CN107122777A - A kind of vehicle analysis system and analysis method based on video file - Google Patents
A kind of vehicle analysis system and analysis method based on video file Download PDFInfo
- Publication number
- CN107122777A CN107122777A CN201710277862.2A CN201710277862A CN107122777A CN 107122777 A CN107122777 A CN 107122777A CN 201710277862 A CN201710277862 A CN 201710277862A CN 107122777 A CN107122777 A CN 107122777A
- Authority
- CN
- China
- Prior art keywords
- image
- character
- mrow
- license plate
- template
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/14—Image acquisition
- G06V30/148—Segmentation of character regions
- G06V30/158—Segmentation of character regions using character size, text spacings or pitch estimation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
- G06V20/625—License plates
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention belongs to technical field of data processing, a kind of vehicle analysis system and analysis method based on video file are disclosed, including：Read in elementary area, image pre-processing unit, License Plate and extraction unit, Character segmentation unit, character recognition unit；Analysis method includes：By technologies such as image gray processing, image enhaucament, edge extracting, binaryzations, license plate image is converted into the binary image for License Plate；License Plate and image information are extracted；By Character segmentation unit, positioning licence plate information character border intercepts all characters；Application template matching algorithm carries out character recognition, recognizes car plate internal digital.The present invention substantially increases the speed of vehicle analysis subsequent treatment, and by the localization method based on chrominance component, realizes the positioning for blue licence plate vehicle.
Description
Technical field
The invention belongs to technical field of data processing, more particularly to a kind of vehicle analysis system based on video file and point
Analysis method.
Background technology
It is the vital task in government's daily management mission safely in the process for building safe city.As city is reported
The construction of alert and monitoring system, the analysis for monitoring data also shows important day, it is necessary to be regarded to emphasis street or crossing collection
Frequency is according to being analyzed.Realization can be counted to the vehicle flow at street or crossing, on the one hand be solved from mass data
The difficulty of middle manual analysis, on the other hand by the statistics of motor vehicles, all has to urban road construction and traffic administration
Great importance.
The automatic identification for realizing license plate for vehicle is the basis of all intelligent traffic administration system activities.How to develop a set of row it
Effective automatic license plate identification system, particularly License Plate subsystem, it is significant.
With the development of graphics technology, the accuracy rate of Car license recognition is gradually improved both at home and abroad at present, still, and one
The complete Vehicle License Plate Recognition System of set is generally made up of triggering, IMAQ, picture recognition module, secondary light source and communication module,
And it is related to the multinomial technology such as optics, electrical equipment, Electronic Control, Digital Image Processing, computation vision, artificial intelligence, wherein, at present
Research for License Plate and image processing problem is still immature, however it remains a series of problems, is mainly reflected in existing
The region of vehicle analysis system car plate in whole image is obtained, and identify that the ability of license plate number is poor.
With extensive use of the intelligent transportation system in field of traffic, image processing techniques turns into intelligent transportation system
Important research field, with extremely important theoretical and application value.Vehicle is carried out using the various technologies of image procossing
Vehicle location license, Location of Vehicle License Plate it is accurate and clear, be to ensure the accurate premise of identification and basis.The present invention proposes a set of
Vehicle analysis system and analysis method based on video, have great meaning for the vehicle identification in road traffic with analysis
Justice, has decisive role to the intelligent management of urban traffic safety and vehicle, while also having expanded graph processing technique to car
The application value of board analysis.
The content of the invention
The problem of existing for prior art, the invention provides a kind of vehicle analysis system based on video file and point
Analysis method.
The present invention is achieved in that a kind of vehicle analysis system based on video file, described based on video file
Vehicle analysis system includes：
Elementary area is read in, is defaulted as inputting license plate image, and by the image of input clearly, without angle of inclination
Image；
Image pre-processing unit, for the image that reading elementary area is inputted, by image gray processing, image enhaucament,
After edge extracting, binaryzation, the binary image for License Plate is converted into；
License Plate and extraction unit, for image information to be positioned and extracted to car plate；
Character segmentation unit, location character border is carried out for the license board information to positioning, and then will be all in car plate
Character intercepts out；
Character recognition unit, the algorithm of application template matching is used for the identification of character, identifies the numeral inside car plate.
Another object of the present invention is to provide a kind of analysis method of the vehicle based on video file, by reading in image list
Member, inputs license plate image, and the image of input is defaulted as into image clearly, without angle of inclination；
By image pre-processing unit, the image that reading elementary area is inputted, by image gray processing, image enhaucament,
After edge extracting, binaryzation, the binary image for License Plate is converted into；
Image information is positioned and extracted to car plate by License Plate and extraction unit；
By Character segmentation unit, the license board information to positioning carries out location character border, and then will be all in car plate
Character intercepts out；
By character recognition unit, the algorithm matched using application template is entered the identification of line character, identified inside car plate
Numeral.
Further, the method for described image gray processing includes：Using weighted mean method, by image gray processing, formula is as follows：
F (i, j)=0.299R (i, j)+0.588G (i, j)+0.114B (i, j).
Further, the method for the edge extracting includes：
Utilize calculus of finite differences edge extracting.
Further, the method for described image binaryzation includes：
Gray level image is processed into bianry image by the binaryzation of image；Suitable threshold value is found to distinguish object and background；
Threshold calculations formula：T=Gmax- (Gmax-Gmin)/3；
Image binaryzation formula：
Further, the license plate locating method includes：
Judge licence plate using the feature of license plate area, license plate area is split from view picture vehicle image；
There is big saltus step in characters on license plate and licence plate background color, on gray value according to the trip point of horizontal direction in image
Changing rule carries out the extraction of the textural characteristics on car plate position, the up-and-down boundary of positioning licence plate；Orient up-and-down boundary it
Afterwards, calculate car plate approximate altitude, according to the wide high proportion of car plate, calculate the width of car plate, draw one it is customized solid
Determine the license plate area of size；In the range of the car plate bound found, the horizontal traversing graph picture of the area size drawn, system are utilized
Prospect points in region are counted, meets and requires to think to find license plate area, otherwise continue search for, until finding satisfactory area
Domain；
The row projection for seeking image and the formula for arranging projection are respectively：
In formula, i and j are respectively the line number and columns of image；fv-sum(i), fh-sum(i) be respectively image level projection and
The statistical information of upright projection.According to projection properties, in conjunction with frequency analysis and transition times, you can find the upper following of car plate
Boundary.
Further, Character segmentation algorithm includes：
In car plate there are certain intervals in intercharacter, then continuous crest and trough occur in the histogram of car plate；At trough
The as boundary of character；License plate image after arrangement statistics binaryzation, calculates the points containing foreground point at least in all row；
Scan image is rearranged, adjacent two row counted containing minimum pixel are once searched, the spacing of two row is calculated, if less than 2,
It is then separator, determines the border of character, and each character is intercepted from license plate image according to border comes out；
According to gray level threshold segmentation method, using statistics with histogram method, the left and right boundary point of each character is found, judgement is
No is character, and right boundary is retained if being, is otherwise searched backward.
Further, Character segmentation algorithm includes character normalization method, is specially：
If the character picture size cut out is M × N, template size is L × K, then it is that L × K is big to normalize character picture
It is small.Main thought is that character picture is uniformly divided into L × K blocks region, the number of character foreground point in each piece is counted respectively, such as
Fruit is more than 50% of pixel count in block, then it is assumed that this block pixel position foreground point corresponding after L × K sizes are converted into；
If the character picture cut out is it is impossible to meet uniformly L × K blocks region is divided into, (M × L) × (N is first defined
× K) matrix, character picture is scaled up as (M × L) × (N × K) sizes, the character picture so no matter cut out
To be much, L × K blocks region is cut into, the normalized of character is realized；
Character segmentation algorithm also includes character thinning methods：
Using OPTA thinning algorithms, the extension that will be normalized above and below ground character picture, right boundary adds 1, the edge of extension
Background is defaulted as, travels through each foreground point pixel from top to bottom, from left to right in character picture, 3 × 3 disappear with what is given
Go template and 4 × 4 reservation template to be compared, meet cancellation template and the point is then set to background dot, meet and retain template
It is then constant, still it is foreground point.
