CN107273832A - Licence plate recognition method and system based on integrating channel feature and convolutional neural networks - Google Patents
Licence plate recognition method and system based on integrating channel feature and convolutional neural networks Download PDFInfo
- Publication number
- CN107273832A CN107273832A CN201710416776.5A CN201710416776A CN107273832A CN 107273832 A CN107273832 A CN 107273832A CN 201710416776 A CN201710416776 A CN 201710416776A CN 107273832 A CN107273832 A CN 107273832A
- Authority
- CN
- China
- Prior art keywords
- mrow
- feature
- msub
- msup
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- 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
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
- G06V20/625—License plates
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
Abstract
The present invention provides a kind of licence plate recognition method and system based on integrating channel feature and convolutional neural networks, and the licence plate recognition method includes：The sample image of license plate image is obtained, convolutional neural networks detector is generated according to sample image；Image to be detected is obtained, the feature pyramid of the different scale of generation image to be detected is calculated；Feature pyramid is detected by a sliding window using convolutional neural networks detector, the object candidate area under different scale is obtained；Enter the differentiation of line character and non-character in object candidate area using the first full articulamentum of convolutional neural networks detector, and enter the identification of line character in object candidate area using the second full articulamentum of convolutional neural networks detector.
Description
Technical field
The present invention is on Computer Vision Recognition technology, specifically on one kind based on integrating channel feature and volume
The licence plate recognition method and system of product neutral net.
Background technology
At the positions such as expressway gateway, parking lot gateway automatic detection, identify the number-plate number of vehicular traffic
It is a kind of important technology in intelligent transportation system.Under normal circumstances due to the extremely difficult condition such as intense light irradiation, big side angle, fuzzy, make
Obtain the detection of car plate becomes relatively difficult with identification.At present, the detection method for car plate has three major types：
The first is the method based on marginal information.This method is by phases such as marginal information and Hough transform, morphological operations
Handled with reference to image, obtain the candidate region of target in picture, then pass through specific priori, such as edges of regions
Density information, length-width ratio, shape etc., these regions are handled, are screened layer by layer, what is finally obtained is exactly the region where car plate.
F.Faradji et al. is in document " F.Faradji, A.H.Rezaie, and M.Ziaratban, A morphological-
based license plate location,in Proc.IEEE Int.Conf.Image Process,pp.57-60,
Sep.-Oct.2007 " proposes a kind of method for being combined marginal information and morphological operation and carrying out car plate detection, this method
By carrying out rim detection to image, morphological operation is then carried out, object candidate region is obtained.Followed by candidate region
Geometric attribute is screened, and finally realizes the positioning of car plate.Its weak point be susceptible to abundant textural characteristics and
The interference of similar shape object, in addition, when car plate is deformed upon, being easily missed.
Second is the method based on colouring information, and such method is made up of the module such as Color Segmentation and target positioning, adopted
Coloured image is split with multi-layer perception (MLP), potential license plate area is then partitioned into by projecting method.Such as W.Jia
People is in document " W.Jia, H.Zhang, X.He, and M.Piccardi, Mean shift for accurate license
In plate localization.in Proc.IEEE Conf.Intell.Trans.Syst.pp.566-571, Sep.2005 "
Coloured image split by mean shift algorithm to obtain some candidate regions, according to geometric attribute, edge density information
These candidate regions are made whether be license plate area classification, finally give testing result.Its deficiency is to illumination variation
Sensitivity, is vulnerable to the interference for having same color characteristic area in image.
The third is the method based on machine learning, and such as L.Dlagnekov et al. is in document " L.Dlagnekov, and
S.Belongie,Recognizing cars.Dept.Comput.Sci.Eng.UCSD,San Diego,
Tech.Rep.CS2005-083,2005 " the middle methods proposed, on the basis of Harr-Like features, is classified with AdaBoost
Device is classified to candidate region, realizes the positioning of car plate.The deficiency of such method is that false alarm rate is high, it may appear that many flase drops,
Detecting completely also is difficult to license plate area, so being not individually fine with the algorithm effect of machine learning.
The content of the invention
The main purpose of the embodiment of the present invention is to provide a kind of car based on integrating channel feature and convolutional neural networks
Board recognition methods and system, to tackle the car plate detection under complex scene and identification problem.
To achieve these goals, the embodiment of the present invention provides a kind of based on integrating channel feature and convolutional neural networks
Licence plate recognition method, described licence plate recognition method includes：The sample image of license plate image is obtained, is given birth to according to the sample image
Into convolution Neural Network Detector；Image to be detected is obtained, the feature gold of the different scale of the described image to be detected of generation is calculated
Word tower；The feature pyramid is detected by a sliding window using the convolutional neural networks detector, obtained not
With the object candidate area under yardstick；Using the first full articulamentum of the convolutional neural networks detector in the target candidate
The differentiation of line character and non-character is entered in region, and uses the second full articulamentum of the convolutional neural networks detector in the mesh
The identification of line character is entered in mark candidate region.
In one embodiment, it is above-mentioned that convolutional neural networks detector is generated according to the sample image, specifically include：Will
The sample image composing training collection, wherein, the sample image includes positive sample and negative sample, and positive sample is to include car plate
Image, negative sample is the background image not comprising car plate；Calculate the integrating channel feature of each sample image in the training set；
Integrating channel feature to the sample image carries out pond processing, generates the pond feature of the sample image；By the pond
Change feature input decision tree forest, using Adaboost algorithm, and pass through distribution function of the spatial distribution probability to Adaboost
Optimize, generate the convolutional neural networks detector.
In one embodiment, the above-mentioned integrating channel feature for calculating each sample in the training set, is specifically included：Step a：
The sample image is transformed into hsv color space, the color characteristic of three passages is calculated：
Wherein, the space coordinate of i, j representative sample image；H, S, V correspond to the value of three different passages respectively, and H represents color, value
For 60-330, red correspondence 0, green correspondence 120, blueness correspondence 240；S represents saturation degree, the brightness of color；V represents tone；
Step b：Calculate the gradient orientation histogram feature of the sample image；Step c：According to the feature and gradient of three passages
The described integrating channel feature of direction histogram feature generation.
Further, above-mentioned step b：The gradient orientation histogram feature of the sample image is calculated, is specifically included：Step
Rapid b1：Calculate the gradient direction value of each pixel in the sample image：Gx(x, y)=H (x+1, y)-H (x-1, y)；Gy(x,y)
=H (x, y+1)-H (x, y-1)；Wherein, H (x, y) represents the pixel value in gray space, Gx(x, y), Gy(x, y), is represented respectively
At (x, y) place, gradient both horizontally and vertically；Step b2：According to the gradient direction value calculate gradient magnitude G (x, y) and
Direction Step b3：According to finger
Fixed cell size, the gradient of each pixel in unit is projected by direction in different intervals, whole unit is generated
Gradient orientation histogram；Wherein, each interval angular range is 360/N, and N is the quantity of gradient direction；Step b4：According to described
Gradient orientation histogram determines the gradient orientation histogram feature：Wherein, i, j are sample
The coordinate of this image；N=1,2 ... N, represent interval number corresponding to different gradient directions.
