CN109508715A - A kind of License Plate and recognition methods based on deep learning - Google Patents

A kind of License Plate and recognition methods based on deep learning Download PDF

Info

Publication number
CN109508715A
CN109508715A CN201811281994.3A CN201811281994A CN109508715A CN 109508715 A CN109508715 A CN 109508715A CN 201811281994 A CN201811281994 A CN 201811281994A CN 109508715 A CN109508715 A CN 109508715A
Authority
CN
China
Prior art keywords
license plate
alexnet
layer
cnn
network model
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
CN201811281994.3A
Other languages
Chinese (zh)
Other versions
CN109508715B (en
Inventor
闵卫东
李祥鹏
赵浩宇
黄杰
韩清
刘瑞康
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.)
Nanchang University
Original Assignee
Nanchang University
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 Nanchang University filed Critical Nanchang University
Priority to CN201811281994.3A priority Critical patent/CN109508715B/en
Publication of CN109508715A publication Critical patent/CN109508715A/en
Application granted granted Critical
Publication of CN109508715B publication Critical patent/CN109508715B/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
    • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G06F18/2193—Validation; Performance evaluation; Active pattern learning techniques based on specific statistical tests
    • 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
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/60—Analysis of geometric attributes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
    • G06V20/625—License plates
    • 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

Landscapes

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

Abstract

The present invention provides a kind of License Plate based on deep learning and recognition methods, including two aspects: License Plate based on Faster R-CNN model and based on the end-to-end Recognition of License Plate Characters of AlexNet-L network model；Step 1, to the license plate area size application k-means++ algorithm process of training set, optimal candidate's size is selected, Faster R-CNN is integrated to；Step 2, using Faster R-CNN training picture, model is obtained；Step 3, it inputs the picture comprising license plate and obtains characteristic pattern after process of convolution；Step 4, the characteristic pattern that will acquire is by RPN network output license plate position and score；Step 5, license plate area is intercepted from original image；Step 6, the license plate area of interception passes through the end-to-end convolutional neural networks of AlexNet-L network model, each character on final output license plate, the present invention carries out License Plate using the method based on Faster R-CNN, using the end-to-end license plate character recognition method of AlexNet-L, the accuracy rate to License Plate and character recognition can be effectively improved.

