Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, completeSite preparation description, it is clear that described embodiments are some of the embodiments of the present invention, instead of all the embodiments.Based on this hairEmbodiment in bright, every other implementation obtained by those of ordinary skill in the art without making creative effortsExample, shall fall within the protection scope of the present invention.
Referring to Fig. 1, Fig. 1 shows the application environment of licence plate recognition method provided in an embodiment of the present invention.The Car license recognitionMethod is applied to be carried out in Car license recognition scene in the license plate for the vehicle taken.The identification scene includes server-side and clientEnd, wherein be attached between server-side and client by network, client will take license plate image and be sent to serviceEnd, server-side receive the license plate image that client is sent and are identified, client specifically can be, but not limited to be hypervelocity camera shootingIt head, the monitoring of day net, electronic police, various personal computers, laptop, smart phone, tablet computer and portable wearsEquipment is worn, server-side can specifically be realized with the server cluster that independent server or multiple servers form.
Referring to Fig. 2, being applied in this way Fig. 2 shows a kind of licence plate recognition method provided in an embodiment of the present invention in Fig. 1In server-side for be illustrated, details are as follows:
S10: initial license plate image is obtained.
Specifically, after client obtains initial license plate image, initial license plate image is sent to by the network transmission protocolServer-side, server-side receive the initial license plate image by the network transmission protocol.
Wherein, client specifically can be the monitoring with shooting function such as hypervelocity camera, the monitoring of day net, electronic policeEquipment, directly shooting obtain initial license plate image, are also possible to various personal computers, laptop, smart phone or flatThe intelligent terminals such as plate computer, the memory space with storing initial license plate image, and network interaction is carried out with server-sideFunction.
Wherein, the network transmission protocol includes but is not limited to: Internet Control Message agreement (Internet ControlMessage Protocol, ICMP), address resolution protocol (ARP Address Resolution Protocol, ARP) and textPart transport protocol (File Transfer Protocol, FTP) etc..
S20: pre-processing initial license plate image, obtains target license plate image.
Specifically, due to the angle of shooting, distance and automobile the factors such as run at high speed influence, get justBeginning license plate image is of low quality, and directly progress recognition correct rate is lower, it is then desired to first be located initial license plate image in advanceReason influences to reduce these factor brings, improves the accuracy rate of subsequent identification, pre-process to initial license plate imageAfterwards, the target license plate image that size is a × b pixel is obtained.
Wherein, pretreatment includes but is not limited to: image cropping, normalization and slant correction etc..
Wherein, a and b is preset numerical value, and basic unit is the width of 1 pixel, using a as horizontal direction side length, i.e.,Using the width of a pixel as horizontal direction side length, using b as vertical direction side length, i.e., using the width of b pixel as vertical sideTo side length, a and b are positive integer, and specific value, which can according to need, to be configured, and is not specifically limited herein.
Preferably, in embodiments of the present invention, the value of a takes the value of 140, b to take 28 to get to 140 × 28 pixel sizesTarget license plate image.
S30: target license plate image is input in preset convolutional neural networks model, wherein preset convolutional NeuralNetwork model includes input layer, convolutional layer, pond layer and full articulamentum.
Specifically, the target license plate image of a × b pixel got in step S20 is input to preset convolution mindThrough being identified, being identified to target license plate image with will pass through preset convolutional neural networks model in network modelAs a result.
Wherein, convolutional neural networks model (Convolutional Neural Network, CNN) is a kind of feedforward mindThrough network, its artificial neuron can respond the surrounding cells in a part of coverage area, can rapidly and efficiently carry out imageProcessing, preset convolutional neural networks model includes input layer, convolutional layer, pond layer and full articulamentum.
Wherein, pond layer be used for after convolutional layer convolution data carry out one-dimensional compression processing, obtain one-dimensional toMeasure feature classifies to the vector characteristics so as to subsequent.
Wherein, the full articulamentum in the present embodiment is used for that treated that vector characteristics are classified to pond layer compression, obtainsTo multiple specification areas, meanwhile, full articulamentum includes multiple preset classifiers, and each classifier carries out a specification areaIdentification identifies that multiple identification operations can execute parallel, is conducive to improve recognition efficiency.