In one embodiment, the feature pyramid of the different scale of above-mentioned described image to be detected of calculating generation, specifically
Including：Step 1：The multi-channel feature of described image to be detected under different scale is calculated by below equation：Wherein, FsThe corresponding features of yardstick s are represented, R represents to use yardstick s resamplings, F=Ω to image
(I) feature of image respective channel is represented, Ω corresponds to different feature passages；Step 2：According to the multi-channel feature, lead to
Cross the characteristic pattern that below equation calculates described image to be detected under generation different scale：
Wherein, Fs′It is characteristic pattern of the image to be detected in yardstick s ', λΩIt is the corresponding scale factor of Ω passages, other yardstick s are corresponding
Feature FsAccording to the dimension scale relation between different scale images and to have calculated characteristic pattern under obtained a certain yardstick approximate
Obtain the characteristic pattern under another yardstick；Step 3：According to characteristic pattern generation feature gold of the described image to be detected under each yardstick
Word tower.
In one embodiment, the corresponding λ of the different passage Ω of calculatingΩProcess include：Statistics collection global feature with
The average of change of scaleBy formulaCan
Obtain μsWith λΩRelational expression it is as follows：Wherein, E [ε] represents the desired value of error, fΩ(Is) lead to be all
The weighted sum in road, i.e. fΩ(I)=∑i,j,kωi,j,kF (i, j, k), ω features are the weights of respective channel, and k represents the sequence of passage
Row.
The embodiment of the present invention also provides a kind of Vehicle License Plate Recognition System based on integrating channel feature and convolutional neural networks, institute
The Vehicle License Plate Recognition System stated includes：Convolutional neural networks detector maturation unit, the sample image for obtaining license plate image, root
Convolutional neural networks detector is generated according to the sample image；Feature pyramid generation unit, for obtaining image to be detected, meter
Calculate the feature pyramid of the different scale of the described image to be detected of generation；Object candidate area acquiring unit, for using described
Convolutional neural networks detector is detected by a sliding window to the feature pyramid, obtains the target under different scale
Candidate region；Recognition unit, for the first full articulamentum using the convolutional neural networks detector in the target candidate
The differentiation of line character and non-character is entered in region, and uses the second full articulamentum of the convolutional neural networks detector in the mesh
The identification of line character is entered in mark candidate region.
In one embodiment, above-mentioned convolutional neural networks detector maturation unit specifically for：By the sample image
Composing training collection, wherein, the sample image includes positive sample and negative sample, and positive sample is the image comprising car plate, negative sample
For the background image not comprising car plate；Calculate the integrating channel feature of each sample image in the training set；To the sample
The integrating channel feature of image carries out pond processing, generates the pond feature of the sample image；By pond feature input
Decision tree forest, is optimized using Adaboost algorithm, and by spatial distribution probability to Adaboost distribution function, raw
Into the convolutional neural networks detector.
In one embodiment, the above-mentioned integrating channel feature for calculating each sample in the training set, is specifically included：Step
a：The sample image is transformed into hsv color space, the color characteristic of three passages is calculated：
Wherein, the space coordinate of i, j representative sample image；H, S, V correspond to the value of three different passages respectively, and H represents color, value
For 60-330, red correspondence 0, green correspondence 120, blueness correspondence 240；S represents saturation degree, the brightness of color；V represents tone；
Step b：Calculate the gradient orientation histogram feature of the sample image；Step c：According to the feature and gradient of three passages
The described integrating channel feature of direction histogram feature generation.
Further, above-mentioned step b：The gradient orientation histogram feature of the sample image is calculated, is specifically included：Step
Rapid b1：Calculate the gradient direction value of each pixel in the sample image：Gx(x, y)=H (x+1, y)-H (x-1, y)；Gy(x,y)
=H (x, y+1)-H (x, y-1)；Wherein, H (x, y) represents the pixel value in gray space, Gx(x, y), Gy(x, y), is represented respectively
At (x, y) place, gradient both horizontally and vertically；Step b2：According to the gradient direction value calculate gradient magnitude G (x, y) and
Direction Step b3：According to specified
Cell size, the gradient of each pixel in unit is projected by direction in different intervals, the ladder of the whole unit of generation
Spend direction histogram；Wherein, each interval angular range is 360/N, and N is the quantity of gradient direction；Step b4：According to the ladder
Degree direction histogram determines the gradient orientation histogram feature：Wherein, i, j are sample
The coordinate of image；N=1,2 ... N, represent interval number corresponding to different gradient directions.
In one embodiment, above-mentioned feature pyramid generation unit specifically for：Step 1：Calculated by below equation
The multi-channel feature of described image to be detected under different scale：Wherein, FsRepresent that yardstick s is corresponding
Feature, R represents to image that using yardstick s resamplings F=Ω (I) represent the feature of image respective channel, and Ω corresponds to different
Feature passage；Step 2：According to the multi-channel feature, the mapping to be checked under generation different scale is calculated by below equation
The characteristic pattern of picture：Wherein, Fs′It is characteristic pattern of the image to be detected in yardstick s ', λΩIt is
The corresponding scale factor of Ω passages, the corresponding feature F of other yardstick ssAccording to the dimension scale relation between different scale images
And the characteristic pattern calculated under obtained a certain yardstick approximately obtains the characteristic pattern under another yardstick；Step 3：According to described to be checked
Characteristic pattern generation feature pyramid of the altimetric image under each yardstick.
In one embodiment, the corresponding λ of the different passage Ω of calculatingΩProcess include：Statistics collection global feature with
The average of change of scaleBy formulaCan
Obtain μsWith λΩRelational expression it is as follows：Wherein, E [ε] represents the desired value of error, fΩ(Is) lead to be all
The weighted sum in road, i.e. fΩ(I)=∑i,j,kωi,j,kF (i, j, k), ω features are the weights of respective channel, and k represents the sequence of passage
Row.
The beneficial effect of the embodiment of the present invention is, using the recognition methods of the embodiment of the present invention, can be in different illumination
Preferably examined under condition (daytime, the different bright dark degree at night) and weather condition (different conditions of fine day, rainy day etc.)
Effect is surveyed, car plate key message can be accurately identified, in addition, both can apply to static schema (for a certain frame figure captured
Picture) under car plate detection and identification, the detection and knowledge of car plate under dynamic mode (be directed to continuous video flowing) can also be applied to
Not, and have the advantages that it is powerful, using the flexible, speed of service is fast, strong adaptability, resource occupation it is few.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, embodiment will be described below
In required for the accompanying drawing that uses be briefly described, it should be apparent that, drawings in the following description are only some of the present invention
Embodiment, for those of ordinary skill in the art, without having to pay creative labor, can also be according to these
Accompanying drawing obtains other accompanying drawings.
