CN101931792A - Multi-plate number locating method in high definition video steaming - Google Patents

Abstract

The invention discloses a multiple plate numbers locating method in a high definition video steaming. In the method, aiming at a video image obtained by a high definition camera with a fixed viewing field, utilizing a background difference method to obtain a moving region and at the same time using morphological filter to remove noisy points; subsequently, carrying out vertical wavelet decomposition on the moving region to analyze the vertical wavelet components of the moving region, and extracting one or more candidate regions of the plate numbers according to multiple rules; and finally, carrying out color clustering on the candidate regions of the plate numbers, and judging whether the candidate region is a real plate number region according to color clustering information. The invention can correctly extract moving regions so that subsequent processing just aims at the moving regions, thereby avoiding full figure search and greatly improving the processing speed of plate number location; since the vertical wavelet components of the moving region are analyzed, most of non-plate number regions are filtered; and a color clustering method further ensures the accuracy of extracting the plate numbers, thereby realizing rapid and accurate location of the multiple plate numbers in the high definition video steaming.

Description

The method of many car plates location in a kind of high definition video steaming

Technical field

The invention belongs to area of pattern recognition, be specifically related to intelligent transportation monitoring technique field, the method for many car plates location in particularly a kind of high definition video steaming.

Background technology

Development along with Modern Transportation Technology, automobile has become indispensable a kind of important vehicles in people's daily life, this must have a higher requirement to traffic administration, along with the automation to urban traffic control, expressway tol lcollection, parking lot management etc., the research of license plate recognition technology becomes heat subject.The car plate location is vital link in the automatic Recognition of License Plate, and the speed of its location and order of accuarcy directly have influence on the performance of Vehicle License Plate Recognition System.Traditional intelligent transportation system based on SD (400,000 pixel) image, exists defectives such as the visual field is little, resolution is low mostly.In recent years, the high-definition camera of mega pixel level (200～5,000,000 pixel) begins to be used for traffic monitoring.Because the high-definition camera visual field is big, resolution is high, can provide strong evidence to law enforcement agency, thereby beginning is applied in intelligent transportation system.Often existing a plurality of car plates in the guarded region of high-definition camera, so the location quick and precisely of many car plates is directly determining the performance of intelligent transportation system, is a field that is worth further investigation.

Propositions such as Choi utilize Hough transformation to seek the method that vertical edge extracts license plate, and robustness is not relatively poor because the distortion of the vertical edge that many automotive front radiators produce and some licence plate frame or some license plate have frame for the method.Barroso etc. think that license plate area has stronger grey scale change characteristic, and image is done the projection of level and vertical direction respectively, judge the horizontal and vertical position of license plate area according to the characteristics of the crest of projection and trough.The method only depends on the gray scale projection characteristics location of car plate, is subjected to the interference in other zones easily, locatees often not accurate enough.Wear high official position etc. and propose a kind ofly based on morphology methods, basic way is: earlier image is carried out edge extracting, again the edge image is carried out morphology operations, only stay in image and treat target-seeking target.But the shortcoming of these class methods is: because need to determine in advance the size of filter operator, so the size of car plate is not had adaptivity.Talk a kind of license plate locating methods of proposition such as Yongxin, yet the texture of the noise of gradient feature road pavement or vehicle body is relatively more responsive, causes wrong report easily based on gradient projection.

Summary of the invention

The objective of the invention is to propose the method for many car plates location in a kind of high definition video steaming (resolution is greater than 2,000,000 pixels).That the method has is fast to the detection speed of car plate in the video flowing, accurate positioning, fail to report wrong report less, advantages such as anti-interference and good reliability, and can locate a plurality of car plates that occur in the frame simultaneously.

Technical scheme of the present invention is: the video image at the fixing high-definition camera of visual field obtains, and at first adopt the background subtraction point-score to obtain the moving region, adopt morphologic filtering to remove noise simultaneously; Then vertical wavelet decomposition is carried out in the moving region that obtains; Analyze the vertical Wavelet Component of moving region then, go out one or morely may be the candidate region of car plate according to multinomial Rule Extraction; At last color cluster is carried out in the candidate region of car plate, judge according to the information of color cluster whether this candidate region is a real license plate area.

Specific implementation steps in sequence of the present invention is:

(1) field scene image acquisition step: high-definition camera is uniformly-spaced to sample to obtain continuous field scene image to live video.

