CN109165643A - A kind of licence plate recognition method based on deep learning - Google Patents

A kind of licence plate recognition method based on deep learning Download PDF

Info

Publication number
CN109165643A
CN109165643A CN201810955796.4A CN201810955796A CN109165643A CN 109165643 A CN109165643 A CN 109165643A CN 201810955796 A CN201810955796 A CN 201810955796A CN 109165643 A CN109165643 A CN 109165643A
Authority
CN
China
Prior art keywords
license plate
indicate
characters
candidate
character
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.)
Granted
Application number
CN201810955796.4A
Other languages
Chinese (zh)
Other versions
CN109165643B (en
Inventor
高飞
蔡益超
葛粟
葛一粟
卢书芳
程振波
陆佳炜
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.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
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 Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN201810955796.4A priority Critical patent/CN109165643B/en
Publication of CN109165643A publication Critical patent/CN109165643A/en
Application granted granted Critical
Publication of CN109165643B publication Critical patent/CN109165643B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/22—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
    • G06V10/23—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition based on positionally close patterns or neighbourhood relationships
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/24—Aligning, centring, orientation detection or correction of the image
    • G06V10/245—Aligning, centring, orientation detection or correction of the image by locating a pattern; Special marks for positioning
    • 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
    • G06V30/153—Segmentation of character regions using recognition of characters or words

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Character Input (AREA)
  • Character Discrimination (AREA)

Abstract

The invention discloses a kind of licence plate recognition methods based on deep learning, include the following steps: step 1: the training one depth convolutional neural networks model M for characters on license plate detection；Construct characters on license plate tag set B；Step 2: the license plate image I that positioning obtains being input to characters on license plate detection network M, exports candidate license plate character set H；Step 3: the candidate license plate character set H that step 2 is obtained, by the upper left corner abscissa h of candidate license plate character minimum circumscribed rectanglei.x ascending that H reorders, obtain set C；Step 4: screening operation further being executed to the set C that step 3 obtains, step 6: the set E obtained to step 5 sequentially traverses set E；Step 7: returning and license plate recognition result L is obtained by step 6.The influence for situations such as the beneficial effects of the present invention are: effectively inhibiting characters on license plate adhesion, fracture, deformation, license plate is stained, and shade is remained on license plate sloped and license plate.