Further, the method that classifier realizes classification includes but is not limited to: logistic regression (LogisticRegression, LR), support vector machines ((Support Vector Machine, SVM), cross entropy (Corss Entropy)With softmax return etc..
It is worth noting that different classifications device in the present embodiment, can identify that specification area is different according to it, be configuredDifferent classification, the set-up mode of this classifier greatly reduce the range of classification, recognition speed not only can be improved,The accuracy rate of identification is improved to a certain extent.
For example, in a specific embodiment, the second specification area corresponds to second character of license plate, and the second of license plateA word is any one in 26 capitalization English letters, thus, the corresponding classifier of second specification area may be configured as wrappingInclude 26 classification, the corresponding capitalization English letter of each classification.
S40: the multi-channel data in target license plate image is extracted by input layer, and multi-channel data is passed into convolutionLayer.
Specifically, by the input layer in preset convolutional neural networks, each image in target license plate image is extractedMulti-channel data, and multi-channel data is passed into convolutional layer.
Wherein, multi-channel data refers to the data in each channel, and channel number can be configured according to the actual situation, hereinIt is not specifically limited, it is preferable that channel number of the embodiment of the present invention is set as 3.
S50: process of convolution is carried out to multi-channel data in convolutional layer, the convolved data after obtaining convolution.
Specifically, it obtains multi-channel data by carrying out process of convolution to multi-channel data in convolutional layer and carries out at convolutionConvolved data after reason, to subsequent feature extraction.
Wherein, convolutional layer (Convolutional layer) is made of several convolution units, the parameter of each convolution unitIt is all to be optimized by back-propagation algorithm.The purpose of process of convolution is to obtain the convolved data for indicating different characteristic,The different characteristic for facilitating subsequent extracted to input, first layer convolutional layer may can only extract some rudimentary features such as edge, lineThe levels such as item and angle, the more network of deep layer grade can from low-level features the more complicated feature of iterative extraction.
It is worth noting that in embodiments of the present invention, there are the convolutional layer of the default number of plies, specific preset quantity can rootIt is determined according to actual conditions, the preset convolutional layer of the embodiment of the present invention is 3 layers as a preferred method, i.e. the present invention is realApplying the convolutional neural networks model in example includes 3 convolutional layers, respectively the first convolutional layer, the second convolutional layer and third convolutionLayer, the convolution kernel size of this 3 convolutional layers is 3 × 3, and the step-length of the first convolutional layer is (2,2), the second convolutional layer and third volumeThe step-length of lamination is (1,1).
Wherein, convolution kernel is to give input picture in image procossing, make to pixel in each zonule of input pictureIt is weighted and averaged with a weight, so that each pixel is pixel in a zonule in input picture in the output imageWeighted average, wherein weight is defined by a function, this function is referred to as convolution kernel.
Wherein, step-length refers to the distance that convolution kernel moves every time, for example, step-length is that (2,2) refer to that each convolution kernel carries outWhen mobile, the distance of two pixels is moved in the horizontal direction, in the distance of mobile two pixels of vertical direction.
S60: feature extraction is carried out to convolved data, obtains characteristic.
Specifically, after progress process of convolution obtains convolved data in step s 50, then feature is carried out to the convolved data and is mentionedIt takes, retains the important feature of needs, abandon inessential information, to obtain the feature that can be used for subsequent behavior predictionData.
Wherein, in embodiments of the present invention, feature extraction is realized by pond layer, pond layer immediately convolutional layer itAfterwards, the amount for compressed data and parameter, so that the information unrelated to behavior prediction and duplicate information are removed, meanwhile, pondOver-fitting can also be reduced by changing layer, be conducive to improve accuracy of identification.
S70: one-dimensional compression processing is carried out to characteristic using pond layer, obtains target figure layer.
Specifically, characteristic is carried out by one-dimensional compression processing by pond layer, obtains target figure layer, this compression sideFormula also is understood as a multi-dimensional matrix drop into one-dimensional matrix.
It is readily appreciated that ground, in such a way that one-dimensional compresses, converts all characteristics to unidirectional one-dimensionalThe target figure layer that vector is constituted is conducive to subsequent divided so that the data in target figure layer can be divided into multiple regionsClass identification.