(2) be gray level image with the color scene image transitions of obtaining.Conversion formula is:

Gray＝0.299R+0.587G+0.114B (1)

Wherein, R, G, B are the redness of a pixel, green, and the value of blue three Color Channels, Gray is the pixel corresponding gray.

(3) adopt the background subtraction point-score to extract the foreground moving zone: for traffic scene, vehicle all move, so we only need detect car plate in the moving region and get final product, and can improve the speed of car plate detection so greatly.

1, the initialization of background image

Adopt the multi-frame mean method to come the initialization background image.With the mean value of the gray level image of initial n frame (n=100～200) as initial background image.

2, the background subtraction point-score extracts the foreground moving zone

With M (i, j, k) represent in the k two field picture foreground moving zone bianry image (i, the value of j) locating, then M (computing formula k) is for i, j:

Wherein (i, j are that the k two field picture of current input is in that ((i, j are that the background image of current k frame correspondence is in that (i, the grey scale pixel value of j) locating, T are given threshold value k) to B for i, the grey scale pixel value of j) locating k) to F.

3, the renewal of background image

Background image need constantly upgrade along with the variation of scene.Among the present invention, for guaranteeing that background image is pure as far as possible, the update strategy of employing is a selective updating, promptly only upgrades the background area, and the foreground moving zone is not upgraded.The expression formula of update strategy is:

Wherein: α ∈ (0,1) is called the study factor, and it has embodied the speed degree that background image upgrades along with the variation of scene.

(4) post-processing step: the region of variation that extracts is carried out morphology operations, to remove noise and the cavity in the foreground moving zone.

The basic operation of morphology operations is defined as respectively for expanding and erosion operation:

(f ⊕ g) (n)=max{f (n-x)+g (x) | (n-x) ∈ D _fAnd x ∈ D _g(4)

(f Θ g) (n)=min{f (n+x)-g (x) | (n+x) ∈ D _fAnd x ∈ D _g(5)

F in the formula (n) is an input signal, and the domain of definition is D _f=0,1,2 ..., N}; G (n) is the one-dimentional structure element sequence, and its domain of definition is D _g=0,1,2 ..., P}; Wherein P and N are integers, N 〉=P; ⊕ represents dilation operation, and Θ represents erosion operation.The definition that can be drawn morphology opening operation, closed operation by basic operation is respectively:

(fоg)(n)＝(fΘg⊕g)(n) (6)

(f·g)(n)＝(f⊕gΘg)(n) (7)

о represents opening operation in the formula, the expression closed operation, and opening operation can suppress normal burst (peak value) noise in the image, and closed operation can suppress negative pulse (low ebb) noise in the image.To the detected foreground moving of background subtraction point-score zone, carry out opening operation earlier one time, carry out a closed operation then.

(5) the vertical wavelet decomposition step of moving region: to carrying out vertical wavelet decomposition through the foreground moving zone behind the morphology operations.

Owing to exist a plurality of characters in the car plate, the vertical edge of character is generally all more clear obviously, so the present invention selects vertical wavelet decomposition is carried out in the moving region.In the present invention, adopt two-dimensional discrete wavelet conversion decomposing to respective regions.

$\left\{\begin{array}{c}\mathrm{\Φ}(x,y)=\mathrm{\φ}\left(x\right)\mathrm{\φ}\left(y\right)\\ {\mathrm{\Ψ}}^H(x,y)=\mathrm{\φ}\left(x\right)\mathrm{\ψ}\left(y\right)\\ {\mathrm{\Ψ}}^V(x,y)=\mathrm{\ψ}\left(x\right)\mathrm{\φ}\left(y\right)\\ {\mathrm{\Ψ}}^D(x,y)=\mathrm{\ψ}\left(x\right)\mathrm{\ψ}\left(y\right)\end{array}\right.---\left(8\right)$

Wherein, the low frequency component of Φ tolerance picture, Ψ ^HTolerance is along the variation (as horizontal edge) of row, Ψ ^HTolerance is along the variation (as vertical edge) of row, Ψ ^DVariation corresponding to diagonal.Original image is every through a two-dimensional discrete wavelet conversion, image all is broken down into the image of four 1/4th sizes, and each in four images all is to be generated by the interval sampling of all carrying out twice through x and y direction again after the inner product of an original image and a wavelet basis image.Through behind two-dimensional discrete wavelet conversion, image is broken down into that a width of cloth low-resolution image and three panel height frequency component subimages comprise horizontal direction (HL), vertical direction (LH) respectively and to the texture information of angular direction (HH).At this, only the Wavelet Component (LH) of vertical direction is analyzed, to extract the candidate region of car plate.