Description

A kind of licence plate recognition method based on deep learning
Technical field
The present invention relates to field of intelligent transportation technology, and in particular to a kind of licence plate recognition method based on deep learning.
Background technique
Over the last couple of decades, license plate recognition technology is all greatly improved on accuracy of identification and efficiency of algorithm. With being constantly progressive for wisdom traffic system the relevant technologies, automatic license plate image recognition is considered as one and possesses mature solution party The settled problem of case.Traffic flow analysis, vehicle speed measuring, vehicle violation detection are many answering based on license plate recognition technology Representative.However, the standard specification of license plate is very more in practice, license plate font, color have notable difference, and characters on license plate is long It spends different.Only just there are blue bottom wrongly written or mispronounced character uniline license plate, yellow bottom black word uniline license plate, yellow bottom black word duplicate rows license plate, black matrix wrongly written or mispronounced character in China The multinomial license plate type of uniline license plate etc. ten.In addition, licence plate recognition method is easy by illumination, resolution ratio, imaged viewing angle, shade etc. The interference of environmental factor.
Computer mould apery identifies license plate, which includes positioning and identification.Therefore, complete automatic Car license recognition system System is made of License Plate and Car license recognition two parts.License Plate is not in the range of this method discussion.So far, it has deposited In many licence plate recognition methods.Method can be classified as to conventional method and deep learning method two major classes.Traditional Car license recognition Process includes two step of License Plate Character Segmentation and Recognition of License Plate Characters.Process can be divided into Character segmentation and word by deep learning method Symbol two steps of identification, end-to-end can also settle at one go.
Traditional licence plate recognition method is Task-decomposing at two subtasks of License Plate Character Segmentation and Recognition of License Plate Characters.License plate Character segmentation is the premise of Car license recognition, Character segmentation it is accurate whether directly influence recognition result.With image procossing skill The continuous development of art, a variety of registration number character dividing methods are suggested, such as: character sciagraphy, and Character mother plate matching method and character connect Logical domain method.Common license plate character recognition method is mainly based upon the OCR of machine learning, such as: the multiword based on support vector machines It accords with classifying identification method and the multiword based on artificial neural network accords with classifying identification method.OCR method based on machine learning is first The feature of character picture is first extracted, then training obtains Classification and Identification model, is eventually used for the identification mission of single character.But Situations such as being that, due to characters on license plate adhesion, fracture, deformation, license plate is stained, shade is remained on license plate sloped and license plate, is very normal See, traditional licence plate recognition method is limited by the selection and calculating of validity feature, cannot reach ideal recognition effect.
Deep learning method effectively inhibits characters on license plate adhesion, fracture, deformation, license plate is dirty by calculating convolution feature Damage, the license plate sloped influence with situations such as remaining shade on license plate.Existing one stream licence plate recognition method is all deep learning side Method.Deep learning method is mainly divided to method and end-to-end two class of method based on segmentation.
Deep learning method based on segmentation still continues to use the process of traditional licence plate recognition method, by Car license recognition task point For two subtasks of Character segmentation and character recognition.By deep learning learn convolution to convolution feature come handle two sons appoint Business.Document (Y.Yang, D.Li and Z.Duan, " Chinese vehicle license plate recognition using kernel-based extreme learning machine with deep convolutional features,"in IET Intelligent Transport Systems,vol.12,no.3,pp.213-219,42018.) The learning machine that transfinites (kernel-based extreme learning machine, KELM) based on core is applied to vehicle for the first time In board character recognition, Softmax is replaced with it to obtain classification results.Author is extracting the KELM classification 31 of depth convolution feature A chinese character (31 Chinese provinces are referred to as), although training speed is than very fast, low in the recognition efficiency of model, resource is accounted for With height.Document (license plate recognition technology research [D] Qingdao of the Zhao Zhenxing based on