CN104298976B - Detection method of license plate based on convolutional neural networks - Google Patents

Detection method of license plate based on convolutional neural networks Download PDF

Info

Publication number
CN104298976B
CN104298976B CN201410549315.1A CN201410549315A CN104298976B CN 104298976 B CN104298976 B CN 104298976B CN 201410549315 A CN201410549315 A CN 201410549315A CN 104298976 B CN104298976 B CN 104298976B
Authority
CN
China
Prior art keywords
car plate
neural networks
convolutional neural
license plate
plate
Prior art date
Application number
CN201410549315.1A
Other languages
Chinese (zh)
Other versions
CN104298976A (en
Inventor
叶茂
王梦伟
郑梦雅
苟群森
彭明超
Original Assignee
电子科技大学
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 电子科技大学 filed Critical 电子科技大学
Priority to CN201410549315.1A priority Critical patent/CN104298976B/en
Publication of CN104298976A publication Critical patent/CN104298976A/en
Application granted granted Critical
Publication of CN104298976B publication Critical patent/CN104298976B/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/00624Recognising scenes, i.e. recognition of a whole field of perception; recognising scene-specific objects
    • G06K9/00791Recognising scenes perceived from the perspective of a land vehicle, e.g. recognising lanes, obstacles or traffic signs on road scenes
    • G06K9/00825Recognition of vehicle or traffic lights
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/20Image acquisition
    • G06K9/34Segmentation of touching or overlapping patterns in the image field
    • G06K9/344Segmentation of touching or overlapping patterns in the image field using recognition of characters or words
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/62Methods or arrangements for recognition using electronic means
    • G06K9/6217Design or setup of recognition systems and techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
    • G06K9/6256Obtaining sets of training patterns; Bootstrap methods, e.g. bagging, boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K2209/00Indexing scheme relating to methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K2209/15Detection and recognition of car license plates

Abstract

The invention discloses a kind of detection method of license plate based on convolutional neural networks;Specifically include and detection acquisition car plate roughing region is carried out to license plate image to be detected by the Adaboost car plate detections device based on Haar features, acquisition car plate final candidate region car plate roughing region is identified by the complete car plate identification model of convolutional neural networks, segmentation is carried out to the final candidate region of car plate by multi-threshold segmentation algorithm and obtains car plate Chinese character, letter and number, pass through Chinese character, letter and number convolutional neural networks identification model is to car plate Chinese character, letter and number, which is identified, obtains license plate recognition result.The present invention can be realized using the complete car plate identification model of Adaboost car plate detections device and convolutional neural networks based on Haar features and the license plate image under the conditions of is accurately identified, segmentation is carried out using multi-threshold segmentation algorithm to character easier to split character picture, there is good result in engineer applied simultaneously.

Description

Detection method of license plate based on convolutional neural networks

Technical field

The invention belongs to Computer Vision Recognition technical field, more particularly to a kind of car plate inspection based on convolutional neural networks Survey method.