Fig. 1 is the licence plate recognition method based on integrating channel feature and convolutional neural networks according to the embodiment of the present invention
Flow chart；
Fig. 2 is the detailed process stream that convolutional neural networks detector is generated according to sample image according to the embodiment of the present invention
Cheng Tu；
Fig. 3 is the specific mistake for calculating the integrating channel feature of each sample image in training set according to the embodiment of the present invention
Journey flow chart；
Fig. 4 is the Vehicle License Plate Recognition System based on integrating channel feature and convolutional neural networks according to the embodiment of the present invention
Structural representation.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Site preparation is described, it is clear that described embodiment is only a part of embodiment of the invention, rather than whole embodiments.It is based on
Embodiment in the present invention, it is every other that those of ordinary skill in the art are obtained under the premise of creative work is not made
Embodiment, belongs to the scope of protection of the invention.
The embodiment of the present invention provides a kind of licence plate recognition method based on integrating channel feature and convolutional neural networks and is
System.Below in conjunction with accompanying drawing, the present invention is described in detail.
The embodiment of the present invention provides a kind of licence plate recognition method based on integrating channel feature and convolutional neural networks, such as schemes
Shown in 1, it is somebody's turn to do the licence plate recognition method based on integrating channel feature and convolutional neural networks and mainly includes the following steps that：
Step S101：The sample image of license plate image is obtained, convolutional neural networks detector is generated according to sample image；
Step S102：Image to be detected is obtained, the feature pyramid of the different scale of generation image to be detected is calculated；
Step S103：Feature pyramid is detected by a sliding window using convolutional neural networks detector, obtained
Take the object candidate area under different scale；
Step S104：Using convolutional neural networks detector the first full articulamentum object candidate area enter line character with
The differentiation of non-character, and enter the knowledge of line character in object candidate area using the second full articulamentum of convolutional neural networks detector
Not.
By above-mentioned step S101 to step S104, the embodiment of the present invention based on integrating channel feature and convolutional Neural
The licence plate recognition method of network, image is realized by learning the integrating channel feature of object, and by the method for sliding window
The solution of characteristic pattern, quickly detects object using the decision tree forest of cascade, the knowledge of car plate is realized with convolutional neural networks
Not.
Below in conjunction with specific embodiment, to the car based on integrating channel feature and convolutional neural networks of the embodiment of the present invention
Each step of board recognition methods is described further.
Above-mentioned step S101, obtains the sample image of license plate image, and convolutional neural networks inspection is generated according to sample image
Survey device.
Specifically, as shown in Fig. 2 step S101 mainly includes procedure below：
Step S1011：By sample image composing training collection, wherein, sample image includes positive sample and negative sample, positive sample
For the image comprising car plate, negative sample is the background image not comprising car plate.In the specific implementation, it can be included in the sample image
The sample of differing tilt angles, different illumination conditions and the lower car plate of different degrees of spot interference is as positive sample, while choosing phase
The region without car plate answered is as negative sample, and both collectively form training set.
Step S1012：Calculate the integrating channel feature of each sample image in training set.As shown in figure 3, calculating training set
In each sample integrating channel feature process, specifically include following steps：
Step S301：Sample image is transformed into hsv color space, the color characteristic of three passages is calculated：
Wherein, the space coordinate of i, j representative sample image；H, S, V correspond to the value of three different passages respectively, and H is represented
Color, value is 60-330, red correspondence 0, green correspondence 120, blueness correspondence 240；S represents saturation degree, the brightness of color；V
Represent tone；
Step S302：Calculate the gradient orientation histogram feature of sample image；Specifically include：
Step b1：Calculate the gradient direction value of each pixel in sample image：
Gx(x, y)=H (x+1, y)-H (x-1, y)；
Gy(x, y)=H (x, y+1)-H (x, y-1)；
Wherein, H (x, y) represents the pixel value in gray space, Gx(x, y), Gy(x, y), is illustrated respectively in (x, y) place, water
The gradient of gentle vertical direction；
Step b2：Gradient magnitude G (x, y) and direction are calculated according to gradient direction value
Step b3：According to specified cell size, the gradient of each pixel in unit is entered by direction in different intervals
Row projection, generates the gradient orientation histogram of whole unit；Wherein, each interval angular range is 360/N, and N is gradient direction
Quantity；
Step b4：Gradient orientation histogram feature is determined according to gradient orientation histogram：
Wherein, i, j are the coordinate of sample image；N=1,2 ... N, represent interval number corresponding to different gradient directions.
After calculating obtains the gradient orientation histogram feature of sample image, pass through step S303：According to three passages
Feature and gradient orientation histogram feature generation integrating channel feature Feature (i, j).
That is sample image i, the integrating channel of j positions is characterized in the combination of color characteristic and Gradient Features.
After calculating obtains the integrating channel feature of each sample image in training set, step S1013 is performed：To sample
The integrating channel feature of image carries out pond processing, generates the pond feature of sample image.
Specifically, above-mentioned pond processing procedure refers to：Region segmentation is carried out to obtained integration feature passage, determined each
The size in block region as pond size.Then max pooling are carried out (with each picture in the image-region to regional
The maximum of plain value as the region result) or average pooling (be averaged with each pixel value in the image-region
It is worth the result as the region), obtain Chi Huahou channel characteristics.
In one embodiment, in the processing procedure of pond, also need to handle the feature before and after pondization, uniform characteristics dimension
Degree, combination, which obtains final feature, is used for training convolutional neural networks detector.Processing procedure described herein be, for example, including
Such as normalization operation, is the conventional processing mode of general features vector.What is such as finally given is 4 × 4 result Feature Mapping
Figure, its actual form is 1 × 16 characteristic vector, passes through normalized so that it meetsUniform characteristics are tieed up
Degree is main to be included carrying out zero-padding to Chi Huahou feature, i.e., zero filling supplement is carried out at its head and the tail two ends, so that pond
Characteristic vector dimension afterwards is identical with before pondization.
Then, step S1014 is performed：Pond feature is inputted into decision tree forest, using Adaboost algorithm, and passes through sky
Between distribution probability Adaboost distribution function is optimized, generate convolutional neural networks detector, and train the convolution god
Through network detector.
General Adaboost is directly several obtained Weak Classifiers will to be trained directly to be added, that is, gives tacit consent to each weak
The weight of grader is all 1.And be exactly by preceding (n- to the optimization that Adaboost distribution function is carried out in this embodiment
1) sample of individual Weak Classifier classification error is used for the training of n-th of Weak Classifier, and in last Weak Classifier anabolic process
In, the weight of these Weak Classifiers is all not 1, but assigns different power according to its performance in training, test process
Weight.
