CN108921162A - Licence plate recognition method and Related product based on deep learning - Google Patents
Licence plate recognition method and Related product based on deep learning Download PDFInfo
- Publication number
- CN108921162A CN108921162A CN201810593633.6A CN201810593633A CN108921162A CN 108921162 A CN108921162 A CN 108921162A CN 201810593633 A CN201810593633 A CN 201810593633A CN 108921162 A CN108921162 A CN 108921162A
- Authority
- CN
- China
- Prior art keywords
- license plate
- network structure
- anchor
- context module
- convolutional layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/14—Image acquisition
- G06V30/148—Segmentation of character regions
- G06V30/153—Segmentation of character regions using recognition of characters or words
-
- 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
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biomedical Technology (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the present application discloses a kind of licence plate recognition method based on deep learning, and this method is applied to electronic device, and this method includes：Original input picture is obtained, the picture element matrix of the original input picture is extracted；Trained network model execution multilayer forward operation is preset into picture element matrix input and obtains forward operation as a result, determining license plate set of coordinates according to the forward operation result；The license plate range in the original input picture is determined according to the license plate set of coordinates, identifies the license plate content in the license plate range.
Description
Technical field
This application involves computer vision recognition technology fields, and in particular to a kind of Car license recognition side based on deep learning
Method and Related product.
Background technique
Car plate detection and identification are sport technique segments important in intelligent transportation system, to other portions of intelligent transportation system
Point, such as vehicle-logo recognition, body color analysis, model analysis, bayonet test, electronic police and other traffic events correlation modules
The quality of processing have an important influence.
It is several to be roughly divided into car plate detection, License Plate, Character segmentation and character recognition etc. for current Vehicle License Plate Recognition System
Step, but network model memory is big, arithmetic speed is slow, is all made of the same network model, effect to all types of license plates
It is single, it can not precisely identify the license plate of all sizes.
Summary of the invention
The embodiment of the present application provides a kind of licence plate recognition method and Related product based on deep learning, to reduce net
The memory of network model improves the speed of Car license recognition.
In a first aspect, the embodiment of the present application provides a kind of licence plate recognition method based on deep learning, the method application
In electronic device, the method includes：
Original input picture is obtained, the picture element matrix of the original input picture is extracted；
Trained network model execution multilayer forward operation is preset into picture element matrix input and obtains forward operation knot
Fruit determines license plate set of coordinates according to the forward operation result；
The license plate range in the original input picture is determined according to the license plate set of coordinates, is identified in the license plate range
License plate content.
Second aspect, the embodiment of the present application provide a kind of electronic device of Car license recognition, and the electronic device includes：
Acquiring unit extracts the picture element matrix of the original input picture for obtaining original input picture；
Arithmetic element is obtained for trained network model execution multilayer forward operation to be preset in picture element matrix input
To forward operation as a result, determining license plate set of coordinates according to the forward operation result；
Determination unit is identified for determining the license plate range in the original input picture according to the license plate set of coordinates
License plate content in the license plate range.
The third aspect, the embodiment of the present application provide a kind of terminal, including one or more processors, one or more storages
Device, one or more transceivers, and one or more programs, one or more of programs are stored in the memory
In, and be configured to be executed by one or more of processors, described program includes for executing as described in relation to the first aspect
The instruction of step in method.
Fourth aspect, the embodiment of the present application provide a kind of computer readable storage medium, and storage is handed over for electronic data
The computer program changed, wherein the computer program makes the method for computer execution as described in relation to the first aspect.
5th aspect, the embodiment of the present application provide a kind of computer program product, and the computer program product includes depositing
The non-transient computer readable storage medium of computer program is stored up, the computer is operable to make computer to execute such as the
Method described in one side.
Implement the embodiment of the present application, has the advantages that：
As can be seen that establishing multiple convolutional layers in the network model of the embodiment of the present application, sufficiently extract in license plate image
Vehicle license plate characteristic, and context module network structure of having connected, for license plate actual size by context module net
The anchor of network structure is designed and sized to 1:2, to accurately determine license plate area, improve network model positioning licence plate coordinate
Accuracy.
Detailed description of the invention
In order to more clearly explain the technical solutions in the embodiments of the present application, make required in being described below to embodiment
Attached drawing is briefly described, it should be apparent that, the accompanying drawings in the following description is some embodiments of the present application, for ability
For the those of ordinary skill of domain, without creative efforts, it can also be obtained according to these attached drawings other attached
Figure.