Background technology

Intelligent transportation system is mainly used in magnitude of traffic flow monitoring, vehicle prison as a part for smart city, digital city Control, freeway toll station management, cell intelligent management, parking lot management, traffic police etc., wherein car plate detection is intelligent friendship The most key part of way system, is to realize the intelligentized important step of traffic administration, is computer vision and pattern-recognition skill Important subject of the art in intelligent transportation system.License plate recognition technology is passed by one's way in vehicle, full-automatic non-parking charge of passing a bridge, Accident automatic telemetry on the measurement of magnitude of traffic flow Con trolling index, automatic vehicle identification, highway, onstream inspection, vehicle Positioning, automobile burglar, inspection and tracking rule-breaking vehicle, illegal activities, safeguard traffic safety and urban public security, prevent traffic from blocking up Plug, improves the service rate of charge road and bridge, all has positive effect in terms of relieving traffic density, is mainly shown as: In terms of social security and social management, intelligent transportation system is divided the video information that monitoring scene is transmitted from each crossing Analysis, obtains the information of vehicles in video, the running orbit of target vehicle is tracked;In charge station, parking lot etc., intelligence is handed over Way system can be realized to be extracted to vehicle license plate, and then realizes the automatic management to vehicle, improves efficiency and promptness;In height Shelves cell, intelligent transportation system can also be managed to the vehicle of cell, recognize district vehicles information, be cell increase safely Chip.Complete license plate recognition technology includes license plate image acquisition, license plate area positioning, License Plate Character Segmentation, characters on license plate knowledge Not etc., the car plate technology used both at home and abroad at present mainly includes the vehicle license location technique based on car plate textural characteristics, based on car plate The vehicle license location technique of color characteristic, the vehicle license location technique based on car plate edge feature, the License Plate based on genetic algorithm The localization methods such as technology, the cutting techniques, the character recognition skill based on character stroke such as the Character segmentation technology based on projecting method Art etc..Main stream approach used in car plate detection product is still that computer vision algorithms make is mutually tied with DSP technologies both at home and abroad at present Close.Due to the difference of domestic and international different regions car plate, so Related product is portable poor, so exploitation is applied to continent The car plate detection in area is current problem urgently to be resolved hurrily with identification technology.The car plate detection that product on existing market is used with Recognition methods mainly faces three major issues:1) when body color and close car plate color, it is impossible to accurately determine license plate area;2) It is relatively complicated based on methods such as character strokes that current Recognition of License Plate Characters is used, and accuracy rate is not high.3) when license plate image illumination Characters on license plate can not preferably be split when uneven.The Deng Haifeng of Tencent Technology (Shenzhen) Co., Ltd. et al. invention《Car plate Localization method》On May 7th, 2012 applies for a patent and got the Green Light to China national Department of Intellectual Property, in public affairs on March 27th, 2013 Open, Publication No.:CN102999753A.A kind of license plate locating method of the disclosure of the invention, comprises the following steps:To the complete of collection Scape figure carries out color detection, and extracts designated color region as license plate candidate area;The edge for extracting license plate candidate area is special Levy, split license plate candidate area using edge feature, generate license plate candidate area collection;The car plate that license plate candidate area is concentrated is waited Favored area is merged according to merging condition set in advance, generates new license plate candidate area.To in license plate candidate area Car plate carries out Slant Rectify;The candidate region that candidate region is concentrated is verified, non-license plate area is rejected.The invention passes through profit Car plate coarse positioning is carried out with the color characteristic of license plate area, then the car plate after coarse positioning is screened using textural characteristics, finally Correction process is carried out to car plate, candidate license plate region is obtained.But actually the invention has three below shortcoming:1) it is simple according to Color and texture space feature is relied to carry out positioning anti-interference to car plate poor;2) during using color space HSV to License Plate, Related Color characteristics parameters numerical value is bad during to License Plate under nonideality holds;3) texture is being carried out to license plate area During Feature Selection, license plate area is screened using the transition times of every a line, when license plate area uneven illumination without preferable Effect.Can not preferably be adapted to the positioning of all license plate areas, especially car plate color in this way and body color is close When, it can not be well used in practice in engineering.The invention of the fertile spark of South China Science & Engineering University et al.《A kind of Car license recognition Method》August is applied for a patent and got the Green Light to China national Department of Intellectual Property on the 30th within 2012, open on January 16th, 2013, Publication No.:CN102880859A.A kind of licence plate recognition method of the disclosure of the invention, key step such as next time:Obtain car plate figure Picture;Image preprocessing;License Plate;Car plate super-resolution rebuilding;Character segmentation;Character recognition.The invention is realized in image Car plate can be accurately identified in the case of unsharp, can be to the identification problem under car plate defect, ambiguity, and do not schemed As the limitation of background, body color etc..But actually the invention is more bothered in super-resolution rebuilding, on the other hand to character Identification on using industry commonly use the recognition methods based on character stroke, the recognition methods based on stroke more bother, and It is particularly not high to characters on license plate cut quality, when image has pollution, recognition effect it cannot be guaranteed that, such as to confusable character In identification.

The content of the invention

In order to solve problem above, the present invention proposes a kind of detection method of license plate based on convolutional neural networks, can It is easier, more accurately realize the license plate image under the conditions of various be identified.

Present disclosure is described for convenience, and following term is explained first:

Term 1:Adaboost graders

Adaboost graders are that different Weak Classifiers are trained for same training set, then these Weak Classifiers Gather one strong classifier of composition.

Term 2:BP algorithm

BP algorithm is a kind of back-propagation algorithm, is divided into forward process, backward process two parts, and wherein forward process refers to The process of final result that obtains is entered data into network, and reverse procedure refers to the difference by forward process and sample actual numerical value Value is as error, the process being updated to network weight.

Term 3:Multi-threshold segmentation algorithm

Multi-threshold segmentation algorithm is the problems such as there may be shade, uneven illumination for license plate image, specific with some Threshold value, which carries out binaryzation to license plate image, can not preferably split all characters and propose in binary image A kind of multi-threshold segmentation algorithm.White point occupies the percentage of license plate area and is used as finger in license plate image using after binaryzation for it Mark, solve corresponding threshold value, then constantly reduce the ratio of white point, and then UNICOM domain where all character zones is divided A kind of method for cutting out, is finally merged, sorting determines its position relationship, wherein to each threshold to all UNICOM domains The UNICOM domain that value is split pixel number zero setting before next threshold value is split.

The technical scheme is that:A kind of detection method of license plate based on convolutional neural networks, comprises the following steps:

S1. prepare car plate positive sample and negative sample, train the Adaboost car plate detection devices based on Haar features;

S2. gray processing processing is carried out to image to be detected, and utilizes the Adaboost based on Haar features in step S1 Car plate detection device carries out sliding window detection, obtains the car plate roughing region in image to be detected;

S3. complete car plate positive sample and negative sample, the complete car plate identification model of training convolutional neural networks are prepared;

S4. the complete car plate identification model of convolutional neural networks in step S3 is utilized to the car plate roughing region in step S2 Screened, obtain the final candidate region of car plate, and preserve the final candidate region copy of car plate;

S5. horizontal tilt angle inspection is carried out to the final candidate region of car plate in step S4 using local horizontal projecting method Survey, and horizontal tilt correction is carried out to the final candidate region of car plate and the final candidate region copy of car plate using rotation correction method;

S6. the final candidate region of car plate and the final candidate of car plate after being corrected using multi-threshold segmentation algorithm to horizontal tilt Region copy is handled, and obtains characters on license plate correspondence image;

S7. Chinese character, letter and number sample, training Chinese character convolutional neural networks identification model, letter and number volume are prepared Product neural network recognization model;

S8. the Chinese character convolutional neural networks identification model in step S7, the identification of letter and number convolutional neural networks are utilized Model carries out Recognition of License Plate Characters to characters on license plate correspondence image in step S6, and counts recognition result.