Further, character identifying method includes：
A) feature extraction；Characteristic vector, which is constituted, using five features carries out template matches；Five characteristic values are respectively：
A number of hits of the row with template and with identification character at template and the width of character picture to be identified 1/5；
A number of hits of the row with template and with identification character at template and the width of character picture to be identified 1/2；
A number of hits of the row with template and with identification character at template and the width of character picture to be identified 4/5；
Template and character picture to be identified 1/4 highly locate number of hits of a line with template and with identification character；Mould
Plate and character picture to be identified 3/4 highly locate number of hits of a line with template and with identification character；
B) template matches；
The characteristic vector of the characteristic vector of character picture to be identified and 10 templates is once compared, end value is minimum
For the result that identifies, return to the numeral representated by corresponding template, it is as follows that characteristic vector compares formula：
Further, in the image of input, the denoising of license plate image need to be carried out, is specifically included：
Using LPF and bandpass filtering come denoising.
Advantages of the present invention and good effect are：
The present invention has carried out system in terms of image preprocessing, License Plate, Character segmentation and character recognition respectively
Analysis.By arranging and summarizing analysis results and developing direction both at home and abroad in terms of License Plate, segmentation, character recognition,
And the inherent feature of China's car plate, and the characteristics of Car license recognition, using the localization method based on Gray Level Jump, image is entered
Row pretreatment and binarization operation, realize correlation function.By dependence test, the main technique effect of the present invention is as follows：
1. this method had both remained the information of license plate area, the interference of noise is reduced again, so as to simplify at binaryzation
Reason process, improves the speed of subsequent treatment；
2. the localization method based on chrominance component, is blueness to car plate with the method based on blue picture element point statistical property
Car plate positioned, experiment shows, the License Plate accuracy rate realized by this method is higher.
Brief description of the drawings
Fig. 1 is the vehicle analysis system schematic provided in an embodiment of the present invention based on video file.
In figure：1st, elementary area is read in；2nd, image pre-processing unit；3rd, License Plate and extraction unit；4th, Character segmentation
Unit；5th, character recognition unit.
Fig. 2 is the vehicle analysis method flow diagram provided in an embodiment of the present invention based on video file.
Fig. 3 is that license plate area provided in an embodiment of the present invention is accurately positioned flow.
Fig. 4 is Character segmentation flow chart provided in an embodiment of the present invention.
Fig. 5 is character identifying method flow chart provided in an embodiment of the present invention.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
The application principle of the present invention is further described below in conjunction with the accompanying drawings.
As shown in figure 1, the vehicle analysis system provided in an embodiment of the present invention based on video file, including：
Elementary area 1 is read in, is defaulted as inputting license plate image, and by the image of input clearly, without inclination angle
The image of degree；
Image pre-processing unit 2, for the image that reading elementary area is inputted, by image gray processing, image enhaucament,
After edge extracting, binaryzation, the binary image for License Plate is converted into；
License Plate and extraction unit 3, for image information to be positioned and extracted to car plate；
Character segmentation unit 4, location character border is carried out for the license board information to positioning, and then will be all in car plate
Character intercepts out；
Character recognition unit 5, the algorithm of application template matching is used for the identification of character, identifies the numeral inside car plate.
As shown in Fig. 2 the analysis method provided in an embodiment of the present invention that the vehicle based on video file is provided, including：
S101:By reading in elementary area, license plate image is inputted, and the image of input is defaulted as clearly, not incline
The image of rake angle；
S102:Pass through image pre-processing unit, the image that reading elementary area is inputted, by image gray processing, image
After enhancing, edge extracting, binaryzation, the binary image for License Plate is converted into；
S103:Image information is positioned and extracted to car plate by License Plate and extraction unit；
S104:By Character segmentation unit, the license board information to positioning carries out location character border, and then by car plate
All characters intercept out；
S105:By character recognition unit, the algorithm matched using application template is entered the identification of line character, identifies car plate
Internal numeral.
The invention will be further described for knot specific embodiment below.
Vehicle analysis method provided in an embodiment of the present invention based on video file, including：
1) image pre-processing method：
Image gray processing：
Using weighted mean method, by image gray processing, formula is as follows：
F (i, j)=0.299R (i, j)+0.588G (i, j)+0.114B (i, j)
Edge extracting：
Using the calculus of finite differences edge extracting in above-mentioned., can be with using calculus of finite differences edge extracting for the image request of system
Good edge extracting effect is reached, and is realized simple, it is readily appreciated that.
Image binaryzation：
In systems because edge extracting is calculus of finite differences, the gray value of whole image, which has, significantly to be declined,
So resetting the 1/3 of the threshold value that binary-state threshold is calculated by formula by experiment statisticses.
Foreground point (i.e. character) is white after image binaryzation, and background dot is black.For convenience of showing and observing, by its turn
Foreground point is changed to for black, background dot is white.
2) Plate searching and localization method：
The starting point of license plate locating method is to judge licence plate using the feature of license plate area, by license plate area from view picture car
Split in image.Car plate itself has many inherent features, and these features are different for different countries.
From the visual angle of people, China's car plate has the feature that can be used for positioning below:
Character wide 45mm, high 90mm, blank character 10mm inside car plate wide 440mm, high 140mm, the ratio of width to height 3.14., each
Character pitch 12mm；
Car plate has a continuous or discontinuous frame due to abrasion；
Car plate background color typically has larger difference with body color, character color, and character and licence plate background color are deposited on gray value
In larger saltus step, there is more trip point to occur；
Numerical character is longitudinally connection.
By the above-mentioned feature of car plate it is recognised that characters on license plate and licence plate background color have larger saltus step on gray value,
Therefore the extraction of the textural characteristics on car plate position can be carried out according to the changing rule of the trip point of horizontal direction in image,
The up-and-down boundary of positioning licence plate.Orient after up-and-down boundary, the approximate altitude of car plate can be calculated, according to the ratio of width to height of car plate
Example, approximate calculation goes out the width of car plate, now can then draw the license plate area of a customized fixed size.What is found
In the range of car plate bound, using the horizontal traversing graph picture of the area size drawn, prospect is counted in statistical regions, meets requirement i.e.
Think to find license plate area, otherwise continue search for, until finding satisfactory region.
3) Character segmentation algorithm
In car plate the characteristics of character, there are certain intervals in each intercharacter, then go out in the histogram of car plate after certainty
Now continuous crest and trough.It is the boundary of character at trough.Therefore, by the license plate image after row statistics binaryzation, meter
Calculate the points containing foreground point at least in all row.Case scan image, once searches the adjacent minimum pixel that contains and counts again
Two row (two borders of character), calculate two row spacing, be then separator, it may be determined that the border of character if less than 2,
And each character is intercepted from license plate image according to border come out.
According to gray level threshold segmentation method, using statistics with histogram method, the left and right boundary point of each character is found, judgement is
No is character, and right boundary is retained if being, is otherwise searched backward.
4) character normalization method：
Main is as follows：
If the character picture size cut out is M × N, template size is L × K, then it is that L × K is big to normalize character picture
It is small.Main thought is that character picture is uniformly divided into L × K blocks region, the number of character foreground point in each piece is counted respectively, such as
Fruit is more than 50% of pixel count in block, then it is assumed that this block pixel position foreground point corresponding after L × K sizes are converted into.
But it is impossible to meet can uniformly be divided into L × K blocks region, therefore the way in system for the character picture being to cut out
It is first to define (M × L) × (a N × K) matrix, character picture is scaled up as (M × L) × (N × K) sizes, this
The character picture no matter sample cuts out is much, L × K blocks region can be cut into, so as to realize the normalized of character.
5) character thinning methods：
Using OPTA thinning algorithms, the extension that will be normalized above and below ground character picture, right boundary adds 1, the edge of extension
Background is defaulted as, travels through each foreground point pixel from top to bottom, from left to right in character picture, 3 × 3 disappear with what is given
Go template and 4 × 4 reservation template to be compared, meet cancellation template and the point is then set to background dot, meet and retain template
It is then constant, still it is foreground point.
6) character identifying method：
Enter the identification of line character in system using the algorithm of template matches.The recognition template used in system is download
20 × 40 gray level image (black matrix wrongly written or mispronounced character) of standard, includes 0 to 90 numerals.Template is also done into micronization processes, to carry out
The extraction of template numerical characteristic.
A) feature extraction
System carries out template matches using five feature composition characteristic vectors.Five characteristic values are respectively：
(1) number of hits of the row with template and with identification character at template and the width of character picture to be identified 1/5；
(2) number of hits of the row with template and with identification character at template and the width of character picture to be identified 1/2；
(3) number of hits of the row with template and with identification character at template and the width of character picture to be identified 4/5；
(4) template and character picture to be identified 1/4 highly locate number of hits of a line with template and with identification character；
(5) template and character picture to be identified 3/4 highly locate number of hits of a line with template and with identification character；
B) template matches
The characteristic vector of the characteristic vector of character picture to be identified and 10 templates is once compared, wherein end value
Minimum is the result identified, returns to the numeral representated by corresponding template, it is as follows that characteristic vector compares formula：
With reference to specific embodiment, the invention will be further described.