Description

A kind of License Plate and recognition methods based on deep learning
Technical field
The present invention relates to field of license plate recognition, specially a kind of License Plate based on deep learning and identification side Method.
Background technique
Vehicle License Plate Recognition System (Vehicle License Plate Recognition, VLPR) refer to be able to detect that by It monitors the vehicle on road surface and automatically extracts vehicle license information (containing chinese character, English alphabet, Arabic numerals and number plate face Color) technology that is handled.Car license recognition is one of the important component in modern intelligent transportation system, and application is very wide It is general.It is based on the technologies such as Digital Image Processing, pattern-recognition, computer vision, to the vehicle image of shot by camera Or video sequence is analyzed, and each unique license plate number of automobile is obtained, to complete identification process.After some Parking lot fee collection management, the measurement of magnitude of traffic flow Con trolling index, vehicle location, automobile burglar, high speed may be implemented in continuous processing means Highway hypervelocity automation supervision, electronic eye used for catching red light runner, toll station etc. function.Maintenance traffic safety and city are controlled Peace prevents traffic jam, realizes that traffic automation management has the meaning of reality.
Registration number is unique " identity " mark of vehicle, and licence plate automatic identification technology can not made any in automobile The automatic registration and verifying of automobile " identity " are realized in the case where change, this technology has been applied to highway toll, parking pipe The various occasions such as reason, weighing system, traffic guidance, traffic administration, highway inspection, vehicle scheduling, vehicle detection.
The existing licence plate recognition method based on License Plate Character Segmentation, overwhelming majority experiment is carried out under ideal environment 's.Situations such as once it is fuzzy to encounter such as license plate, being illuminated by the light influences greatly, license plate sloped, the accuracy rate of Car license recognition will be significantly Ground reduces.In addition to this, characters on license plate can not correctly divide the effect that will also directly affect Recognition of License Plate Characters.
Summary of the invention
The purpose of the present invention is to provide a kind of License Plate based on deep learning and recognition methods, to solve above-mentioned back The problem of proposing in scape technology, the present invention selects best license plate area size using k-means++ algorithm, in conjunction with Faster R-CNN improves AlexNet network model to carry out License Plate, designs end and end based on AlexNet-L License plate character recognition method.
To achieve the above object, the invention provides the following technical scheme: a kind of License Plate and knowledge based on deep learning Other method, including two aspects: the License Plate based on Faster R-CNN model and the end based on AlexNet-L network model The Recognition of License Plate Characters of opposite end；
Step 1, to the license plate area size application k-means++ algorithm process of training set, optimal candidate's size is selected, It is integrated to Faster R-CNN；
Step 2, using Faster R-CNN training picture, model is obtained；
Step 3, it inputs the picture comprising license plate and obtains characteristic pattern after process of convolution；
Step 4, resulting characteristic pattern is obtained into candidate frame, the spy that will acquire after full convolutional neural networks RPN processing Sign figure and candidate frame are classified after ROI pooling processing together, export license plate position and score；
Step 5, license plate area is intercepted from original image；
Step 6, the license plate area of interception passes through the end-to-end convolutional neural networks of AlexNet-L network model, finally Export each character on license plate.
It further, is approximately 3: 1, in order to definitely reflect since the license plate area ratio of China is 440cm*140cm The size of license plate area in various pictures, the length-width ratio of 3 ratios is selected using k-means++ algorithm；
K-means++ algorithm needs to use the overlapping rate (IoU) of the candidate frame of model prediction and the candidate frame of label to refer to Mark, the calculation of IoU are as follows:
Wherein, SgroundTruthIndicate true candidate frame, SanchorBoxIndicate the candidate frame of prediction；
K-means++ algorithm obtain the algorithm of initial candidate frame the following steps are included:
The input of S1:k-means++ algorithm is the length and width of license plate
C={ d1(x1, y1), d2(x2, y2) ..., dn(xn, yn) and k candidate frame length-width ratio；
S2: it is c that a sample is randomly selected from C1(c1∈C)；
S3: for each sample in C, each sample is calculated to c1Distance:
d(bi, c1)=1-Io U (bi, c1) (2)
Wherein i ∈ (1,2,3 ..., n)；
S4: the probability that each sample is selected as next mass center is calculated:
S5: S is definedi:
S6: the random number r between one 0 to 1 is generated, judges that r belongs to region { si-1, si, then bi(xi, yi) it is second Mass center:
S7: step S3~S6 is repeated, until obtaining k mass center.
Further, the RPN network in the Faster R-CNN model, is substantially in convolutional neural networks CNN On the basis of increase Quan Juan base cls and reg layers, wherein cls layers be for judging that candidate frame is prospect or background, and Reg layers are for finely tuning candidate frame；
The loss function of Faster R-CNN are as follows:
Wherein, i indicates an index of anchor, piIt is the prediction probability of i-th of anchor, if anchor is positive, Value be 1, conversely,Value be 0, tiIndicate 4 parameter coordinates of predicted boundary frame,It indicates corresponding with positive anchor The coordinate vector of groud-truth box, Classification Loss LclsIt is that the logarithm of 2 classifications (target and non-targeted) loses.
Further, the AlexNet-L network model is an improvement in AlexNet network models, AlexNet-L network model has nine layers of structure, and first and second layer contains convolution, pond layer and normalization；With AlexNet net Unlike network model, the sequence of pond layer and normalization operation is different in the first layer of AlexNet-L network model； The third layer of AlexNet-L network model is to three identical convolution operations of layer 5；Layer 6 used convolutional layer and Pond layer, layer 7 are a full articulamentums, and the 8th and the 9th layer, using seven full articulamentums arranged side by side, is respectively intended to license plate Each character identified.
Further, the AlexNet-L network model carries out following aspects in AlexNet network models Improvement:
1. the improvement of first layer and the second layer:
By normalization and the mutual reversed order of pond layer in first layer in AlexNet and the second layer, i.e., will return in AlexNet One change 1, pond layer 1 and normalization 2,2 reversed order of pond layer；
2. increasing convolutional layer to improve classifying quality:
Increase a same convolutional layer behind the convolutional layer 3 of AlexNet, convolutional layer 4；
3. the improvement of full articulamentum:
The end-to-end character recognition of license plate is 7 characters, and the full articulamentum of the layer 7 of AlexNet is changed to 7 simultaneously The full articulamentum of column obtains the feature vector of 7 characters respectively；
4. the improvement of output layer:
Since the characters on license plate of China has 7 characters, final output should be 7 labels, by the of AlexNet network Eight full articulamentums are changed to 7 full articulamentums arranged side by side, and are respectively connected with 7 full articulamentums arranged side by side of preceding layer, for The last one full articulamentum, the corresponding number of tags of each classification be not unique.
Further, the output layer of the AlexNet-L network model is classified using Softmax regression function, Soffmax Formula are as follows:
G represents classification number, and d is the training function of g；
The calculating process of convolution operation in AlexNet-L network model are as follows:
O=(I+2 × P-K)/S+1 (7)
Wherein, O is the size (length or width of picture) of output data, and I is the size (length of picture of input data Or width), P pad represents whether in width and height both sides filler pixels, K kernel_size, indicates convolution sum Size, S stride, indicate step-length；
The operation calculating process of pond layer are as follows:
O=(I-K)/S+1 (8)
Wherein, the meaning of each parameter is in formula (7)；
In AlexNet-L network, the activation primitive being connected with each convolutional layer is selected with ReLu:
ReLu (x)=Max (x, 0) (9)
Compared with prior art, the beneficial effects of the present invention are:
The present invention carries out License Plate using the method based on Faster R-CNN, is chosen using k-means++ algorithm Best license plate area carries out License Plate, the accuracy rate of Lai Tigao License Plate, using being based in conjunction with Faster R-CNN On the one hand the end-to-end license plate character recognition method of the improved AlexNet-L of AlexNet can be reduced existing based on license plate License Plate Character Segmentation result in the licence plate recognition method of Character segmentation influences Recognition of License Plate Characters；It on the other hand, can be with Improve the accuracy rate of Car license recognition.
Detailed description of the invention
Fig. 1 is that the present invention is based on the License Plates of deep learning and identification framework figure；
Fig. 2 is Faster R-CNN illustraton of model of the present invention；
Fig. 3 is that k-means++ algorithm combination license plate of the present invention obtains anchors procedure chart；
Fig. 4 is 9 anchors dimensional drawings of the invention；
Fig. 5 is AlexNet network architecture figure of the present invention；
Fig. 6 is that the present invention is based on the Recognition of License Plate Characters frame diagrams of AlexNet-L.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right The present invention is described in further detail.The specific embodiments are only for explaining the present invention technical solution described herein, and It is not limited to the present invention.
The present invention provides a kind of technical solution: a kind of License Plate and recognition methods based on deep learning, including two Aspect: the License Plate based on Faster R-CNN model and the end-to-end characters on license plate based on AlexNet-L network model Identification, total algorithm frame are as shown in Figure 1；
Step 1, to the license plate area size application k-means++ algorithm process of training set, optimal candidate's size is selected, It is integrated to Faster R-CNN；
Step 2, using Faster R-CNN training picture, model is obtained；
Step 3, it inputs the picture comprising license plate and obtains characteristic pattern after process of convolution；
Step 4, resulting characteristic pattern is obtained into candidate frame, the spy that will acquire after full convolutional neural networks RPN processing Sign figure and candidate frame are classified after ROI pooling processing together, export license plate position and score；
Step 5, license plate area is intercepted from original image；
Step 6, the license plate area of interception passes through the end-to-end convolutional neural networks of AlexNet-L network model, finally Export each character on license plate.
Existing Faster R-CNN model is not related to the related application of license plate, gets the bid first from data set thus Determine license plate area, then FasterR-CNN is applied among the license plate area data set got, finally obtains every license plate The candidate region of picture.
Further, the present invention analyzes the size of license plate area, convergence rate and License Plate when in order to improve trained Accuracy rate, license plate area size is integrated in Faster R-CNN, due to China license plate area ratio be 440cm* 140cm is approximately 3: 1, in order to definitely reflect the size of license plate area in various pictures, is selected using k-means++ algorithm The length-width ratio of fixed 3 ratios；
K-means++ algorithm needs to use the overlapping rate (IoU) of the candidate frame of model prediction and the candidate frame of label to refer to Mark, the calculation of IoU are as follows:
Wherein, SgroundTruthIndicate true candidate frame, SanchorBoxIndicate the candidate frame of prediction；
The process of the processing of k-means++ algorithm combination license plate is as shown in Figure 3.
K-means++ algorithm obtain the algorithm of initial candidate frame the following steps are included:
The input of S1:k-means++ algorithm is the length and width of license plate
C={ d1(x1, y1), d2(x2, y2) ..., dn(xn, yn) and k candidate frame length-width ratio；
S2: it is c that a sample is randomly selected from C1(c1∈C)；
S3: for each sample in C, each sample is calculated to c1Distance:
d(bi, c1)=1-Io U (bi, c1) (2)
Wherein i ∈ (1,2,3 ..., n)；
S4: the probability that each sample is selected as next mass center is calculated:
S5: S is definedi:
S6: the random number r between one 0 to 1 is generated, judges that r belongs to region { si-1, si, then bi(xi, yi) it is second Mass center:
S7: step S3~S6 is repeated, until obtaining k mass center.
By experiment, selecting k is 3, using the size of k-means++ algorithm picks license plate area, obtain three it is different Length-width ratio (2.2,3.2,4.6), the length-width ratio after rounding are (2,3,5), and the anchor size that benchmark is arranged is 8*8, use Scale is [16,24,40], 9 anchors of each sliding window generation, as shown in Figure 4.
The advantage of Faster R-CNN is not only in that candidate frame to extract partially to be put on GPU and runs, also region candidate The extraction part of frame is embedded into inside network outside network, and the characteristic pattern after convolution can be used to obtain region candidate frame.
Further, the RPN network in the FasterR-CNN model, is substantially in convolutional neural networks CNN On the basis of increase Quan Juan base cls and reg layers, wherein cls layers are and the reg for judging that candidate frame is prospect or background Layer is for finely tuning candidate frame；Faster R-CNN model is as shown in Figure 2.
The loss function of Faster R-CNN are as follows:
Wherein, i indicates an index of anchor, piIt is the prediction probability of i-th of anchor, if anchor is positive, Value be 1, conversely,Value be 0, tiIndicate 4 parameter coordinates of predicted boundary frame,It indicates corresponding with positive anchor The coordinate vector of groud-truth box, Classification Loss LclsIt is that the logarithm of 2 classifications (target and non-targeted) loses.