Further, the step S1 prepares car plate positive sample and negative sample, trains the Adaboost based on Haar features Car plate detection device, be specially:The car plate positive sample and negative sample of preparation are normalized, generated based on Haar features Adaboost car plate detection devices;Wherein, car plate positive sample includes the license plate grey level image sample under different illumination conditions, and car plate is born Sample includes non-license plate area gray level image sample.

Further, the step S3 prepares complete car plate positive sample and negative sample, training convolutional neural networks entire vehicle Board identification model, be specially:The complete car plate positive sample and negative sample of preparation are normalized and name processing, and designed Vehicle board convolutional neural networks structure, generates the complete car plate identification model of convolutional neural networks;Wherein, complete car plate positive sample bag Complete car plate coloured image sample is included, complete car plate negative sample includes non-license plate area coloured image sample.

Further, the step S5 carries out horizontal tilt using local horizontal projecting method to the final candidate region of car plate Angle is detected, and carries out horizontal tilt to the final candidate region of car plate and the final candidate region copy of car plate using rotation correction method Correction, be specially:Choose the final candidate region of car plate and carry out gray processing processing, and the imagery exploitation Da-Jin algorithm after gray processing is entered Row binary conversion treatment;The image after binaryzation is rotated in the range of set angle again, and solves postrotational image water Flat projection values;5 groups of numerical value larger in the floor projection numerical value under each angle are finally chosen to be summed, it is relatively more different Summation numerical value under angle and using the corresponding angle of maximum summation numerical value as the angle of inclination of car plate, and according to obtained car plate Angle of inclination carries out rotation correction to the final candidate region of car plate and the final candidate region copy of car plate.

Further, the final candidate regions of car plate after the step S6 is corrected using multi-threshold segmentation algorithm to horizontal tilt Domain and the final candidate region copy of car plate are handled, and obtain characters on license plate correspondence image, are specially:To the final candidate regions of car plate Domain carries out gray processing processing, and statistics accounts for the ratio of the total number of pixels of license plate area image more than the pixel count of a certain gray values Example;One group of candidate thresholds is determined, and binaryzation is carried out to car plate final candidate region image successively to each threshold value Processing;Solve UNICOM domain to the final candidate region image of car plate after binary conversion treatment each time, and by all overlapping UNICOMs Domain is merged, and the UNICOM domain after merging is ranked up, and obtains the UNICOM domain with characters on license plate sequence consensus;According to what is obtained The positional information of UNICOM domain is split to the final candidate region duplicate pictures of corresponding car plate, obtains characters on license plate corresponding diagram Picture.

Further, the step S7 prepares Chinese character, letter and number sample, training Chinese character convolutional neural networks identification mould Type, letter and number convolutional neural networks identification model, be specially:Chinese character, letter and number sample to preparation carry out normalizing Change is handled, and separately designs the convolutional neural networks structure of Chinese character, letter and number;Using BP algorithm to Chinese character convolutional Neural net Network, letter and number convolutional neural networks are trained, generation Chinese character convolutional neural networks identification model, letter and number convolution Neural network recognization model.

The beneficial effects of the invention are as follows:The detection method of license plate based on convolutional neural networks of the present invention, which is utilized, is based on Haar The Adaboost car plate detections device of feature carries out roughing to license plate image to be detected, then complete using the convolutional neural networks proposed Vehicle board identification model is screened to license plate area, obtains final license plate area, is realized to the license plate image under different condition Accurately identify;Recycle multi-threshold segmentation algorithm to carry out Character segmentation to final license plate image, realize easier to character It is identified;Finally using Chinese character convolutional neural networks identification model, Chinese character and alphabetical convolutional neural networks identification model to right Character picture is answered to be identified;The present invention has good effect on engineer applied.

Brief description of the drawings

Fig. 1 is the detection method of license plate schematic flow sheet based on convolutional neural networks of the present invention.

Fig. 2 is the complete car plate convolutional neural networks structural representation of the present invention.

Fig. 3 is the car plate figure before the horizontal tilt correction of the present invention.

Fig. 4 is the car plate figure after the horizontal tilt correction of the present invention.

Fig. 5 is the multi-threshold segmentation algorithm schematic flow sheet of the present invention.

Fig. 6 is the design sketch of the multi-threshold segmentation algorithm of the present invention.

Fig. 7 is the Chinese character of the present invention, letter and number convolutional neural networks structural representation.

Fig. 8 is the mapping to be checked in the embodiment of the present invention 1.

Fig. 9 is the license plate area testing result figure in the embodiment of the present invention 1.

Figure 10 is the Character segmentation result figure in the embodiment of the present invention 1.

Figure 11 is the recognition result figure in the embodiment of the present invention 1.

Figure 12 is the mapping to be checked in the embodiment of the present invention 2.

Figure 13 is the license plate area testing result figure in the embodiment of the present invention 2.

Figure 14 is the Character segmentation result figure in the embodiment of the present invention 2.