1st, the storage and display of digital picture provided in an embodiment of the present invention：
1) IMAQ, into computer, is to be stored in the form of image file, when needing to use, by program to this
A little files are carried out and information processing.
The file structure of picture：
Header file, for depositing image essential information, including various features parameter, pixel data and toning version data in text
Position in part, and textual annotation etc.；
Pixel data, is deposited in the form of a bitmap, the color value on every pixel data correspondence image relevant position, specific face
The definition of colour is provided by palette data；
Palette data, refers to two-value, 16 colors, the tone data of 256 color color images, the pixel data of true color image
In existing color component, therefore no longer need palette information.
It should be noted that the scanning direction (storage order) of image, the scanning direction that most of image files are taken is certainly
Under above (such as TIF and jpeg file), minority is from bottom to top (such as BMP files).
2) image is shown：
Image display process is carried out according to picture storing process opposite direction.
2nd, the gray processing method of image：
Coloured image is converted into the gray processing of gray level image.The each pixel of coloured image includes R, G, B tri- colors point
Amount, each color component has 256 intermediate values can use.And gray level image is the equal special face of tri- color component values of R, G, B
Color image.
In RGB models, if during R=G=B, colour represents a kind of greyscale color, and wherein R=G=B value is ash
Angle value, therefore, each pixel of gray level image only need a byte storage gray value (also known as intensity level, brightness value), tonal range
For 0 to 255.It is general to there is following four method to carry out gray processing to coloured image：
Component method：
Using the three-component brightness in coloured image as the gray value of three gray level images, it can be needed to choose according to application
A kind of gray level image.
f1(i, j)=R (i, j) f2(i, j)=G (i, j) f3(i, j)=B (i, j)
Wherein fk(i, j) (k=1,2,3) is gray value of the gray level image after conversion at (i, j) place.
Maximum value process：
Using the maximum of the three-component brightness in coloured image as gray-scale map gray value.
F (i, j)=maxR (i, j) G (i, j) B (i, j)
Mean value method：
Three-component brightness in coloured image is averaged and obtains a gray-scale map.
F (i, j)=(R (i, j)+G (i, j)+B (i, j))/3
Weighted mean method：
According to importance and other indexs, three components are weighted with different weights average.Because human eye is to green
The sensitive highest of color, it is minimum to blue-sensitive, therefore, average energy is weighted to RGB three-components as the following formula and obtains more rational
Gray level image.
F (i, j)=0.299R (i, j)+0.588G (i, j)+0.114B (i, j).
3rd, image binaryzation method：
The binaryzation of image is the process for gray level image being processed into bianry image.The key of binaryzation is to find properly
Threshold value distinguish object and background.Binary Sketch of Grey Scale Image can be substantially reduced the capacity of data storage, and can reduce follow-up place
The complexity of reason.According to the difference of Research on threshold selection, binarization method is broadly divided into three classes：Global thresholding, local threshold
Method and dynamic thresholding method.
A) global threshold binarization method：
Global threshold binarization method is to be distributed to determine a threshold value according to the histogram or gray space of image, and according to
This threshold value realizes gray level image to the conversion of binary image.Typical Global thresholding includes Ostu methods, maximum entropy method
Deng.The advantage of global threshold method is that algorithm is simple, target and background is clearly separated, histogram is bimodal image effect
Well, situations such as but to input picture quantizing noise or uneven illumination, resistance was poor, using being restricted.
B) local threshold binarization method：
The method for the local gray level characteristic threshold value put around the gray value f (i, j) and pixel of pixel (i, j) is referred to as
Local threshold back-and-forth method.Although the intensity profile of situations such as inhomogeneous illumination condition influence general image does not influence local figure but
As property.The regulation of neighborhood and the selection of neighborhood calculation template are all the key factors for determining algorithm effect.Local thresholding method
There are problems that shortcoming and, such as realize that speed is slow, it is impossible to ensure that character stroke is connective, and easily artifact phenomenon etc. occur.
C) dynamic threshold binarization method：
When threshold value selection depends not only on the pixel threshold and the gray value of its surrounding pixel, but also sat with the pixel
When cursor position is relevant, referred to as dynamical threshold selection method.Due to having taken into full account the feature of each neighborhood of pixels, can preferably it dash forward
Go out the border of background and target, adhesion phenomenon will not be produced by making the two lines of close proximity, low quality can be handled even unimodal
Value histogram image.But, the amount of calculation of this method is very big, and arithmetic speed is typically slow.
Threshold calculations formula：T=Gmax- (Gmax-Gmin)/3；
Image binaryzation formula：
4th, the denoising method of license plate image：
In the picture, because the equipment for obtaining image can produce noise, can also there is noise in image transmitting process, and
Car plate also has noise in itself, and this allows for our images of acquisition and is all mingled with more or less noise.Whether need to scheme
As carrying out denoising, this depends on the influence degree of noise on image.In general we are only carried out at denoising to image
The segmentation and identification of car plate could be preferably realized after reason.The noise in image is typically divided into two kinds：One kind is additive noise,
Another is multiplicative noise.Additive noise and original image are the relations added, can be represented with expression formula (1).And multiplicative noise and
Original digital image data is the relation being multiplied, and available expression (2) is represented.
G (x, y)=f (x, y)+η (x, y) (1)
G (x, y)=f (x, y) * η (x, y) (2).
The noise of generation can show different characteristics because of the difference of producing cause.Common noise type has：Uniform point
The noise types such as cloth noise, Gaussian noise, impulsive noise, exponential distribution noise, rayleigh noise, gamma noise.In denoising, I
Need to take corresponding method denoising according to the characteristics of noise.Two kinds can be generally divided into, one kind is gone in time domain
Make an uproar, another is to carry out denoising in frequency domain.Both approaches are to carry out denoising according to different thought, and one is to utilize signal
Carry out denoising with the temporal signatures of noise, another is the frequency domain character using signal and noise come denoising.Either any side
Method, we are provided to signal and noise separation so as to obtain a view data for containing a small amount of noise.Get in time domain
The method made an uproar has many kinds, such as：The filtering methods such as mean filter, Order Filtering, adaptive-filtering.And on frequency domain denoising master
If using the different so as to be isolated of frequency band shared by noise and signal, according to frequency band different where noise, we can be with
Using LPF and bandpass filtering come denoising.
5th, the Enhancement Method of image：
Useful information in image enhaucament, i.e. enhancing image, the process of usually one distortion, its purpose is to strengthen to regard
Feel effect, original unsharp image is apparent from or emphasized some features interested, suppresses uninterested feature, makes
Improvement picture quality, abundant information amount, strengthen image interpretation and recognition effect.
The purpose of image enhaucament is to improve the visual effect of image, for the application scenario of given image, purposefully by force
Change the line map the entirety or local characteristicses of picture, the difference in expanded view picture between different objects feature meets the need of some special analysis
Will.Its method is that some information or conversion data are added to original image by certain means, selectively protrudes in image and feels emerging
The feature of interest suppresses some unwanted features in image, image is matched with eye response characteristic.
Image enhaucament is divided into frequency domain method and space domain method by method therefor.(only low frequency is allowed to believe using LPF
Number pass through) method, the noise in figure can be removed；It using high-pass filtering method, then can strengthen the high-frequency signals such as edge, make fuzzy picture
It is apparent from.Representative spatial-domain algorithm has local averaging method and medium filtering (to take the centre in local neighborhood
Pixel value) method etc., they can be used for removing or weakening noise.
With reference to the analysis to vehicle license feature and the comparison to various localization methods, the system is used based on row
Scan the license plate locating method of Gray Level Jump analysis.This method is combined based on analysis of texture and based on edge detection analysis
The characteristics of method, have the advantages that speed is fast and accuracy is high.
6th, edge detection method：
The edge of image is the most basic feature of image.So-called edge refers to that surrounding pixel gray scale has Spline smoothing or room
Push up the set of those pixels of change.The edge of object is reflected by gray scale discontinuity.Edge is widely present in object
It is that image segmentation, texture feature extraction and shape facility are carried between background, between object and object, between primitive and primitive
The important foundation of graphical analysis such as take.The first step of graphical analysis and understanding is often rim detection.