(6) license plate candidate area extraction step: the vertical Wavelet Component subgraph (LH) that previous step is calculated in rapid scans with the regional P that gives sizing, and the area of regional P should be slightly larger than the area of actual license plate.Below satisfy in the zone 3 when regular, think that then this zone is a license plate candidate area.Decision rule is as follows:

The vertical Wavelet Component average of rule 1, zoning P

Formula is:

$\stackrel{\‾}{M_P}=\frac{1}{n}\underset{(x,y)\∈P}{\mathrm{\Σ}}\left|\mathrm{LH}(x,y)\right|---\left(9\right)$

Wherein n is the total pixel number in the regional P.Because the vertical texture of character is outstanding on the car plate, so require the vertical Wavelet Component average of regional P to satisfy

T ₁Be given threshold value.

The vertical Wavelet Component variance V of rule 2, zoning P _P, formula is:

$V_P=\frac{1}{n}\underset{(x,y)\∈P}{\mathrm{\Σ}}{|\mathrm{LH}(x,y)-\stackrel{\‾}{M_P}|}^2---\left(10\right)$

Because the distribution of character is uniformly on the car plate, so require the vertical Wavelet Component variance of regional P to satisfy V _P＜T ₂, T ₂Be given threshold value.

The rule 3, in a bigger regional W who with regional P is the center, the average of regional P

Should

Be maximum, the area of regional W is generally got 9 times of regional P.

(7) car plate is confirmed step: the license plate candidate area that obtains is carried out color cluster, when decision rule satisfies, judge that this candidate region is real car plate, exports its coordinate information simultaneously; Otherwise, judge that then this candidate region is not real car plate, skip to the field scene image acquisition step, circulation is carried out.

Color is one of important information of differentiating by car plate, adopts the K means clustering algorithm that license plate candidate area is carried out color cluster among the present invention, to obtain this regional color characteristic.K means clustering algorithm principle is: accept input variable K, then n data object is divided into K cluster so that make the cluster that is obtained satisfy: the object similarity in the same cluster is higher; And the object similarity in the different clusters is less.The cluster similarity is to utilize the average of object in each cluster to obtain one " center object " to calculate.The course of work of K means clustering algorithm is described as follows: at first select K initial cluster center; And for by the data object of cluster, then, respectively they are distributed to (cluster centre representative) cluster the most similar to it according to the similarity (distance) of they and these cluster centres; And then calculate the cluster centre (average of all objects in this cluster) of each new cluster that obtains; Constantly repeat this process till the canonical measure function begins convergence.Generally all adopt mean square deviation to have following characteristics as .K cluster of canonical measure function: each cluster itself is compact as much as possible, and separates as much as possible between each cluster.

Common car plate color mainly is blue board (wrongly written or mispronounced character of the blue end) and yellow card (yellow end surplus) at present, so clusters number K is chosen as 4, initial cluster center is chosen as blueness (0,0,255), white (255,255,255), yellow (255,255,0), black (0,0,0).After treating the convergence of K means clustering algorithm, add up 4 class pixel number of color: blue N _Blue, white N _White, yellow N _YellowWith black N _BlackFor candidate region total pixel number N _P, if satisfy N _Blue＞0.3N _PAnd N _White＞0.3N _P, think that then the candidate region is blue car plate; If satisfy N _Yellow＞0.3N _PAnd N _Black＞0.3N _P, think that then the candidate region is yellow car plate; If above-mentioned two are not satisfied, think that then the candidate region is not real car plate.

The general employing of license plate locating method commonly used is at present carried out method for scanning to whole figure and is detected car plate, causes the car plate locating speed in high definition video steaming not fully up to expectations.The present invention adopts the background subtraction point-score to extract the foreground moving zone, makes follow-up processing only at the moving region, avoids carrying out the full figure search, has improved the processing speed of car plate location greatly.Background image utilizes selective updating strategy real-time update simultaneously, thereby can accurately extract the moving region, and outstanding performance is also arranged on antijamming capability.The gradient feature of traditional common analysis image of license plate locating method is located car plate, yet the texture of the noise of gradient feature road pavement or vehicle body is relatively more responsive, causes wrong report easily.The present invention carries out vertical wavelet decomposition to the moving region, analyzes the vertical Wavelet Component that obtains, and can filter out the non-license plate area of the overwhelming majority, has avoided the shortcoming of gradient feature.Adopt the method for color cluster further to guarantee to extract the accuracy of car plate at last.The present invention has realized the location quick and precisely to a plurality of car plates in the high definition video steaming, and all within very low scope, this is that former technology institute is inaccessiable for leakage, wrong report.