S80: in full articulamentum according to preset tag along sort, Classification and Identification is carried out to target figure layer, obtains Car license recognitionAs a result.
Specifically, step S70 by feature data compression at the target figure layer of one-dimensional after, in full articulamentum by presetTag along sort classifies to the target figure layer, i.e., is divided into the target figure layer of the one-dimensional according to preset tag along sort moreA region, the data characteristics of the corresponding characters on license plate in each region, and then Recognition of License Plate Characters is carried out to each region, thusObtain the recognition result of the characters on license plate in each region, by the recognition result of the characters on license plate in each region merge to getTo license plate recognition result.
Wherein, tag along sort refers to the classification model that region division is carried out for the target figure layer to one-dimensional, specifically may be usedIt is configured according to actual needs, herein without limitation.
In the present embodiment, by obtaining initial license plate image, and initial license plate image is pre-processed, obtains targetLicense plate image, and then target license plate image is input in preset convolutional neural networks model and is identified, obtain license plate knowledgeNot as a result, the preset convolutional neural networks model includes: input layer, convolutional layer, pond layer and full articulamentum, pass through input layerThe multi-channel data in target license plate image is extracted, and multi-channel data is passed into convolutional layer, convolutional layer is to multi-channel dataCarry out process of convolution, obtain convolved data, and then extract the characteristic of convolved data, and using pond layer to characteristic intoThe compression of row one-dimensional obtains target figure layer, and carries out the synchronous identification in subregion to target figure layer using full articulamentum, obtains license plateRecognition result, it is this that the one-dimensional data characteristics of target license plate image is divided into multiple classification using convolutional neural networks, andKnowledge method for distinguishing is carried out to each classification simultaneously, improves the efficiency of Car license recognition, and is had to the license plate of different size logicalThe property used.
In one embodiment, pond layer is Flatten layers, for the characteristic to be carried out one-dimensional compression processing.
Wherein, Flatten layers are one-dimensional pond layer, and Flatten layers between convolutional layer and full articulamentum, are used to inputThe data of one-dimensional that is, the input data one-dimensional of multidimensional, and then are input to full articulamentum and identified by data " pressing ".
Specifically, by Flatten layers input be w × h × c characteristic readjust at another size be (w× h × c) × 1 × 1 vector, i.e., characteristic is subjected to one-dimensional compression processing, obtains target figure layer.
Wherein, w refers to that the width of characteristic, h refer to the height of characteristic, and c refers to the number of channels of characteristic.
Wherein, readjustment can be realized by reshape function, the function can quickly readjust matrix line number,Columns, dimension.
For example, in a specific embodiment, the characteristic for 8 × 1 × 64 is inputted, after Flatten layer compression,Change into the one-dimensional vector for 512 × 1 × 1.
It is worth noting that the speed of Flatten layers of progress one-dimensional compression faster, makes relative to common pond layerThe time for obtaining one-dimensional compression is shorter.
In the present embodiment, it is realized by using Flatten layers and quickly the one-dimensional of characteristic is compressed, improve oneThe efficiency of dimensionization compression.
On the basis of the corresponding embodiment of Fig. 1, below by a specific embodiment come to being mentioned in step S0And in full articulamentum according to preset tag along sort, Classification and Identification is carried out to target figure layer, obtains the tool of license plate recognition resultBody implementation method is described in detail.
Referring to Fig. 3, Fig. 3 shows the specific implementation flow of step S80 provided in an embodiment of the present invention, details are as follows:
S81: according to preset division template, target figure layer is in turn divided into n specification area from left to right, wherein nFor preset positive integer.
Specifically, it is preset with division template in server-side, which is used for according to preset region division mode, willTarget image is in turn divided into n specification area from left to right, wherein the quantity n of specification area is preset in division templatePositive integer.
Preferably, n is set as 7 in the present embodiment, i.e., target image is divided into 7 specification areas from left to right.
S82: the corresponding tag along sort mark of each specification area is obtained, wherein each tag along sort mark is one correspondingDefault classifier, total n default classifiers.