AlexNet is the convolutional neural networks model based on classification that Krizhevsky et al. is proposed, and obtains 2012 The champion of year ImageNet match.AlexNet structure is as shown in Figure 5.
AlexNet has eight layers of structure, and the 1st, 2 layer has used convolutional layer (conv), pond layer (pool) and normalization behaviour Make (norm)；3rd, 4 layer the same, contains convolutional layer；5th layer uses convolutional layer and pond layer；6-8 layers apply and connect entirely Meet layer (fc).
Although AlexNet network significant effect on target classification, it is simultaneously not applied among Recognition of License Plate Characters, There is no the modellings for having specific meanings；The present invention improves and structure again on the basis of AlexNet network model It builds, proposes a kind of convolutional neural networks model AlexNet-L dedicated for Recognition of License Plate Characters of enhancing.
Further, the AlexNet-L network model is an improvement in AlexNet network models, is had Conducive to the accuracy for the identification for improving characters on license plate；AlexNet-L network model has nine layers of structure, and first and second layer contains Convolution, pond layer and normalization；Unlike AlexNet network model, pond in the first layer of AlexNet-L network model The sequence of layer and normalization operation is different；The third layer of AlexNet-L network model is to three identical volumes of layer 5 Product operation；Layer 6 has used convolutional layer and pond layer, and layer 7 is a full articulamentum, and the 8th and the 9th layer uses side by side Seven full articulamentums are respectively intended to identify that overall framework is as shown in Figure 6 to each character of license plate.
Further, the AlexNet-L network model carries out following aspects in AlexNet network models Improvement:
1. the improvement of first layer and the second layer:
By normalization and the mutual reversed order of pond layer in first layer in AlexNet and the second layer, identification essence can be improved Degree, while but also normalization operation, which reduces, calculates time and memory, therefore 1 will be normalized in AlexNet, pond layer 1 with And normalization 2,2 reversed order of pond layer；
2. increasing convolutional layer to improve classifying quality:
In order to improve the classifying quality of last Recognition of License Plate Characters, increase by one on the convolutional layer 3 of AlexNet, convolutional layer 4 The identical convolutional layer of layer；Due to convolutional layer 3, convolutional layer 4 is better than other layers in terms of classification, after convolutional layer 3, convolutional layer 4 Face increases a same convolutional layer, to find the expression of more vehicle license plate characteristics；
3. the improvement of full articulamentum:
The end-to-end character recognition of license plate is 7 characters, and the full articulamentum of the layer 7 of AlexNet is changed to 7 simultaneously The full articulamentum of column obtains the feature vector of 7 characters respectively；Because of the Chinese character in characters on license plate, letter and number feature is complete Complete different, processing, which is equivalent to, so separately handles the feature of each character, is conducive to improve the accurate of Recognition of License Plate Characters Rate；
4. the improvement of output layer:
Since the characters on license plate of China has 7 characters, final output should be 7 labels, by the of AlexNet network Eight full articulamentums are changed to 7 full articulamentums arranged side by side, and are respectively connected with 7 full articulamentums arranged side by side of preceding layer, for The last one full articulamentum, the corresponding number of tags of each classification be not unique.
In order to verify each validity improved to Recognition of License Plate Characters in AlexNet-L network model, for Corresponding control variate method analysis has been done in the improvement each time of AlexNet-L, does following groups experiment respectively:
First group of experiment: by AlexNet improve in 4. as one group of experiment；
Second group of experiment: by AlexNet improve in be 1. 4. combined together as one group of experiment；
The experiment of third group: by AlexNet improve in the be 2. 4. combined together as one group of experiment；
4th group of experiment: by AlexNet improve in be 3. 4. combined together as one group of experiment；
1. 2. 3. 4. 5th group of experiment: will be combined together as one group of experiment, the experiment of this group is also AlexNet-L net Network.
Experimental result is as shown in table 1, and Chinese character represents the 1st character of license plate, and letter represents the 2nd character of license plate, word Mother+number represents the remaining character of license plate, accuracy rate=correct number/test number.
Table 1 is directed to several groups of comparative experimentss of AlexNet network improvement
From table 1 it follows that the improvement in second group to the 4th group experiment for AlexNet network, to characters on license plate The result of identification is better than first group of experimental result, illustrates the improvement to AlexNet network model to Recognition of License Plate Characters Accuracy rate increases；The accuracy rate of 5th group of experimental result is significantly better than other four groups of experiments, illustrates AlexNet-L network mould Type is conducive to improve the accuracy rate of Recognition of License Plate Characters to the improvement of AlexNet.
Car license recognition of the present invention is based primarily upon Chinese license plate (inland and Taiwan), is all made of 7 last two layers of AlexNet-L Full articulamentum arranged side by side respectively corresponds the 1-7 character of inland license plate, rather than by the 1-7 character of license plate using same The recognition accuracy of single classification character can be improved in one full articulamentum in this way.
The present invention has also done comparative experiments on common data sets Road Patrol (RP).Because what RP data set was collected The license plate of TaiWan, China, TaiWan, China license plate shares 6 characters, and wherein preceding 4 characters of license plate are number, most latter two Character is letter, so being changed to 6 full articulamentums arranged side by side for last two layers of AlexNet-L network.
Further, the output layer of the AlexNet-L network model is classified using Softmax regression function, Soffmax Function can solve polytypic problem, and classical Sigmoid regression function can only solve the problems, such as two classification, Soffmax Formula are as follows:
G represents classification number, and d is the training function of g；
The calculating process of convolution operation in AlexNet-L network model are as follows:
O=(I+2 × P-K)/S+1 (7)
Wherein, O is the size (length or width of picture) of output data, and I is the size (length of picture of input data Or width), P pad represents whether in width and height both sides filler pixels, K kemel_size, indicates convolution sum Size, S stride, indicate step-length；
The operation calculating process of pond layer are as follows:
O=(I-K)/S+1 (8)
Wherein, the meaning of each parameter is in formula (7)；
