CN103745197B - A kind of detection method of license plate and device - Google Patents

A kind of detection method of license plate and device Download PDF

Info

Publication number
CN103745197B
CN103745197B CN201310739338.4A CN201310739338A CN103745197B CN 103745197 B CN103745197 B CN 103745197B CN 201310739338 A CN201310739338 A CN 201310739338A CN 103745197 B CN103745197 B CN 103745197B
Authority
CN
China
Prior art keywords
image
square frame
hog
models
edge
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201310739338.4A
Other languages
Chinese (zh)
Other versions
CN103745197A (en
Inventor
唐健
李昕
李锐
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Jieshun Science and Technology Industry Co Ltd
Original Assignee
Shenzhen Jieshun Science and Technology Industry Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen Jieshun Science and Technology Industry Co Ltd filed Critical Shenzhen Jieshun Science and Technology Industry Co Ltd
Priority to CN201310739338.4A priority Critical patent/CN103745197B/en
Publication of CN103745197A publication Critical patent/CN103745197A/en
Application granted granted Critical
Publication of CN103745197B publication Critical patent/CN103745197B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Abstract

The present invention is applied to image processing field there is provided a kind of detection method of license plate and device, and methods described comprises the steps:Multiple scale detecting is carried out to image using HOG models, and extracts the boxed area by HOG model inspections;Edge calculations are carried out to image, and calculate the integrogram of the image after edge calculations, the edge integrogram of image is acquired;The boxed area by HOG model inspections is obtained on the edge integrogram, the marginal density of the boxed area by HOG model inspections is calculated, the square frame that marginal density is more than or equal to threshold value is extracted;The square frame that marginal density is more than or equal to threshold value is detected by Haar models, detection obtains accurate license plate area.The present invention is detected by HOG features to image, then carries out rim detection and marginal density detection, license plate area is gone out finally by Haar model inspections.