Specifically, according to preset division template, it may be determined that the tag along sort of each specification area identifies.
Wherein, for identifying a specification area, each tag along sort is identified with uniquely to be corresponding to it tag along sort markClassifier, the data characteristics of the classifier specification area of tag along sort Identification for identification.
For example, in a specific embodiment, tag along sort mark includes: tag along sort 1, tag along sort 2, tag along sort3, tag along sort 4, tag along sort 5, tag along sort 6 and tag along sort 7, wherein tag along sort 1 corresponds to the 1st word of license plate numberAccord with corresponding region A, the corresponding region B of the 2nd character of the corresponding license plate number of tag along sort 2, the corresponding license plate number of tag along sort 3The corresponding region C of the 3rd character of code, the corresponding region D of the 4th character of the corresponding license plate number of tag along sort 4, tag along sortThe corresponding region E of the 5th character of 5 corresponding license plate numbers, the 6th corresponding area of character of the corresponding license plate number of tag along sort 6Domain F, the corresponding region G of the 7th character of the corresponding license plate number of tag along sort 7.
S83: identifying corresponding default classifier using each tag along sort, identifies corresponding specification area to tag along sortIt is identified, obtains n region recognition result.
Specifically, for each specification area, corresponding default point is identified using the tag along sort for identifying the specification areaClass device identifies the specification area, obtains n region recognition result.
In embodiments of the present invention, the corresponding classification of each classifier is a variety of, of the invention as a preferred method,Embodiment returns to realize identification of the classifier to multiple classification using softmax.
It is to be appreciated that the content that different classifications device needs to identify is different, thus, it as a preferred method, can rootClassifier is further limited according to identification content and needs the classification for including, so that recognition speed is faster.
It is said so that the first tag along sort for identifying first specification area identifies corresponding first classifier as an example belowBright, in civilian license plate number, first characters on license plate is Chinese character abbreviation, this referred to as may be any one in 37 Chinese charactersIt is a, this 37 Chinese characters be " capital, saliva, Shanghai, Chongqing, Ji, Henan, cloud, the Liao Dynasty, black, Hunan, Anhui, Fujian, Shandong, new, Soviet Union, Zhejiang, Jiangxi, Hubei Province, osmanthus, it is sweet,Shanxi, illiteracy, Shan, Ji, expensive, Guangdong, blueness, hiding, river, it is peaceful, fine jade, port, Australia, make, lead, learn and warn ", thus, can will be in the first classifierClassification and Identification is set as this 37 Chinese characters, so that the content of Classification and Identification greatly reduces, is conducive to rapidly and accurately to firstSpecification area is identified.
It is worth noting that since each classifier only identifies corresponding specification area, so that this nIt is simultaneously and concurrently to execute that classifier, which carries out Classification and Identification, greatly accelerates recognition speed, improves the efficiency of identification.
S84: according to the sequence of specification area, n region recognition result is combined, license plate recognition result is obtained.
Specifically, according to the sequence from left to right of specification area in step S351, successively by the corresponding area of specification areaDomain recognition result is combined, and obtains license plate recognition result.
For example, in a specific embodiment, according to sequence from left to right, the corresponding region recognition result of specification areaRespectively " Shanghai ", " A ", " 3 ", " 7 ", " F ", " 6 " and " 6 " successively combines these region recognition results, obtains license plate knowledgeOther result is " Shanghai A37F66 ".
In the present embodiment, according to preset division template, target figure layer is in turn divided into n classification area from left to rightDomain, and the corresponding tag along sort mark of each specification area is obtained, and then identify corresponding default point using each tag along sortClass device identifies corresponding specification area to tag along sort and identifies, obtains n region recognition as a result, according still further to specification areaSequence, n region recognition result is combined, license plate recognition result is obtained, it is this to be divided into multiple specification areas, respectivelyKnown otherwise using corresponding classifier, accelerate recognition speed, improves the efficiency of identification.
On the basis of the corresponding embodiment of Fig. 2, below by a specific embodiment come to being mentioned in step S20And initial license plate image is pre-processed, the concrete methods of realizing for obtaining target license plate image is described in detail.