When it is implemented, mainly point four-stage carries out the training of whole detector.First stage is 64 decision trees,
(decision tree is produced based on prior art, is a tree-shaped decision diagram of additional probability result, is intuitively with statistics
The figure method of probability analysis, each of which non-leaf nodes represents the test on a characteristic attribute, and each branch represents this feature
Output of the attribute in some codomain, and each leaf node deposits a classification.Using decision tree carry out decision-making process be exactly
Since root node, corresponding characteristic attribute in item to be sorted is tested, and output branch is selected according to its value, until reaching leaf
Node, the classification that leaf node is deposited is used as the result of decision), the decision tree forest that training data is inputted into the first stage (is
One include multiple decision trees grader, be between the decision tree in forest do not have it is related.When data enter decision forest
When, allow each decision tree to carry out classification and see which kind of this sample should belong to, finally take classification results in all decision trees
That most classes are used as final result), its testing result is then seen, the data of classification error are regard as " important training number
According to " next stage, i.e. second stage are put into, second stage has 128 decision trees ... by that analogy, until four ranks
The training of section terminates.The decision tree number of wherein four-stage is divided into for 64,128,256,1024, and these data are established a capital really is
With reference to the conventional parameter in various applications.
Above-mentioned step S102, obtains image to be detected, calculates the golden word of feature of the different scale of generation image to be detected
Tower.
In one embodiment, step S102 specifically includes procedure below：
Step 1：The multi-channel feature of image to be detected under different scale is calculated by below equation：
Wherein, FsThe corresponding features of yardstick s are represented, R represents to use image yardstick s resamplings, and F=Ω (I) represent figure
As the feature of respective channel, Ω corresponds to different feature passages；
Step 2：According to multi-channel feature, the feature of image to be detected under generation different scale is calculated by below equation
Figure：
Wherein, Fs′It is characteristic pattern of the image to be detected in yardstick s ', λΩIt is the corresponding scale factor of Ω passages, other yardsticks
The corresponding feature F of ssAccording to the dimension scale relation between different scale images and feature under obtained a certain yardstick is calculated
Figure approximately obtains the characteristic pattern under another yardstick.
Wherein, for the corresponding λ of different passage Ω in above-mentioned formula (1-1) and (1-2)ΩCalculating process it is as follows：
Statistics collection global feature is needed with the mean μ of change of scales：
By formula：
It can obtain：
E [ε] represents the desired value of error, f in above formulaΩ(Is) be all passages weighted sum, i.e.,：
fΩ(I)=∑i,j,kωi,j,kF (i, j, k) (1-6),
ω features are the weights of respective channel, and k represents the sequence of passage.Can by simultaneous above formula (1-3) to (1-6)
In the hope of λΩ。
Step 3：Feature pyramid is generated according to characteristic pattern of the image to be detected under each yardstick.
Above-mentioned step S103, is examined using convolutional neural networks detector by a sliding window to feature pyramid
Survey, obtain the object candidate area under different scale.Using features described above pyramid calculation method, image to be detected is obtained different
The feature of yardstick, is detected in each yardstick by sliding window using the detector trained, obtains object candidate area.
In one embodiment, for obtained object candidate area, using (the overlapping inspection to having of non-maxima suppression method
Survey result candidate region and detect that score is ranked up by it, obtain point highest as testing result region, remove other areas
Domain), take suitable threshold value to filter out final detection zone.In non-maxima suppression, the bounding box obtained for detection, meter
Calculate their Duplication：
What overlap=intersection (bbs), intersection (bbs) were calculated is the detection that phase mutual is occured simultaneously
The intersecting rate of results area, two detection blocks provided with common factor are respectively bb1 and bb2, and inter (bb1, bb2) represents two squares
The area of the common factor of shape frame, union (bb1, bb2) represents the union area of two rectangle frames, then intersection (bbs) is fixed
Justice is inter (bb1, bb2)/union (bb1, bb2).
If overlap is more than threshold tau, the highest that only keeps score bbsi.τ value is set to 0.5 in the present invention.bbs
Score obtained by the node corresponding threshold value summation of decision tree.
Above-mentioned step S104, is carried out using the first full articulamentum of convolutional neural networks detector in object candidate area
The differentiation of character and non-character, and use the second full articulamentum of convolutional neural networks detector to carry out word in object candidate area
The identification of symbol.
Utilize the positive and negative samples in data set, training convolutional neural networks, for carrying out license plate area in testing result
Feature extraction；Train the full articulamentum for distinguishing seven character features, the differentiation for seven character features in license plate area；
The full articulamentum for character recognition is trained, is used for the classification into line character, the identification of character in license plate area is realized.
Wherein, the full articulamentum of above-mentioned two substantially contributes to what is classified, the two full articulamentums be it is in parallel,
Obtaining feature map in front conventional part, (feature map described herein, refer to that image to be detected passes through convolution
(it is two-dimensional matrix in a computer to be similar to image to several numerical matrixs obtained after the processing of each layer of neutral net
Form), because in convolutional neural networks, convolution layer segment in front is mainly used for feature extraction, and the full articulamentum of back
Part is mainly classified, so the data obtained by conventional part are referred to as feature map, as conventional part from
The feature extracted in image to be detected) after, two full articulamentums in parallel.First feature map have been marked
(bounding box information refers to the markup information in training data to license plate area rectangle frame to bounding box information, right
Train picture in each width, there is a corresponding xml document to correspond to therewith, the content of file be mark picture in car plate
Rectangle frame (such as upper left angle point, the coordinate of bottom right angle point) and the corresponding character string of car plate content, training convolutional neural networks
Guidance is used as by above-mentioned markup information, allows convolutional neural networks to determine that what desired result is, so that pre- according to network
The difference surveyed between the labeled data of result and reality carries out the adjustment of network parameter, makes predicting the outcome for network most with real data
Possible identical, i.e., how training convolutional neural networks go identification) it is input to the first full articulamentum and is trained, only need to herein
Carry out car plate and the classification of non-car plate；The training of second full articulamentum in addition to needing feature map and bounding box,
The markup information of car plate content is also needed to, progress is each character recognition, i.e., polytypic process, both when training
Can be parallel, because of the difference and the no dependence of training of classification task；But, it is necessary to by feature when test
Map first passes through the first full articulamentum, further determines that and this result is input into the second full articulamentum again after license plate area, enter
The identification of line character, the i.e. full articulamentum of test phase second are to rely on the result of the first full articulamentum.
The licence plate recognition method based on integrating channel feature and convolutional neural networks of the embodiment of the present invention, by integration
The utilization of channel characteristics, fully excavates the information of car plate under different conditions, it is obtained more accurate feature representation, pass through pond
The influence of deformation band can effectively be tackled by changing operation.Using the decision tree forest detector of cascade, it can realize to target
Quick detection.Space where image, the region that positioning licence plate occurs fully are searched for using sliding window.Obtained by study
Model can be preferably extended into different scenes, while the shadow of the other factorses such as different illumination variations can be tackled preferably
Ring.
Using the recognition methods of the embodiment of the present invention, can different illumination conditions (daytime, night different bright dark journeys
Degree) and weather condition (different conditions of fine day, rainy day etc.) under obtain preferable Detection results, can accurately identify car plate key
Information, in addition, the detection and identification of car plate under static schema (for a certain two field picture captured) had both been can apply to, can also
Applied to the detection and identification of car plate under dynamic mode (be directed to continuous video flowing), and with it is powerful, using flexibly,
The advantage that the speed of service is fast, strong adaptability, resource occupation are few.