Description of drawings

The system construction drawing of many license plate locating methods in Fig. 1, the high definition video steaming of the present invention

The operational flowchart of many license plate locating methods in Fig. 2, the high definition video steaming of the present invention

Embodiment

Embodiment 1

The system that Fig. 1 has provided many license plate locating methods in the high definition video steaming of the present invention constitutes schematic diagram: the video image of the road area C that high definition camera D is captured is sent to embedded high definition car plate detector A, and embedded high definition car plate detector A utilizes the embedded video image analysis program that the operational flowchart of many license plate locating methods is write in the high definition video steaming according to the present invention that the video image that captures is carried out real-time analysis.Have car plate in the road area if detect, embedded high definition car plate detector A is sent to information such as car plate position coordinates and sectional drawing the monitor supervision platform B of rear end automatically; If road area does not detect car plate, then the first step of Returning process is restarted circulation.

Fig. 2 has provided the program flow diagram of many license plate locating methods in the high definition video steaming of the present invention.Concrete steps are: 1, obtain the current frame image data; 2, the coloured image with present frame is converted to gray level image; 3, obtain the foreground moving zone by the background subtraction point-score; 4, background image updating; 5, morphologic filtering is carried out in the foreground moving zone that obtains and remove noise; 6, vertical wavelet decomposition is carried out in the foreground moving zone after the denoising; 7, analyze the vertical Wavelet Component of moving region, go out one or morely may be the candidate region of car plate according to multinomial Rule Extraction; 8, adopt the color cluster algorithm to differentiate license plate candidate area, confirm as true car plate and then export the car plate position coordinates, return step 1 and circulate again.

The concrete operations step of many license plate locating methods is as follows in the high definition video steaming of the present invention:

1, high definition video steaming car plate detection system hardware platform builds

Monitoring camera requires the camera for the output high definition video steaming, according to shown in Figure 1 high definition video steaming is inserted embedded high definition car plate detector.Simultaneously, embedded high definition car plate detector can connect the monitor supervision platform of rear end by public network, uploads information such as car plate position coordinates and sectional drawing in real time.

2, obtain current frame image and be converted to gray level image

With the color scene image transitions of obtaining is gray level image.Conversion formula is:

Gray＝0.299R+0.587G+0.114B (1)

Wherein, R, G, B are the redness of a pixel, green, and the value of blue three Color Channels, Gray is the pixel corresponding gray.

3, the initialization of background image

When system start-up, system default is vehicle not occur in the monitored road scene.Adopt the multi-frame mean method to come the initialization background image.With the mean value of the gray level image of initial n frame (n=100～200) as initial background image.

4, adopt the background subtraction point-score to extract the foreground moving zone

With M (i, j, k) represent in the k two field picture foreground moving zone bianry image (i, the value of j) locating, then M (computing formula k) is for i, j:

Wherein (i, j are that the k two field picture of current input is in that ((i, j are that the background image of current k frame correspondence is in that (i, the grey scale pixel value of j) locating, T are given threshold value k) to B for i, the grey scale pixel value of j) locating k) to F.

5, the renewal of background image

It is selective updating that background image upgrades the update strategy that adopts, and promptly only upgrades the background area, and the foreground moving zone is not upgraded.The expression formula of update strategy is:

Wherein: α ∈ (0,1) is called the study factor, and it has embodied the speed degree that background image upgrades along with the variation of scene.

6, morphologic filtering

To the foreground moving zone that the background subtraction point-score detects, carry out opening operation earlier one time, carry out a closed operation then, to remove noise and the cavity in the foreground moving zone.

7, the vertical wavelet decomposition of moving region

To carrying out vertical wavelet decomposition through the foreground moving zone behind the morphology operations.The present invention adopts two-dimensional discrete wavelet conversion decomposing respective regions.Original image is every through a two-dimensional discrete wavelet conversion, image all is broken down into the image of four 1/4th sizes, and each in four images all is to be generated by the interval sampling of all carrying out twice through x and y direction again after the inner product of an original image and a wavelet basis image.Through behind two-dimensional discrete wavelet conversion, image is broken down into that a width of cloth low-resolution image and three panel height frequency component subimages comprise horizontal direction (HL), vertical direction (LH) respectively and to the texture information of angular direction (HH).At this, only the Wavelet Component (LH) of vertical direction is analyzed, to extract the candidate region of car plate.