Figure 15 is the recognition result figure in the embodiment of the present invention 2.

Embodiment

In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not For limiting the present invention.

As shown in figure 1, being the detection method of license plate schematic flow sheet based on convolutional neural networks of the present invention.One kind is based on The detection method of license plate of convolutional neural networks, comprises the following steps:

S1. prepare car plate positive sample and negative sample, train the Adaboost car plate detection devices based on Haar features.

For the positioning licence plate candidate region from license plate image to be detected, the present invention prepares 2000 positive samples and 5000 Negative sample is used to train the Adaboost car plate detection devices based on Haar features.Wherein, car plate positive sample includes different illumination bars License plate grey level image sample under part, car plate negative sample includes the non-license plate area gray-scale maps such as vehicle body, windshield, exhaust screen Decent.Image training series in car plate positive sample collection and car plate negative sample collection is 15 grades.By car plate positive sample and car plate Negative sample is normalized to 45 × 15 pixel sizes, generates the Adaboost car plate detection devices based on Haar features.

S2. gray processing processing is carried out to image to be detected, and utilizes the Adaboost based on Haar features in step S1 Car plate detection device carries out sliding window detection, obtains the car plate roughing region in image to be detected.

Here image to be detected is the image detected the need for obtaining in advance.Car plate roughing region generally comprises car Region is easily obscured in board region and deep bead etc..Handled by gray processing and sliding window is detected, car plate color and car can be directed to The close situation of body color accurately recognizes license plate area.

S3. complete car plate positive sample and negative sample, the complete car plate identification model of training convolutional neural networks are prepared.

In order to isolate complete license plate image from car plate roughing region, the present invention prepare 10000 positive samples and 10000 negative samples are used for the complete car plate identification model of training convolutional neural networks.Wherein, complete car plate positive sample includes complete Car plate coloured image sample, complete car plate negative sample includes non-license plate area coloured image sample.Positive sample and negative sample are returned One turns to 48 × 16 pixel sizes, and is named processing;By designing complete car plate convolutional neural networks structure, training is obtained The complete car plate identification model of convolutional neural networks.Here positive sample is named with 1 beginning, negative sample is named with 0 beginning, realized The corresponding correct label of sample is determined according to the name of positive sample and negative sample.According to label and the difference of network output valve, profit Network weight is updated with BP algorithm.As shown in Fig. 2 being the complete car plate convolutional neural networks structural representation of the present invention Figure.Wherein, conv1 is that 1, conv2 of convolution operation layer be that convolution operation layer 2, pool1 is that pond operation layer 1, pool2 are pond Operation layer 2, feature map are the characteristic pattern obtained after convolution operation, and fc is the full articulamentum in neutral net;Conv1 layers There are 4 feature map, conv2 layer there are 4 feature map, between input layer and conv1 layers, pool1 layers and conv2 layers Between, pool2 layers between full fc layers of connection be full connection;Between conv1 layers and pool1 layers, conv2 layers and pool2 layers Between be local connection.

S4. the complete car plate identification model of convolutional neural networks in step S3 is utilized to the car plate roughing region in step S2 Screened, obtain the final candidate region of car plate, and preserve the final candidate region copy of car plate.

Car plate roughing region is screened using convolutional neural networks complete car plate identification model, pseudo- region is removed, obtains The final candidate region of pick-up board, and preserve the final candidate region copy of car plate.The final candidate region copy of car plate is used to determine word Character is cut after symbol particular location.

S5. horizontal tilt angle inspection is carried out to the final candidate region of car plate in step S4 using local horizontal projecting method Survey, and horizontal tilt correction is carried out to the final candidate region of car plate and the final candidate region copy of car plate using rotation correction method.

For the final candidate region of car plate it is possible that it is inaccurate the problem of, the present invention be respectively adopted local horizontal projection Method and rotation correction method carry out horizontal tilt correction to the final candidate region of car plate and the final candidate region copy of car plate.Here Partial projection method is used for the angle of inclination for detecting license plate candidate area;Rotation correction method is used for according to the car plate candidate regions detected Domain angle angle of inclination, horizontal tilt correction is carried out to the final candidate region of car plate and the final candidate region copy of car plate.Such as Fig. 3 It is shown, it is the car plate figure before the horizontal tilt of the present invention is corrected.As shown in figure 4, being the car after the horizontal tilt correction of the present invention Board figure.0.8 × 0.8 pixel region for choosing the final candidate region center of car plate carries out gray processing processing, and to gray processing Area image afterwards carries out binary conversion treatment using big law;Again to the image after binaryzation in (- 30 °, 30 °) angular range Each angle is rotated respectively, and solves the floor projection numerical value of postrotational image;Finally choose each angle Under floor projection numerical value in larger 5 groups of numerical value summed, the summation numerical value under relatively more different angles is simultaneously by maximum summation The corresponding angle of numerical value as car plate angle of inclination, and according to obtained license plate sloped angle to the final candidate region of car plate and The final candidate region copy of car plate carries out rotation correction.

S6. the final candidate region of car plate and the final candidate of car plate after being corrected using multi-threshold segmentation algorithm to horizontal tilt Region copy is handled, and obtains characters on license plate correspondence image.