Fig. 1 is a kind of flow diagram of licence plate recognition method based on deep learning provided by the embodiments of the present application；
Fig. 2 is a kind of structural schematic diagram of deep learning network model provided by the embodiments of the present application；
Fig. 3 is the knot of the inception network structure in a kind of deep learning network model provided by the embodiments of the present application
Structure schematic diagram；
Fig. 4 is the context module network structure in a kind of deep learning network model provided by the embodiments of the present application
Structural schematic diagram；
Fig. 5 is a kind of flow diagram of convolution algorithm disclosed in the embodiment of the present application；
Fig. 6 is the flow diagram of another kind convolution algorithm disclosed in the embodiment of the present application；
Fig. 7 is a kind of functional structure of the Car license recognition electronic device based on deep learning disclosed in the embodiment of the present application
Figure.
Specific embodiment
Below in conjunction with the attached drawing in the embodiment of the present application, technical solutions in the embodiments of the present application carries out clear, complete
Site preparation description, it is clear that described embodiment is some embodiments of the present application, instead of all the embodiments.Based on this Shen
Please in embodiment, every other implementation obtained by those of ordinary skill in the art without making creative efforts
Example, shall fall in the protection scope of this application.
The description and claims of this application and term " first ", " second ", " third " and " in the attached drawing
Four " etc. are not use to describe a particular order for distinguishing different objects.In addition, term " includes " and " having " and it
Any deformation, it is intended that cover and non-exclusive include.Such as it contains the process, method of a series of steps or units, be
System, product or equipment are not limited to listed step or unit, but optionally further comprising the step of not listing or list
Member, or optionally further comprising other step or units intrinsic for these process, methods, product or equipment.
Referenced herein " embodiment " is it is meant that the special characteristic, result or the characteristic that describe can wrap in conjunction with the embodiments
It is contained at least one embodiment of the application.Each position in the description occur the phrase might not each mean it is identical
Embodiment, nor the independent or alternative embodiment with other embodiments mutual exclusion.Those skilled in the art explicitly and
Implicitly understand, embodiment described herein can be combined with other embodiments.
Electronic device involved by the embodiment of the present application may include the various handheld devices with wireless communication function,
Mobile unit, wearable device calculate equipment or are connected to other processing equipments and various forms of radio modem
User equipment (User Equipment, UE), mobile station (Mobile Station, MS), terminal device (terminal
Device) etc..For convenience of description, apparatus mentioned above is referred to as terminal.Operation system involved by the embodiment of the present invention
System is the software systems for being managed collectively to hardware resource, and providing a user business interface.
Refering to fig. 1, Fig. 1 is that a kind of process of licence plate recognition method based on deep learning provided by the embodiments of the present application is shown
It is intended to, the method is applied to electronic device, the method includes：
Step S101, original input picture is obtained, the picture element matrix of the original input picture is extracted.
Optionally, after getting the original input picture, this is extracted to image progress gray proces and is originally inputted figure
The tonal gradation of each pixel as in, group become the picture element matrix of the original input picture.Certainly, as original input picture is
Color image can also extract the original input picture rgb pixel value, i.e. R picture element matrix, G picture element matrix and B picture element matrix,
R picture element matrix, G picture element matrix and B picture element matrix group are become into triple channel input model.
Step S102, trained network model execution multilayer forward operation is preset in picture element matrix input to obtain just
To operation result, license plate set of coordinates is determined according to the forward operation result.
Optionally, before trained network model is preset in picture element matrix input, initial depth is trained first
Learning network model construction this preset trained network model.
Optionally, as shown in Fig. 2, the initial deep learning network model includes：First convolutional layer Conv1, the first pond
Change layer Pool1, the second convolutional layer Conv2, the second pond layer Pool2, the first inception network structure, the 2nd inception
Network structure, the 3rd inception network structure, third convolutional layer Conv3, Volume Four lamination Conv4, the first context
Module network structure and the 2nd context module network structure, first convolutional layer Conv1, the first pond layer Pool1,
Second convolutional layer Conv2, the second pond layer Pool2, the first inception network structure, the 2nd inception network structure,
3rd inception network structure, third convolutional layer Conv3 and Volume Four lamination Conv4 are sequentially connected in series, this first
Context module network structure and the third convolutional layer Conv3 are connected in series, the 2nd context module network knot
Structure and the Volume Four lamination Conv4 are connected in series.As shown in Fig. 2, the step-length stride=4 of the first convolutional layer Conv1；It is specific next
The convolution kernel of the first convolutional layer Conv1 is said having a size of 7*7, the number of convolution kernel is 24；The stride=of first pond layer Pool1
2, convolution kernel size can be 3*3；The stride=2 of second convolutional layer Conv2, convolution kernel are having a size of 5*5, convolution kernel number
64；The stride=2 of second pond layer Pool2, convolution kernel is having a size of 3*3；The stride=1 of third convolutional layer Conv3, convolution
For core having a size of 1*1, convolution kernel number is 128；The stride=2 of Volume Four lamination Conv4, convolution kernel is having a size of 3*3, convolution kernel
Number is 256, it can be seen that total step-length of first convolutional layer, the first pond layer, the second convolutional layer and the second pond layer
Stride=4*2*2*2=32, therefore in convolution algorithm, 32 times of picture element matrix rapid drop of original input picture are compared
Current car plate detection network improves the speed for extracting vehicle license plate characteristic.