Description

A kind of detection method of license plate and device
Technical field
The invention belongs to image processing field, more particularly to a kind of detection method of license plate and device.
Background technology
Car license recognition has been widely used in parking lot gateway, the place such as mall gateway.Car license recognition typically divides For car plate detection, segmentation of the characters and their identification and car plate are voted this three step.Car plate detection, i.e., detect car plate from video, and Ticket detection is again a very important step in Car license recognition, is also a most time-consuming step.Car plate detection is exactly in vehicle image The position of middle detection car plate.Because car license recognition equipment is typically all that, in outdoor application, the image-forming condition of car plate and background is not It is controllable, illumination, complex background, the setting angle of video camera and apart from etc. bring difficulty to car plate detection.License plate area Shared ratio is smaller in entire image, to detect that license plate area must be in substantial amounts of background information from entire image In scan for.There is the limitation to the time in actual life to the identification application scenarios of car plate, therefore more need quick inspection Measure car plate.Detection method of license plate based on video has a lot, including the angle detection in the binary image based on line template is calculated Method, car plate is detected using genetic algorithm, also has Adaboost detections based on Haar features etc..Wherein based on Haar features Adaboost training patterns carry out full figure search.This method verification and measurement ratio is high, but speed is slow, and full figure, which carries out Haar detections, accurately to be examined Car plate is measured, but operand is big, is wanted if applied in embedded device, the arithmetic speed of embedded device is difficult to reach Ask.Other Haar detections angle requirement is high, and certain missing inspection occurs if car plate angle has exceeded 30 degree.
The content of the invention
The purpose of the embodiment of the present invention is to provide a kind of detection method of license plate and device, it is desirable to provide a kind of car plate detection Method, is detected by HOG features to image, rim detection and marginal density detection is then carried out, finally by Haar moulds Type detects license plate area.
The embodiment of the present invention is achieved in that a kind of detection method of license plate, and methods described comprises the steps:
Multiple scale detecting is carried out to image using HOG models, and extracts the boxed area by HOG model inspections;
Edge calculations are carried out to image, and calculate the integrogram of the image after edge calculations, image is acquired Edge integrogram;
The boxed area by HOG model inspections is obtained on the edge integrogram, is calculated by HOG model inspections The marginal density of boxed area, extracts the square frame that marginal density is more than or equal to threshold value;
To marginal density be more than or equal to threshold value square frame divide equally, and calculating divide equally rear each square frame average edge it is close Degree, average edge density variation is obtained according to average edge density, and square frame of the calibration offset in threshold range is similar car plate Square frame;
The square frame that marginal density is more than or equal to threshold value is detected and with reference to the similar car plate side by Haar models Frame, detection obtains accurate license plate area.
Further, methods described includes the step of carrying out multiple scale detecting to image using HOG models before:
Image is changed, gray-scale map is converted the image into;
Image is zoomed in and out.
Further, methods described is changed to image, is also included before the step of converting the image into gray-scale map:
According to the positive negative sample of collection, HOG models are set up.
Further, it is described to be detected and combined institute to the square frame that marginal density is more than or equal to threshold value by Haar models Similar car plate square frame is stated, the step of detection obtains accurate license plate area includes:
Haar features are trained by the Adaboost algorithm of cascade and obtain Haar models;
The square frame that marginal density is more than or equal to threshold value is detected and with reference to the similar car plate side by Haar models Frame, detection obtains accurate license plate area.
The another object of the embodiment of the present invention is to provide a kind of car plate detection device, and described device includes:
Boxed area extraction unit, HOG moulds are passed through for carrying out multiple scale detecting to image using HOG models, and extracting The boxed area of type detection;
Integrogram computing unit, for carrying out edge calculations to image, and calculates the product of the image after edge calculations Component, acquires the edge integrogram of image;
Square frame extraction unit, for obtaining the boxed area by HOG model inspections on the edge integrogram, is calculated By the marginal density of the boxed area of HOG model inspections, the square frame that marginal density is more than or equal to threshold value is extracted;
Unit is demarcated, the square frame for being more than or equal to threshold value to marginal density is divided equally, and rear each square frame is divided in calculating equally Average edge density, average edge density variation, side of the calibration offset in threshold range are obtained according to average edge density Frame is similar car plate square frame;
License plate area detection unit, for being detected by Haar models to the square frame that marginal density is more than or equal to threshold value And with reference to the similar car plate square frame, detection obtains accurate license plate area.
Further, described device also includes:
Image changes unit for scaling, for being changed to image, converts the image into gray-scale map;Image is contracted Put.Further, described device also includes:
Image conversion unit, for being changed to image, converts the image into gray-scale map.
Further, the license plate area detection unit also includes:
Model training unit, is trained to Haar features for the Adaboost algorithm by cascade and obtains Haar moulds Type;
Accurate detection unit, for the square frame that marginal density is more than or equal to threshold value to be detected and tied by Haar models The similar car plate square frame is closed, detection obtains accurate license plate area.