In AlexNet-L network, the activation primitive that is connected with each convolutional layer selection ReLu, due to Sigmoid and Tanh activation primitive there are speed in gradient descent procedures slow, the problem more than the number of iterations, using ReLu function as activating Speed and efficiency can be improved in function, and the period is greatly shortened, ReLu function are as follows:
ReLu (x)=Max (x, 0) (9)
Input data size for AlexNet-L network is 227 × 227 × 3, by level 1 volume product core number: 96, Size: 11, step-length: 4, the size of the characteristic pattern after convolution is 55 × 55 × 96, and after ReLU activation operation, size of data is not Become, using convolution kernel size: 3, step-length: 2, the size of the characteristic pattern of acquisition is 27 × 27 × 96, special by normalized The size for levying figure is constant, so final input data size is 27 × 27 × 96.
The 2nd layer of AlexNet-L is similar with the 1st layer, and the input of 256 5 × 5 convolution kernels is 27 × 27 × 96 features Figure, further extracts feature, because level 2 volume product is all filled with 2 pixels in width and height both sides, the spy after convolution The size for levying figure is 27 × 27 × 256, and after ReLU activation operation, size of data is constant, collects core size using rolling up: 3, step It is long: 2, the size of the characteristic pattern of acquisition is 13 × 13 × 256, and by normalized, the size of characteristic pattern is constant, so finally Input data size be 13 × 13 × 256.
3-5 layers the same, activates and operates by convolution sum ReLU, and layer 6 is by convolution kernel and ReLU activation operation Afterwards, it joined pond layer, by 3-6, the size of data of output is 6 × 6 × 256, using 7-8 layers of arranged side by side 7 Full articulamentum, the convolution kernel number of the last each label of output layer from front to back are followed successively by 31,26,36,36,36,36,36.
In present invention experiment, the data set of test is from the data set and common data sets RP collected, from the data collected Collection is all Chinese inland license plate.
In order to verify the validity of Faster R-CNN combination VGG network, the present invention is directed to Faster R-CNN combination VGG Do the comparative experiments of License Plate respectively with ZF network.In Faster R-CNN training process, the number of iterations is set as 200000.In changing preceding 160000, Study rate parameter is set as in 0.001,40000 iteration, learning rate 0.0001, Momentum parameter is set as 0.9, and weight attenuation parameter is set as 0.0005.It is that 2440 pictures are tested in test set, It is resulting that the results are shown in Table 2.Wherein, recall rate meaning is that algorithm is determined as license plate and determines that correct license plate number accounts for survey The ratio of lump picture number is tried, accuracy rate meaning is that algorithm is determined as license plate and determines that correct license plate number accounts for algorithm judgement For the ratio of license plate number, the time spent in time is characters on license plate in one picture of average test.
License Plate comparative experiments of the table 2 based on Faster R-CNN
Algorithm Recall rate Accuracy rate Time (/s)
VGG 0.9909 0.9887 0.0321
ZF 0.9762 0.9778 0.0143
Although from Table 2, it can be seen that Faster R-CNN combination VGG network test the time spent in ratio Faster per second The time spent in R-CNN combination ZF network, is long, but accuracy rate of the Faster R-CNN combination VGG network in terms of License Plate Much higher than Faster R-CNN combination ZF network, so the present invention uses the license plate based on Faster R-CNN combination VGG network Localization method.
Table 3 is algorithms of different from the comparative experiments of license plate locating accuracy on the data set and RP data set collected.
License plate locating accuracy comparison (%) on 3 different data collection of table
From table 3 it is observed that the License Plate accuracy rate of inventive algorithm on different data sets is above other calculations Method.
To sum up, fixed in various natural scenes and the enterprising driving board of different data collection, inventive algorithm is in License Plate effect Be significantly better than that other several algorithms, overall performance are superior in accuracy rate.
In the experiment of AlexNet-L network model, as follows, maximum number of iterations 200000 is arranged in network parameter, initially Learning rate is 0.001.Then when iteration 20000 times, learning rate is changed to 0.1 times of last learning rate, Momentum parameter is set as 0.9, and weight attenuation parameter is set as 0.0005.In addition to this, the present invention also provides a comparison of license plate word Symbol segmentation combines the experiment of ANN neural metwork training characters on license plate, and License Plate Character Segmentation adds the Recognition of License Plate Characters of template matching Algorithm, License Plate Character Segmentation add the Recognition of License Plate Characters algorithm of SVM.These types of algorithm is respectively in the test set and RP number from collection It is tested according to collection, experimental result is as shown in table 4 and table 5.In table 4, Chinese character represents the 1st character of license plate, and letter represents 2nd character of license plate, letter+number represent the remaining character of license plate.Accuracy rate=correct number/test number, total character Refer to the result that all characters that license plate includes are correctly validated.Time indicates the mean time of average single license plate test Between.In table 5, total character represents 6 characters in RP data set and identifies correct result.
Table 4 compares (%) from the accuracy rate collected on data set
Algorithm Chinese character Letter Letter+number Total character Time (s)
The present invention 0.952 9 0.973 8 0.969 2 0.950 8 0.019 2
Artificial neural network 0.891 8 0.927 0 0.932 0 0.876 2 0.037 8
Character mother plate matching 0.914 3 0.944 7 0.936 9 0.909 4 0.029 6
SVM 0.950 0 0.966 4 0.955 3 0.946 0 0.064 3
Accuracy rate on 5 RP data set of table compares (%)
Algorithm Total character Time (s)
The present invention 0.9774 0.0136
Artificial neural network 0.9516 0.0289
Character mother plate matching 0.9583 0.0187
SVM 0.9651 0.0421
As can be seen that the single character of the license plate totality recognition accuracy of inventive algorithm and license plate from table 4 and table 5 Recognition accuracy is above other comparison algorithms.In addition to this, when test needed for the average Recognition of License Plate Characters of inventive algorithm Between less than other comparison algorithm.Show that the Recognition of License Plate Characters algorithm of end-to-end AlexNet-L proposed by the invention exists There is superiority in its relevant application.
The above only expresses the preferred embodiment of the present invention, and the description thereof is more specific and detailed, but can not be because This and be interpreted as limitations on the scope of the patent of the present invention.It should be pointed out that for those of ordinary skill in the art, Under the premise of not departing from present inventive concept, several deformations can also be made, improves and substitutes, these belong to protection of the invention Range.Therefore, the scope of protection of the patent of the invention shall be subject to the appended claims.