Optionally, the convolution kernel size in each convolutional layer in the initial deep learning network model can for 7*7,
5*5,3*3 or other values, above-mentioned convolution kernel size are to do for example, but not limiting the tool of convolution kernel in each convolutional layer
Body size.
Optionally, the first inception network structure, the 2nd inception network structure and the third
Inception network structure is all made of inception-v1Network structure, due to the first inception network structure, second
Inception network structure and the 3rd inception network structure are all made of inception-v1Network structure, therefore with first
The composition of inception network structure is illustrated for inception network structure.As shown in figure 3, this first
Inception network structure includes four branches, wherein the first branch only includes a convolutional layer, and the convolution of the convolutional layer
For core having a size of 1*1, convolution kernel number is 32；Second branch includes a pond layer and a convolutional layer, and the volume of the pond layer
Product core is having a size of 3*3, and for the convolution kernel of the convolutional layer having a size of 1*1, convolution kernel number is 32；Third branch includes two convolution
Layer, the convolution kernel of first convolutional layer is having a size of 1*1, and convolution nuclear volume is 24, and the convolution kernel of second convolutional layer is having a size of 3*
3, convolution kernel number is 32；4th branch includes three convolutional layers, and the convolution kernel of first convolutional layer is having a size of 1*1, convolution
Core number is 24, and the convolution kernel of second convolutional layer is having a size of 3*3, and the number of convolution kernel is 32, the convolution of third convolutional layer
For core having a size of 3*3, convolution kernel number is 32.It is understood that total step-length of this four branches is identical, i.e., each branch's output
Characteristic pattern dimension it is identical.For example, as the first branch convolutional layer stride=4, then the pond layer of the second branch with
The stride of convolutional layer is respectively 2 and 2, i.e., total step-length is 4, and the step-length of two convolutional layers of third branch may be alternatively provided as 2 Hes
2, three convolutional layers of the 4th branch can be set to 1,2 and 2, that is, guarantee that total step-length of each branch in four branches is 4.
Spy of the concat function for exporting four branches in further, in the first inception network structure
Sign figure carries out dimension splicing, in order to increase the width width of network model.
Optionally, the first context module network structure and the 2nd context module network structure are main
For multiple characteristic patterns of the convolutional layer output of front to be carried out full attended operation, a various dimensions feature vector is obtained, that is, is given birth to
At a characteristic pattern.
Optionally, the first context module network structure and the 2nd context module network structure are adopted
With the building region candidate network RPN of the convolutional network FAST R-CNN based on region (Region Proposal Network,
Referred to as：RPN strategy) constructs respective RPN to obtain respective anchor coordinate, due to the first context module
Network structure and the 2nd context module network structure, which are all made of in FAST R-CNN, constructs the strategy of RPN to construct
The RPN of first context module network structure and the 2nd context module network structure, therefore the first context
Module network structure and the 2nd context module network structure structure having the same, below with the first context
The structure of context module network structure is illustrated for module network structure.As shown in figure 4, this first
Context module network structure includes Liang Ge branch, wherein and first branch only includes the convolutional layer of a 3*3, and second
A branch includes the convolutional layer of a 3*3, and the convolutional layer of the 3*3 includes Kuo Liangge sub-branch again, and first sub-branch includes
The convolutional layer of one 3*3, second sub-branch include the convolutional layer of two concatenated 3*3.Such as above-mentioned inception network knot
Total step-length of structure, the first context module Liang Ge branch is identical, no longer describes herein.