The extracting method at general edge is the change of each pixel gray scale in some field of image under consideration, utilizes edge
Neighbouring single order or Second order directional changing rule, choose appropriate operator, and edge is detected with the method for convolution.Edge extracting
, largest benefit be exactly can projecting edge region, and make unrelated part in background image dim.Car plate part is edge phase
To the region of concentration, after splitting to its edge, it can be seen that car plate part clearly protrude.
Image Edge-Detection significantly reduces data volume, and eliminates it is considered that incoherent information, retains
Image important structure attribute.Conventional edge detection method has Sobel operators, Roberts operators, Prewitt operators etc..
One) Roberts operators：
Roberts operators are one of most ancient operators, are a kind of cross-differential operators.Because it only uses current pixel
2 × 2 neighborhoods, be simplest gradient operator, thus calculate it is very simple.The pixel one that Roberts operators are utilized when calculating
4 are had, can be added and obtained with 4 pixels of template correspondence element multiplication corresponding with template.
The positioning of Roberts operators edge is accurate, and major defect is its high susceptibility to noise, and reason is only used
Seldom several pixels carry out approximate gradient.It is practically applicable to the image segmentation that edge is obvious and noise is less.Therefore, generally with 3 × 3
Neighborhood calculates Grad.
Two) Sobel operators：
Using gradient differential sharpening image, while can be strengthened noise, striped etc., Sobel operators are then in certain journey
This problem is overcome on degree.Sobel operators are with the differences of two pixels unlike common gradient operator, and which results in two
Individual advantage：
Due to introducing equilibrating factor, thus there is certain smoothing effect to the random noise in image；
Because it is the difference that is separated by two rows or two row, therefore both sides of edges element is strengthened, and edge seems thick and bright.
Three) Prewitt operators：
Prewitt operators will use 9 pixels when being calculated.For the gradient in each direction, template correspondence can be used
Corresponding with the template element multiplication of 9 pixels be added and obtain, its calculating process is similar with Roberts operators.
Prewitt operators have inhibitory action to noise, and the principle for suppressing noise is average by pixel.But pixel is average
LPF is carried out equivalent to image, so positioning of the Prewitt operators to edge is not so good as Roberts operators.
Four) Canny operators：
Canny operators are the edge detection algorithms of the best results based on image first derivative characteristic.Canny algorithms are in fact
It is the thought that the connection of some edges is introduced during rim detection, so its effect is more satisfactory.
The general principle of Canny operators is to obtain edge by searching the method for the local maximum of image gradient intensity
Information, the calculating of image gradient uses Gauss wave filters.For the precision for eliminating noise jamming, improving rim detection, Canny
Operator extracts marginal point using dual threshold.High threshold is used to extract the stronger edge pixel point of contrast.If gradient intensity
The testing conditions of high threshold are unsatisfactory for, but are connected with the relatively strong edge point detected, and meet Low threshold condition, then
It is defined as weak marginal point.Dual-threshold voltage makes the marginal point of Canny operator extractions have more robustness.
It is to choose appropriate Gauss filter parameters and dual threshold using the key at Canny operator detection images edge.
The standard deviation sigma parameter of Gauss wave filters affects the yardstick of edge extracting, and σ is smaller, and the positioning precision at edge is higher, but not
False edge and noise can effectively be suppressed；σ is bigger, and the suppression to small yardstick edge and noise spot is stronger, while reducing positioning
Precision.
Five) Laplacian operators：
In order to protrude enhancing image in isolated point, isolated line or isolated end points, in some practical uses frequently with
Laplacian operators, this operator is invariable rotary operator.Four kinds of operators of the above are single order operator, and Laplacian operators are
Second-order Operator.Laplacian operators are the popular approach of the approximate second dervative for only providing gradient magnitude.Usually using 3 × 3
Mask, can be divided into 4 neighborhoods and 8 neighborhoods according to neighborhood difference.
Laplacian operators are more sensitive for edge.General enhancing technology is for precipitous edge and slow edge
It is difficult to determine the position of its edge line, this operator can be determined with the zero crossing between second differential posivtive spike and negative peak.
Laplacian operators are more sensitive to noise, so image typically first passes through smoothing processing.Therefore Laplacian
Operator generally is used in combination to split image with smoothing operator.One of Laplacian operators has the disadvantage it to image
Some of edge produce double-response.
7th, feature extracting method：
1) texture characteristic extracting method：
Texture concept conventional in being graphical analysis and recognizing, but the definition formal to it is there is no at present, generally may be used
To think that texture is made up of many elements close to each other, mutually working out, and Chang Fuyou is periodically.It is also assumed that being ash
Spend the pattern produced in space with certain form change.
Main method is statistics of histogram method.During histogram is image window, the pixel distribution of a variety of different gray scales
Probability statistics.Texture is visual performance of the pixel grayscale change with spatial regularity.It therefore, it can the rule according to gray scale
Rule property changes to find the textural characteristics in image.
2) shape and architectural feature extracting method：
For an image identification system, the shape of object is the key character of an identification of relying.One image
Shape and architectural feature have two kinds of forms, one kind is numerical characteristic, mainly include geometric attribute and topological attribute；It is another to be
Syntax-language.
Due to it is interested be the shape and architectural feature of image, so its half-tone information can often be ignored, as long as energy
Target and background is separated, a kind of conventional technology is exactly the binaryzation of image.Part interested is marked with maximum ash
Level is spent, background is marked with minimal gray level.
With reference to localization method, the invention will be further described.
1st, position：
Under natural environment, automobile image background is complicated, and uneven illumination is even, and license plate area is accurately determined in natural background
It is the key in whole image identification process.A wide range of relevant search is carried out to the image collected first, finds and meets automobile
Then some regions of licence plate feature are further analyzed to these candidate regions, judge as candidate regions.Finally select one
Best region splits it as license plate area from image.To consider license plate sloped problem simultaneously.Algorithm flow is such as
Under：
(1) extracted region is carried out to bianry image, calculated and comparison domain characteristic parameter, extract license plate area.
(2) calculate the minimum comprising institute's marked region wide and high, and according to prior knowledge, extract and show closer car
The subvalue of board two schemes.
(3) license plate sloped problem is solved by calculating the car plate anglec of rotation.Cause drop shadow effect peak not due to license plate sloped
Substantially, car plate correction process is needed.Take linear fit method.Calculate car plate top or following image value straight for l point fitting
The angle of line and trunnion axis.With the rotation car plate picture function Imrotate of MATLAB functions.Calculate the car plate anglec of rotation and warp
Car plate two-value subgraph after rotation, binaryzation.
The present invention have found based on image difference sciagraphy after a large amount of practices, so that the Car license recognition time be shortened to
Within 100 milliseconds.Its principle is that by horizontal direction vehicle gray image is asked into difference diagram, difference is then sought in a vertical direction, most
Differentiated vehicle image is both horizontally and vertically being projected respectively afterwards, being found out according to given car plate size range can be like car
Board region.
License plate area is accurately positioned using the synthesis license plate locating method based on car plate texture and colouring information, it is realized
Flow is illustrated in fig. 3 shown below.
Assuming that input color image is I, I is converted into by gray scale by Gray=0.229 × R+0.587 × G+0.114 × B
Scheme Gray, ask the method for threshold value to obtain optimal threshold T using iteration, carrying out binary conversion treatment to Gray according to T obtains image B；
Rim detection is carried out with Sobel vertical operators, edge image S is obtained；To S, using smoothing method to its de-noising, caustic solution is used
Some isolated points and thin protrusion are removed, with the curved hole of slot in expanding method fused images and duck eye, image F are obtained；In F
Connected region formation car plate candidate region Rect [i] is searched for, pseudo- car plate area is removed according to colouring information for each Rect [i]
Domain, then passes through area row saltus step Rect [i] jumpnum, upright projection Rect [i] to the license plate area after correction
And floor projection Rect [i] .vproject is accurately positioned to license plate area .hproject.
1.1 car plate rough localization methods：
Car plate coarse positioning is screened to license plate area.Because image background is often more complicated, other areas of car body
Domain has disturbing factor, therefore the license plate candidate area that License Plate is extracted is multiple, it is necessary to further to this toward accommodating
A little candidate regions are screened, and are rejected interference region, are oriented real car plate.