8, extract candidate's license plate area

The vertical Wavelet Component subgraph (LH) that previous step is calculated in rapid scans with the regional P that gives sizing, and the area of regional P should be slightly larger than the area of actual license plate.Below satisfy in the zone 3 when regular, think that then this zone is a license plate candidate area.Decision rule is as follows:

The vertical Wavelet Component average of rule 1, zoning P

Formula is:

$\stackrel{\‾}{M_P}=\frac{1}{n}\underset{(x,y)\∈P}{\mathrm{\Σ}}\left|\mathrm{LH}(x,y)\right|---\left(4\right)$

Wherein n is the total pixel number in the regional P.Because the vertical texture of character is outstanding on the car plate, so require the vertical Wavelet Component average of regional P to satisfy

T ₁Be given threshold value.

The vertical Wavelet Component variance V of rule 2, zoning P _P, formula is:

$V_P=\frac{1}{n}\underset{(x,y)\∈P}{\mathrm{\Σ}}{|\mathrm{LH}(x,y)-\stackrel{\‾}{M_P}|}^2---\left(5\right)$

Because the distribution of character is uniformly on the car plate, so require the vertical Wavelet Component variance of regional P to satisfy V _P＜T ₂, T ₂Be given threshold value.

The rule 3, in a bigger regional W who with regional P is the center, the average of regional P

Should be maximum, the area of regional W is generally got 9 times of regional P.

9, car plate is confirmed

Color is one of important information of differentiating by car plate, so the last colouring information that adopts is confirmed the candidate's license plate area that obtains.Adopt the K means clustering algorithm that license plate candidate area is carried out color cluster among the present invention, to obtain this regional color characteristic.K means clustering algorithm principle is: accept input variable K, then n data object is divided into K cluster so that make the cluster that is obtained satisfy: the object similarity in the same cluster is higher; And the object similarity in the different clusters is less.The cluster similarity is to utilize the average of object in each cluster to obtain one " center object " to calculate.The course of work of K means clustering algorithm is described as follows: at first select K initial cluster center; And for by the data object of cluster, then, respectively they are distributed to (cluster centre representative) cluster the most similar to it according to the similarity (distance) of they and these cluster centres; And then calculate the cluster centre (average of all objects in this cluster) of each new cluster that obtains; Constantly repeat this process till the canonical measure function begins convergence.Generally all adopt mean square deviation to have following characteristics as .K cluster of canonical measure function: each cluster itself is compact as much as possible, and separates as much as possible between each cluster.

Common car plate color mainly is blue board (wrongly written or mispronounced character of the blue end) and yellow card (yellow end surplus) at present, so clusters number K is chosen as 4, initial cluster center is chosen as blueness (0,0,255), white (255,255,255), yellow (255,255,0), black (0,0,0).After treating the convergence of K means clustering algorithm, add up 4 class pixel number of color: blue N _Blue, white N _White, yellow N _YellowWith black N _BlackFor candidate region total pixel number N _P, if satisfy N _Blue＞0.3N _PAnd N _White＞0.3N _P, think that then the candidate region is blue car plate; If satisfy N _Yellow＞0.3N _PAnd N _Black＞0.3N _P, think that then the candidate region is yellow car plate; If above-mentioned two are not satisfied, think that then the candidate region is not real car plate.

After candidate's license plate area was confirmed to be real car plate, embedded high definition car plate detector can be with information such as the position coordinates of car plate and sectional drawings be uploaded to Surveillance center during by network implementation.

Claims (2)

1. the method for many car plates location in the high definition video steaming is characterized in that: the video image at the fixing high-definition camera of visual field obtains, and at first adopt background subtraction point-score method to obtain the moving region, adopt morphologic filtering removal noise simultaneously; Then vertical wavelet decomposition is carried out in the moving region that obtains; Analyze the vertical Wavelet Component of moving region then, go out one or morely may be the candidate region of car plate according to multinomial Rule Extraction; At last color cluster is carried out in the candidate region of car plate, judge according to the information of color cluster whether this candidate region is a real license plate area.