Claims (6)

1. a kind of License Plate and recognition methods based on deep learning, it is characterised in that: including two aspects: being based on Faster The License Plate of R-CNN model and end-to-end Recognition of License Plate Characters based on AlexNet-L network model；
Step 1, to the license plate area size application k-means++ algorithm process of training set, optimal candidate's size is selected, in conjunction with To Faster R-CNN；
Step 2, using Faster R-CNN training picture, model is obtained；
Step 3, it inputs the picture comprising license plate and obtains characteristic pattern after process of convolution；
Step 4, resulting characteristic pattern is obtained into candidate frame, the characteristic pattern that will acquire after full convolutional neural networks RPN processing Classify after ROI pooling processing together with candidate frame, exports license plate position and score；
Step 5, license plate area is intercepted from original image；
Step 6, the license plate area of interception passes through the end-to-end convolutional neural networks of AlexNet-L network model, final output Each character on license plate.
2. a kind of License Plate and recognition methods based on deep learning according to claim 1, it is characterised in that: due to The license plate area ratio of China is 440cm*140cm, is approximately 3: 1, in order to definitely reflect license plate area in various pictures Size selectes the length-width ratio of 3 ratios using k-means++ algorithm；
K-means++ algorithm needs to use the overlapping rate (IoU) of the candidate frame of model prediction and the candidate frame of label for index, The calculation of IoU are as follows:
Wherein, SgroundTruthIndicate true candidate frame, SanchorBoxIndicate the candidate frame of prediction；
K-means++ algorithm obtain the algorithm of initial candidate frame the following steps are included:
The input of S1:k-means++ algorithm is the length and width C={ d of license plate1(x1, y1), d2(x2, y2) ..., dn(xn, yn) and k candidate frame length-width ratio；
S2: it is c that a sample is randomly selected from C1(c1∈C)；
S3: for each sample in C, each sample is calculated to c1Distance:
d(bi, c1)=1-IoU (bi, c1) (2)
Wherein i ∈ (1,2,3 ..., n)；
S4: the probability that each sample is selected as next mass center is calculated:
S5: S is definedi:
S6: the random number r between one 0 to 1 is generated, judges that r belongs to region { si-1, si, then bi(xi, yi) it is second mass center；
S7: step S3~S6 is repeated, until obtaining k mass center.
3. a kind of License Plate and recognition methods based on deep learning according to claim 1, it is characterised in that: described RPN network in Faster R-CNN model is substantially to increase Quan Juan base on the basis of convolutional neural networks CNN Cls and reg layers, wherein cls layers are for judging that candidate frame is prospect or background, and reg layers are for finely tuning candidate frame；
The loss function of Faster R-CNN are as follows:
Wherein, i indicates an index of anchor, piIt is the prediction probability of i-th of anchor, if anchor is positive,Value be 1, conversely,Value be 0, tiIndicate 4 parameter coordinates of predicted boundary frame,Indicate groud- corresponding with positive anchor The coordinate vector of truth box, Classification Loss LclsIt is that the logarithm of 2 classifications (target and non-targeted) loses.
4. a kind of License Plate and recognition methods based on deep learning according to claim 1, it is characterised in that: described AlexNet-L network model is an improvement in AlexNet network models, and AlexNet-L network model has nine layers Structure, first and second layer contains convolution, pond layer and normalization；Unlike AlexNet network model, AlexNet-L The sequence of pond layer and normalization operation is different in the first layer of network model；The third layer of AlexNet-L network model arrives Three identical convolution operations of layer 5；Layer 6 has used convolutional layer and pond layer, and layer 7 is a full articulamentum, 8th and the 9th layer, using seven full articulamentums arranged side by side, is respectively intended to identify each character of license plate.
5. a kind of License Plate and recognition methods based on deep learning according to claim 4, it is characterised in that: described AlexNet-L network model carries out improvement following aspects in AlexNet network models:
1. the improvement of first layer and the second layer:
By normalization and the mutual reversed order of pond layer in first layer in AlexNet and the second layer, i.e., will be normalized in AlexNet 1, pond layer 1 and normalization 2,2 reversed order of pond layer；
2. increasing convolutional layer to improve classifying quality:
Increase a same convolutional layer behind the convolutional layer 3 of AlexNet, convolutional layer 4；
3. the improvement of full articulamentum:
The end-to-end character recognition of license plate is 7 characters, by the full articulamentum of the layer 7 of AlexNet be changed to 7 it is arranged side by side Full articulamentum obtains the feature vector of 7 characters respectively；
4. the improvement of output layer:
Since the characters on license plate of China has 7 characters, final output should be 7 labels, by the 8th of AlexNet network Full articulamentum is changed to 7 full articulamentums arranged side by side, and is respectively connected with 7 full articulamentums arranged side by side of preceding layer, for last One full articulamentum, the corresponding number of tags of each classification be not unique.
6. a kind of License Plate and recognition methods based on deep learning according to claim 4 or 5, it is characterised in that: The output layer of the AlexNet-L network model is classified using Softmax regression function, Softmax formula are as follows:
G represents classification number, and d is the training function of g；
The calculating process of convolution operation in AlexNet-L network model are as follows:
O=(I+2 × P-K)/S+1 (7)
Wherein, O is the size (length or width of picture) of output data, and I is the size (length or width of picture of input data Degree), P pad is represented whether in width and height both sides filler pixels, and K kemel_size indicates the big of convolution sum Small, S stride indicates step-length；
The operation calculating process of pond layer are as follows:
O=(I-K)/S+1 (8)
Wherein, the meaning of each parameter is in formula (7)；
In AlexNet-L network, the activation primitive being connected with each convolutional layer is selected with ReLu:
ReLu (x)=Max (x, 0) (9).
CN201811281994.3A 2018-10-30 2018-10-30 License plate positioning and identifying method based on deep learning Active CN109508715B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811281994.3A CN109508715B (en) 2018-10-30 2018-10-30 License plate positioning and identifying method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811281994.3A CN109508715B (en) 2018-10-30 2018-10-30 License plate positioning and identifying method based on deep learning