The embodiment of the present invention also provides a kind of Vehicle License Plate Recognition System based on integrating channel feature and convolutional neural networks, such as
Shown in Fig. 4, being somebody's turn to do the Vehicle License Plate Recognition System based on integrating channel feature and convolutional neural networks mainly includes：Convolutional neural networks are examined
Survey device generation unit 1, feature pyramid generation unit 2, object candidate area acquiring unit 3 and recognition unit 4.
Wherein, convolutional neural networks detector maturation unit 1 is used for the sample image for obtaining license plate image, according to sample graph
As generation convolutional neural networks detector；Feature pyramid generation unit 2 is used to obtain image to be detected, calculates generation to be detected
The feature pyramid of the different scale of image；Object candidate area acquiring unit 3 is used to lead to using convolutional neural networks detector
Cross a sliding window to detect feature pyramid, obtain the object candidate area under different scale；Recognition unit 4 is used to make
Enter the differentiation of line character and non-character in object candidate area with the first full articulamentum of convolutional neural networks detector, and use
Second full articulamentum of convolutional neural networks detector enters the identification of line character in object candidate area.
By the collaborative work between each above-mentioned part, the embodiment of the present invention based on integrating channel feature with
The Vehicle License Plate Recognition System of convolutional neural networks, by learning the integrating channel feature of object, and the method for passing through sliding window
The solution of characteristics of image figure is realized, object is quickly detected using the decision tree forest of cascade, is realized with convolutional neural networks
The identification of car plate.
Below in conjunction with specific embodiment, to the car based on integrating channel feature and convolutional neural networks of the embodiment of the present invention
Each part and its function of board identifying system are described further.
Above-mentioned convolutional neural networks detector maturation unit 1, the sample image for obtaining license plate image, according to sample
Image generates convolutional neural networks detector.
Specifically, as shown in Fig. 2 the convolutional neural networks detector maturation unit 1 mainly performs procedure below：
Step S1011：By sample image composing training collection, wherein, sample image includes positive sample and negative sample, positive sample
For the image comprising car plate, negative sample is the background image not comprising car plate.In the specific implementation, it can be included in the sample image
The sample of differing tilt angles, different illumination conditions and the lower car plate of different degrees of spot interference is as positive sample, while choosing phase
The region without car plate answered is as negative sample, and both collectively form training set.
Step S1012 calculates the integrating channel feature of each sample image in training set.As shown in figure 3, calculating training set
In each sample integrating channel feature process, specifically include following steps：
Step S301：Sample image is transformed into hsv color space, the color characteristic of three passages is calculated：
Wherein, the space coordinate of i, j representative sample image；H, S, V correspond to the value of three different passages respectively, and H is represented
Color, value is 60-330, red correspondence 0, green correspondence 120, blueness correspondence 240；S represents saturation degree, the brightness of color；V
Represent tone；
Step S302：Calculate the gradient orientation histogram feature of sample image；Specifically include：
Step b1：Calculate the gradient direction value of each pixel in sample image：
Gx(x, y)=H (x+1, y)-H (x-1, y)；
Gy(x, y)=H (x, y+1)-H (x, y-1)；
Wherein, H (x, y) represents the pixel value in gray space, Gx(x, y), Gy(x, y), is illustrated respectively in (x, y) place, water
The gradient of gentle vertical direction；
Step b2：Gradient magnitude G (x, y) and direction are calculated according to gradient direction value
Step b3：According to specified cell size, the gradient of each pixel in unit is entered by direction in different intervals
Row projection, generates the gradient orientation histogram of whole unit；Wherein, each interval angular range is 360/N, and N is gradient direction
Quantity；
Step b4：Gradient orientation histogram feature is determined according to gradient orientation histogram：
Wherein, i, j are the coordinate of sample image；N=1,2 ... N, represent interval number corresponding to different gradient directions.
After calculating obtains the gradient orientation histogram feature of sample image, pass through step S303：According to three passages
Feature and gradient orientation histogram feature generation integrating channel feature Feature (i, j).
That is sample image i, the integrating channel of j positions is characterized in the combination of color characteristic and Gradient Features.
After calculating obtains the integrating channel feature of each sample image in training set, the life of convolutional neural networks detector
Step S1013 is performed into unit 1：Integrating channel feature to sample image carries out pond processing, generates the pond of sample image
Feature.
Specifically, above-mentioned pond processing procedure refers to：Region segmentation is carried out to obtained integration feature passage, determined each
The size in block region as pond size.Then max pooling are carried out (with each picture in the image-region to regional
The maximum of plain value as the region result) or average pooling (be averaged with each pixel value in the image-region
It is worth the result as the region), obtain Chi Huahou channel characteristics.
In one embodiment, in the processing procedure of pond, also need to handle the feature before and after pondization, uniform characteristics dimension
Degree, combination, which obtains final feature, is used for training convolutional neural networks detector.Processing procedure described herein be, for example, including
Such as normalization operation, is the conventional processing mode of general features vector.What is such as finally given is 4 × 4 result Feature Mapping
Figure, its actual form is 1 × 16 characteristic vector, passes through normalized so that it meetsUniform characteristics are tieed up
Degree is main to be included carrying out zero-padding to Chi Huahou feature, i.e., zero filling supplement is carried out at its head and the tail two ends, so that pond
Characteristic vector dimension afterwards is identical with before pondization.
Then, convolutional neural networks detector maturation unit 1 performs step S1014：Pond feature input decision tree is gloomy
Woods, is optimized using Adaboost algorithm, and by spatial distribution probability to Adaboost distribution function, generation convolution god
Through network detector, and train the convolutional neural networks detector.
General Adaboost is directly several obtained Weak Classifiers will to be trained directly to be added, that is, gives tacit consent to each weak
The weight of grader is all 1.And be exactly by preceding (n- to the optimization that Adaboost distribution function is carried out in this embodiment
1) sample of individual Weak Classifier classification error is used for the training of n-th of Weak Classifier, and in last Weak Classifier anabolic process
In, the weight of these Weak Classifiers is all not 1, but assigns different power according to its performance in training, test process
Weight.
When it is implemented, mainly point four-stage carries out the training of whole detector.First stage is 64 decision trees,
(decision tree is produced based on prior art, is a tree-shaped decision diagram of additional probability result, is intuitively with statistics
The figure method of probability analysis, each of which non-leaf nodes represents the test on a characteristic attribute, and each branch represents this feature
Output of the attribute in some codomain, and each leaf node deposits a classification.Using decision tree carry out decision-making process be exactly
Since root node, corresponding characteristic attribute in item to be sorted is tested, and output branch is selected according to its value, until reaching leaf
Node, the classification that leaf node is deposited is used as the result of decision), the decision tree forest that training data is inputted into the first stage (is
One include multiple decision trees grader, be between the decision tree in forest do not have it is related.When data enter decision forest
When, allow each decision tree to carry out classification and see which kind of this sample should belong to, finally take classification results in all decision trees
That most classes are used as final result), its testing result is then seen, the data of classification error are regard as " important training number
According to " next stage, i.e. second stage are put into, second stage has 128 decision trees ... by that analogy, until four ranks
The training of section terminates.The decision tree number of wherein four-stage is divided into for 64,128,256,1024, and these data are established a capital really is
With reference to the conventional parameter in various applications.