Assuming that a height of h of i-th of candidate region, a width of W, then candidate region Gao Yukuan ratio R and area A is respectively
Ri=ωi/hi, Ai=ωihi；
Assume again that i-th of candidate region be yellow, blue, white three-color pixel number respectively Si,y, Si,bAnd Si,w；Transition times
For Si,j；Jth row white points are n after binaryzationij,w。
1) car plate generally has certain length and width, and length and width have certain ratio, domestic single file car plate length and width
Than for 440/140=22/7.Using this feature of car plate, if 300 pixels≤Ai≤ 1200 pixels, or 3≤Ri≤ 4, just
It is license plate area to think i-th of candidate region, otherwise removes it.Usual candidate region can be bigger than actual license plate area, such as
Fruit is strictly judged that license plate area may result in wrong screening by the length-width ratio of standard license plate area.For this in coarse positioning according to
Actual conditions relax the limitation of candidate region, to prevent the generation of " mistake screening ".
2) license plate area also has an obvious feature to be exactly that its background color has fixed collocation, the standard of such as China with character
Car plate has 4 kinds of regular collocations：Blue bottom wrongly written or mispronounced character, yellow bottom surplus, white gravoply, with black engraved characters and white background The Scarlet Letter.If Si,y≥Ai/ 4, or Si,b
≥Ai/ 4, or Si,w≥Ai/ 4, and Si,j>=14, this region is regarded as license plate area, is otherwise exactly pseudo- license plate area.
3) it is the speed of further raising License Plate, it is contemplated that white point number often capable compares after license plate area binaryzation
It is many, when system is realized, set a threshold value T.If nij,w≤ T, then candidate region J rows just filter out.It can so reduce
Amount of calculation during License Plate, so as to reduce the quantity of candidate license plate connected region.
Real car plate can be filtered out in the candidate region of car plate by 3 conditions above.
The Slant Rectify method of 1.2 images：
The influence of the factor such as situation of camera lens and the angle of licence plate, the motion of vehicle and road surface, such as car during due to shooting
Position of the board in image is caught is not fixed, and headstock or camera lens occur to swing when catching image and car plate inherently hangs askew
Or road conditions are poor, all the license plate image photographed may be made to have certain gradient, in order to which correct recognize needs to carry out gradient
Correction, otherwise will be unable to carry out the correct segmentation of single character, the error rate of character recognition will rise.If but solid with some
Fixed empirical value is unified to all licence plates to carry out rotation processing, and normal licence plate inclination originally can be made again, causes new mistake.Cause
This is necessary to extract its angle of inclination for specific License Plate Image, then is subject to corresponding rotation processing.How from licence plate from
Dynamic its gradient of extracting is a difficult point in preprocessing process, general to be come pair using the straight line in Hough transform detection image
The inclination of image is corrected accordingly.
Specific algorithm is determined according to the average height of the black picture element of the right and left on image.The figure of character composition
Near should being horizontal as the character pixel height of its right and left, if the average bit of both sides character pixels
It is equipped with and is risen and fallen than larger, that just illustrates that image is present and tilted, it is necessary to be adjusted.First have to calculate the left one side of something of image and right half
The average height of the pixel on side, then seeks slope, and image, that is, reflecting from new images to old image are reorganized according to slope
Penetrate.If the pixel-map in new images puts the pixel in new images to scope during old image beyond old image
Into white.
The minimizing technology of 1.3 car plate frames and Liu Ding：
It is extremely complex in face of the licence plate character background to be recognized in actual treatment, there is larger interference, noise.When
The bianry image of car plate is unified for black matrix wrongly written or mispronounced character and after sharpening to image gradient and removing noise, also has bumper
With a region of two rivet interference on the incomplete image and licence plate of licence plate four edges frame.By priori, it can adopt
Image processing method is taken to remove interference from above-mentioned complex background.License plate image is carried out after binaryzation, image is only black, white two
Value.White pixel point (gray value 25 5) takes l, and black pixel point (gray value 0) takes 0, white gravoply, with black engraved characters mould is used here
Formula.The removal of upper and lower side frame and rivet is critically important in license plate image, does not remove the license plate image of frame line and rivet, warp
Often there is the phenomenon of rivet and character and frame line adhesion, the segmentation to follow-up characters on license plate causes very big difficulty.In car plate
, generally there are four rivets the inner side of frame line, they to some extent with the 2nd character or the 6th Characters Stuck, if do not gone
Except rivet, difficulty will be caused in the identification of character to the 2nd and the 6th.
For standard car plate, intercharacter spacing is 12 ㎜, the 2nd, 3 intercharacter spacing be 34 ㎜, wherein, middle dot
L0 ㎜ are wide, and dot and the 2nd, 3 intercharacter spacing are respectively 12 ㎜.According to these prioris, license plate image is entered line by line
Formula is scanned row from inside to outside, when scanning is into license plate image a line, (first when the width of white pixel point is more than a certain threshold values
Individual qualified row), then it is assumed that it is the edge of characters on license plate, cuts off all rows of this line above and below.To remove
Exemplified by upper side frame and the rivet interference of car plate, its algorithm idea is：If dimension of picture to be processed is m × n, wherein m is row, n
For row, take empirical value 9n/l0 as the threshold value for direction projection of being expert at, up progressively scanned since i=2m/3 rows, if the row
Projection j in the row direction>9n/l0, is considered as have found the coboundary of character at i+l rows, and all rows since i are complete
Portion is cut off.If not finding qualified row, then it is assumed that without upper side frame interference.
1.4 car plate accurate positioning methods：
After the success of coarse positioning car plate, there is some other background letter in addition to license plate image in obtained license plate area
Breath, it is necessary to which these background informations are filtered out, obtains the image of only characters on license plate and could carry out characters on license plate more accurately dividing
Cut and recognize.After Slant Rectify, car plate border is searched for by sciagraphy.
1) car plate up-and-down boundary is searched for
Frequency method and sciagraphy are used in combination with by the extraction of car plate up-and-down boundary.So-called sciagraphy is exactly to analyze image
Direction in length and breadth projection value, to find the border of car plate.The row projection for seeking image and the formula for arranging projection are respectively：
In formula, i and j are respectively the line number and columns of image；fv-sum(i), fh-sum(i) be respectively image level projection and
The statistical information of upright projection.According to projection properties, in conjunction with frequency analysis and transition times, you can find the upper following of car plate
Boundary.
2) car plate right boundary method is searched for：
License plate area shows different statistical natures in upright projection：The upright projection of true license plate area has significantly
The phenomenon at the peak of one paddy of peak one；Because characters on license plate is 7, so crest number is generally higher than 7 (assuming that character grey level is higher than car
Board background)；Value between crest and crest, trough and trough is more or less the same.According to projection properties, in conjunction with frequency analysis, i.e.,
The right boundary of car plate can be found.
With reference to license plate area processing, the invention will be further described.
License plate area processing：
Because the license plate image of the system is in outdoor shooting, therefore inevitably by available light, season
Etc. the influence of factor.Remove these interference first must pre-process to license plate image.The data of original image typically compare
Greatly, the time handled it is general also longer, and due to the requirement of real-time, the extraction of car plate needs disposable processing just
Most feature extractions can be come out.Obtained image is needed to be pre-processed.The process of pretreatment needs image to turn
Bianry image is turned to, can so greatly reduce data volume, a certain degree of ensure is provided for real-time.Car plate after binaryzation will
Former character picture can be reproduced, stroke fracture and adhesion phenomenon are occurred without substantially, the feature of former character is not lost as far as possible.Through experiment,
Characters on license plate is the clearest when using threshold value near 0.2, and miscellaneous point is minimum.
Image preprocessing mainly includes：Gray processing, binaryzation, Grads Sharp, noise reduction, segmentation, normalization.
With reference to character segmentation method, the invention will be further described.
As shown in figure 4, Character segmentation：
Character is split using Gray Projection method.The thought of this algorithm is to be based on image projection histogram i.e. binary map
With there is Wave crest and wave trough fluctuating change in the often statistical relationship of the appearance number of the white pixel value of row or each column in the row or column as in
Change, wherein trough may be exactly the interval of adjacent character.According to priori, the vehicle license of standard has 7 characters, the first
To save name referred to as (Chinese character), secondary position is English alphabet, and last position is Chinese character (extension) or English alphabet or numeral, and other positions are English
Literary letter or number.It is possible that certain trough after in view of being projected between Chinese radical and radical, and Chinese character is with making an uproar
Also there can be trough between sound, therefore directly be split using sciagraphy.One threshold value is set, by trough between projection apart from small
Merged in the projection of this threshold value, so that the projection of Chinese radical and radical is merged.After merging, calculate its first
The width of individual character and last character, if calculated value is more than 1.5 times of characters on license plate mean breadth, is taken as mistake
Merge, then revocation is merged and merged with debug, it is to avoid merge the projection of car plate frame in character.For other 5 words
Symbol, due to being numeral and letter, the characteristics of its structure has connective, so being entered using domain method is connected in mathematical morphology to it
Row segmentation.