Publications (2)

Publication Number Publication Date
CN109508715A true CN109508715A (en) 2019-03-22
CN109508715B CN109508715B (en) 2022-11-08

Family

ID=65747145

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811281994.3A Active CN109508715B (en) 2018-10-30 2018-10-30 License plate positioning and identifying method based on deep learning

Country Status (1)

Country Link
CN (1) CN109508715B (en)

Cited By (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993138A (en) * 2019-04-08 2019-07-09 北京易华录信息技术股份有限公司 A kind of car plate detection and recognition methods and device
CN110097044A (en) * 2019-05-13 2019-08-06 苏州大学 Stage car plate detection recognition methods based on deep learning
CN110210475A (en) * 2019-05-06 2019-09-06 浙江大学 A kind of characters on license plate image partition method of non-binaryzation and edge detection
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
CN110390330A (en) * 2019-07-25 2019-10-29 网链科技集团有限公司 Electric bicycle license plate identifying system and method
CN110502655A (en) * 2019-07-31 2019-11-26 武汉大学 A kind of image nature descriptive statement generation method being embedded in scene text information
CN110598693A (en) * 2019-08-12 2019-12-20 浙江工业大学 Ship plate identification method based on fast-RCNN
CN110674821A (en) * 2019-09-24 2020-01-10 浙江工商大学 License plate recognition method for non-motor vehicle
CN110766017A (en) * 2019-10-22 2020-02-07 国网新疆电力有限公司信息通信公司 Mobile terminal character recognition method and system based on deep learning
CN110969164A (en) * 2019-12-20 2020-04-07 湖南千视通信息科技有限公司 Low-illumination imaging license plate recognition method and device based on deep learning end-to-end
CN110991444A (en) * 2019-11-19 2020-04-10 复旦大学 Complex scene-oriented license plate recognition method and device
CN111104937A (en) * 2019-11-25 2020-05-05 上海眼控科技股份有限公司 Vehicle door information detection method and device, computer equipment and storage medium
CN111563515A (en) * 2020-05-18 2020-08-21 哈尔滨理工大学 Improved Faster-RCNN-based license plate recognition method
CN112200007A (en) * 2020-09-15 2021-01-08 青岛邃智信息科技有限公司 License plate detection and identification method under community monitoring scene
CN112308092A (en) * 2020-11-20 2021-02-02 福州大学 Light-weight license plate detection and identification method based on multi-scale attention mechanism
CN112464938A (en) * 2020-12-18 2021-03-09 深圳赛安特技术服务有限公司 License plate detection and identification method, device, equipment and storage medium
CN112560856A (en) * 2020-12-18 2021-03-26 深圳赛安特技术服务有限公司 License plate detection and identification method, device, equipment and storage medium
CN112766273A (en) * 2021-01-20 2021-05-07 湘潭大学 License plate recognition method
WO2021103897A1 (en) * 2019-11-29 2021-06-03 深圳云天励飞技术股份有限公司 License plate number recognition method and device, electronic device and storage medium
CN112906429A (en) * 2019-11-19 2021-06-04 英业达科技有限公司 Image-text image identification system and method
CN112950954A (en) * 2021-02-24 2021-06-11 电子科技大学 Intelligent parking license plate recognition method based on high-position camera
CN113191220A (en) * 2021-04-15 2021-07-30 广州紫为云科技有限公司 Deep learning-based double-layer license plate recognition method
US20220284723A1 (en) * 2020-12-23 2022-09-08 Abbyy Development Inc. Zero-footprint image capture by mobile device
CN115713763A (en) * 2022-11-25 2023-02-24 青海卓旺智慧信息科技有限公司 Potato image recognition system based on deep learning

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107016677A (en) * 2017-03-24 2017-08-04 北京工业大学 A kind of cloud atlas dividing method based on FCN and CNN
CN107358596A (en) * 2017-04-11 2017-11-17 阿里巴巴集团控股有限公司 A kind of car damage identification method based on image, device, electronic equipment and system
CN107563385A (en) * 2017-09-02 2018-01-09 西安电子科技大学 License plate character recognition method based on depth convolution production confrontation network
CN107657233A (en) * 2017-09-28 2018-02-02 东华大学 Static sign language real-time identification method based on modified single multi-target detection device
US20180089593A1 (en) * 2016-09-26 2018-03-29 Acusense Technologies, Inc. Method and system for an end-to-end artificial intelligence workflow
CN107977671A (en) * 2017-10-27 2018-05-01 浙江工业大学 A kind of tongue picture sorting technique based on multitask convolutional neural networks
WO2018088170A1 (en) * 2016-11-09 2018-05-17 パナソニックＩｐマネジメント株式会社 Information processing method, information processing device, and program
US20180165551A1 (en) * 2016-12-08 2018-06-14 Intel Corporation Technologies for improved object detection accuracy with multi-scale representation and training
CN108171191A (en) * 2018-01-05 2018-06-15 百度在线网络技术（北京）有限公司 For detecting the method and apparatus of face
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

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180089593A1 (en) * 2016-09-26 2018-03-29 Acusense Technologies, Inc. Method and system for an end-to-end artificial intelligence workflow
WO2018088170A1 (en) * 2016-11-09 2018-05-17 パナソニックＩｐマネジメント株式会社 Information processing method, information processing device, and program
US20180165551A1 (en) * 2016-12-08 2018-06-14 Intel Corporation Technologies for improved object detection accuracy with multi-scale representation and training
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
CN107016677A (en) * 2017-03-24 2017-08-04 北京工业大学 A kind of cloud atlas dividing method based on FCN and CNN
CN107358596A (en) * 2017-04-11 2017-11-17 阿里巴巴集团控股有限公司 A kind of car damage identification method based on image, device, electronic equipment and system
CN107563385A (en) * 2017-09-02 2018-01-09 西安电子科技大学 License plate character recognition method based on depth convolution production confrontation network
CN107657233A (en) * 2017-09-28 2018-02-02 东华大学 Static sign language real-time identification method based on modified single multi-target detection device
CN107977671A (en) * 2017-10-27 2018-05-01 浙江工业大学 A kind of tongue picture sorting technique based on multitask convolutional neural networks
CN108171191A (en) * 2018-01-05 2018-06-15 百度在线网络技术（北京）有限公司 For detecting the method and apparatus of face

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
MIN WEIDONG: ""A New Approach to Track Multiple Vehicles With the Combination of Robust Detection and Two Classifiers"", 《IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS》 *
XIAO LISHA: ""Scene classification with improved AlexNet model"", 《2017 12TH INTERNATIONAL CONFERENCE ON INTELLIGENT SYSTEMS AND KNOWLEDGE ENGINEERING》 *
卢洪涛 等: ""深度卷积神经网络在计算机视觉中的应用研究综述"", 《数据采集与处理》 *
王玮 等: ""基于择优检测和多尺度匹配的实时人脸识别"", 《计算机工程与设计》 *