2. the method for many car plates location in a kind of high definition video steaming according to claim 1, its steps in sequence is:

1) obtains current frame image and be converted to gray level image

With the color scene image transitions of obtaining is gray level image; Conversion formula is:

Gray＝0.299R+0.587G+0.114B (1)

Wherein, R, G, B are the redness of a pixel, green, and the value of blue three Color Channels, Gray is the pixel corresponding gray;

2) initialization of background image

When system start-up, system default is vehicle not occur in the monitored road scene; Adopt the multi-frame mean method to come the initialization background image; With the mean value of the gray level image of initial n frame as initial background image, wherein, n=100～200;

3) adopt the background subtraction point-score to extract the foreground moving zone

With M (i, j, k) represent in the k two field picture foreground moving zone bianry image (i, the value of j) locating, then M (computing formula k) is for i, j:

Wherein F (i, j, k) be current input the k two field picture (i, the grey scale pixel value of j) locating, B (i, j, k) be current k frame correspondence background image (i, the grey scale pixel value of j) locating, T are given threshold value;

4) renewal of background image

It is selective updating that background image upgrades the update strategy that adopts, and promptly only upgrades the background area, and the foreground moving zone is not upgraded; The expression formula of update strategy is:

Wherein: α ∈ (0,1) is called the study factor, and it has embodied the speed degree that background image upgrades along with the variation of scene;

5) morphologic filtering

To the foreground moving zone that the background subtraction point-score detects, carry out opening operation earlier one time, carry out a closed operation then, to remove noise and the cavity in the foreground moving zone;

6) the vertical wavelet decomposition of moving region

To carrying out vertical wavelet decomposition through the foreground moving zone behind the morphology operations.The present invention adopts two-dimensional discrete wavelet conversion decomposing respective regions.Original image is every through a two-dimensional discrete wavelet conversion, image all is broken down into the image of four 1/4th sizes, and each in four images all is to be generated by the interval sampling of all carrying out twice through x and y direction again after the inner product of an original image and a wavelet basis image.Through behind two-dimensional discrete wavelet conversion, image is broken down into that a width of cloth low-resolution image and three panel height frequency component subimages comprise horizontal direction HL, vertical direction LH respectively and to the texture information of angular direction HH.At this, only the Wavelet Component LH to vertical direction analyzes, to extract the candidate region of car plate;

7) extract candidate's license plate area

The vertical Wavelet Component subgraph that previous step is calculated in rapid scans with the regional P that gives sizing, and the area of regional P should be slightly larger than the area of actual license plate; Below satisfy in the zone 3 when regular, think that then this zone is a license plate candidate area; Decision rule is as follows:

The vertical Wavelet Component average of rule 1, zoning P Formula is:

$\stackrel{\‾}{M_P}=\frac{1}{n}\underset{(x,y)\∈P}{\mathrm{\Σ}}\left|\mathrm{LH}(x,y)\right|---\left(4\right)$

Wherein n is the total pixel number in the regional P; Because the vertical texture of character is outstanding on the car plate, so require the vertical Wavelet Component average of regional P to satisfy

T ₁Be given threshold value;

The vertical Wavelet Component variance V of rule 2, zoning P _P, formula is:

$V_P=\frac{1}{n}\underset{(x,y)\∈P}{\mathrm{\Σ}}{|\mathrm{LH}(x,y)-\stackrel{\‾}{M_P}|}^2---\left(5\right)$

Because the distribution of character is uniformly on the car plate, so require the vertical Wavelet Component variance of regional P to satisfy V _P＜T ₂, T ₂Be given threshold value;

The rule 3, in a bigger regional W who with regional P is the center, the average of regional P

Should be maximum, the area of regional W is got 9 times of regional P;

8) car plate is confirmed

Adopt colouring information that the candidate's license plate area that obtains is confirmed at last; Adopt the K means clustering algorithm that license plate candidate area is carried out color cluster among the present invention, to obtain this regional color characteristic;

After treating the convergence of K means clustering algorithm, add up 4 class pixel number of color: blue pixel N _Blue, white pixel N _White, yellow pixel N _YellowWith black picture element N _Black

For candidate region total pixel number N _P, if satisfy N _Blue＞0.3N _PAnd N _White＞0.3N _P, think that then the candidate region is blue car plate; If satisfy N _Yellow＞0.3N _PAnd N _Black＞0.3N _P, think that then the candidate region is yellow car plate; If above-mentioned two are not satisfied, think that then the candidate region is not real car plate.

Images