The embodiment of the present invention acquires some sides by HOG features first by a kind of detection method of license plate and device Frame region, while edge calculations are carried out to image calculates the edge integrogram for obtaining image, passes through obtained boxed area and side Edge integrogram accurately extracts the square frame for meeting preparatory condition again, goes out license plate area finally by Haar model inspections, due to head First employing the progress detection of HOG features can realize that wide-angle is detected, while the model of HOG features is to light robust so that most Probability afterwards by Haar model inspections to car plate is higher.
Brief description of the drawings
Fig. 1 is a kind of implementation process figure for detection method of license plate that first embodiment of the invention is provided;
Fig. 2 is a kind of implementation process figure for detection method of license plate that second embodiment of the invention is provided;
Fig. 3 is a kind of structure chart for car plate detection device that second embodiment of the invention is provided;And
Fig. 4 is a kind of structure chart for car plate detection device that second embodiment of the invention is provided.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and It is not used in the restriction present invention.
Implementing for the present invention is described in detail below in conjunction with specific embodiment:
Embodiment one:
Fig. 1 shows a kind of implementation process for detection method of license plate that first embodiment of the invention is provided, and details are as follows:
S101, carries out multiple scale detecting to image, and extract the boxed area by HOG model inspections using HOG models.
Using HOG (Histogram of oriented gradients) gradient orientation histogram feature to each width just, Negative sample is characterized, and forms HOG characteristic vectors.The partial gradient amplitude and direction character of HOG picture engravings, it is special based on gradient Levy and the characteristic vector of block is normalized.It is overlapped between HOG permission blocks, therefore to illumination variation and in a small amount inclined Move and insensitive, can effectively depict edge feature.It is good for the big car plate detection effect of angle.Pass through formation afterwards HOG characteristic vectors set up HOG models, multiple scale detecting are carried out by HOG models, to what can each be verified in image by HOG Region, is all shown with a boxed area, extracts the boxed area by HOG model inspections.
S102, edge calculations are carried out to image, and calculate the integrogram of the image after edge calculations, acquire figure The edge integrogram of picture.
Rule of thumb, car plate is typically at edge than comparatively dense and the region that is evenly distributed, calculates the side of image full figure Edge, for the convenience of search, then calculates the integrogram of the full figure after edge calculations, referred to herein as edge integrogram, obtains Obtain the edge integrogram of image.
S103, the boxed area by HOG model inspections is obtained on the edge integrogram, calculates and passes through HOG models The marginal density of the boxed area of detection, extracts the square frame that marginal density is more than or equal to threshold value.
Suspected license plate area is outlined on edge integrogram using the boxed area of acquisition, then these are calculated by integrogram The marginal density in region, deletes the square frame that marginal density in gray-scale map is less than threshold value, extracts marginal density and is more than or equal to threshold value Square frame.
S104, detects, detection obtains accurate by Haar models to the square frame that marginal density is more than or equal to threshold value License plate area.
Detection carries out cluster operation after terminating to remaining square frame, and all square frame occupied areas of each class are linked up to be formed Block, carries out detection to remaining piece by Haar models afterwards and obtains accurate car plate, detect accurate license plate area.
The embodiment of the present invention is by the above-mentioned means, a kind of detection method of license plate, is acquired by HOG features first Boxed area, the edge integrogram of image is obtained while carrying out edge calculations to image and calculating, by obtained boxed area with Edge integrogram accurately extracts the square frame for meeting preparatory condition again, and license plate area is gone out finally by Haar model inspections, due to Employing the progress detection of HOG features first can realize that wide-angle is detected, while the model of HOG features is to light robust so that Probability finally by Haar model inspections to car plate is higher.
Embodiment two:
Fig. 2 shows a kind of implementation process for detection method of license plate that second embodiment of the invention is provided, and details are as follows:
S201, according to the positive negative sample of collection, sets up HOG models.
A large amount of car plate pictures and non-car plate picture are selected as the positive and negative samples of training set, the process of selection can lead to Cross artificial mode to be chosen, each width positive and negative samples are characterized using HOG features afterwards, form characteristic vector, build Vertical HOG models.
S202, is changed to image, converts the image into gray-scale map;Image is zoomed in and out.
Camera is subjected to image conversion process in the image captured first, gray level image is converted the image into, passes through Convert the image into gray level image, it is possible to reduce to the amount of calculation of image procossing;Image after conversion is zoomed in and out, specifically The width and height of picture can respectively be reduced half in implementation process.Can further it be reduced after image is zoomed in and out to figure The amount of calculation of picture, improves the speed for detecting license plate area.
S203, carries out multiple scale detecting to image, and extract the boxed area by HOG model inspections using HOG models.
Multiple scale detecting is carried out to image using HOG models, the wherein acquisition of HOG models can be by camera at each The video of a large amount of parking lot gateways is shot and stored under period and weather condition.In these videos, each is manually intercepted out The license plate image of car, license plate image as positive sample, manually intercept out be not wherein car plate part, the non-car plate figure intercepted out Piece is used as negative sample.Each width positive and negative samples are characterized using HOG features, HOG characteristic vectors are formed.HOG picture engravings Partial gradient amplitude and direction character, the characteristic vector of block is normalized based on Gradient Features.HOG allow block it Between it is overlapped, therefore to illumination variation and a small amount of skews and insensitive, can effectively depict edge feature.For angle Big car plate detection effect is good.HOG features are the gradient statistical informations of gray-scale map, and gradient is primarily present in the place at edge. Calculate the First-order Gradient of image, gradient magnitude: Gradient direction:Ang (x, y)=arccos (I (x+1, y)-I (x-1, y)/R).Specifically can be with gradient calculation, using it is one-dimensional from Dissipate differential template [- 10 1] in the horizontal direction to handle sample image, obtain horizontal gradient image;Using it is one-dimensional from Dissipate differential template [- 10 1]TSample image is handled in vertical direction, vertical gradient image is obtained.Histogram direction It can be specifically 9, the one-dimensional histogram of gradients of all pixels in each piecemeal is added to wherein, is formed final Feature.HOG models are set up finally by final feature, HOG features are trained using the Adaboost algorithm of cascade, most 5 layers of cascade car plate strong classifier model is obtained eventually, reuses HOG models and multiple scale detecting is carried out to image, extract and pass through HOG The boxed area of model inspection.