For license plate image it is possible that shade or uneven illumination the problem of, the present invention utilizes multi-threshold segmentation algorithm Candidate region image final to car plate carries out Character segmentation, can more easily realize Character segmentation.As shown in figure 5, being this hair Bright multi-threshold segmentation algorithm schematic flow sheet.Ash is carried out to the final candidate region image of car plate after slant correction first Degreeization processing, counts the corresponding pixel number of each gray value, and white after binaryzation according to each gray value is carried out Color dot accounting determines one group of threshold value.When carrying out binarization operation, the grey scale pixel value that will be greater than a certain gray value is revised as 255, That is white point;Gray value less than a certain threshold value is revised as 0, i.e. black color dots.Binaryzation behaviour is being carried out for some gray values After work, image effect is close, i.e., white point accounting difference is less than a certain numerical value, then need to only choose wherein a certain gray value. Successively using each group of threshold value to the final candidate region image of car plate carry out binary conversion treatment, due to the sets of threshold values of determination be according to Order arrangement from small to large, so the image after binaryzation is in gradual manner.As shown in fig. 6, being the multi thresholds point of the present invention Cut the design sketch of algorithm.UNICOM domain is solved to the final candidate region image of car plate after binary conversion treatment each time, and will be all Overlapping UNICOM domain merge, the UNICOM domain after merging is ranked up based on x coordinate, obtained and characters on license plate sequence consensus UNICOM domain.Theoretical here according to Computer Image Processing, the horizontal respective coordinates in image are x-axis, i.e. width;Image Middle longitudinal respective coordinates are y-axis, i.e. short transverse;Car plate transverse direction correspondence x-axis, longitudinal direction correspondence y-axis utilizes each UNICOM domain upper left The x coordinate size at angle determines the left-right relation of each UNICOM domain.Transition times, candidate region further according to image outline region are big Whether small, candidate region height ratio information, it is character to determine profile;If profile is character, character information is kept;If profile It is not character, then continues to carry out binary conversion treatment to the final candidate region image of car plate using next group of threshold value successively.Preserve word Accord with after information, by the profile domain zero setting of the final candidate region image of car plate, that is, be revised as black, so utilizing next group of threshold value Carry out that during binarization operation the region will not be detected again;Continue successively using next group of threshold value to the final candidate region image of car plate Carry out binary conversion treatment.The final candidate region duplicate pictures of corresponding car plate are carried out according to the positional information of obtained UNICOM domain Segmentation, obtains the correspondence image of 7 characters of car plate.

S7. Chinese character, letter and number sample, training Chinese character convolutional neural networks identification model, letter and number volume are prepared Product neural network recognization model.

The present invention is using Chinese character convolutional neural networks identification model, letter and number convolutional neural networks identification model to car Board character correspondence image is identified, and reduces identification difficulty, while also reducing the requirement to the quality of character.It is of the invention accurate Being ready for use on training Chinese character convolutional neural networks identification model, the sample of letter and number convolutional neural networks identification model includes 8000 each province abbreviation Hanzi specimen images, wherein Chinese character totally 31;8000 letter and number sample images, wherein word Main bag includes A, B ... Z, digital to include 0,1 ... 9;Particularly, the present invention has been also prepared for port, Australia, police, sample image.Distinguish first It is 16 × 32 pixel sizes to Chinese character, letter and number samples normalization;Correspondence Chinese character convolutional neural networks knot is separately designed again Structure, letter and number convolutional neural networks structure;Using BP algorithm respectively to Chinese character convolutional neural networks structure, letter and number Convolutional neural networks structure is trained, Chinese character convolutional neural networks structure here and letter and number convolutional neural networks knot Structure is identical, ultimately produces Chinese character convolutional neural networks identification model, letter and number convolutional neural networks identification model.Such as Fig. 7 It is shown, it is Chinese character, the letter and number convolutional neural networks structural representation of the present invention.Conv1 layers, conv2 layers have 16 It is full connection between feature map, input layer and conv1 layers, pool1 layers and conv2 layers, pool2 layers and fc layers, It is local connection between conv1 layers and pool1 layers, conv2 layers and pool2 layers.

S8. the Chinese character convolutional neural networks identification model in step S7, the identification of letter and number convolutional neural networks are utilized Model carries out Recognition of License Plate Characters to characters on license plate correspondence image in step S6, and counts recognition result.

7 character correspondence images of car plate are detected respectively, final recognition result is obtained.

The present invention gathers image to be detected at different conditions respectively, and using the present invention based on convolutional neural networks Image to be detected under 2 kinds of different conditions is identified respectively for detection method of license plate.As shown in figure 8, being the embodiment of the present invention 1 In mapping to be checked.As shown in figure 9, being the license plate area testing result figure in the embodiment of the present invention 1.As shown in Figure 10, it is this Character segmentation result figure in inventive embodiments 1.As shown in figure 11, it is recognition result figure in the embodiment of the present invention 1.Such as Figure 12 It is shown, it is the car plate figure to be detected in the embodiment of the present invention 2.As shown in figure 13, it is license plate area inspection in the embodiment of the present invention 2 Survey result figure.As shown in figure 14, it is Character segmentation result figure in the embodiment of the present invention 2.As shown in figure 15, it is implementation of the present invention Recognition result figure in example 2.As can be seen that the practical implementation of the present invention at different conditions from above-mentioned 2 embodiments In, with good recognition effect.