deep learning: Qingdao University of Science and Technology, 2017) base Complete Car license recognition task in two stages in SSD: the first stage is characters on license plate detection, i.e. segmentation simultaneously and identification license plate word Symbol；Second stage is that characters on license plate process of aggregation is obtained by the characters on license plate set permutation and combination for obtaining identification that reorders Final license plate recognition result.Although this method can handle uniline license plate and duplicate rows license plate simultaneously, license plate resolution ratio not Gao Shi, this method first stage, there are erroneous detection and missing inspection, the permutation and combination of second stage can not solve the problems, such as the first stage.Text Offer (Hefei research [D] of Recognition Algorithm of License Plate under soup jade Yao complex background: China Science & Technology University, 2016.) it is being based on Connected domain method is divided after obtaining characters on license plate, after obtaining disaggregated model with deep learning training, one by one to the character divided Identification, obtains final recognition result.This method has a defect that traditional Character segmentation based on connected domain is ineffective, into And influence subsequent identification.
End-to-end licence plate recognition method no longer needs individual License Plate Character Segmentation, inputs a license plate image, output License plate recognition result.Document (Hong-Hyun Kim, Je-Kang Park, Joo-Hee Oh, Dong-Joong Kang.Multi-task convolutional neural network system for license plate recognition[J].International Journal ofControl,Automation and Systems,2017,15 (6): 2942-2949 the end-to-end Car license recognition of single phase) is realized based on depth convolutional neural networks, but the network mechanism It is fixed, it can only identify South Korea's license plate of fixed format, computational efficiency is not high.Document (Jakub Spanhel, Jakub Sochor, Roman Juranek,Adam Herout,Lukas Marsik,Pavel Zemcik.Holistic recognition of low quality license plates by CNN using track annotated data.14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), Lecce, Italy, Aug 29-Sep 01,2017.) propose the neural network comprising 8 full connection branches, energy Enough identification contains up to the low resolution license plate of 8 characters, and each branch is responsible for character machining and the identification in a region.But It is that the network only considered uniline license plate, does not account for the license plate of other standards.
Deep learning method based on segmentation has that error accumulation, the accuracy of Character segmentation affect subsequent Character recognition, and the diversity of license plate standard increases the difficulty of Character segmentation.Although end-to-end licence plate recognition method exists It is dominant simultaneously in efficiency and effect, but this kind of network is high to the required precision of License Plate, and depth network design difficulty limits The license plate standard that can be applicable in is made.
In conclusion there are the following shortcomings for licence plate recognition method at present: 1) license plate standard is varied, and Car license recognition is difficult To take into account the license plate of all standards；2) there is mistake in the method based on segmentation (including conventional method and based on the method for deep learning) The segmentation error that poor accumulating questions, i.e. segmentation stage generate can accumulate the subsequent Recognition of License Plate Characters stage, reduce whole knowledge Not rate；3) although end-to-end Car license recognition network can have been able to the multiple types of processing variable character length in design License plate, but the license plate with similar standard can only be still handled, the network suitable for uniline license plate cannot take on duplicate rows license plate Ideal effect is obtained, and such network needs premised on accurate License Plate, condition is harsh；4) a variety of vehicles can be handled simultaneously The licence plate recognition method of board standard requires license plate imaging definition high, it is difficult to successfully handle low-resolution image；5) environment Factor (such as: illumination, noise) has a certain impact to the identification of license plate image.
Summary of the invention
For the disadvantages mentioned above for overcoming the prior art, the present invention proposes a kind of licence plate recognition method based on deep learning, After deep learning detects characters on license plate, first sorts to set element, then set element is screened, finally reconfigure To license plate recognition result character string.
Technical scheme is as follows:
A kind of licence plate recognition method based on deep learning, which comprises the steps of:
Step 1: the training one depth convolutional neural networks model M for characters on license plate detection；Construct characters on license plate mark Sign set B={ bi| i=1,2 ..., n, n=67 }=' 0 ', ' 1 ', ' 2 ', ' 3 ', ' 4 ', ' 5 ', ' 6 ', ' 7 ', ' 8 ', ' 9 ', ‘A’,‘B’,‘C’,‘D’,‘E’,‘F’,‘G’,‘H’,‘I’,‘J’,‘K’,‘L’,‘M’,‘N’,‘O’,‘P’,‘Q’,‘R’,‘S’, ' T ', ' U ', ' V ', ' W ', ' X ', ' Y ', ' Z ', ' capital ', ' saliva ', ' Ji ', ' Shanxi ', ' illiteracy ', ' the Liao Dynasty ', ' Ji ', ' black ', ' Shanghai ', ' Soviet Union ', ' Zhejiang ', ' Anhui ', ' Fujian ', ' Jiangxi ', ' Shandong ', ' Henan ', ' Hubei Province ', ' Hunan ', ' Guangdong ', ' osmanthus ', ' fine jade ', ' Chongqing ', ' river ', ' expensive ', ' cloud ', ' hiding ', ' Shan ', ' sweet ', ' blueness ', ' peaceful ', ' new ' }；
Step 2: the license plate image I that positioning obtains being input to characters on license plate detection network M, exports candidate license plate character set Close H={ hi| i=1,2,3 ..., nH, wherein nHIndicate the element number of set H, hiIndicate i-th of candidate license plate of set H Character, hiIt is the triple being made of (b, q, r), b indicates that the class label of candidate characters, b ∈ B, q indicate setting for candidate characters Reliability, q ∈ [0,1], r are the four-tuples being made of (x, y, w, h), and r indicates that the boundary rectangle frame of candidate characters, x, y, w and h divide Not Biao Shi rectangle frame upper left corner abscissa, upper left corner ordinate, width and height；
Step 3: the candidate license plate character set H that step 2 is obtained, by the upper left of candidate license plate character minimum circumscribed rectangle Angle abscissa hi.x ascending that H reorders, obtain set C={ ci| i=1,2,3 ..., nC, ciIndicate the i-th of set C A candidate license plate character, ciIt is the triple being made of (b, q, r), nCIndicate the element number of set C；
Step 4: screening operation, specific steps are further executed to the set C that step 3 obtains are as follows:
Step 4.1: C is sequentially traversed, it is rightI=1,2 ..., nCIf ci.q<Tq, then by ciSet D is added, and will ciIt is deleted from set C；Otherwise by ciStay in set C；Finally, obtaining D={ di| i=1,2 ..., nD, diIt is by (b, q, r) structure At triple, nDIndicate the quantity of element in set D；Wherein, ci.q candidate characters c is indicatediConfidence level, TqIndicate confidence Spend threshold value；
Step 4.2: the set C that sequentially traversal step 4.1 obtains is rightI=1,2 ..., nC- 1, by formula (1) A is calculated to delete element a from set C if a ≠ φ；Otherwise, operation is not executed；
Wherein, ci.r candidate characters c is indicatediBoundary rectangle frame, ci+1.r candidate characters c is indicatedi+1Boundary rectangle frame, ToverIndicate the anti-eclipse threshold of rectangle frame, | | indicate the area of rectangle frame, ∩ operation indicates to seek the intersection of two rectangle frames；
Step 4.3: the set D that the set C and step 4.1 obtain to step 4.2 is obtained is rightI=1,2 ..., nC- 1, ifMeet formula (2), records so that djMeet the i value of formula (2), and by djSet C is inserted into as the (i+1) a element；Wherein, djIndicate j-th of element of set D, j=1,2 ..., nD；
Wherein, ci.r.x、ci.r.y、ciAnd c .r.wi.r.h candidate characters c is respectively indicatediThe upper left corner of boundary rectangle frame is horizontal Coordinate, upper left corner ordinate, width and height；dj.r.x、dj.r.y、djAnd d .r.wj.r.h candidate characters d is respectively indicatedjIt is external Upper left corner abscissa, upper left corner ordinate, width and the height of rectangle frame；
Step 5: the set C that step 4.3 is obtained, firstly, set C is sequentially traversed, ifI=1,2 ..., nC, Meet formula (3), by ciSet E is added, and by ciIt is deleted from C；Then, the surplus element for traversing set C, is added sequentially to Set E；Finally obtain E={ ei| i=1,2,3 ..., nE, eiIndicate i-th of candidate license plate character of set E, nEIndicate set E Element number；
Step 6: the set E obtained to step 5 sequentially traverses set E, the label of all elements in set E is connected into License plate recognition resultWherein, ei.b ∈ B, ei.b the label of i-th of candidate license plate character of set E, Σ table are indicated Show character serial operation；
Step 7: algorithm terminates.
The beneficial effects of the present invention are:
1) the present invention is based on deep learning realize characters on license plate detection and identification, effectively inhibit characters on license plate adhesion, The influence for situations such as fracture, deformation, license plate are stained, and shade is remained on license plate sloped and license plate.
2) present invention completes Car license recognition by subsequent screening operation and permutation and combination, does not limit the license plate that can be identified Character length is not high to license plate resolution requirement suitable for the uniline license plate and duplicate rows license plate of a variety of license plate standards.
3) compared to the licence plate recognition method based on segmentation, the present invention is avoided segmentation step bring error accumulation to knowledge Other step, accuracy of identification are higher.
4) compared to end-to-end licence plate recognition method, the present invention is lower to the required precision of License Plate, and not receiving end Limitation in correspondent network design, can identify the license plate of more multi-standard.