Further, the anchor dimension scale of the first context module network structure is respectively 16*32,32*
64 and 64*128, wherein the anchor number of dimension scale 16*32,32*64 and 64*128 are respectively 16,4 and 1, this second
The anchor ratio of context module network structure is respectively 128*256 and 256*512, wherein dimension scale 128*
The anchor number of 256 and 256*512 is respectively 1 and 1, further, dimension scale 16*32,32*64,64*128,128*
The anchor separation of 256 and 256*512 is respectively 8,16,32,64 and 64.
Optionally, the first context module network structure and the 2nd context module network structure is complete
Articulamentum loss function is：
Wherein, LclsIt is the objective function of classification task, using Cross Entropy function, is mainly used for estimation range
It is divided into contexts region；LregIt is the loss function of Detection task, uses smooth L1 function；pi, tiIt is predicted value respectively
With prediction coordinate, pi *, ti *It is ground-truth label and actual coordinate respectively.
Optionally, the initial deep learning network model specific method of training includes：Training sample is obtained, by the training sample
Originally it is input to initial network model execution multilayer forward operation and obtains the prediction license plate set of coordinates of the training sample, according to preset
The anchor ratio-dependent prediction license plate set of coordinates corresponding multiple prediction license plate frame regions in the training sample, it is more to obtain this
Degree of overlapping IOU (Intersection over Union, the abbreviation of a prediction license plate frame region and actual license plate frame region：IOU)
Sample greater than preset threshold is positive sample, and the sample less than the preset threshold is negative sample, according to the positive sample and the negative sample
This determination exports result gradient, and the output result gradient execution reversed operation of multilayer is updated the initial deep learning network mould
The weight gradient of every layer of convolutional layer in type obtains this and presets trained network model.
Wherein, which is specifically as follows 0.35,0.5,0.7 or other values, is usually set to 0.35 or 0.5.
Optionally, if the picture element matrix of original input picture is single channel input data (the corresponding pixel of such as gray level image
Matrix or the corresponding picture element matrix of bianry image), for convolution operation, not due to convolution kernel size and the picture element matrix
Picture element matrix need to be cut into multiple basic granularities when executing convolution operation and carry out convolution with convolution kernel respectively by matching.Such as Fig. 5
Shown, Fig. 5 shows the picture element matrix that input data is 9*9, and convolution kernel is having a size of 3*3, when stride=2, generates 4*4 output
The convolution process of matrix.
Optionally, as picture element matrix be multichannel input data (such as rgb pixel matrix of color image), in every secondary volume
When product operation, the characteristic pattern that each channel obtains is overlapped and generates the final characteristic pattern of the secondary convolution operation, therefore characteristic pattern
Quantity it is related with the quantity of convolution kernel, it is unrelated with input channel number.It is rolled up as shown in fig. 6, Fig. 6 shows RGB triple channel
Product operation is to generate the detailed process of characteristic pattern, wherein three picture element matrixs input size of RGB is 9*9, convolution kernel size
3*3*3,3*3*3 refers to the convolution kernel for having a 3*3 on each channel here, without referring to three-dimensional data, therefore the volume in each channel
A 4*4 output matrix is generated after product core and respective picture element matrix convolution, three 4*4 output matrixes is superimposed to obtain final
4*4 characteristic pattern.
In the application, illustrated so that picture element matrix is single channel input data as an example.
Optionally, it which is inputted this presets trained network model and executes multilayer forward operation and obtain forward direction
Operation result determines license plate set of coordinates according to the forward operation result, specifically includes：The picture element matrix is inputted into the first volume
Lamination carries out convolution algorithm and exports 24 fisrt feature figures, which is input to the second pond layer and carries out pond
Change processing 24 the first ponds of output as a result, this 24 the first pond results, which are input to second convolutional layer, carries out convolution algorithm
64 second feature figures are exported, which is input to the second pond layer carries out pondization processing and export 64 the
Two ponds as a result, this 64 the second pond results are input to the first inception network structure, by this first
The convolution algorithm of inception network structure, the 2nd inception network structure and the 3rd inception network structure
128 third feature figures are exported afterwards, which is input to the third convolutional layer and carries out convolution algorithm output
128 fourth feature figures are inputted the first context module network structure respectively and rolled up by 128 fourth feature figures
Product operation obtains fifth feature figure, obtains N number of first anchor coordinate of N number of characteristic point in the fifth feature figure, according to this
Anchor dimension scale in one context module network structure determines N number of first anchor coordinate corresponding N number of
One license plate range extracts M the first license plate ranges that IOU in N number of first license plate range is greater than the preset threshold, determines the M
Corresponding M the first anchor coordinate of a first license plate range is the license plate set of coordinates, wherein N, the M are positive integer, M≤N.