S204, edge calculations are carried out to image, and calculate the integrogram of the image after edge calculations, acquire figure The edge integrogram of picture.
Edge calculations are carried out to image, the integrogram of the image after edge calculations is calculated, edge calculations are passed through in calculating The integrogram of image afterwards, acquires the edge integrogram of image.
S205, the boxed area by HOG model inspections is obtained on the edge integrogram, calculates and passes through HOG models The marginal density of the boxed area of detection, extracts the square frame that marginal density is more than or equal to threshold value.
Doubtful car plate area is outlined on edge integrogram by the coordinate of the boxed area of the HOG model inspections of acquisition afterwards Domain, the marginal density in these regions is calculated by integrogram, is deleted the square frame that marginal density in gray-scale map is less than threshold value, is extracted side Edge density is more than or equal to the square frame of threshold value.The calculating of marginal density can wherein be accelerated using integrogram.In order to avoid a side Frame computing repeatedly of being added of marginal value a little, integrogram has been used in the algorithm.Each point (x, y) bag on integrogram The marginal value from point (0,0) to all pixels of point (x, y) is contained.If representing the gray-scale map after edge calculations with I (x, y), II (x, y) represents the integrogram calculated, then:An arbitrary rectangle can use following Mode calculate:If the upper left corner of rectangle is (xlt,ylt), the coordinate in the lower right corner is (xrb,yrb), then the integration of the rectangle Figure can be calculated with following formula:SUMD=II (xrb,yrb)-II(xlt,yrb)-II(xrb,ylt)+II(xlt,ylt), wherein for threshold value Setting it is different and variant according to specific application scenarios and the sample of training, the threshold value set in the present embodiment is 7. Extract the square frame that marginal density is more than or equal to threshold value.
S206, divides equally to the square frame that marginal density is more than or equal to threshold value, and calculates the average side for dividing rear each square frame equally Edge density, average edge density variation is obtained according to average edge density, and square frame of the calibration offset in threshold range is similar Car plate square frame.
There are some rectangle objects in background and vehicle body, feature is similar with car plate, but edge distribution is uneven.For example some Car light can be car plate by flase drop together with the region on side and marginal density is high, but its Kernel density estimate is uneven.We need The square frame that marginal density is more than or equal to threshold value is divided equally, specifically the square frame detected can be divided into 5 pieces, point This 5 pieces of average edge density is not calculated.Then the deviation of this 5 pieces of averag density is calculated, if deviation is in threshold value, is represented The edge distribution formula in this region is the region of similar car plate in a balanced way, can also be using square frame to the local of similar car plate Cluster, specific is the method using arest neighbors, as long as coincidence between any two reaches that 60% is considered as the two frames and belongs to One class.Deleted after cluster according to default deletion condition with part square frame:Rule of thumb, a license plate area searches out what is come Square frame can be a lot, therefore just all square frames of whole class are deleted if the square frame that classification is included is less than 10 after cluster.Such as The position that really multiple scale are detected is overlapped, and only retains the scale of centre, deletes the square frame that others scale is detected.Together One scale is translated in detection process, if the multiple square frames detected two neighboring large area occur and overlapped, is only protected Stay center square frame, and the square frame up and down, left and right directionss expand 10% respectively.Finally all square frame institutes of each class Account for area to link up to form block, finally demarcation is linked up to form block for similar car plate square frame.
S207, detects, detection obtains accurate by Haar models to the square frame that marginal density is more than or equal to threshold value License plate area.
Remaining piece is detected finally by Haar models, obtain accurate car plate position, wherein Haar models are Haar features are trained using the Adaboost algorithm of cascade, Haar models are obtained by training, specifically can be by people Work selects a large amount of car plate pictures and non-car plate picture as the positive and negative samples of training set.Using Haar features to each width just, Negative sample is characterized, and forms characteristic vector, obtains 20 layers of cascade car plate strong classifier.Haar models are formed, pass through Haar Model detects that detection obtains accurate license plate area to the square frame that marginal density is more than or equal to threshold value.
The embodiment of the present invention is by the above-mentioned means, a kind of detection method of license plate, is carried out complete first with 5 layers of HOG models Figure detection, then carries out marginal density detection to suspected license plate area and marginal density equilibrium is detected.Then using cluster Method is close region clustering into suspected license plate area.Finally to the Haar models of remaining several 20 layers of license plate areas progress The method of detection.This method carries out Haar detection speeds soon than traditional full figure, and utilizes HOG characteristic energy detection angles Big car plate, the results show, the verification and measurement ratio of this method is high, and to light robust.
Embodiment three:
Fig. 3 shows a kind of structure chart for car plate detection device that third embodiment of the invention is provided, for convenience of description, It illustrate only the part related to the embodiment of the present invention.