One of ordinary skill in the art will be appreciated that embodiment described here is to aid in reader and understands this hair Bright principle, it should be understood that protection scope of the present invention is not limited to such especially statement and embodiment.This area Those of ordinary skill can make according to these technical inspirations disclosed by the invention various does not depart from the other each of essence of the invention Plant specific deformation and combine, these deformations and combination are still within the scope of the present invention.

Claims (12)

1. a kind of detection method of license plate based on convolutional neural networks, it is characterised in that comprise the following steps:
S1. prepare car plate positive sample and negative sample, train the Adaboost car plate detection devices based on Haar features;
S2. gray processing processing is carried out to image to be detected, and utilizes the Adaboost car plates based on Haar features in step S1 Detector carries out sliding window detection, obtains the car plate roughing region in image to be detected;
S3. complete car plate positive sample and negative sample, the complete car plate identification model of training convolutional neural networks are prepared;
S4. the car plate roughing region in step S2 is carried out using the convolutional neural networks complete car plate identification model in step S3 Screening, obtains the final candidate region of car plate, and preserve the final candidate region copy of car plate;
S5. horizontal tilt angle detection is carried out to the final candidate region of car plate in step S4 using local horizontal projecting method, And horizontal tilt correction is carried out to the final candidate region of car plate and the final candidate region copy of car plate using rotation correction method, specifically For:Choose the final candidate region of car plate and carry out gray processing processing, and binaryzation is carried out to the imagery exploitation Da-Jin algorithm after gray processing Processing;The image after binaryzation is rotated in the range of set angle again, and solves postrotational image level projection number Value;Finally choose 5 groups of numerical value larger in the floor projection numerical value under each angle to be summed, under relatively more different angles Sum numerical value and using the corresponding angle of maximum summation numerical value as the angle of inclination of car plate, and according to obtained license plate sloped angle Rotation correction is carried out to the final candidate region of car plate and the final candidate region copy of car plate;
S6. the final candidate region of car plate and the final candidate region of car plate after being corrected using multi-threshold segmentation algorithm to horizontal tilt Copy is handled, and obtains characters on license plate correspondence image;
S7. Chinese character, letter and number sample, training Chinese character convolutional neural networks identification model, letter and number convolution god are prepared Through Network Recognition model;
S8. Chinese character convolutional neural networks identification model, the letter and number convolutional neural networks identification model in step S7 are utilized Recognition of License Plate Characters is carried out to characters on license plate correspondence image in step S6, and counts recognition result.
2. the detection method of license plate as claimed in claim 1 based on convolutional neural networks, it is characterised in that the step S1 is accurate Standby car plate positive sample and negative sample, train the Adaboost car plate detection devices based on Haar features, are specially:By the car plate of preparation Positive sample and negative sample are normalized, and generate the Adaboost car plate detection devices based on Haar features;Wherein, car plate is being just Sample includes the license plate grey level image sample under different illumination conditions, and it is decent that car plate negative sample includes non-license plate area gray-scale map This.
3. the detection method of license plate as claimed in claim 1 based on convolutional neural networks, it is characterised in that the step S3 is accurate Standby complete car plate positive sample and negative sample, the complete car plate identification model of training convolutional neural networks, specially:By the complete of preparation Car plate positive sample and negative sample are normalized and name processing, and design complete car plate convolutional neural networks structure, generation volume The complete car plate identification model of product neutral net;Wherein, complete car plate positive sample includes complete car plate coloured image sample, entire vehicle Board negative sample includes non-license plate area coloured image sample.
4. the detection method of license plate as claimed in claim 1 based on convolutional neural networks, it is characterised in that the step S6 profits The final candidate region of car plate and the final candidate region copy of car plate after being corrected with multi-threshold segmentation algorithm to horizontal tilt are carried out Processing, obtains characters on license plate correspondence image, is specially:Candidate region final to car plate carries out gray processing processing, and statistics is more than The pixel count of a certain gray values accounts for the ratio of the total number of pixels of license plate area image;One group of candidate thresholds is determined, and to wherein Each threshold value successively to the final candidate region image of car plate carry out binary conversion treatment;To the car after binary conversion treatment each time The final candidate region image of board solves UNICOM domain, and all overlapping UNICOM domains are merged, and the UNICOM domain after merging is entered Row sequence, obtains the UNICOM domain with characters on license plate sequence consensus;According to the positional information of obtained UNICOM domain to corresponding car plate Final candidate region duplicate pictures are split, and obtain characters on license plate correspondence image.
5. the detection method of license plate as claimed in claim 1 based on convolutional neural networks, it is characterised in that the step S7 is accurate Standby Chinese character, letter and number sample, training Chinese character convolutional neural networks identification model, the identification of letter and number convolutional neural networks Model, be specially:Chinese character, letter and number sample to preparation are normalized, and separately design Chinese character, alphabetical sum The convolutional neural networks structure of word;Chinese character convolutional neural networks, letter and number convolutional neural networks are carried out using BP algorithm Training, generation Chinese character convolutional neural networks identification model, letter and number convolutional neural networks identification model.
CN201410549315.1A 2014-10-16 2014-10-16 Detection method of license plate based on convolutional neural networks CN104298976B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201410549315.1A CN104298976B (en) 2014-10-16 2014-10-16 Detection method of license plate based on convolutional neural networks