Above-mentioned feature pyramid generation unit 2 is used to obtain image to be detected, calculates the difference of generation image to be detected
The feature pyramid of yardstick.
This feature pyramid generation unit 2 specifically performs procedure below：
Step 1：The multi-channel feature of image to be detected under different scale is calculated by below equation：
Wherein, FsThe corresponding features of yardstick s are represented, R represents to use image yardstick s resamplings, and F=Ω (I) represent figure
As the feature of respective channel, Ω corresponds to different feature passages；
Step 2：According to multi-channel feature, the feature of image to be detected under generation different scale is calculated by below equation
Figure：
Wherein, Fs′It is characteristic pattern of the image to be detected in yardstick s ', λΩIt is the corresponding scale factor of Ω passages, other yardsticks
The corresponding feature F of ssAccording to the dimension scale relation between different scale images and feature under obtained a certain yardstick is calculated
Figure approximately obtains the characteristic pattern under another yardstick.
Wherein, for the corresponding λ of different passage Ω in above-mentioned formula (1-1) and (1-2)ΩCalculating process it is as follows：
Statistics collection global feature is needed with the mean μ of change of scales：
By formula：
It can obtain：
E [ε] represents the desired value of error, f in above formulaΩ(Is) be all passages weighted sum, i.e.,：
fΩ(I)=∑i,j,kωi,j,kF (i, j, k) (1-6),
ω features are the weights of respective channel, and k represents the sequence of passage.Can by simultaneous above formula (1-3) to (1-6)
In the hope of λΩ。
Step 3：Feature pyramid is generated according to characteristic pattern of the image to be detected under each yardstick.
Above-mentioned object candidate area acquiring unit 3, for passing through a sliding window using convolutional neural networks detector
Feature pyramid is detected, the object candidate area under different scale is obtained.Using features described above pyramid calculation method,
The feature of image to be detected different scale is obtained, is examined using the detector trained by sliding window in each yardstick
Survey, obtain object candidate area.
In one embodiment, for obtained object candidate area, using (the overlapping inspection to having of non-maxima suppression method
Survey result candidate region and detect that score is ranked up by it, obtain point highest as testing result region, remove other areas
Domain), take suitable threshold value to filter out final detection zone.In non-maxima suppression, the bounding box obtained for detection, meter
Calculate their Duplication：
What overlap=intersection (bbs), intersection (bbs) were calculated is the detection that phase mutual is occured simultaneously
The intersecting rate of results area, two detection blocks provided with common factor are respectively bb1 and bb2, and inter (bb1, bb2) represents two squares
The area of the common factor of shape frame, union (bb1, bb2) represents the union area of two rectangle frames, then intersection (bbs) is fixed
Justice is inter (bb1, bb2)/union (bb1, bb2).
If overlap is more than threshold tau, the highest that only keeps score bbsi.τ value is set to 0.5 in the present invention.bbs
Score obtained by the node corresponding threshold value summation of decision tree.
Above-mentioned recognition unit 4, for the first full articulamentum using convolutional neural networks detector in target candidate area
The differentiation of line character and non-character is entered in domain, and uses the second full articulamentum of convolutional neural networks detector in object candidate area
Enter the identification of line character.
Utilize the positive and negative samples in data set, training convolutional neural networks, for carrying out license plate area in testing result
Feature extraction；Train the full articulamentum for distinguishing seven character features, the differentiation for seven character features in license plate area；
The full articulamentum for character recognition is trained, is used for the classification into line character, the identification of character in license plate area is realized.
Wherein, the full articulamentum of above-mentioned two substantially contributes to what is classified, the two full articulamentums be it is in parallel,
Obtaining feature map in front conventional part, (feature map described herein, refer to that image to be detected passes through convolution
(it is two-dimensional matrix in a computer to be similar to image to several numerical matrixs obtained after the processing of each layer of neutral net
Form), because in convolutional neural networks, convolution layer segment in front is mainly used for feature extraction, and the full articulamentum of back
Part is mainly classified, so the data obtained by conventional part are referred to as feature map, as conventional part from
The feature extracted in image to be detected) after, two full articulamentums in parallel.First feature map have been marked
(bounding box information refers to the markup information in training data to license plate area rectangle frame to bounding box information, right
Train picture in each width, there is a corresponding xml document to correspond to therewith, the content of file be mark picture in car plate
Rectangle frame (such as upper left angle point, the coordinate of bottom right angle point) and the corresponding character string of car plate content, training convolutional neural networks
Guidance is used as by above-mentioned markup information, allows convolutional neural networks to determine that what desired result is, so that pre- according to network
The difference surveyed between the labeled data of result and reality carries out the adjustment of network parameter, makes predicting the outcome for network most with real data
Possible identical, i.e., how training convolutional neural networks go identification) it is input to the first full articulamentum and is trained, only need to herein
Carry out car plate and the classification of non-car plate；The training of second full articulamentum in addition to needing feature map and bounding box,
The markup information of car plate content is also needed to, progress is each character recognition, i.e., polytypic process, both when training
Can be parallel, because of the difference and the no dependence of training of classification task；But, it is necessary to by feature when test
Map first passes through the first full articulamentum, further determines that and this result is input into the second full articulamentum again after license plate area, enter
The identification of line character, the i.e. full articulamentum of test phase second are to rely on the result of the first full articulamentum.
The Vehicle License Plate Recognition System based on integrating channel feature and convolutional neural networks of the embodiment of the present invention, by integration
The utilization of channel characteristics, fully excavates the information of car plate under different conditions, it is obtained more accurate feature representation, pass through pond
The influence of deformation band can effectively be tackled by changing operation.Using the decision tree forest detector of cascade, it can realize to target
Quick detection.Space where image, the region that positioning licence plate occurs fully are searched for using sliding window.Obtained by study
Model can be preferably extended into different scenes, while the shadow of the other factorses such as different illumination variations can be tackled preferably
Ring.
Using the identifying system of the embodiment of the present invention, can different illumination conditions (daytime, night different bright dark journeys
Degree) and weather condition (different conditions of fine day, rainy day etc.) under obtain preferable Detection results, can accurately identify car plate key
Information, in addition, the detection and identification of car plate under static schema (for a certain two field picture captured) had both been can apply to, can also
Applied to the detection and identification of car plate under dynamic mode (be directed to continuous video flowing), and with it is powerful, using flexibly,
The advantage that the speed of service is fast, strong adaptability, resource occupation are few.
One of ordinary skill in the art will appreciate that realizing that all or part of step in above-described embodiment method can lead to
Cross program to instruct the hardware of correlation to complete, the program can be stored in a computer read/write memory medium, such as
ROM/RAM, magnetic disc, CD etc..