Alternatively, 128 fourth feature figures are inputted Volume Four product 256 sixth feature figures of output, by 256 sixth feature figures
The 2nd context module network structure progress convolution algorithm is inputted respectively and obtains seventh feature figure, obtains the seventh feature
P the 2nd anchor coordinates of P characteristic point in figure, according to the anchor ruler in the 2nd context module network structure
It is big to extract IOU in the P the second license plate ranges for corresponding P the second license plate range of very little the 2nd anchor coordinate of ratio-dependent
In Q the second license plate ranges of the preset threshold, determine that corresponding Q the 2nd anchor coordinate of Q the second license plate range is to be somebody's turn to do
License plate set of coordinates, P, the Q are positive integer, P≤Q.
The first context network structure and the 2nd context network structure is arranged in size based on license plate, it can be seen that
The anchor ratio of 2nd context network structure is 2 times of the first context network structure, therefore the first context network structure
Mainly for small size license plate, the 2nd context network structure is mainly for large-sized license plate, therefore when identifying small license plate,
It obtains and predicts license plate range identification license plate content in the first context network structure, when identifying big license plate, obtain second
Predict that license plate range identifies license plate content in context network structure,
Step S103, the license plate range in the original input picture is determined according to the license plate set of coordinates, described in identification
License plate content in license plate range.
Based on the license plate set of coordinates in step S102, license plate range is determined according to anchor dimension scale, to license plate range
Interior license plate image carries out Character segmentation, identifies the corresponding license plate content of the license plate image.
As can be seen that three inception structures of having connected in the network model of the embodiment of the present application, pass through concat letter
Number increases the width of network model, and the network model is made to adapt to the license plate of various sizes, improves network model identification license plate
Precision, in addition, the actual ratio based on license plate constructs 1:2 anchor, improve identification license plate range efficiency and
Precision, and network model extraction vehicle license plate characteristic speed is fast, and memory is small.
It is consistent with above-mentioned Fig. 1 embodiment, referring to Fig. 7, Fig. 7 is provided by the embodiments of the present application a kind of based on depth
The possible functional unit of the electronic device 700 of habit forms block diagram, which includes：Acquiring unit 710, operation list
First 720, determination unit 730, wherein；
Acquiring unit 710 extracts the picture element matrix of the original input picture for obtaining original input picture；
Arithmetic element 720 executes multilayer forward direction fortune for trained network model to be preset in picture element matrix input
Calculation obtains forward operation as a result, determining license plate set of coordinates according to the forward operation result；
Determination unit 730 is known for determining the license plate range in the original input picture according to the license plate set of coordinates
License plate content in the not described license plate range.
The embodiment of the present application also provides a kind of computer storage medium, wherein computer storage medium storage is for electricity
The computer program of subdata exchange, it is as any in recorded in above method embodiment which execute computer
A kind of some or all of the licence plate recognition method based on deep learning step.
The embodiment of the present application also provides a kind of computer program product, and the computer program product includes storing calculating
The non-transient computer readable storage medium of machine program, the computer program are operable to that computer is made to execute such as above-mentioned side
Some or all of any licence plate recognition method based on deep learning recorded in method embodiment step.
It should be noted that for the various method embodiments described above, for simple description, therefore, it is stated as a series of
Combination of actions, but those skilled in the art should understand that, the application is not limited by the described action sequence because
According to the application, some steps may be performed in other sequences or simultaneously.Secondly, those skilled in the art should also know
It knows, embodiment described in this description belongs to alternative embodiment, related actions and modules not necessarily the application
It is necessary.
In the above-described embodiments, it all emphasizes particularly on different fields to the description of each embodiment, there is no the portion being described in detail in some embodiment
Point, reference can be made to the related descriptions of other embodiments.
In several embodiments provided herein, it should be understood that disclosed device, it can be by another way
It realizes.For example, the apparatus embodiments described above are merely exemplary, such as the division of the unit, it is only a kind of
Logical function partition, there may be another division manner in actual implementation, such as multiple units or components can combine or can
To be integrated into another system, or some features can be ignored or not executed.Another point, shown or discussed is mutual
Coupling, direct-coupling or communication connection can be through some interfaces, the indirect coupling or communication connection of device or unit,
It can be electrical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme
's.
It, can also be in addition, each functional unit in each embodiment of the application can integrate in one processing unit
It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list
Member both can take the form of hardware realization, can also be realized in the form of software program module.