Detailed description of the invention
Fig. 1 is the license plate image of input of the invention.
Fig. 2 is the testing result visual image of the invention by depth convolutional neural networks to input license plate.
Fig. 3 is the result visualization image of the invention handled by step 4.
Fig. 4 is the final result visual image of the invention handled by step 7.
Specific embodiment
The specific implementation of the licence plate recognition method of the invention based on deep learning is elaborated below with reference to embodiment Mode.
Step 1: the training one depth convolutional neural networks model M for characters on license plate detection；Construct characters on license plate mark Sign set B={ bi| i=1,2 ..., n, n=67 }=' 0 ', ' 1 ', ' 2 ', ' 3 ', ' 4 ', ' 5 ', ' 6 ', ' 7 ', ' 8 ', ' 9 ', ‘A’,‘B’,‘C’,‘D’,‘E’,‘F’,‘G’,‘H’,‘I’,‘J’,‘K’,‘L’,‘M’,‘N’,‘O’,‘P’,‘Q’,‘R’,‘S’, ' T ', ' U ', ' V ', ' W ', ' X ', ' Y ', ' Z ', ' capital ', ' saliva ', ' Ji ', ' Shanxi ', ' illiteracy ', ' the Liao Dynasty ', ' Ji ', ' black ', ' Shanghai ', ' Soviet Union ', ' Zhejiang ', ' Anhui ', ' Fujian ', ' Jiangxi ', ' Shandong ', ' Henan ', ' Hubei Province ', ' Hunan ', ' Guangdong ', ' osmanthus ', ' fine jade ', ' Chongqing ', ' river ', ' expensive ', ' cloud ', ' hiding ', ' Shan ', ' sweet ', ' blueness ', ' peaceful ', ' new ' }；
Step 2: the license plate image I that positioning obtains being input to characters on license plate detection network M, exports candidate license plate character set Close H={ hi| i=1,2,3 ..., nH, wherein nHIndicate the element number of set H, hiIndicate i-th of candidate license plate of set H Character, hiIt is the triple being made of (b, q, r), b indicates that the class label of candidate characters, b ∈ B, q indicate setting for candidate characters Reliability, q ∈ [0,1], r are the four-tuples being made of (x, y, w, h), and r indicates that the boundary rectangle frame of candidate characters, x, y, w and h divide Not Biao Shi rectangle frame upper left corner abscissa, upper left corner ordinate, width and height；In this example, license plate image is tested such as Shown in Fig. 1, deep neural network M testing result is as shown in Figure 2；
Step 3: the candidate license plate character set H that step 2 is obtained, by the upper left of candidate license plate character minimum circumscribed rectangle Angle abscissa hi.x ascending that H reorders, obtain set C={ ci| i=1,2,3 ..., nC, ciIndicate the i-th of set C A candidate license plate character, ciIt is the triple being made of (b, q, r), nCIndicate the element number of set C；
Step 4: screening operation, specific steps are further executed to the set C that step 3 obtains are as follows:
Step 4.1: C is sequentially traversed, it is rightI=1,2 ..., nCIf ci.q<Tq, then by ciSet D is added, and will ciIt is deleted from set C；Otherwise by ciStay in set C；Finally, obtaining D={ di| i=1,2 ..., nD, diIt is by (b, q, r) structure At triple, nDIndicate the quantity of element in set D；Wherein, ci.q candidate characters c is indicatediConfidence level, TqIndicate confidence Spend threshold value；In this example, Tq=0.5；
Step 4.2: the set C that sequentially traversal step 4.1 obtains is rightI=1,2 ..., nC- 1, by formula (1) A is calculated to delete element a from set C if a ≠ φ；Otherwise, operation is not executed；
Wherein, ci.r candidate characters c is indicatediBoundary rectangle frame, ci+1.r candidate characters c is indicatedi+1Boundary rectangle frame, ToverIndicate the anti-eclipse threshold of rectangle frame, | | indicate the area of rectangle frame, ∩ operation indicates to seek the intersection of two rectangle frames； In this example, Tover=0.6；
Step 4.3: the set D that the set C and step 4.1 obtain to step 4.2 is obtained is rightI=1,2 ..., nC- 1, ifMeet formula (2), records so that djMeet the i value of formula (2), and by djSet C is inserted into as the (i+1) a element；Wherein, djIndicate j-th of element of set D, j=1,2 ..., nD；In this example, the knot of final set C Fruit visual image is as shown in Figure 3；
Wherein, ci.r.x、ci.r.y、ciAnd c .r.wi.r.h candidate characters c is respectively indicatediThe upper left corner of boundary rectangle frame is horizontal Coordinate, upper left corner ordinate, width and height；dj.r.x、dj.r.y、djAnd d .r.wj.r.h candidate characters d is respectively indicatedjIt is external Upper left corner abscissa, upper left corner ordinate, width and the height of rectangle frame；
Step 5: the set C that step 4.3 is obtained, firstly, set C is sequentially traversed, ifI=1,2 ..., nC, Meet formula (3), by ciSet E is added, and by ciIt is deleted from C；Then, the surplus element for traversing set C, is added sequentially to Set E；Finally obtain E={ ei| i=1,2,3 ..., nE, eiIndicate i-th of candidate license plate character of set E, nEIndicate set E Element number；
Step 6: the set E obtained to step 5 sequentially traverses set E, the label of all elements in set E is connected into License plate recognition resultWherein, ei.b ∈ B, ei.b the label of i-th of candidate license plate character of set E, Σ table are indicated Show character serial operation；In this example, L is printed upon in original image and obtains final result visual image, as shown in Figure 4；
Step 7: algorithm terminates.