Cited By (33)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993138A (en) * 2019-04-08 2019-07-09 北京易华录信息技术股份有限公司 A kind of car plate detection and recognition methods and device
CN110210475B (en) * 2019-05-06 2021-05-18 浙江大学 License plate character image segmentation method based on non-binarization and edge detection
CN110210475A (en) * 2019-05-06 2019-09-06 浙江大学 A kind of characters on license plate image partition method of non-binaryzation and edge detection
CN110097044A (en) * 2019-05-13 2019-08-06 苏州大学 Stage car plate detection recognition methods based on deep learning
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
CN110390330A (en) * 2019-07-25 2019-10-29 网链科技集团有限公司 Electric bicycle license plate identifying system and method
CN110502655A (en) * 2019-07-31 2019-11-26 武汉大学 A kind of image nature descriptive statement generation method being embedded in scene text information
CN110502655B (en) * 2019-07-31 2022-04-01 武汉大学 Method for generating image natural description sentences embedded with scene character information
CN110598693A (en) * 2019-08-12 2019-12-20 浙江工业大学 Ship plate identification method based on fast-RCNN
CN110674821A (en) * 2019-09-24 2020-01-10 浙江工商大学 License plate recognition method for non-motor vehicle
CN110674821B (en) * 2019-09-24 2022-05-03 浙江工商大学 License plate recognition method for non-motor vehicle
CN110766017A (en) * 2019-10-22 2020-02-07 国网新疆电力有限公司信息通信公司 Mobile terminal character recognition method and system based on deep learning
CN110991444B (en) * 2019-11-19 2023-08-29 复旦大学 License plate recognition method and device for complex scene
CN110991444A (en) * 2019-11-19 2020-04-10 复旦大学 Complex scene-oriented license plate recognition method and device
CN112906429A (en) * 2019-11-19 2021-06-04 英业达科技有限公司 Image-text image identification system and method
CN111104937A (en) * 2019-11-25 2020-05-05 上海眼控科技股份有限公司 Vehicle door information detection method and device, computer equipment and storage medium
WO2021103897A1 (en) * 2019-11-29 2021-06-03 深圳云天励飞技术股份有限公司 License plate number recognition method and device, electronic device and storage medium
CN110969164A (en) * 2019-12-20 2020-04-07 湖南千视通信息科技有限公司 Low-illumination imaging license plate recognition method and device based on deep learning end-to-end
CN111563515A (en) * 2020-05-18 2020-08-21 哈尔滨理工大学 Improved Faster-RCNN-based license plate recognition method
CN112200007A (en) * 2020-09-15 2021-01-08 青岛邃智信息科技有限公司 License plate detection and identification method under community monitoring scene
CN112308092B (en) * 2020-11-20 2023-02-28 福州大学 Light-weight license plate detection and identification method based on multi-scale attention mechanism
CN112308092A (en) * 2020-11-20 2021-02-02 福州大学 Light-weight license plate detection and identification method based on multi-scale attention mechanism
CN112560856B (en) * 2020-12-18 2024-04-12 深圳赛安特技术服务有限公司 License plate detection and identification method, device, equipment and storage medium
CN112464938B (en) * 2020-12-18 2024-04-12 深圳赛安特技术服务有限公司 License plate detection and identification method, device, equipment and storage medium
CN112464938A (en) * 2020-12-18 2021-03-09 深圳赛安特技术服务有限公司 License plate detection and identification method, device, equipment and storage medium
CN112560856A (en) * 2020-12-18 2021-03-26 深圳赛安特技术服务有限公司 License plate detection and identification method, device, equipment and storage medium
US20220284723A1 (en) * 2020-12-23 2022-09-08 Abbyy Development Inc. Zero-footprint image capture by mobile device
US11948385B2 (en) * 2020-12-23 2024-04-02 Abbyy Development Inc. Zero-footprint image capture by mobile device
CN112766273A (en) * 2021-01-20 2021-05-07 湘潭大学 License plate recognition method
CN112950954A (en) * 2021-02-24 2021-06-11 电子科技大学 Intelligent parking license plate recognition method based on high-position camera
CN112950954B (en) * 2021-02-24 2022-05-20 电子科技大学 Intelligent parking license plate recognition method based on high-position camera
CN113191220A (en) * 2021-04-15 2021-07-30 广州紫为云科技有限公司 Deep learning-based double-layer license plate recognition method
CN115713763A (en) * 2022-11-25 2023-02-24 青海卓旺智慧信息科技有限公司 Potato image recognition system based on deep learning

Also Published As

Publication number Publication date
CN109508715B (en) 2022-11-08

Similar Documents

Publication Publication Date Title
CN109508715A (en) A kind of License Plate and recognition methods based on deep learning
CN109919108B (en) Remote sensing image rapid target detection method based on deep hash auxiliary network
CN111553201B (en) Traffic light detection method based on YOLOv3 optimization algorithm
CN109447034A (en) Traffic mark detection method in automatic Pilot based on YOLOv3 network
CN111091105A (en) Remote sensing image target detection method based on new frame regression loss function
CN106650731B (en) Robust license plate and vehicle logo recognition method
CN106096602A (en) Chinese license plate recognition method based on convolutional neural network
CN106504233A (en) Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN
CN109447033A (en) Vehicle front obstacle detection method based on YOLO
CN107103317A (en) Fuzzy license plate image recognition algorithm based on image co-registration and blind deconvolution
CN111079602A (en) Vehicle fine granularity identification method and device based on multi-scale regional feature constraint
CN108171112A (en) Vehicle identification and tracking based on convolutional neural networks
CN105354568A (en) Convolutional neural network based vehicle logo identification method
CN109190444A (en) A kind of implementation method of the lane in which the drivers should pay fees vehicle feature recognition system based on video
CN105825212A (en) Distributed license plate recognition method based on Hadoop
CN108416270B (en) Traffic sign identification method based on multi-attribute combined characteristics
CN104978567A (en) Vehicle detection method based on scenario classification
CN108681696A (en) Merge the infrared target robustness recognition methods of multiple features dimensionality reduction and transfer learning
Xiang et al. Lightweight fully convolutional network for license plate detection
CN110008899B (en) Method for extracting and classifying candidate targets of visible light remote sensing image
CN111126459A (en) Method and device for identifying fine granularity of vehicle
CN110032952A (en) A kind of road boundary point detecting method based on deep learning
CN112084890A (en) Multi-scale traffic signal sign identification method based on GMM and CQFL
CN108960175A (en) A kind of licence plate recognition method based on deep learning
CN115170611A (en) Complex intersection vehicle driving track analysis method, system and application

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