Referring to Fig. 4, Fig. 4 shows the specific implementation flow of step S20 provided in an embodiment of the present invention, details are as follows:
S21: by edge detection algorithm, the lower boundary of the coboundary of license plate and license plate in initial license plate image is obtained.
Specifically, since the factors such as the angle and distance of shooting influence, the license plate image taken is generally in addition to license plate areaExcept domain, it will also include some non-license plate images outside license plate area, these non-license plate images can make subsequent Car license recognitionIt thus in order to improve the accuracy rate of subsequent Car license recognition, needs to find out initial license plate figure by edge inspection algorithms at interferenceThe coboundary of license plate and the lower boundary of license plate as in, so that it is determined that the license plate range in initial license plate image.
Wherein, edge detection algorithm is the algorithm for carrying out edge detection, and edge detection is image procossing and computerBasic problem in vision, in embodiments of the present invention, the purpose of edge detection are that brightness change is obvious in mark license plate imagePoint, i.e. the point on license plate boundary, the significant changes in image attributes usually reflect the critical event and variation of attribute, these figuresAs attribute includes but is not limited to: discontinuous, surface direction in depth is discontinuous, material property variation and scene lighting variationDeng.
Wherein, the edge of image refers to that region jumpy occurs for gray scale in image, and the situation of change of ganmma controller canTo be reflected with the gradient of intensity profile.
Common edge detection algorithm includes but is not limited to: Sobel Operator (Sobel operator) edge detection is calculatedMethod, Gauss-Laplace edge detection algorithm, Luo Baici overlapping edges detection (Roberts Cross operator) are calculatedMethod and Canny multistage edge detection algorithm etc..
Preferably, the edge detection algorithm used by inventive embodiments is Canny multistage edge detection algorithm.
It is worth noting that the coboundary of obtained license plate and the lower boundary of license plate are two sections according to edge detection algorithmLine segment.That is, the coboundary for obtaining license plate includes two vertex in left and right, the lower boundary of obtained license plate includes two vertex in left and right.
S22: according to the lower boundary of the coboundary of license plate and license plate, the range image of license plate is determined.
Specifically, the lower boundary of the coboundary of license plate obtained in step S21 and license plate is connected, that is, by license plateThe left side vertex of coboundary be connected with the left side vertex of the lower boundary of license plate, by the right vertex of the coboundary of license plate and license plateLower boundary the right vertex be connected, so that a quadrangle is obtained, using the image in this square range as license plateRange image.
S23: slant correction is carried out to range image using Radon transform, the base image after being corrected.
Specifically, due to the influence of the angle and distance of shooting, the initial license plate image got can have inclination, in turnThere is also inclinations for obtained range image, in order to improve the accuracy rate of subsequent Car license recognition, by Radon transform to range imageSlant correction is carried out, thus the base image after being corrected.
Wherein, Radon transform (Radon transform) is that one kind is superimposed by determining direction projection, finds maximal projection valueWhen angle, so that it is determined that image inclination angle, and then be corrected, the method for the image after being corrected.
S24: centered on the center of gravity of base image, cutting base image, obtains target license plate image.
Specifically, right using b as vertical direction side length using a as horizontal direction side length centered on the center of gravity of base imageThe base image that step S23 is obtained is cut, and the target license plate image of the rectangle of an a × b pixel size is obtained,In, a and b are positive integer, and value can be preset according to actual needs.
Preferably, the preset a of the embodiment of the present invention is 140, b 28, that is, the target of 140 × 28 pixels finally obtainedLicense plate image.
For example, in a specific embodiment, the coordinate of the center of gravity of base image is (82,21), and horizontal direction side length is140, vertical direction side length is 28, then obtains top left corner apex coordinate (12,7), upper right corner apex coordinate (152,7), the lower left cornerApex coordinate (12,35), lower right corner apex coordinate (152,35), by the rectangle for 140 × 28 pixels that this four vertex form,Along this clipping rectangle base image, the image in the rectangular extent, i.e. target license plate image are obtained.
Preferably, after obtaining target license plate image, the embodiment of the present invention also carries out mean value to target license plate imageAnd normalization, eliminate the difference in target license plate image between different dimensions image data.