If the integrated unit is realized in the form of software program module and sells or use as independent product
When, it can store in a computer-readable access to memory.Based on this understanding, the technical solution of the application substantially or
Person says that all or part of the part that contributes to existing technology or the technical solution can body in the form of software products
Reveal and, which is stored in a memory, including some instructions are used so that a computer equipment
(can be personal computer, server or network equipment etc.) executes all or part of each embodiment the method for the application
Step.And memory above-mentioned includes：USB flash disk, read-only memory (ROM, Read-Only Memory), random access memory
The various media that can store program code such as (RAM, Random Access Memory), mobile hard disk, magnetic or disk.
Those of ordinary skill in the art will appreciate that all or part of the steps in the various methods of above-described embodiment is can
It is completed with instructing relevant hardware by program, which can store in a computer-readable memory, memory
May include：Flash disk, read-only memory (English：Read-Only Memory, referred to as：ROM), random access device (English：
Random Access Memory, referred to as：RAM), disk or CD etc..
The embodiment of the present application is described in detail above, specific case used herein to the principle of the application and
Embodiment is expounded, the description of the example is only used to help understand the method for the present application and its core ideas；
At the same time, for those skilled in the art can in specific embodiments and applications according to the thought of the application
There is change place, in conclusion the contents of this specification should not be construed as limiting the present application.
Claims (10)
1. a kind of licence plate recognition method based on deep learning, the method is applied to electronic device, which is characterized in that the side
Method includes：
Original input picture is obtained, the picture element matrix of the original input picture is extracted；
Trained network model execution multilayer forward operation is preset into picture element matrix input and obtains forward operation as a result, root
License plate set of coordinates is determined according to the forward operation result；
The license plate range in the original input picture is determined according to the license plate set of coordinates, identifies the vehicle in the license plate range
Board content.
2. the method is also wrapped the method according to claim 1, wherein described before obtaining original image
It includes：
Training sample is obtained, the training sample is input to initial network model execution multilayer forward operation and obtains the training
The prediction license plate set of coordinates of sample, prediction license plate set of coordinates is in the training sample according to preset anchor ratio-dependent
In corresponding multiple prediction license plate frame regions, obtain it is the multiple prediction license plate frame region and actual license plate frame region degree of overlapping
The sample that IOU is greater than preset threshold is positive sample, and the sample less than preset threshold is negative sample, according to the positive sample and described
Negative sample determines output result gradient, and the output result gradient is executed the reversed operation of multilayer and updates the initial network model
In the weight gradient of every layer of convolutional layer obtain described presetting trained network model.
3. according to the method described in claim 2, it is characterized in that, the initial network model includes：First convolutional layer, first
Pond layer, the second convolutional layer, the second pond layer, the first inception network structure, the 2nd inception network structure, third
Inception network structure, third convolutional layer, Volume Four lamination, the first context module network structure and second
Context module network structure, first convolutional layer, first pond layer, second convolutional layer, described second
Pond layer, the first inception network structure, the 2nd inception network structure, the 3rd inception
Network structure, the third convolutional layer and the Volume Four lamination are sequentially connected in series, the first context module net
Network structure and the third convolutional layer are connected in series, the 2nd context module network structure and the Volume Four lamination
It is connected in series.
4. according to the method described in claim 3, it is characterized in that, the first inception network structure, described second
Inception network structure and the 3rd inception network structure are all made of inception-v1Network structure.
5. according to the method described in claim 2, it is characterized in that, first convolutional layer, first pond layer, described
Two convolutional layers, second pond layer, the third convolutional layer and the Volume Four lamination step-length be respectively 4,2,2,2,1 and
2, the convolution kernel number of first convolutional layer is 24, and the convolution kernel number of second convolutional layer is 64, the third convolution
The convolution kernel number of layer is 128, and the convolution kernel number of the Volume Four lamination is 256.
6. according to the method described in claim 5, it is characterized in that, the first context module network structure and described
2nd context module network structure is using building region candidate network in the convolutional network FAST R-CNN based on region
The respective RPN of the construction of strategy of RPN is to obtain respective anchor coordinate, and the first context module network knot
The anchor dimension scale of structure and the 2nd context module network structure is set as 1:2, and described first
Context module network structure is different with the anchor size of the 2nd context module network structure.