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201410549315.1A CN104298976B (en) 2014-10-16 2014-10-16 Detection method of license plate based on convolutional neural networks

Publications (2)

Publication Number Publication Date
CN104298976A CN104298976A (en) 2015-01-21
CN104298976B true CN104298976B (en) 2017-09-26

Family

ID=52318697

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201410549315.1A CN104298976B (en) 2014-10-16 2014-10-16 Detection method of license plate based on convolutional neural networks

Country Status (1)

Country Link
CN (1) CN104298976B (en)

Families Citing this family (41)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104636309B (en) * 2015-02-06 2018-01-05 南京理工大学 Matrix recognition methods based on machine vision
CN104680149B (en) * 2015-03-10 2018-07-03 苏州科达科技股份有限公司 A kind of object type recognition methods and system
CN104809443B (en) * 2015-05-05 2018-12-28 上海交通大学 Detection method of license plate and system based on convolutional neural networks
CN104933420B (en) * 2015-07-03 2018-11-27 北京旷视科技有限公司 A kind of scene image recognition methods and scene image identify equipment
CN105279488B (en) * 2015-10-16 2018-12-25 西南交通大学 A kind of mileage Sign Board automatic identifying method of highway video routing inspection
CN105335710A (en) * 2015-10-22 2016-02-17 合肥工业大学 Fine vehicle model identification method based on multi-stage classifier
CN105335743A (en) * 2015-10-28 2016-02-17 重庆邮电大学 Vehicle license plate recognition method
CN105528607B (en) * 2015-10-30 2019-02-15 小米科技有限责任公司 Method for extracting region, model training method and device
CN105279556B (en) * 2015-11-05 2017-11-07 国家卫星海洋应用中心 A kind of Enteromorpha detection method and device
CN105469047B (en) * 2015-11-23 2019-02-22 上海交通大学 Chinese detection method and system based on unsupervised learning deep learning network
CN105354572B (en) * 2015-12-10 2018-10-12 苏州大学 A kind of automatic license plate identification system based on simplified convolutional neural networks
CN105608454B (en) * 2015-12-21 2019-08-09 上海交通大学 Character detecting method and system based on text structure component detection neural network
CN105590102A (en) * 2015-12-30 2016-05-18 中通服公众信息产业股份有限公司 Front car face identification method based on deep learning
CN107025452A (en) * 2016-01-29 2017-08-08 富士通株式会社 Image-recognizing method and image recognition apparatus
CN106156767A (en) 2016-03-02 2016-11-23 平安科技(深圳)有限公司 Driving license effect duration extraction method, server and terminal
CN106127248A (en) * 2016-06-24 2016-11-16 平安科技(深圳)有限公司 Car plate sorting technique based on degree of depth study and system
CN106446784A (en) * 2016-08-30 2017-02-22 东软集团股份有限公司 Image detection method and apparatus
CN106469304A (en) * 2016-09-22 2017-03-01 西安理工大学 Handwritten signature location positioning method in bill based on depth convolutional neural networks
CN106548155A (en) * 2016-10-28 2017-03-29 安徽四创电子股份有限公司 A kind of detection method of license plate based on depth belief network
CN106557747B (en) * 2016-11-15 2018-06-22 平安科技(深圳)有限公司 The method and device of identification insurance single numbers
CN106650748A (en) * 2016-11-16 2017-05-10 武汉工程大学 Chinese character recognition method based on convolution neural network
CN106778737B (en) * 2016-11-24 2019-06-18 北京文安智能技术股份有限公司 A kind of license plate antidote, device and a kind of video acquisition device
CN106407981B (en) * 2016-11-24 2019-04-23 北京文安智能技术股份有限公司 A kind of licence plate recognition method, apparatus and system
CN108121984A (en) * 2016-11-30 2018-06-05 杭州海康威视数字技术股份有限公司 A kind of character identifying method and device
US10296794B2 (en) 2016-12-20 2019-05-21 Jayant Rtti On-demand artificial intelligence and roadway stewardship system
CN106778745A (en) * 2016-12-23 2017-05-31 深圳先进技术研究院 A kind of licence plate recognition method and device, user equipment
WO2018112900A1 (en) * 2016-12-23 2018-06-28 深圳先进技术研究院 License plate recognition method and apparatus, and user equipment
CN106778650A (en) * 2016-12-26 2017-05-31 深圳极视角科技有限公司 Scene adaptive pedestrian detection method and system based on polymorphic type information fusion
CN106778659A (en) * 2016-12-28 2017-05-31 深圳市捷顺科技实业股份有限公司 A kind of licence plate recognition method and device
CN106845487A (en) * 2016-12-30 2017-06-13 佳都新太科技股份有限公司 A kind of licence plate recognition method end to end
CN106874902A (en) * 2017-01-19 2017-06-20 博康智能信息技术有限公司北京海淀分公司 A kind of license board information recognition methods and device
CN106897770A (en) * 2017-01-19 2017-06-27 博康智能信息技术有限公司北京海淀分公司 A kind of method and device for setting up Car license recognition model
CN106874907B (en) * 2017-01-19 2019-08-27 博康智能信息技术有限公司北京海淀分公司 A kind of method and device for establishing Car license recognition model
CN106886778B (en) * 2017-04-25 2020-02-07 福州大学 License plate character segmentation and recognition method in monitoring scene
CN107203754B (en) * 2017-05-26 2019-04-12 北京邮电大学 A kind of license plate locating method and device based on deep learning
CN107392201A (en) * 2017-06-09 2017-11-24 中国科学院自动化研究所 The pillar recognition methods of catenary mast, storage medium, processing equipment
CN107705323A (en) * 2017-10-13 2018-02-16 北京理工大学 A kind of level set target tracking method based on convolutional neural networks
CN107506765A (en) * 2017-10-13 2017-12-22 厦门大学 A kind of method of the license plate sloped correction based on neutral net
CN108109414A (en) * 2017-12-19 2018-06-01 武汉极目智能技术有限公司 A kind of preventing road monitoring system and method based on vehicle assistant drive
CN108416348A (en) * 2018-01-29 2018-08-17 重庆邮电大学 Plate location recognition method based on support vector machines and convolutional neural networks
CN108515967B (en) * 2018-04-25 2020-02-07 深圳卡安兴科技有限公司 Anti-collision early warning method and device based on license plate recognition