Claims (1)

1. a kind of licence plate recognition method based on deep learning, which comprises the steps of:
Step 1: the training one depth convolutional neural networks model M for characters on license plate detection；Construct characters on license plate tally set Close B={ bi| i=1,2 ..., n, n=67 }=' 0 ', ' 1 ', ' 2 ', ' 3 ', ' 4 ', ' 5 ', ' 6 ', ' 7 ', ' 8 ', ' 9 ', ' A ', ‘B’,‘C’,‘D’,‘E’,‘F’,‘G’,‘H’,‘I’,‘J’,‘K’,‘L’,‘M’,‘N’,‘O’,‘P’,‘Q’,‘R’,‘S’,‘T’, ' U ', ' V ', ' W ', ' X ', ' Y ', ' Z ', ' capital ', ' saliva ', ' Ji ', ' Shanxi ', ' illiteracy ', ' the Liao Dynasty ', ' Ji ', ' black ', ' Shanghai ', ' Soviet Union ', ' Zhejiang ', ' Anhui ', ' Fujian ', ' Jiangxi ', ' Shandong ', ' Henan ', ' Hubei Province ', ' Hunan ', ' Guangdong ', ' osmanthus ', ' fine jade ', ' Chongqing ', ' river ', ' expensive ', ' cloud ', ' hiding ', ' Shan ', ' sweet ', ' blueness ', ' peaceful ', ' new ' }；
Step 2: the license plate image I that positioning obtains being input to characters on license plate detection network M, exports candidate license plate character set H ={ hi| i=1,2,3 ..., nH, wherein nHIndicate the element number of set H, hiIndicate i-th of candidate license plate word of set H Symbol, hiIt is the triple being made of (b, q, r), b indicates that the class label of candidate characters, b ∈ B, q indicate the confidence of candidate characters Degree, q ∈ [0,1], r are the four-tuples being made of (x, y, w, h), and r indicates the boundary rectangle frame of candidate characters, x, y, w and h difference Indicate upper left corner abscissa, upper left corner ordinate, width and the height of rectangle frame；
Step 3: the candidate license plate character set H that step 2 is obtained, it is horizontal by the upper left corner of candidate license plate character minimum circumscribed rectangle Coordinate hi.x ascending that H reorders, obtain set C={ ci| i=1,2,3 ..., nC, ciIndicate i-th of time of set C Select characters on license plate, ciIt is the triple being made of (b, q, r), nCIndicate the element number of set C；
Step 4: screening operation, specific steps are further executed to the set C that step 3 obtains are as follows:
Step 4.1: C is sequentially traversed, it is rightIf ci.q<Tq, then by ciSet D is added, and by ciFrom It is deleted in set C；Otherwise by ciStay in set C；Finally, obtaining D={ di| i=1,2 ..., nD, diIt is to be made of (b, q, r) Triple, nDIndicate the quantity of element in set D；Wherein, ci.q candidate characters c is indicatediConfidence level, TqIndicate confidence level Threshold value；
Step 4.2: the set C that sequentially traversal step 4.1 obtains is rightA is calculated by formula (1), If a ≠ φ, element a is deleted from set C；Otherwise, operation is not executed；
Wherein, ci.r candidate characters c is indicatediBoundary rectangle frame, ci+1.r candidate characters c is indicatedi+1Boundary rectangle frame, Tover Indicate the anti-eclipse threshold of rectangle frame, | | indicate the area of rectangle frame, ∩ operation indicates to seek the intersection of two rectangle frames；
Step 4.3: the set D that the set C and step 4.1 obtained according to step 4.2 is obtained is right IfMeet formula (2), records so that djMeet the i value of formula (2), and by djSet C is inserted into as (i+1) A element；Wherein, djIndicate j-th of element of set D, j=1,2 ..., nD；
Wherein, ci.r.x、ci.r.y、ciAnd c .r.wi.r.h candidate characters c is respectively indicatediThe horizontal seat in the upper left corner of boundary rectangle frame Mark, upper left corner ordinate, width and height；dj.r.x、dj.r.y、djAnd d .r.wj.r.h candidate characters d is respectively indicatedjExternal square Upper left corner abscissa, upper left corner ordinate, width and the height of shape frame；
Step 5: the set C that step 4.3 is obtained, firstly, set C is sequentially traversed, if Meet public Formula (3), by ciSet E is added, and by ciIt is deleted from C；Then, the surplus element for traversing set C, is added sequentially to set E； Finally obtain E={ ei| i=1,2,3 ..., nE, eiIndicate i-th of candidate license plate character of set E, nEIndicate the element of set E Number；
Step 6: the set E obtained to step 5 sequentially traverses set E, the label of all elements in set E is connected into license plate Recognition resultWherein, ei.b ∈ B, ei.b indicate that the label of i-th of candidate license plate character of set E, Σ indicate word Accord with serial operation；
Step 7: algorithm terminates.
CN201810955796.4A 2018-08-21 2018-08-21 License plate recognition method based on deep learning Active CN109165643B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810955796.4A CN109165643B (en) 2018-08-21 2018-08-21 License plate recognition method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810955796.4A CN109165643B (en) 2018-08-21 2018-08-21 License plate recognition method based on deep learning