7. according to the method described in claim 6, it is characterized in that, the first context module network structure
Anchor dimension scale is respectively 16*32,32*64 and 64*128, wherein dimension scale 16*32,32*64 and 64*128's
Anchor number is respectively 16,4 and 1, and the anchor dimension scale of the 2nd context module network structure is respectively
128*256 and 256*512, wherein the anchor number that dimension scale is 128*256 and 256*512 is respectively 1 and 1, the ruler
Very little ratio is that the anchor separation of 16*32,32*64,64*128,128*256 and 256*512 are respectively 8,16,32,64 and
64。
8. method as claimed in claim 2, which is characterized in that described that trained network mould is preset in picture element matrix input
Type executes multilayer forward operation and obtains forward operation as a result, determining license plate set of coordinates according to the forward operation result, including：
The picture element matrix is inputted into first convolutional layer and exports 24 fisrt feature figures, 24 fisrt feature figures is defeated
Enter to second pond layer and exports 24 the first ponds as a result, 24 first pond results are input to the volume Two
Lamination exports 64 second feature figures, and 64 second feature figures are input to second pond layer and export 64 the second ponds
Change as a result, 64 second pond results are input to the first inception network structure, by described first
The convolution of inception network structure, the 2nd inception network structure and the 3rd inception network structure
128 third feature figures are exported after operation, and 128 third feature figures are input to the third convolutional layer and export 128
128 fourth feature figures are inputted the first context module network structure respectively and obtain by fourth feature figure
Five characteristic patterns obtain N number of first anchor coordinate of N number of characteristic point in the fifth feature figure, according to the first context
Anchor dimension scale in module network structure determines the corresponding N number of first license plate model of N number of first anchor coordinate
It encloses, extracts M the first license plate ranges that IOU in N number of first license plate range is greater than the preset threshold, determine the M
Corresponding M the first anchor coordinate of first license plate range is the license plate set of coordinates, and described N, M are positive integer, M≤N.
9. according to the method described in claim 8, it is characterized in that, the method also includes：
128 fourth feature figures are inputted into Volume Four product 256 sixth feature figures of output, by described 256 the 6th
Characteristic pattern inputs the 2nd context module network structure respectively and obtains seventh feature figure, obtains the seventh feature figure
P the 2nd anchor coordinates of middle P characteristic point, according to the anchor ratio in the 2nd context module network structure
Example determines corresponding P the second license plate range of P the 2nd anchor coordinate, extracts IOU in the P the second license plate ranges
Greater than Q the second license plate ranges of the preset threshold, determine that corresponding Q the 2nd anchor of the second license plate of Q range is sat
It is designated as the license plate set of coordinates, described P, Q are positive integer, P≤Q.
10. a kind of Car license recognition electronic device based on deep learning, which is characterized in that the electronic device includes：
Acquiring unit extracts the picture element matrix of the original input picture for obtaining original input picture；
Arithmetic element obtains just for trained network model execution multilayer forward operation to be preset in picture element matrix input
To operation result, license plate set of coordinates is determined according to the forward operation result；
Determination unit, for determining the license plate range in the original input picture according to the license plate set of coordinates, described in identification
License plate content in license plate range.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810593633.6A CN108921162A (en) | 2018-06-11 | 2018-06-11 | Licence plate recognition method and Related product based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810593633.6A CN108921162A (en) | 2018-06-11 | 2018-06-11 | Licence plate recognition method and Related product based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108921162A true CN108921162A (en) | 2018-11-30 |
Family
ID=64418043
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810593633.6A Pending CN108921162A (en) | 2018-06-11 | 2018-06-11 | Licence plate recognition method and Related product based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108921162A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TWI670610B (en) * | 2018-12-06 | 2019-09-01 | 中華電信股份有限公司 | Customizable method for detecting specific style objects |
| CN110378245A (en) * | 2019-06-26 | 2019-10-25 | 平安科技（深圳）有限公司 | Football match Activity recognition method, apparatus and terminal device based on deep learning |
| CN111028160A (en) * | 2019-11-21 | 2020-04-17 | 西北工业大学 | Remote sensing image noise suppression method based on convolutional neural network |
| CN111460996A (en) * | 2020-03-31 | 2020-07-28 | 东南大学 | Truck number plate identification method based on extreme learning deep network fusion model |