Wherein, normalization refers to the characteristics of image amplitude normalization in target license plate image to same range, even ifThe standard deviation that all characteristics of image are removed with each characteristics of image, using obtained result as the image after characteristics of image normalizationFeature.
Wherein, it goes mean value to refer to and each dimension of characteristics of image all centers in target license plate image is turned to 0, i.e., by targetThe central point of license plate image is withdrawn on coordinate origin.
In the present embodiment, by edge detection algorithm, obtain in initial license plate image under the coboundary and license plate of license plateBoundary, and then according to the lower boundary of the coboundary of license plate and license plate, it determines the range image of license plate, reuses Radon transform to modelIt encloses image and carries out slant correction, the base image after being corrected, then centered on the center of gravity of base image, to base imageIt is cut, obtains target license plate image, when so that subsequent use base image being identified, avoided outside due to license plate rangeImage interference and Tilt factor caused by identification error, enhance the quality of target license plate image, be conducive to improve subsequentThe accuracy rate of Car license recognition.
On the basis of the corresponding embodiment of Fig. 4, below by a specific embodiment come to being mentioned in step S21And by edge detection algorithm, obtain the specific implementation side of the lower boundary of the coboundary and license plate of license plate in initial license plate imageMethod is described in detail.
Referring to Fig. 5, Fig. 5 shows the specific implementation flow of step S21 provided in an embodiment of the present invention, details are as follows:
S211: noise removal is carried out to initial license plate image by Gaussian Blur, obtains denoising license plate image.
Specifically, the edge of license plate image is high-frequency signal, but the noise of image also focuses on high-frequency signal, it is easy to quiltIt is mistakenly identified as edge, it is then desired to remove the noise of image, avoids the noise of image to determining edge bring interference.At thisIn inventive embodiments, noise removal is carried out to initial license plate image using Gaussian Blur, obtains denoising license plate image.
Wherein, the edge of license plate image refers to that gray scale occurs sharply in the image of license plate area and non-license plate area intersectionThe region of variation.The situation of change of ganmma controller can be reflected with the gradient of intensity profile.
Wherein, the noise of the noise of image namely image, refer to be present in it is unnecessary or extra in image dataInterference information, the presence of noise have seriously affected the quality of image, therefore before image enhancement processing and classification processing, it is necessary toCorrected, in embodiments of the present invention, the correcting method used for use Gaussian Blur remove noise.
Wherein, Gaussian Blur (Gaussian Blur), is also Gaussian smoothing, it is a kind of image fuzzy filter, it is usedNormal distribution calculates the transformation of each pixel in image, and picture noise is usually reduced with it and reduces level of detail.
It is worth noting that image border and noise are high-frequency signal, thus the radius selection of Gaussian Blur is critically important,Excessive radius is easy to allow some weak endpoint detections less than the specific setting of radius can be adjusted according to the actual situationIt is whole, herein with no restriction.
S212: the gradient value and vertical direction of the horizontal direction of denoising license plate image are calculated using preset gradient operatorGradient value obtains Initial Gradient value set.
Specifically, the edge of image can be pointed in different directions, and by using preset operator, calculate denoising license plate imageGradient value horizontally and vertically, obtain Initial Gradient value set.
Wherein, digital picture is exactly discrete point value spectrum, can also be two-dimensional discrete function, the gradient of image is exactly thisThe derivation of two-dimensional discrete function as a result, gradient operator, that is, the method for being used to calculate gradient.
Wherein, preset gradient operator includes but is not limited to: Sobel Operator (Sobel operator), Prewitt are calculatedSon, Luo Baici operator (Roberts operator) and Canny operator.
Preferably, gradient operator used in the embodiment of the present invention is Canny operator.
S213: edge thinning processing is carried out to Initial Gradient value set by the way of the inhibition of non-maximum value, obtains widthFor the gradient edge of single pixel.
Specifically, edge thinning processing is carried out to Initial Gradient value set by the way of the inhibition of non-maximum value, obtains widthDegree is the gradient edge of single pixel.
Wherein, non-maximum value inhibit (Non Maximum Suppression, NMS) be inhibit be not maximum element,It can be regarded as part and carry out maximum value search, to help to retain local maxima gradient and inhibit every other gradient value, this meaningTaste only remain position most sharp keen in change of gradient.