With reference to character identifying method, the invention will be further described.
Character identifying method：The character being partitioned into is known using the strong BP neural network of adaptivity and learning ability
Not, its identification process is illustrated in fig. 5 shown below.
With reference to the design of neural network classifier and using the invention will be further described.
The distribution of characters on license plate is regular.The regularity being distributed according to character on car plate, designs " Chinese character in systems
Network ", 3 identification networks such as " numeral, alphabetical network " and " Chinese character, letter, digital network ".
Using when, the 1st separating character is sent to " Chinese character network ", the 7th word is sent to " Chinese character, letter, digital network
Network ", other characters are sent to " numeral, alphabetical network ".
The design of 1.1BP neutral nets：
1) selection of initial weight.Initial weight directly affects the convergent speed of neutral net, if selecting bad, nerve net
Network will cause to be difficult to restrain or be absorbed in local minimum in saturation.In systems, initial weight uniformly divides between choosing 0~1
The random value of cloth.
2) determination of input layer number.Separating character is normalized to the character picture of the dot matrix size of 2O × 16, thus it is defeated
Enter a layer neuron number and take 320.
3) determination of output layer nodes.Output layer nodes are relevant with character number to be identified, need to know in system
Other 50 Chinese characters, 1O numeral and 25 letters (not including alphabetical I, because characters on license plate is without alphabetical I).Character is used in system
8421 yards come to identification character encode, so output layer nodes take 7.
4) determination of node in hidden layer.Node in hidden layer is few, and learning process does not restrain；Node in hidden layer is more
, the study of network and Reasoning Efficiency are deteriorated.Hidden node number selection empirical equation beWherein, N
For hidden neuron number, n is input layer number, and m is output layer neuron number, and a is the integer between 1~10.
Take, n=320, m=7.Further accordance with conventional experience, and substantial amounts of experimental result is analyzed, compared, know hidden layer node
Recognition effect is relatively good when number is 30, therefore takes N=30.
5) feature normalization.The character picture of input is normalized to the character picture of 20 × 16 dot matrix sizes, system will
Character picture binaryzation after normalization, is then uniformly converted into black matrix wrongly written or mispronounced character character as defeated people's feature of neutral net.
The selection of 1.2 training samples：
In character recognition, the training sample selected by each class character must be able to truly reflect the common of such character
Feature.License Plate Character Segmentation result occurs that separating character is nonstandard and seen as if in order to pursue character recognition when undesirable
By character lack of standardization as training sample, network is as a result often set although to have learned the character lack of standardization, but to real calibration
The discrimination of the character of rule but have dropped.Therefore, the image after Character segmentation is shown in systems, if segmentation effect
Neutral net is reused during relatively good but incorrect identification to be trained, and can so reach system compatible other common car plates point
Cut the purpose of character recognition.Certainly, it is contemplated that the other factors of system identification, as vehicle is split in static or motion
Character effect is variant, and the function of preserving and read heterogeneous networks connection weight is provided in systems, to recognize different mesh
Training file is switched over when marking the character being partitioned into come to the accurate identification of separating character.
The identification process of 1.3BP networks：
For a network trained, the identification process weights that exactly a utilization is trained in fact complete once saequential transmission
Process is broadcast, then a recognition result is obtained in output layer.Output layer has 7 nodes, the output result of each node in system
Coding all with a character to be identified is corresponding.Character code uses 8421 yards, if output and the character code phase of node
Together, the character for being considered as input is the corresponding character of the node.Due to the output for the activation primitive Sigmoid functions that system is used
0 or 1 is never equally likely to, while in order to improve system recognition rate, definition：When output valve≤0.1,0 is output as；And work as
1 is output as during output valve >=O.9.
With reference to CR recognition methods, the invention will be further described.
OCR is recognized：
Image is inputted：
For different picture formats, there are different storage formats, different compress mode, for different forms
Picture is parsed, and is obtained picture pixels content, is then proceeded to following operation.
Binaryzation：
The picture shot to camera, most of is coloured image, and coloured image information contained amount is huge, for picture
Content, we can simply be divided into prospect and background, in order to allow computer faster, and preferably identification word is, it is necessary to first right
Cromogram is handled, and makes picture foreground information and background information, can simply define foreground information for black, background letter
Cease for white, here it is binary picture.
Noise remove is searched for：
For different documents, the definition to noise can be different, according to the feature of dry sound go dry, are just called noise
Remove.
Tilt calibration：
Due to general user, when taking pictures document, all relatively arbitrarily, therefore the picture for taking pictures out is inevitably produced
Tilt, this is accomplished by software for discerning characters and carries out calibration.
Printed page analysis：
Document picture is paragraphed, the process of branch is just called printed page analysis, it is complicated due to the diversity of actual document
Property, therefore, Slicing Model for Foreign fixed there is presently no one, optimal.
Character segmentation：
Due to the limitation of photographical condition, Characters Stuck is often resulted in, break pen, therefore strongly limit the property of identifying system
Can, this, which is accomplished by software for discerning characters, Character segmentation function.
Character recognition：
Based on feature extraction, due to the displacement of word, the thickness of stroke, break pen, adhesion, the influence of the factor such as rotation,
The difficulty of the extraction of extreme influence feature.
The space of a whole page recovers：
It is desirable to the word after recognizing, remain as original text shelves picture and arrange like that, paragraph is constant, and position is constant, suitable
Sequence is unchangeably output to word document, pdf documents etc..
Post processing, check and correction：
According to the relation of specific Linguistic context, calibration is carried out to recognition result.
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 (10)
1. a kind of vehicle analysis system based on video file, it is characterised in that the vehicle analysis system based on video file
System includes：
Elementary area is read in, figure clearly, without angle of inclination is defaulted as inputting license plate image, and by the image of input
Picture；
Image pre-processing unit, for the image that reading elementary area is inputted, by image gray processing, image enhaucament, edge
Extract, after binaryzation, be converted into the binary image for License Plate；
License Plate and extraction unit, for image information to be positioned and extracted to car plate；
Character segmentation unit, for carrying out location character border to the license board information of positioning, and then by all characters in car plate
Interception comes out；
Character recognition unit, the algorithm of application template matching is used for the identification of character, identifies the numeral inside car plate.
2. a kind of analysis method of the vehicle analysis system based on video file as claimed in claim 1, it is characterised in that described
The analysis method of vehicle based on video file, by reading in elementary area, inputs license plate image, and the image of input is given tacit consent to
For image clearly, without angle of inclination；
Pass through image pre-processing unit, the image that reading elementary area is inputted, by image gray processing, image enhaucament, edge
Extract, after binaryzation, be converted into the binary image for License Plate；
Image information is positioned and extracted to car plate by License Plate and extraction unit；
By Character segmentation unit, location character border is carried out to the license board information of positioning, and then by all characters in car plate
Interception comes out；
By character recognition unit, the algorithm matched using application template enters the identification of line character, identifies the number inside car plate
Word.
3. the vehicle analysis method as claimed in claim 2 based on video file, it is characterised in that
The method of described image gray processing includes：Using weighted mean method, by image gray processing, formula is as follows：
F (i, j)=0.299R (i, j)+0.588G (i, j)+0.114B (i, j).
4. the vehicle analysis method as claimed in claim 2 based on video file, it is characterised in that
The method of the edge extracting includes：
Utilize calculus of finite differences edge extracting.