Boxed area extraction unit 31, HOG is passed through for carrying out multiple scale detecting to image using HOG models, and extracting The boxed area of model inspection.
Boxed area extraction unit is used to characterize each width positive and negative samples using HOG features, forms HOG features Vector.The partial gradient amplitude and direction character of HOG picture engravings, the characteristic vector of block is normalized based on Gradient Features Processing.HOG allows overlapped between block, therefore to illumination variation and a small amount of skews and insensitive, can effectively depict Edge feature.It is good for the big car plate detection effect of angle.HOG models are set up by the HOG characteristic vectors of formation afterwards, passed through HOG models carry out multiple scale detecting, to the region that can be each verified in image by HOG, are all shown, carried with a boxed area Take the boxed area by HOG model inspections.
Integrogram computing unit 32, for carrying out edge calculations to image, and calculates the image after edge calculations Integrogram, acquires the edge integrogram of image.
Rule of thumb, car plate is typically at edge than comparatively dense and the region that is evenly distributed, passes through integrogram and calculates single Member calculates the edge of image full figure, for the convenience of search, then calculates the integrogram of the full figure after edge calculations, referred to here as For edge integrogram, the edge integrogram of image is acquired.
Square frame extraction unit 33, for obtaining the boxed area by HOG model inspections, meter on the edge integrogram The marginal density of the boxed area by HOG model inspections is calculated, the square frame that marginal density is more than or equal to threshold value is extracted.
Square frame extraction unit is used to outline suspected license plate area on edge integrogram using the boxed area of acquisition, then leads to The marginal density that integrogram calculates these regions is crossed, the square frame that marginal density in gray-scale map is less than threshold value is deleted, extracts edge close Square frame of the degree more than or equal to threshold value.
License plate area detection unit 34, for being examined by Haar models to the square frame that marginal density is more than or equal to threshold value Survey, detection obtains accurate license plate area.
License plate area detection unit carries out cluster operation after terminating for detection to remaining square frame, by all sides of each class Frame occupied area links up to form block, and detection is carried out to remaining piece by Haar models afterwards obtains accurate car plate, detection Go out accurate license plate area.
The embodiment of the present invention is by the above-mentioned means, a kind of car plate detection device, is acquired by HOG features first Boxed area, the edge integrogram of image is obtained while carrying out edge calculations to image and calculating, by obtained boxed area with Edge integrogram accurately extracts the square frame for meeting preparatory condition again, and license plate area is gone out finally by Haar model inspections, due to Employing the progress detection of HOG features first can realize that wide-angle is detected, while the model of HOG features is to light robust so that Probability finally by Haar model inspections to car plate is higher.
Example IV:
Fig. 4 shows a kind of structure chart for car plate detection device that fourth embodiment of the invention is provided, for convenience of description, It illustrate only the part related to the embodiment of the present invention.
Model sets up unit 41, according to the positive negative sample of collection, sets up HOG models.
Model sets up unit according to selecting the positive and negative samples of a large amount of car plate pictures and non-car plate picture as training set, The process of selection can be chosen by artificial mode, carry out table to each width positive and negative samples using HOG features afterwards Levy, form characteristic vector, set up HOG models.
Image changes unit for scaling 42, for being changed to image, converts the image into gray-scale map;Image is contracted Put.
Image conversion unit for scaling can specifically include:Image conversion unit and image scaling unit, wherein image are changed Unit is used to camera is carried out into image conversion process in the image captured first, converts the image into gray level image, passes through Convert the image into gray level image, it is possible to reduce to the amount of calculation of image procossing;Wherein image scaling unit will be used for after changing Image zoom in and out, the width and height of picture can respectively be reduced half in specific implementation process.Image is contracted The amount of calculation to image can be further reduced after putting, the speed for detecting license plate area is improved.
Boxed area extraction unit 43, HOG is passed through for carrying out multiple scale detecting to image using HOG models, and extracting The boxed area of model inspection.
Boxed area extraction unit is used for the acquisition for carrying out multiple scale detecting, wherein HOG models to image using HOG models The video of a large amount of parking lot gateways can be shot and stored under each period and weather condition by camera.Regarded at these In frequency, manually intercept out the license plate image of each car, license plate image as positive sample, manually intercept out be not wherein car plate portion Point, the non-car plate picture intercepted out is as negative sample.Each width positive and negative samples are characterized using HOG features, HOG is formed Characteristic vector.The partial gradient amplitude and direction character of HOG picture engravings, based on characteristic vector of the Gradient Features to block It is normalized.It is overlapped between HOG permission blocks, therefore to illumination variation and a small amount of skews not Sensitivity, can effectively depict edge feature.It is good for the big car plate detection effect of angle.HOG features are gray-scale maps Gradient statistical information, and gradient is primarily present in the place at edge.Calculate the First-order Gradient of image, gradient magnitude:Gradient direction:Ang (x, y)=arccos (I(x+1,y)-I(x-1,y)/R).Specifically can be with using one-dimensional discrete differential template [- 10 1] in level side in gradient calculation Sample image is handled upwards, horizontal gradient image is obtained;Use one-dimensional discrete differential template [- 10 1]TIn Vertical Square Sample image is handled upwards, vertical gradient image is obtained.Histogram direction can be specifically 9, by each piecemeal The one-dimensional histogram of gradients of middle all pixels is added to wherein, is formed final feature.Built finally by final feature Vertical HOG models, are trained, the cascade car plate for finally giving 5 layers is classified by force using the Adaboost algorithm of cascade to HOG features Device model, reuses HOG models and carries out multiple scale detecting to image, extract the boxed area by HOG model inspections.