For example, in a specific embodiment, in vertical direction, thering is the gradient value of 4 pixel wides to constitute an officePortion searches out the maximum picture of gradient value in the gradient value of this part in lid part by the way of the inhibition of non-maximum valueVegetarian refreshments, as gradient edge, to realize edge thinning.
S214: using the weak marginal point in preset dual threshold filter gradient edge, the strong edge in gradient edge is obtainedPoint.
Specifically, edge pixel is distinguished by the way that dual threshold, i.e. a high threshold values and a low valve valve is arranged.If edgePixel gradient value is greater than high threshold values, then is considered as strong edge point, if edge gradient value is less than high threshold values, is greater than low valveValue is then labeled as weak marginal point, and the point less than low valve valve is then suppressed, and strong edge point can be confirmed to be genuine edge, butWeak marginal point then may be genuine edge, it is also possible to accurate as a result, in this hair to obtain caused by noise or color changeIn bright embodiment, weak marginal point is also suppressed.
S215: according to strong edge point, the coboundary of license plate and the lower boundary of license plate are determined.
Specifically, it is obtained under the coboundary and license plate of license plate according to strong edge point by image procossings such as corrosion extensionsBoundary.
Wherein, corrosion can eliminate independent point in strong edge point, and extension can be by adjacent but disjunct strong edge pointIt connects.
In the present embodiment, noise removal is carried out to initial license plate image by Gaussian Blur, obtains denoising license plate image, andThe gradient value horizontally and vertically that denoising license plate image is calculated using preset gradient operator, obtains Initial GradientValue set, and then edge thinning is carried out to Initial Gradient value set by the way of the inhibition of non-maximum value, it is wide to obtain a pixelGradient edge obtain the strong edge point in gradient edge using the weak marginal point in preset dual threshold filter gradient edge,Then according to strong edge point, the coboundary of license plate and the lower boundary of license plate are determined, the accuracy rate of license plate edge detection is improved, hasConducive to the subsequent determination to license plate range.
It should be understood that the size of the serial number of each step is not meant that the order of the execution order in above-described embodiment, each processExecution sequence should be determined by its function and internal logic, the implementation process without coping with the embodiment of the present invention constitutes any limitIt is fixed.
Fig. 6 shows the functional block diagram with the one-to-one license plate recognition device of above-described embodiment licence plate recognition method.Such as Fig. 6Shown, which includes data acquisition module 10, preprocessing module 20, input module 30, transfer module 40, convolutionModule 50, extraction module 60, compression module 70 and identification module 80.Detailed description are as follows for each functional module:
Module 10 is obtained, for obtaining initial license plate image；
Preprocessing module 20 obtains target license plate image for pre-processing to initial license plate image；
Input module 30, for target license plate image to be input in preset convolutional neural networks model, wherein defaultConvolutional neural networks model include input layer, convolutional layer, pond layer and full articulamentum；
Transfer module 40, for extracting the multi-channel data in target license plate image by input layer, and by multichannel numberAccording to passing to convolutional layer；
Convolution module 50, for carrying out process of convolution to multi-channel data in convolutional layer, the convolved data after obtaining convolution；
Extraction module 60 obtains characteristic for carrying out feature extraction to convolved data；
Compression module 70 obtains target figure layer for carrying out one-dimensional compression processing to characteristic using pond layer；
Identification module 80, for, according to preset tag along sort, carrying out Classification and Identification in full articulamentum to target figure layer, obtainingTo license plate recognition result.
Further, identification module 80 includes:
Region division subelement, for according to preset division template, target figure layer to be in turn divided into n from left to rightSpecification area, wherein n is preset positive integer；
Mark obtains subelement, for obtaining the corresponding tag along sort mark of each specification area, wherein each contingency tableLabel identify a corresponding default classifier, total n default classifiers；
Region recognition subelement, for identifying corresponding default classifier using each tag along sort, to tag along sort markKnow corresponding specification area to be identified, obtains n region recognition result；
As a result subelement is combined, for the sequence according to specification area, n region recognition result is combined, is obtainedLicense plate recognition result.