Particular embodiments described above, has been carried out further in detail to the purpose of the present invention, technical scheme and beneficial effect
Describe in detail it is bright, should be understood that the foregoing is only the present invention specific embodiment, the guarantor being not intended to limit the present invention
Scope is protected, within the spirit and principles of the invention, any modification, equivalent substitution and improvements done etc. should be included in this
Within the protection domain of invention.
Claims (12)
1. a kind of licence plate recognition method based on integrating channel feature and convolutional neural networks, it is characterised in that described car plate
Recognition methods includes：
The sample image of license plate image is obtained, convolutional neural networks detector is generated according to the sample image；
Image to be detected is obtained, the feature pyramid of the different scale of the described image to be detected of generation is calculated；
The feature pyramid is detected by a sliding window using the convolutional neural networks detector, obtains different
Object candidate area under yardstick；
Enter line character and non-word in the object candidate area using the first full articulamentum of the convolutional neural networks detector
The differentiation of symbol, and in the object candidate area enter line character using the second full articulamentum of the convolutional neural networks detector
Identification.
2. the licence plate recognition method according to claim 1 based on integrating channel feature and convolutional neural networks, its feature
It is, convolutional neural networks detector is generated according to the sample image, specifically included：
By the sample image composing training collection, wherein, the sample image include positive sample and negative sample, positive sample be comprising
The image of car plate, negative sample is the background image not comprising car plate；
Calculate the integrating channel feature of each sample image in the training set；
Integrating channel feature to the sample image carries out pond processing, generates the pond feature of the sample image；
The pond feature is inputted into decision tree forest, using Adaboost algorithm, and passes through spatial distribution probability pair
Adaboost distribution function is optimized, and generates the convolutional neural networks detector.
3. the licence plate recognition method according to claim 2 based on integrating channel feature and convolutional neural networks, its feature
It is, calculates the integrating channel feature of each sample in the training set, specifically include：
Step a：The sample image is transformed into hsv color space, the color characteristic of three passages is calculated：
<mrow>
<mi>c</mi>
<mi>o</mi>
<mi>l</mi>
<mi>o</mi>
<mi>r</mi>
<mi>f</mi>
<mi>e</mi>
<mi>a</mi>
<mi>t</mi>
<mi>u</mi>
<mi>r</mi>
<mi>e</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<mi>H</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mi>S</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mi>V</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>,</mo>
</mrow>
Wherein, the space coordinate of i, j representative sample image；H, S, V correspond to the value of three different passages respectively, and H represents color,
Value is 60-330, red correspondence 0, green correspondence 120, blueness correspondence 240；S represents saturation degree, the brightness of color；V represents color
Adjust；
Step b：Calculate the gradient orientation histogram feature of the sample image；
Step c：According to the feature of three passages and the described integrating channel feature of gradient orientation histogram feature generation.
4. the licence plate recognition method according to claim 3 based on integrating channel feature and convolutional neural networks, its feature
It is, step b：The gradient orientation histogram feature of the sample image is calculated, is specifically included：
Step b1：Calculate the gradient direction value of each pixel in the sample image：
Gx(x, y)=H (x+1, y)-H (x-1, y)；
Gy(x, y)=H (x, y+1)-H (x, y-1)；
Wherein, H (x, y) represents the pixel value in gray space, Gx(x, y), Gy(x, y), is illustrated respectively in (x, y) place, level and
The gradient of vertical direction；
Step b2：Gradient magnitude G (x, y) and direction are calculated according to the gradient direction value
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>=</mo>
<msqrt>
<mrow>
<msub>
<mi>G</mi>
<mi>x</mi>
</msub>
<msup>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
<mo>+</mo>
<msub>
<mi>G</mi>
<mi>y</mi>
</msub>
<msup>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
</mrow>
</msqrt>
<mo>;</mo>
</mrow>
<mrow>
<mo>&part;</mo>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>=</mo>
<msup>
<mi>tan</mi>
<mrow>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msup>
<mrow>
<mo>(</mo>
<mfrac>
<mrow>
<msub>
<mi>G</mi>
<mi>y</mi>
</msub>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<msub>
<mi>G</mi>
<mi>x</mi>
</msub>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<mo>;</mo>
</mrow>
1
Step b3：According to specified cell size, the gradient of each pixel in unit is thrown by direction in different intervals
Shadow, generates the gradient orientation histogram of whole unit；Wherein, each interval angular range is 360/N, and N is the number of gradient direction
Amount；
Step b4：The gradient orientation histogram feature is determined according to the gradient orientation histogram：
<mrow>
<mi>g</mi>
<mi>r</mi>
<mi>a</mi>
<mi>d</mi>
<mi>f</mi>
<mi>e</mi>
<mi>a</mi>
<mi>t</mi>
<mi>u</mi>
<mi>r</mi>
<mi>e</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>&part;</mo>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>,</mo>
<mi>n</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>;</mo>
</mrow>
Wherein, i, j are the coordinate of sample image；N=1,2 ... N, represent interval number corresponding to different gradient directions.
5. the licence plate recognition method according to claim 4 based on integrating channel feature and convolutional neural networks, its feature
It is, calculates the feature pyramid of the different scale of the described image to be detected of generation, specifically include：
Step 1：The multi-channel feature of described image to be detected under different scale is calculated by below equation：
<mrow>
<msub>
<mi>F</mi>
<mi>s</mi>
</msub>
<mo>=</mo>
<mi>R</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>F</mi>
<mo>,</mo>
<mi>s</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>&times;</mo>
<msup>
<mi>s</mi>
<mrow>
<mo>-</mo>
<msub>
<mi>&lambda;</mi>
<mi>&Omega;</mi>
</msub>
</mrow>
</msup>
<mo>,</mo>
</mrow>
Wherein, FsThe corresponding features of yardstick s are represented, R represents to use image yardstick s resamplings, and F=Ω (I) represent image correspondence
The feature of passage, Ω corresponds to different feature passages；
Step 2：According to the multi-channel feature, the spy of described image to be detected under generation different scale is calculated by below equation
Levy figure：
<mrow>
<msub>
<mi>F</mi>
<mi>s</mi>
</msub>
<mo>&ap;</mo>
<mi>R</mi>
<mrow>
<mo>(</mo>
<mrow>
<msub>
<mi>F</mi>
<msup>
<mi>s</mi>
<mo>&prime;</mo>
</msup>
</msub>
<mo>,</mo>
<mi>s</mi>
<mo>/</mo>
<msup>
<mi>s</mi>
<mo>&prime;</mo>
</msup>
</mrow>
<mo>)</mo>
</mrow>
<mo>&times;</mo>
<msup>
<mrow>
<mo>(</mo>
<mrow>
<mi>s</mi>
<mo>/</mo>
<msup>
<mi>s</mi>
<mo>&prime;</mo>
</msup>
</mrow>
<mo>)</mo>
</mrow>
<mrow>
<mo>-</mo>
<msub>
<mi>&lambda;</mi>
<mi>&Omega;</mi>
</msub>
</mrow>
</msup>
<mo>,</mo>
</mrow>
Wherein, Fs′It is characteristic pattern of the image to be detected in yardstick s ', λΩIt is the corresponding scale factor of Ω passages, other s pairs of yardsticks
The feature F answeredsAccording to the dimension scale relation between different scale images and characteristic pattern under obtained a certain yardstick is calculated
Approximately obtain the characteristic pattern under another yardstick；
Step 3：Feature pyramid is generated according to characteristic pattern of the described image to be detected under each yardstick.
