, y) after corresponding n width low-pass filtering, wherein, i ∈ [1, n], n is positive integer:

$F{(x,y)}_{{\mathrm{\σ}}_i}=\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_i}\exp (-\frac{(x^2+y^2)}{2{\mathrm{\σ}}_i})$

$D_i(x,y)=I(x,y)*F{(x,y)}_{{\mathrm{\σ}}_i}$

In formula,

expression yardstick is σ _igaussian filter function;

(1.4) in log-domain, with the image D after original image I (x, y) and every width low-pass filtering _i(x, y) subtracts each other, and just can obtain the image of n width high frequency enhancement:

R _i(x,y)=LogI(x,y)-LogD _i(x,y)

(1.5) to the n width high frequency enhancement image weighted sum obtaining in (1.4), the vehicle image I1 (x, y) after being enhanced:

$I1(x,y)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ω}}_i{R}_i(x,y)$

Wherein, ω _ibe the weight of i width high frequency enhancement image,

In described step (2), the detailed process that the color value of each point in vehicle image I1 (x, y) is revised is:

(2.1) image is transformed into HSV space from rgb space;

(2.2) choose black, white, blue, yellow 4 kinds of colors as benchmark color, in calculating I1 (x, y), the distance of each pixel color value and 4 kinds of benchmark colors, is apart from minimum benchmark color the color value that this pixel is new with this pixel;

After having upgraded, described pixel color value obtains revised vehicle image I2 (x, y);

In described step (3), extraction license plate candidate area detailed process is:

(3.1) revised vehicle image I2 (x, y) is lined by line scan, record the line segment position for car plate background color color in each row, the pixel separation of two points of color of the same race is less than distance threshold and is considered as continuous point;

(3.2) every of analytic record line segment if line segment length is less than length threshold, is deleted this line segment from record;

(3.3) with record line segment initialization rectangle, each line segment is initialized as a rectangle, horizontal two line numbers that limit is line segment place of rectangle, rectangular longitudinal to two limits be the row number at two some places of line segment;

(3.4) merge adjacent rectangle, merge into size not in specified scope, delete.

Images