Further, preprocessing module 20 includes:
Detection unit, for by edge detection algorithm, obtaining the coboundary of license plate and license plate in initial license plate imageLower boundary；
Determination unit, for determining the range image of license plate according to the coboundary of license plate and the lower boundary of license plate；
Unit is corrected, for carrying out slant correction to range image using Radon transform, the base image after being corrected；
It cuts unit and obtains target license plate figure for being cut centered on the center of gravity of base image to base imagePicture.
Further, detection unit includes:
Subelement is denoised, for carrying out noise removal to initial license plate image by Gaussian Blur, obtains denoising license plate figurePicture；
Computation subunit, for use preset gradient operator calculate denoising license plate image horizontal direction gradient value andThe gradient value of vertical direction obtains Initial Gradient value set；
Subelement is refined, for carrying out at edge thinning by the way of the inhibition of non-maximum value to Initial Gradient value setReason obtains the gradient edge that width is single pixel；
It filters subelement and obtains gradient edge for using the weak marginal point in preset dual threshold filter gradient edgeIn strong edge point；
Subelement is delimited, for determining the coboundary of license plate and the lower boundary of license plate according to strong edge point.
Specific about license plate recognition device limits the restriction that may refer to above for licence plate recognition method, herein notIt repeats again.Modules in above-mentioned license plate recognition device can be realized fully or partially through software, hardware and combinations thereof.OnStating each module can be embedded in the form of hardware or independently of in the processor in computer equipment, can also store in a software formIn memory in computer equipment, the corresponding operation of the above modules is executed in order to which processor calls.
Fig. 7 is the schematic diagram for the computer equipment that one embodiment of the invention provides.The computer equipment internal structure chart canWith as shown in Figure 7.The computer equipment includes processor, memory, network interface and the database connected by system bus.Wherein, the processor of the computer equipment is for providing calculating and control ability.The memory of the computer equipment includes non-easyThe property lost storage medium, built-in storage.The non-volatile memory medium is stored with operating system, computer program and database.It shouldBuilt-in storage provides environment for the operation of operating system and computer program in non-volatile memory medium.The computer equipmentDatabase for storing preset convolutional neural networks model.The network interface of the computer equipment is used for and external terminalIt is communicated by network connection.To realize a kind of licence plate recognition method when the computer program is executed by processor.
In one embodiment, a kind of computer equipment is provided, including memory, processor and storage are on a memoryAnd the computer program that can be run on a processor, processor realize above-described embodiment Car license recognition side when executing computer programThe step of method, such as step S10 shown in Fig. 2 is to step 80.Alternatively, processor realizes above-mentioned implementation when executing computer programThe function of each module/unit of example license plate recognition device, such as module shown in fig. 6 10 is to the function of module 80.To avoid weightMultiple, which is not described herein again.
It is apparent to those skilled in the art that for convenience of description and succinctly, only with above-mentioned each functionCan unit, module division progress for example, in practical application, can according to need and by above-mentioned function distribution by differentFunctional unit, module are completed, i.e., the internal structure of described device is divided into different functional unit or module, more than completingThe all or part of function of description.
In one embodiment, a computer readable storage medium is provided, meter is stored on the computer readable storage mediumThe step of calculation machine program, which realizes above-described embodiment licence plate recognition method when being executed by processor, alternatively, the meterCalculation machine program realizes the function of each module/unit in above-described embodiment license plate recognition device when being executed by processor.To avoid weightMultiple, which is not described herein again.
It is to be appreciated that the computer readable storage medium may include: that can carry the computer program codeAny entity or device, recording medium, USB flash disk, mobile hard disk, magnetic disk, CD, computer storage, read-only memory(Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), electric carrier signal andTelecommunication signal etc..
Embodiment described above is merely illustrative of the technical solution of the present invention, rather than its limitations；Although referring to aforementioned realityApplying example, invention is explained in detail, those skilled in the art should understand that: it still can be to aforementioned eachTechnical solution documented by embodiment is modified or equivalent replacement of some of the technical features；And these are modifiedOr replacement, the spirit and scope for technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution should allIt is included within protection scope of the present invention.