Integrogram computing unit 44, for carrying out edge calculations to image, and calculates the image after edge calculations Integrogram, acquires the edge integrogram of image.
Integrogram computing unit is used to carry out edge calculations to image, calculates the integration of the image after edge calculations Figure, calculates the integrogram of the image after edge calculations, acquires the edge integrogram of image.
Square frame extraction unit 45, for obtaining the boxed area by HOG model inspections, meter on the edge integrogram The marginal density of the boxed area by HOG model inspections is calculated, the square frame that marginal density is more than or equal to threshold value is extracted.
Square frame extraction unit is used for the coordinate of the boxed area of the HOG model inspections by obtaining in edge integrogram upper ledge Go out suspected license plate area, the marginal density in these regions is calculated by integrogram, delete marginal density in gray-scale map and be less than threshold value Square frame, extract marginal density be more than or equal to threshold value square frame.The calculating of marginal density can wherein be accelerated using integrogram.For Avoid square frame computing repeatedly of being added of marginal value a little, integrogram has been used in the algorithm.On integrogram Each point (x, y) contains the marginal value from point (0,0) to all pixels of point (x, y).If representing edge meter with I (x, y) Gray-scale map after calculation, II (x, y) represents the integrogram calculated, then:An arbitrary square Shape can use following mode to calculate:If the upper left corner of rectangle is (xlt,ylt), the coordinate in the lower right corner is (xrb,yrb), that The integrogram of the rectangle can be calculated with following formula:SUMD=II (xrb,yrb)-II(xlt,yrb)-II(xrb,ylt)+II(xlt, ylt), wherein the setting for threshold value is variant according to specific application scenarios and the sample of training difference, in the present embodiment The middle threshold value set is 7.Extract the square frame that marginal density is more than or equal to threshold value.
Unit 46 is demarcated, the square frame for being more than or equal to threshold value to marginal density is divided equally, and rear each side is divided in calculating equally The average edge density of frame, obtains average edge density variation, calibration offset is in threshold range according to average edge density Square frame is similar car plate square frame.
Demarcating unit is used for there are some rectangle objects in background and vehicle body, and feature is similar with car plate, but edge distribution is not Uniformly.Such as some car lights can be car plate by flase drop together with the region on side and marginal density is high, but its Kernel density estimate It is uneven.We need to divide the square frame that marginal density is more than or equal to threshold value equally, specifically can be the square frame detected 5 pieces are divided into, this 5 pieces of average edge density is calculated respectively.Then the deviation of this 5 pieces of averag density is calculated, if partially Difference is in threshold value, and the edge distribution formula for representing this region is the region of similar car plate in a balanced way, to the local of similar car plate also The cluster of square frame can be used, specific is the method using arest neighbors, as long as coincidence between any two reaches that 60% is considered as The two frames belong to a class.Deleted after cluster according to default deletion condition with part square frame:Rule of thumb, a car plate The square frame that range searching comes out can be a lot, therefore if the square frame that classification is included is less than 10 after cluster, just whole class All square frames are deleted.If the position that multiple scale are detected is overlapped, only retain the scale of centre, delete others scale The square frame detected.Same scale is translated in detection process, if the multiple square frames appearance detected is two neighboring Large area overlap, only retain center square frame, and the square frame up and down, left and right directionss expand 10% respectively.Finally every All square frame occupied areas of one class link up to form block, and finally demarcation is linked up to form block for similar car plate square frame.
License plate area detection unit 47, for being examined by Haar models to the square frame that marginal density is more than or equal to threshold value Survey, detection obtains accurate license plate area.
License plate area detection unit includes:Model training unit and accurate detection unit, model training unit are used to pass through The Adaboost algorithm of cascade is trained to Haar features and obtains Haar models;Accurate detection unit is used to pass through Haar models The square frame that marginal density is more than or equal to threshold value is detected, detection obtains accurate license plate area.It is specifically license plate area Detection unit obtains accurate car plate position, wherein Haar moulds for being detected finally by Haar models to remaining piece Type is the model that is trained to Haar features of Adaboost algorithm using cascade, will manually select a large amount of car plate pictures and Non- car plate picture as training set positive and negative samples.Each width positive and negative samples are characterized using Haar features, form special Vector is levied, 20 layers of cascade car plate strong classifier is obtained.Haar models are formed, marginal density is more than or equal to by Haar models The square frame of threshold value is detected that detection obtains accurate license plate area.
The embodiment of the present invention is by the above-mentioned means, a kind of car plate detection device, is carried out complete first with 5 layers of HOG models Figure detection, then carries out marginal density detection to suspected license plate area and marginal density equilibrium is detected.Then using cluster Method is close region clustering into suspected license plate area.Finally to the Haar models of remaining several 20 layers of license plate areas progress The method of detection.This method carries out Haar detection speeds soon than traditional full figure, and utilizes HOG characteristic energy detection angles Big car plate, the results show, the verification and measurement ratio of this method is high, and to light robust.
Can be with one of ordinary skill in the art will appreciate that realizing that all or part of step in above-described embodiment method is The hardware of correlation is instructed to complete by program, described program can be stored in computer read/write memory medium, institute The storage medium stated, such as ROM/RAM, disk, CD.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention Any modifications, equivalent substitutions and improvements made within refreshing and principle etc., should be included in the scope of the protection.