| CN113139579A (en) * | 2021-03-23 | 2021-07-20 | 广东省科学院智能制造研究所 | Image classification method and system based on image feature adaptive convolution network |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107679531A (en) * | 2017-06-23 | 2018-02-09 | 平安科技（深圳）有限公司 | Licence plate recognition method, device, equipment and storage medium based on deep learning |
-
2018
- 2018-06-11 CN CN201810593633.6A patent/CN108921162A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107679531A (en) * | 2017-06-23 | 2018-02-09 | 平安科技（深圳）有限公司 | Licence plate recognition method, device, equipment and storage medium based on deep learning |
Non-Patent Citations (4)
| Title |
|---|
| CHRISTIAN SZEGEDY ET AL.: "Rethinking the Inception Architecture for Computer Vision", 《ARXIV》 * |
| CHRISTIAN SZEGEDY: "Going deeper with convolutions", 《2015 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| SHAOQING REN: "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| 赵振兴: "基于深度学习的车牌识别技术研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TWI670610B (en) * | 2018-12-06 | 2019-09-01 | 中華電信股份有限公司 | Customizable method for detecting specific style objects |
| CN110378245A (en) * | 2019-06-26 | 2019-10-25 | 平安科技（深圳）有限公司 | Football match Activity recognition method, apparatus and terminal device based on deep learning |
| CN110378245B (en) * | 2019-06-26 | 2023-07-21 | 平安科技（深圳）有限公司 | Football match behavior recognition method and device based on deep learning and terminal equipment |
| CN111028160A (en) * | 2019-11-21 | 2020-04-17 | 西北工业大学 | Remote sensing image noise suppression method based on convolutional neural network |
| CN111460996A (en) * | 2020-03-31 | 2020-07-28 | 东南大学 | Truck number plate identification method based on extreme learning deep network fusion model |
| CN111460996B (en) * | 2020-03-31 | 2022-02-15 | 东南大学 | Truck number plate identification method based on extreme learning deep network fusion model |
| CN113139579A (en) * | 2021-03-23 | 2021-07-20 | 广东省科学院智能制造研究所 | Image classification method and system based on image feature adaptive convolution network |
| CN113139579B (en) * | 2021-03-23 | 2024-02-02 | 广东省科学院智能制造研究所 | Image classification method and system based on image feature self-adaptive convolution network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108921162A (en) | Licence plate recognition method and Related product based on deep learning | |
| CN109522874B (en) | Human body action recognition method and device, terminal equipment and storage medium | |
| CN113822209B (en) | Hyperspectral image recognition method and device, electronic equipment and readable storage medium | |
| CN108830211A (en) | Face identification method and Related product based on deep learning | |
| CN105930402A (en) | Convolutional neural network based video retrieval method and system | |
| CN111626184B (en) | Crowd density estimation method and system | |
| CN112163601B (en) | Image classification method, system, computer device and storage medium | |
| CN108846404B (en) | Image significance detection method and device based on related constraint graph sorting | |
| CN110807379B (en) | Semantic recognition method, semantic recognition device and computer storage medium | |
| CN108875826B (en) | Multi-branch object detection method based on coarse and fine granularity composite convolution | |
| CN104699781B (en) | SAR image search method based on double-deck anchor figure hash | |
| CN107506350A (en) | A kind of method and apparatus of identification information | |
| CN114783021A (en) | Intelligent detection method, device, equipment and medium for wearing of mask | |
| CN109146891A (en) | A kind of hippocampus dividing method, device and electronic equipment applied to MRI | |
| CN111709415B (en) | Target detection method, device, computer equipment and storage medium | |
| CN114358252A (en) | Operation execution method and device in target neural network model and storage medium | |
| CN106503386A (en) | The good and bad method and device of assessment luminous power prediction algorithm performance | |
| CN117884379A (en) | Ore sorting method and system | |
| CN117951308A (en) | Zero sample knowledge graph completion method and device | |
| CN111027670B (en) | Feature map processing method and device, electronic equipment and storage medium | |
| CN110659631A (en) | License plate recognition method and terminal equipment | |
| CN113239693A (en) | Method, device and equipment for training intention recognition model and storage medium | |
| CN116227573B (en) | Segmentation model training method, image segmentation device and related media | |
| CN114494782B (en) | Image processing method, model training method, related device and electronic equipment | |
| CN113139617B (en) | Power transmission line autonomous positioning method and device and terminal equipment |
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 | ||
| CB02 | Change of applicant information |
Address after: Room 1301, No.132 Fengqi Road, phase III, software park, Xiamen City, Fujian Province Applicant after: Xiamen Entropy Technology Co., Ltd Address before: 361000, Xiamen three software park, Fujian Province, 8 North Street, room 2001 Applicant before: XIAMEN ZKTECO BIOMETRIC IDENTIFICATION TECHNOLOGY Co.,Ltd. |
|
| CB02 | Change of applicant information | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181130 |
|
| RJ01 | Rejection of invention patent application after publication |