5. the vehicle analysis method as claimed in claim 2 based on video file, it is characterised in that described image binaryzation
Method includes：
Gray level image is processed into bianry image by the binaryzation of image；Suitable threshold value is found to distinguish object and background；
Threshold calculations formula：T=Gmax- (Gmax-Gmin)/3；
Image binaryzation formula：
<mrow>
<mo>{</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mtable>
<mtr>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mo>&GreaterEqual;</mo>
<mi>T</mi>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mo><</mo>
<mi>T</mi>
</mrow>
</mtd>
</mtr>
</mtable>
<mo>.</mo>
</mrow>
6. the vehicle analysis method as claimed in claim 2 based on video file, it is characterised in that the license plate locating method
Including：
Judge licence plate using the feature of license plate area, license plate area is split from view picture vehicle image；
There is big saltus step in characters on license plate and licence plate background color, on gray value according to the change of the trip point of horizontal direction in image
Rule carries out the extraction of the textural characteristics on car plate position, the up-and-down boundary of positioning licence plate；Orient after up-and-down boundary, count
The approximate altitude of car plate is calculated, according to the wide high proportion of car plate, the width of car plate is calculated, draws a customized fixed size
License plate area；In the range of the car plate bound found, the horizontal traversing graph picture of the area size drawn, statistical regions are utilized
Interior prospect points, meet and require to think to find license plate area, otherwise continue search for, until finding satisfactory region；
The row projection for seeking image and the formula for arranging projection are respectively：
<mrow>
<msub>
<mi>f</mi>
<mrow>
<mi>v</mi>
<mo>-</mo>
<mi>s</mi>
<mi>u</mi>
<mi>m</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
</munder>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
<mrow>
<msub>
<mi>f</mi>
<mrow>
<mi>h</mi>
<mo>-</mo>
<mi>s</mi>
<mi>u</mi>
<mi>m</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
</munder>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
In formula, i and j are respectively the line number and columns of image；fv-sum(i), fh-sum(i) it is respectively image level projection and vertical throwing
The statistical information of shadow.According to projection properties, in conjunction with frequency analysis and transition times, you can find the up-and-down boundary of car plate.
7. the vehicle analysis method as claimed in claim 2 based on video file, it is characterised in that Character segmentation algorithm bag
Include：
In car plate there are certain intervals in intercharacter, then continuous crest and trough occur in the histogram of car plate；It is at trough
The boundary of character；License plate image after arrangement statistics binaryzation, calculates the points containing foreground point at least in all row；Arrange again
Column scan image, once searches adjacent two row counted containing minimum pixel, calculates the spacing of two row, if less than 2, is then
Separator, determines the border of character, and each character is intercepted from license plate image according to border comes out；
According to gray level threshold segmentation method, using statistics with histogram method, the left and right boundary point of each character is found, is determined whether
Character, retains right boundary, otherwise searches backward if being.
8. the vehicle analysis method as claimed in claim 7 based on video file, it is characterised in that Character segmentation algorithm includes
Character normalization method, be specially：
If the character picture size cut out is M × N, template size is L × K, then it is L × K sizes to normalize character picture；Will
Character picture is uniformly divided into L × K blocks region, and the number of character foreground point in each piece is counted respectively, if greater than pixel count in block
50%, then it is assumed that this block corresponding pixel position foreground point after L × K sizes are converted into；
If the character picture cut out is it is impossible to meet uniformly L × K blocks region is divided into, (M × L) × (a N × K) is first defined
Matrix, character picture is scaled up as (M × L) × (N × K) sizes, the character picture so no matter cut out is many
Greatly, L × K blocks region is cut into, the normalized of character is realized；
Character segmentation algorithm also includes character thinning methods：
Using OPTA thinning algorithms, the extension that will be normalized above and below ground character picture, right boundary adds 1, the edge acquiescence of extension
For background, each foreground point pixel is traveled through from top to bottom, from left to right in character picture, with 3 × 3 cancellation mould given
Plate and 4 × 4 reservation template are compared, and meet cancellation template and the point then is set into background dot, meet retain template then not
Become, be still foreground point.
9. the vehicle analysis method as claimed in claim 2 based on video file, it is characterised in that character identifying method bag
Include：
A) feature extraction；Characteristic vector, which is constituted, using five features carries out template matches；Five characteristic values are respectively：
A number of hits of the row with template and with identification character at template and the width of character picture to be identified 1/5；
A number of hits of the row with template and with identification character at template and the width of character picture to be identified 1/2；
A number of hits of the row with template and with identification character at template and the width of character picture to be identified 4/5；
Template and character picture to be identified 1/4 highly locate number of hits of a line with template and with identification character；
Template and character picture to be identified 3/4 highly locate number of hits of a line with template and with identification character；
B) template matches；
The characteristic vector of the characteristic vector of character picture to be identified and 10 templates is once compared, end value it is minimum be
The result identified, returns to the numeral representated by corresponding template, it is as follows that characteristic vector compares formula：
<mrow>
<mi>D</mi>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mn>4</mn>
</munderover>
<msup>
<mrow>
<mo>(</mo>
<mi>s</mi>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
<mo>-</mo>
<mi>t</mi>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
<mo>.</mo>
</mrow>
10. the vehicle analysis method as claimed in claim 2 based on video file, it is characterised in that in the image of input, is needed
The denoising of license plate image is carried out, is specifically included：
Using LPF and bandpass filtering come denoising.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710277862.2A CN107122777A (en) | 2017-04-25 | 2017-04-25 | A kind of vehicle analysis system and analysis method based on video file |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710277862.2A CN107122777A (en) | 2017-04-25 | 2017-04-25 | A kind of vehicle analysis system and analysis method based on video file |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107122777A true CN107122777A (en) | 2017-09-01 |
Family
ID=59725740
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710277862.2A Pending CN107122777A (en) | 2017-04-25 | 2017-04-25 | A kind of vehicle analysis system and analysis method based on video file |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107122777A (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108682018A (en) * | 2018-05-26 | 2018-10-19 | 任阿毛 | The instant detection platform of equipment degree of lacking |
| CN109614945A (en) * | 2018-12-18 | 2019-04-12 | 杭州匹知共创科技有限公司 | A kind of vehicle feature recognition method filtering environmental disturbances |
| CN110706171A (en) * | 2019-09-26 | 2020-01-17 | 中国电子科技集团公司第十一研究所 | Image noise reduction method and device |
| CN110728687A (en) * | 2019-10-15 | 2020-01-24 | 卓尔智联(武汉)研究院有限公司 | File image segmentation method and device, computer equipment and storage medium |
| CN111008634A (en) * | 2019-11-22 | 2020-04-14 | 北京精英路通科技有限公司 | Character recognition method and character recognition device based on example segmentation |
| CN111095265A (en) * | 2017-09-07 | 2020-05-01 | 奥迪股份公司 | Method for evaluating an optical representation in a vehicle environment and vehicle |
| CN111178219A (en) * | 2019-12-24 | 2020-05-19 | 泰康保险集团股份有限公司 | Bill identification management method and device, storage medium and electronic equipment |
| CN111652230A (en) * | 2020-05-25 | 2020-09-11 | 浙江大华技术股份有限公司 | License plate recognition method, electronic device and storage medium |
| CN111914747A (en) * | 2020-07-31 | 2020-11-10 | 长江三峡通航管理局 | Ship lock miter gate gap video monitoring system and method |
| CN112232237A (en) * | 2020-10-20 | 2021-01-15 | 城云科技（中国）有限公司 | Vehicle flow monitoring method, system, computer device and storage medium |