Claims (8)

1. a kind of detection method of license plate, it is characterised in that methods described comprises the steps:
Multiple scale detecting is carried out to image using HOG models, and extracts the boxed area by HOG model inspections;
Edge calculations are carried out to image, and calculate the integrogram of the image after edge calculations, the edge of image is acquired Integrogram;
The boxed area by HOG model inspections is obtained on the edge integrogram, the square frame by HOG model inspections is calculated The marginal density in region, extracts the square frame that marginal density is more than or equal to threshold value;
The square frame that marginal density is more than or equal to threshold value is divided equally, and calculates the average edge density for dividing rear each square frame equally, root Average edge density variation is obtained according to average edge density, square frame of the calibration offset in threshold range is similar car plate square frame;
The square frame that marginal density is more than or equal to threshold value is detected by Haar models and with reference to the similar car plate square frame, inspection Measure accurate license plate area.
2. the method as described in claim 1, it is characterised in that methods described is multiple dimensioned to image progress using HOG models Included before the step of detection:
Image is changed, gray-scale map is converted the image into;
Image is zoomed in and out.
3. method as claimed in claim 2, it is characterised in that methods described is changed to image, is converted the image into Also included before the step of gray-scale map:
According to the positive negative sample of collection, HOG models are set up.
4. the method as described in claim 1, it is characterised in that described that threshold value is more than or equal to marginal density by Haar models Square frame detected and that with reference to the similar car plate square frame the step of detection obtains accurate license plate area includes:
Haar features are trained by the Adaboost algorithm of cascade and obtain Haar models;
The square frame that marginal density is more than or equal to threshold value is detected by Haar models and with reference to the similar car plate square frame, inspection Measure accurate license plate area.
5. a kind of car plate detection device, it is characterised in that described device includes:
Boxed area extraction unit, is examined for carrying out multiple scale detecting to image using HOG models, and extracting by HOG models The boxed area of survey;
Integrogram computing unit, for carrying out edge calculations, and the integrogram of image of the calculating after edge calculations to image, Acquire the edge integrogram of image;
Square frame extraction unit, for obtaining the boxed area by HOG model inspections on the edge integrogram, calculating passes through The marginal density of the boxed area of HOG model inspections, extracts the square frame that marginal density is more than or equal to threshold value;
Unit is demarcated, the square frame for being more than or equal to threshold value to marginal density is divided equally, and the flat of rear each square frame is divided in calculating equally Equal marginal density, obtains average edge density variation, square frame of the calibration offset in threshold range is according to average edge density Similar car plate square frame;
License plate area detection unit, for the square frame that marginal density is more than or equal to threshold value to be detected and tied by Haar models The similar car plate square frame is closed, detection obtains accurate license plate area.
6. device as claimed in claim 5, it is characterised in that described device also includes:
Image changes unit for scaling, for being changed to image, converts the image into gray-scale map;Image is zoomed in and out.
7. device as claimed in claim 5, it is characterised in that described device also includes:
Model sets up unit, according to the positive negative sample of collection, sets up HOG models.
8. device as claimed in claim 5, it is characterised in that the license plate area detection unit also includes:
Model training unit, is trained to Haar features for the Adaboost algorithm by cascade and obtains Haar models;
Accurate detection unit, for being detected and being combined institute to the square frame that marginal density is more than or equal to threshold value by Haar models Similar car plate square frame is stated, detection obtains accurate license plate area.
CN201310739338.4A 2013-12-27 2013-12-27 A kind of detection method of license plate and device Active CN103745197B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201310739338.4A CN103745197B (en) 2013-12-27 2013-12-27 A kind of detection method of license plate and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201310739338.4A CN103745197B (en) 2013-12-27 2013-12-27 A kind of detection method of license plate and device