Publications (2)

Publication Number Publication Date
CN109165643A true CN109165643A (en) 2019-01-08
CN109165643B CN109165643B (en) 2021-11-16

Family

ID=64896372

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810955796.4A Active CN109165643B (en) 2018-08-21 2018-08-21 License plate recognition method based on deep learning

Country Status (1)

Country Link
CN (1) CN109165643B (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110020650A (en) * 2019-03-26 2019-07-16 武汉大学 A kind of construction method, recognition methods and the device of the deep learning identification model for inclination license plate
CN110287959A (en) * 2019-06-27 2019-09-27 浙江工业大学 A kind of licence plate recognition method based on recognition strategy again
CN110288031A (en) * 2019-06-27 2019-09-27 浙江工业大学 A kind of licence plate recognition method based on Sequence Learning
CN110427937A (en) * 2019-07-18 2019-11-08 浙江大学 A kind of correction of inclination license plate and random length licence plate recognition method based on deep learning
CN110717493A (en) * 2019-09-16 2020-01-21 浙江大学 License plate recognition method containing stacked characters based on deep learning
CN110852324A (en) * 2019-08-23 2020-02-28 上海撬动网络科技有限公司 Deep neural network-based container number detection method
CN111461112A (en) * 2020-03-03 2020-07-28 华南理工大学 License plate character recognition method based on double-cycle transcription network
CN113723408A (en) * 2021-11-02 2021-11-30 上海仙工智能科技有限公司 License plate recognition method and system and readable storage medium

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102663377A (en) * 2012-03-15 2012-09-12 华中科技大学 Character recognition method based on template matching
WO2015147764A1 (en) * 2014-03-28 2015-10-01 Kisa Mustafa A method for vehicle recognition, measurement of relative speed and distance with a single camera
US20150307278A1 (en) * 2014-04-29 2015-10-29 Flextronics Ap, Llc Method and system for order fulfillment
CN106650731A (en) * 2016-12-23 2017-05-10 中山大学 Robust license plate and logo recognition method
CN107292898A (en) * 2017-05-04 2017-10-24 浙江工业大学 A kind of car plate shadow Detection and minimizing technology based on HSV

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102663377A (en) * 2012-03-15 2012-09-12 华中科技大学 Character recognition method based on template matching
WO2015147764A1 (en) * 2014-03-28 2015-10-01 Kisa Mustafa A method for vehicle recognition, measurement of relative speed and distance with a single camera
US20150307278A1 (en) * 2014-04-29 2015-10-29 Flextronics Ap, Llc Method and system for order fulfillment
CN106650731A (en) * 2016-12-23 2017-05-10 中山大学 Robust license plate and logo recognition method
CN107292898A (en) * 2017-05-04 2017-10-24 浙江工业大学 A kind of car plate shadow Detection and minimizing technology based on HSV

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
JIAN W. FANG: ""The research for license plate recognition using sub-image fast independent component analysis"", 《2011 CHINESE CONTROL AND DECISION CONFERENCE (CCDC)》 *
YASUKO TAKAHASHI等: ""License plate recognition using gray image template matching with noise reduction filters and character alignment"", 《SYSTEMS AND COMPUTERS IN JAPAN》 *
王洪元等: ""基于多特征组合的车牌识别算法"", 《常州大学学报（自然科学版）》 *