| CN112307842A (en) * | 2019-07-31 | 2021-02-02 | 株洲中车时代电气股份有限公司 | Video identification system and method for matching of train operation monitoring record files |
| CN112801232A (en) * | 2021-04-09 | 2021-05-14 | 苏州艾隆科技股份有限公司 | Scanning identification method and system applied to prescription entry |
| CN113255657A (en) * | 2020-12-31 | 2021-08-13 | 深圳怡化电脑股份有限公司 | Method and device for detecting scratch on surface of bill, electronic equipment and machine-readable medium |
| CN113658302A (en) * | 2021-08-23 | 2021-11-16 | 李帮音 | Three-dimensional animation data processing method and device |
| CN114627653A (en) * | 2022-05-12 | 2022-06-14 | 浙江电马云车科技有限公司 | 5G intelligent barrier gate management system based on binocular recognition |
| CN114934467A (en) * | 2022-07-08 | 2022-08-23 | 江苏永达电力金具有限公司 | Parking space barrier gate control method, parking space barrier gate system and medium |
| WO2023061173A1 (en) * | 2021-10-13 | 2023-04-20 | 北京字节跳动网络技术有限公司 | Image processing method and apparatus |
| CN117011507A (en) * | 2023-09-04 | 2023-11-07 | 南京诺源医疗器械有限公司 | 2CMOS image recognition data analysis method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090185236A1 (en) * | 2008-01-17 | 2009-07-23 | Zeng Yijun | Image binarization using dynamic sub-image division |
| CN101833859A (en) * | 2010-05-14 | 2010-09-15 | 山东大学 | Self-triggering license plate identification method based on virtual coil |
| CN102750525A (en) * | 2012-06-20 | 2012-10-24 | 太仓博天网络科技有限公司 | License plate recognition system under real-time traffic |
-
2017
- 2017-04-25 CN CN201710277862.2A patent/CN107122777A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090185236A1 (en) * | 2008-01-17 | 2009-07-23 | Zeng Yijun | Image binarization using dynamic sub-image division |
| CN101833859A (en) * | 2010-05-14 | 2010-09-15 | 山东大学 | Self-triggering license plate identification method based on virtual coil |
| CN102750525A (en) * | 2012-06-20 | 2012-10-24 | 太仓博天网络科技有限公司 | License plate recognition system under real-time traffic |
Non-Patent Citations (4)
| Title |
|---|
| ZW11111111: "毕业设计（论文）-基于数字识别的车牌号码识别系统", 《道客巴巴：HTTP://WWW.DOC88.COM/P-9738231810535.HTML》 * |
| 吴李汉 等: "车牌自动识别系统的设计与实现", 《可编程控制器与工厂自动化》 * |
| 李坤 等: "车牌自动识别系统的研究与实现", 《青岛大学学报（工程技术版）》 * |
| 王刚 等: "基于MATLAB的车牌识别系统的研究", 《电子设计工程》 * |
Cited By (30)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111095265A (en) * | 2017-09-07 | 2020-05-01 | 奥迪股份公司 | Method for evaluating an optical representation in a vehicle environment and vehicle |
| CN111095265B (en) * | 2017-09-07 | 2023-09-05 | 奥迪股份公司 | Method for evaluating an optical representation in a vehicle environment and vehicle |
| CN108682018B (en) * | 2018-05-26 | 2019-05-07 | 上海尚雍电子技术工程有限公司 | The instant detection platform of equipment degree of lacking |
| CN108682018A (en) * | 2018-05-26 | 2018-10-19 | 任阿毛 | The instant detection platform of equipment degree of lacking |
| CN109614945A (en) * | 2018-12-18 | 2019-04-12 | 杭州匹知共创科技有限公司 | A kind of vehicle feature recognition method filtering environmental disturbances |
| CN112307842A (en) * | 2019-07-31 | 2021-02-02 | 株洲中车时代电气股份有限公司 | Video identification system and method for matching of train operation monitoring record files |
| CN110706171A (en) * | 2019-09-26 | 2020-01-17 | 中国电子科技集团公司第十一研究所 | Image noise reduction method and device |
| CN110706171B (en) * | 2019-09-26 | 2024-04-26 | 中国电子科技集团公司第十一研究所 | Image noise reduction method and device |
| CN110728687B (en) * | 2019-10-15 | 2022-08-02 | 卓尔智联(武汉)研究院有限公司 | File image segmentation method and device, computer equipment and storage medium |
| CN110728687A (en) * | 2019-10-15 | 2020-01-24 | 卓尔智联(武汉)研究院有限公司 | File image segmentation method and device, computer equipment and storage medium |
| CN111008634A (en) * | 2019-11-22 | 2020-04-14 | 北京精英路通科技有限公司 | Character recognition method and character recognition device based on example segmentation |
| CN111008634B (en) * | 2019-11-22 | 2023-08-22 | 北京精英路通科技有限公司 | Character recognition method and character recognition device based on instance segmentation |
| CN111178219A (en) * | 2019-12-24 | 2020-05-19 | 泰康保险集团股份有限公司 | Bill identification management method and device, storage medium and electronic equipment |
| CN111652230B (en) * | 2020-05-25 | 2023-05-12 | 浙江大华技术股份有限公司 | License plate recognition method, electronic device and storage medium |
| CN111652230A (en) * | 2020-05-25 | 2020-09-11 | 浙江大华技术股份有限公司 | License plate recognition method, electronic device and storage medium |
| CN111914747A (en) * | 2020-07-31 | 2020-11-10 | 长江三峡通航管理局 | Ship lock miter gate gap video monitoring system and method |
| CN112232237B (en) * | 2020-10-20 | 2024-03-12 | 城云科技（中国）有限公司 | Method, system, computer device and storage medium for monitoring vehicle flow |
| CN112232237A (en) * | 2020-10-20 | 2021-01-15 | 城云科技（中国）有限公司 | Vehicle flow monitoring method, system, computer device and storage medium |
| CN113255657B (en) * | 2020-12-31 | 2024-04-05 | 深圳怡化电脑股份有限公司 | Method and device for detecting scratch on bill surface, electronic equipment and machine-readable medium |
| CN113255657A (en) * | 2020-12-31 | 2021-08-13 | 深圳怡化电脑股份有限公司 | Method and device for detecting scratch on surface of bill, electronic equipment and machine-readable medium |
| CN112801232A (en) * | 2021-04-09 | 2021-05-14 | 苏州艾隆科技股份有限公司 | Scanning identification method and system applied to prescription entry |
| CN113658302B (en) * | 2021-08-23 | 2024-04-12 | 麦应俊 | Three-dimensional animation data processing method and device |
| CN113658302A (en) * | 2021-08-23 | 2021-11-16 | 李帮音 | Three-dimensional animation data processing method and device |
| WO2023061173A1 (en) * | 2021-10-13 | 2023-04-20 | 北京字节跳动网络技术有限公司 | Image processing method and apparatus |
| CN114627653B (en) * | 2022-05-12 | 2022-08-02 | 浙江电马云车科技有限公司 | 5G intelligent barrier gate management system based on binocular recognition |
| CN114627653A (en) * | 2022-05-12 | 2022-06-14 | 浙江电马云车科技有限公司 | 5G intelligent barrier gate management system based on binocular recognition |
| CN114934467A (en) * | 2022-07-08 | 2022-08-23 | 江苏永达电力金具有限公司 | Parking space barrier gate control method, parking space barrier gate system and medium |
| CN114934467B (en) * | 2022-07-08 | 2024-04-30 | 江苏永达电力金具有限公司 | Parking space barrier control method, parking space barrier system and medium |
| CN117011507A (en) * | 2023-09-04 | 2023-11-07 | 南京诺源医疗器械有限公司 | 2CMOS image recognition data analysis method and system |
| CN117011507B (en) * | 2023-09-04 | 2024-01-26 | 南京诺源医疗器械有限公司 | 2CMOS image recognition data analysis method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107122777A (en) | A kind of vehicle analysis system and analysis method based on video file | |
| CN102915438B (en) | The extracting method of a kind of video caption and device | |
| CN103116751B (en) | A kind of Method of Automatic Recognition for Character of Lcecse Plate | |
| CN109255350B (en) | New energy license plate detection method based on video monitoring | |
| CN106650553A (en) | License plate recognition method and system | |
| CN102043945B (en) | License plate character recognition method based on real-time vehicle tracking and binary index classification | |
| CN107273896A (en) | A kind of car plate detection recognition methods based on image recognition | |
| CN104156731B (en) | Vehicle License Plate Recognition System and method based on artificial neural network | |
| CN104751142A (en) | Natural scene text detection algorithm based on stroke features | |
| CN107103317A (en) | Fuzzy license plate image recognition algorithm based on image co-registration and blind deconvolution | |
| CN106845408A (en) | A kind of street refuse recognition methods under complex environment | |
| CN108491797A (en) | A kind of vehicle image precise search method based on big data | |
| CN107066933A (en) | A kind of road sign recognition methods and system | |
| CN103903018A (en) | Method and system for positioning license plate in complex scene | |
| CN101398894A (en) | Automobile license plate automatic recognition method and implementing device thereof | |
| CN106056751A (en) | Prefix number identification method and system | |
| CN105335716A (en) | Improved UDN joint-feature extraction-based pedestrian detection method | |
| Li et al. | Research on vehicle license plate location based on neural networks | |
| CN110298376A (en) | A kind of bank money image classification method based on improvement B-CNN | |
| CN104809433A (en) | Zebra stripe detection method based on maximum stable region and random sampling | |
| CN110008900B (en) | Method for extracting candidate target from visible light remote sensing image from region to target | |
| Azad et al. | New method for optimization of license plate recognition system with use of edge detection and connected component | |
| Shaikh et al. | A novel approach for automatic number plate recognition | |
| CN111832461A (en) | Non-motor vehicle riding personnel helmet wearing detection method based on video stream | |
| CN104463138A (en) | Text positioning method and system based on visual structure attribute |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170901 |