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103605953A (en) * 2013-10-31 2014-02-26 电子科技大学 Vehicle interest target detection method based on sliding window search
CN103679209A (en) * 2013-11-29 2014-03-26 广东领域安防有限公司 Sparse theory based character recognition method
CN103761531A (en) * 2014-01-20 2014-04-30 西安理工大学 Sparse-coding license plate character recognition method based on shape and contour features
CN103824091A (en) * 2014-02-27 2014-05-28 中国石油大学(华东) Vehicle license plate recognition method for intelligent transportation system

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TWI384408B (en) * 2009-04-30 2013-02-01 Ind Tech Res Inst Method and system for identifying image and outputting identification result

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103605953A (en) * 2013-10-31 2014-02-26 电子科技大学 Vehicle interest target detection method based on sliding window search
CN103679209A (en) * 2013-11-29 2014-03-26 广东领域安防有限公司 Sparse theory based character recognition method
CN103761531A (en) * 2014-01-20 2014-04-30 西安理工大学 Sparse-coding license plate character recognition method based on shape and contour features
CN103824091A (en) * 2014-02-27 2014-05-28 中国石油大学(华东) Vehicle license plate recognition method for intelligent transportation system

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
"卷积神经网络的研究及其在车牌识别系统中的应用";陆璐;《中国优秀博硕士学位论文全文数据库 (硕士) 信息科技辑》;20060915;论文正文第15页第4段-第16页第3段、第27页第3段-第36页第4段,附图4.5-4.15,表4.1-4.5 *
"运动模糊车牌自动识别算法研究与实现";高巍;《中国优秀硕士学位论文全文数据库》;20120415;论文正文第28页第4段-第37页第3段、第46页第3段-第48页第2段,附图3.12-3.15 *
"面向车牌检测Adaboost算法的FPGA实现";王亮;《中国优秀硕士学位论文全文数据库 信息科技辑》;20130315;论文正文第9页第2段-第12页第2段、第15页第1段-第27页第5段,附图2.1-2.5、附图3.2、3.12 *

Also Published As

Publication number Publication date
CN104298976A (en) 2015-01-21

Similar Documents

Publication Publication Date Title
CN105373794B (en) A kind of licence plate recognition method
CN103077384B (en) A kind of method and system of vehicle-logo location identification
Deb et al. Vehicle License Plate Detection Method Based on Sliding Concentric Windows and Histogram.
CN101929867B (en) Clear path detection using road model
Wu et al. Applying a functional neurofuzzy network to real-time lane detection and front-vehicle distance measurement
RU2708422C1 (en) Atm management system and method
CN102163284B (en) Chinese environment-oriented complex scene text positioning method
CN103035013B (en) A kind of precise motion shadow detection method based on multi-feature fusion
CN103198332B (en) A kind of far infrared vehicle-mounted pedestrian detection method of real-time robust
CN102054274B (en) Method for full automatic extraction of water remote sensing information in coastal zone
Yang et al. A novel approach for license plate character segmentation
CN103065138B (en) Recognition method of license plate number of motor vehicle
CN102682292B (en) Method based on monocular vision for detecting and roughly positioning edge of road
CN104392212B (en) The road information detection and front vehicles recognition methods of a kind of view-based access control model
CN103714538B (en) road edge detection method, device and vehicle
CN101398894B (en) Automobile license plate automatic recognition method and implementing device thereof
CN105260699A (en) Lane line data processing method and lane line data processing device
CN105184271A (en) Automatic vehicle detection method based on deep learning
CN101408942B (en) Method for locating license plate under a complicated background
CN102722704B (en) Method and system for recognizing vehicle license plate by integrating video dynamic tracking
CN105550665B (en) A kind of pilotless automobile based on binocular vision can lead to method for detecting area
CN102855622B (en) A kind of infrared remote sensing image sea ship detection method based on significance analysis
CN106022232A (en) License plate detection method based on deep learning
CN104392205B (en) A kind of recognition methods and system of improper license plate
KR20070027768A (en) Method for traffic sign detection

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20170926

Termination date: 20181016