Publications (2)

Publication Number Publication Date
CN103745197A CN103745197A (en) 2014-04-23
CN103745197B true CN103745197B (en) 2017-08-29

Family

ID=50502214

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201310739338.4A Active CN103745197B (en) 2013-12-27 2013-12-27 A kind of detection method of license plate and device

Country Status (1)

Country Link
CN (1) CN103745197B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104268509B (en) * 2014-09-17 2018-05-29 深圳市捷顺科技实业股份有限公司 The method and system of dump truck car plate detection
CN104182769B (en) * 2014-09-17 2018-04-27 深圳市捷顺科技实业股份有限公司 A kind of detection method of license plate and system
CN104318225B (en) * 2014-11-19 2019-02-22 深圳市捷顺科技实业股份有限公司 Detection method of license plate and device
CN104463220A (en) * 2014-12-19 2015-03-25 深圳市捷顺科技实业股份有限公司 License plate detection method and system
CN105184291B (en) * 2015-08-26 2019-07-30 深圳市捷顺科技实业股份有限公司 A kind of polymorphic type detection method of license plate and system
CN106651774B (en) * 2016-12-27 2020-12-04 深圳市捷顺科技实业股份有限公司 License plate super-resolution model reconstruction method and device
CN106846353B (en) * 2017-01-04 2020-02-21 美的集团股份有限公司 Density-based image processing method, image processing apparatus and device

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102129562A (en) * 2010-01-15 2011-07-20 富士通株式会社 Method and device for identifying icons

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7203363B2 (en) * 2003-06-27 2007-04-10 Chunghwa Telecom Co., Ltd. String extraction method for image based on multiple stroke width patterns matching
AU2003249223A1 (en) * 2003-07-10 2005-02-25 James Simon Autonomous wide-angle license plate recognition
CN101246551A (en) * 2008-03-07 2008-08-20 北京航空航天大学 Fast license plate locating method

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102129562A (en) * 2010-01-15 2011-07-20 富士通株式会社 Method and device for identifying icons

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
基于灰度方差和边缘密度的车牌定位算法;张浩鹏 等;《仪器仪表学报》;20110531;第32卷(第5期);第1095-1102页 *
基于边缘与颜色信息的车牌精确定位算法;曾丽华 等;《北京航空航天大学学报》;20070930;第33卷(第9期);第2.2.1节 *
用于车牌定位的分类器设计与相关算法研究;赵元兴;《中国优秀硕士学位论文全文数据库 信息科技辑》;20120715;第2012年卷(第07期);第31页第三章 *

Also Published As

Publication number Publication date
CN103745197A (en) 2014-04-23

Similar Documents

Publication Publication Date Title
CN103745197B (en) A kind of detection method of license plate and device
CN106874894B (en) Human body target detection method based on regional full convolution neural network
Liu et al. Fast multiclass vehicle detection on aerial images
CN103699905B (en) Method and device for positioning license plate
CN104318225B (en) Detection method of license plate and device
CN102803991B (en) Object detection device
CN104933414B (en) A kind of living body faces detection method based on WLD-TOP
CN106023257B (en) A kind of method for tracking target based on rotor wing unmanned aerial vehicle platform
CN104166841A (en) Rapid detection identification method for specified pedestrian or vehicle in video monitoring network
Chen et al. Human shadow removal with unknown light source
US9740965B2 (en) Information processing apparatus and control method thereof
CN104036284A (en) Adaboost algorithm based multi-scale pedestrian detection method
CN104978567B (en) Vehicle checking method based on scene classification
CN104517095B (en) A kind of number of people dividing method based on depth image
CN103617426A (en) Pedestrian target detection method under interference by natural environment and shelter
CN104866616A (en) Method for searching monitor video target
CN104463240B (en) A kind of instrument localization method and device
CN103310194A (en) Method for detecting head and shoulders of pedestrian in video based on overhead pixel gradient direction
CN104951793B (en) A kind of Human bodys' response method based on STDF features
CN102855758A (en) Detection method for vehicle in breach of traffic rules
CN103218604A (en) Method for detecting pedestrians in traffic scene based on road surface extraction
CN105404886B (en) Characteristic model generation method and characteristic model generating means
CN104463220A (en) License plate detection method and system
CN103530638A (en) Method for matching pedestrians under multiple cameras
CN102819733A (en) Rapid detection fuzzy method of face in street view image

Legal Events

Date Code Title Description
PB01 Publication
C06 Publication
SE01 Entry into force of request for substantive examination
C10 Entry into substantive examination
GR01 Patent grant
GR01 Patent grant