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110020650A (en) * 2019-03-26 2019-07-16 武汉大学 A kind of construction method, recognition methods and the device of the deep learning identification model for inclination license plate
CN110020650B (en) * 2019-03-26 2021-08-03 武汉大学 Inclined license plate recognition method and device based on deep learning recognition model
CN110287959B (en) * 2019-06-27 2021-06-29 浙江工业大学 License plate recognition method based on re-recognition strategy
CN110287959A (en) * 2019-06-27 2019-09-27 浙江工业大学 A kind of licence plate recognition method based on recognition strategy again
CN110288031A (en) * 2019-06-27 2019-09-27 浙江工业大学 A kind of licence plate recognition method based on Sequence Learning
CN110288031B (en) * 2019-06-27 2021-07-27 浙江工业大学 License plate recognition method based on sequence learning
CN110427937A (en) * 2019-07-18 2019-11-08 浙江大学 A kind of correction of inclination license plate and random length licence plate recognition method based on deep learning
CN110427937B (en) * 2019-07-18 2022-03-22 浙江大学 Inclined license plate correction and indefinite-length license plate identification method based on deep learning
CN110852324A (en) * 2019-08-23 2020-02-28 上海撬动网络科技有限公司 Deep neural network-based container number detection method
CN110717493A (en) * 2019-09-16 2020-01-21 浙江大学 License plate recognition method containing stacked characters based on deep learning
CN110717493B (en) * 2019-09-16 2022-04-01 浙江大学 License plate recognition method containing stacked characters based on deep learning
CN111461112A (en) * 2020-03-03 2020-07-28 华南理工大学 License plate character recognition method based on double-cycle transcription network
CN111461112B (en) * 2020-03-03 2023-02-14 华南理工大学 License plate character recognition method based on dual-cycle transcription network
CN113723408A (en) * 2021-11-02 2021-11-30 上海仙工智能科技有限公司 License plate recognition method and system and readable storage medium
CN113723408B (en) * 2021-11-02 2022-02-25 上海仙工智能科技有限公司 License plate recognition method and system and readable storage medium

Also Published As

Publication number Publication date
CN109165643B (en) 2021-11-16

Similar Documents

Publication Publication Date Title
CN109165643A (en) A kind of licence plate recognition method based on deep learning
CN105046196B (en) Front truck information of vehicles structuring output method based on concatenated convolutional neutral net
CN105608456B (en) A kind of multi-direction Method for text detection based on full convolutional network
WO2017031716A1 (en) Method for analyzing and recognizing handwritten mathematical formula structure in natural scene image
CN100357988C (en) Identification method of mobile number plate based on three-channel parallel artificial nerve network
CN113344857B (en) Defect detection network training method, defect detection method and storage medium
CN104778470B (en) Text detection based on component tree and Hough forest and recognition methods
CN105205488B (en) Word area detection method based on Harris angle points and stroke width
CN103577817A (en) Method and device for identifying forms
CN109190625B (en) Large-angle perspective deformation container number identification method
CN102880859A (en) Method for recognizing number plate
CN112381175A (en) Circuit board identification and analysis method based on image processing
CN111967313A (en) Unmanned aerial vehicle image annotation method assisted by deep learning target detection algorithm
CN113096085A (en) Container surface damage detection method based on two-stage convolutional neural network
CN111860570A (en) Cloud particle image extraction and classification method
CN111340032A (en) Character recognition method based on application scene in financial field
CN110516673A (en) Ancient Books in Yi Language character detection method based on connected component and regression equation character segmentation
CN114387592B (en) Character positioning and identifying method under complex background
CN107392115B (en) Traffic sign identification method based on hierarchical feature extraction
CN110991374A (en) Fingerprint singular point detection method based on RCNN
CN111414938A (en) Target detection method for bubbles in plate heat exchanger
CN109508714B (en) Low-cost multi-channel real-time digital instrument panel visual identification method and system
Sharma et al. Pincode detection using deep CNN for postal automation
CN112884741B (en) Printing apparent defect detection method based on image similarity comparison
CN116862883A (en) Concrete slump detection method based on image semantic segmentation

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
GR01 Patent grant
GR01 Patent grant