6. the licence plate recognition method according to claim 5 based on integrating channel feature and convolutional neural networks, its feature
It is, the corresponding λ of the different passage Ω of calculatingΩProcess include：
Statistics collection global feature is with the mean μ of change of scales：
<mrow>
<msub>
<mi>&mu;</mi>
<mi>s</mi>
</msub>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>N</mi>
</mfrac>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</msubsup>
<msub>
<mi>f</mi>
<mi>&Omega;</mi>
</msub>
<mrow>
<mo>(</mo>
<msubsup>
<mi>I</mi>
<mi>s</mi>
<mi>i</mi>
</msubsup>
<mo>)</mo>
</mrow>
<mo>/</mo>
<msub>
<mi>f</mi>
<mi>&Omega;</mi>
</msub>
<mrow>
<mo>(</mo>
<msup>
<mi>I</mi>
<mi>i</mi>
</msup>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
By formulaμ can be obtainedsWith λΩRelational expression it is as follows：
<mrow>
<msub>
<mi>&mu;</mi>
<mi>s</mi>
</msub>
<mo>=</mo>
<msup>
<mi>s</mi>
<mrow>
<mo>-</mo>
<msub>
<mi>&lambda;</mi>
<mi>&Omega;</mi>
</msub>
</mrow>
</msup>
<mo>+</mo>
<mi>E</mi>
<mrow>
<mo>&lsqb;</mo>
<mi>&epsiv;</mi>
<mo>&rsqb;</mo>
</mrow>
<mo>,</mo>
</mrow>
Wherein, E [ε] represents the desired value of error, fΩ(Is) be all passages weighted sum, i.e. fΩ(I)=∑i,j,kωi,j,kF
(i, j, k), ω features are the weights of respective channel, and k represents the sequence of passage.
7. a kind of Vehicle License Plate Recognition System based on integrating channel feature and convolutional neural networks, it is characterised in that described car plate
Identifying system includes：
Convolutional neural networks detector maturation unit, the sample image for obtaining license plate image is given birth to according to the sample image
Into convolution Neural Network Detector；
Feature pyramid generation unit, for obtaining image to be detected, calculates the different scale of the described image to be detected of generation
Feature pyramid；
Object candidate area acquiring unit, for using the convolutional neural networks detector by a sliding window to the spy
Levy pyramid to be detected, obtain the object candidate area under different scale；
Recognition unit, for being entered using the first full articulamentum of the convolutional neural networks detector in the object candidate area
The differentiation of line character and non-character, and the second full articulamentum of the convolutional neural networks detector is used in the target candidate
The identification of line character is entered in region.
8. the Vehicle License Plate Recognition System according to claim 7 based on integrating channel feature and convolutional neural networks, its feature
Be, the convolutional neural networks detector maturation unit specifically for：
By the sample image composing training collection, wherein, the sample image include positive sample and negative sample, positive sample be comprising
The image of car plate, negative sample is the background image not comprising car plate；
Calculate the integrating channel feature of each sample image in the training set；
Integrating channel feature to the sample image carries out pond processing, generates the pond feature of the sample image；
The pond feature is inputted into decision tree forest, using Adaboost algorithm, and passes through spatial distribution probability pair
Adaboost distribution function is optimized, and generates the convolutional neural networks detector.
9. the Vehicle License Plate Recognition System according to claim 8 based on integrating channel feature and convolutional neural networks, its feature
It is, calculates the integrating channel feature of each sample in the training set, specifically include：
Step a：The sample image is transformed into hsv color space, the color characteristic of three passages is calculated：
<mrow>
<mi>c</mi>
<mi>o</mi>
<mi>l</mi>
<mi>o</mi>
<mi>r</mi>
<mi>f</mi>
<mi>e</mi>
<mi>a</mi>
<mi>t</mi>
<mi>u</mi>
<mi>r</mi>
<mi>e</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<mi>H</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mi>S</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mi>V</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>,</mo>
</mrow>
Wherein, the space coordinate of i, j representative sample image；H, S, V correspond to the value of three different passages respectively, and H represents color,
Value is 60-330, red correspondence 0, green correspondence 120, blueness correspondence 240；S represents saturation degree, the brightness of color；V represents color
Adjust；
Step b：Calculate the gradient orientation histogram feature of the sample image；
Step c：According to the feature of three passages and the described integrating channel feature of gradient orientation histogram feature generation.
10. the Vehicle License Plate Recognition System according to claim 9 based on integrating channel feature and convolutional neural networks, its feature
It is, step b：The gradient orientation histogram feature of the sample image is calculated, is specifically included：
Step b1：Calculate the gradient direction value of each pixel in the sample image：
Gx(x, y)=H (x+1, y)-H (x-1, y)；
Gy(x, y)=H (x, y+1)-H (x, y-1)；
Wherein, H (x, y) represents the pixel value in gray space, Gx(x, y), Gy(x, y), is illustrated respectively in (x, y) place, level and
The gradient of vertical direction；
Step b2：Gradient magnitude G (x, y) and direction are calculated according to the gradient direction value
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>=</mo>
<msqrt>
<mrow>
<msub>
<mi>G</mi>
<mi>x</mi>
</msub>
<msup>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
<mo>+</mo>
<msub>
<mi>G</mi>
<mi>y</mi>
</msub>
<msup>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
</mrow>
</msqrt>
<mo>;</mo>
</mrow>
<mrow>
<mo>&part;</mo>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>=</mo>
<msup>
<mi>tan</mi>
<mrow>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msup>
<mrow>
<mo>(</mo>
<mfrac>
<mrow>
<msub>
<mi>G</mi>
<mi>y</mi>
</msub>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<msub>
<mi>G</mi>
<mi>x</mi>
</msub>
<mrow>
<mo>(</mo>
<mrow>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<mo>;</mo>
</mrow>
Step b3：According to specified cell size, the gradient of each pixel in unit is thrown by direction in different intervals
Shadow, generates the gradient orientation histogram of whole unit；Wherein, each interval angular range is 360/N, and N is the number of gradient direction
Amount；
Step b4：The gradient orientation histogram feature is determined according to the gradient orientation histogram：
<mrow>
<mi>g</mi>
<mi>r</mi>
<mi>a</mi>
<mi>d</mi>
<mi>f</mi>
<mi>e</mi>
<mi>a</mi>
<mi>t</mi>
<mi>u</mi>
<mi>r</mi>
<mi>e</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>&part;</mo>
<mrow>
<mo>(</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>,</mo>
<mi>n</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>;</mo>
</mrow>
Wherein, i, j are the coordinate of sample image；N=1,2 ... N, represent interval number corresponding to different gradient directions.
11. the Vehicle License Plate Recognition System according to claim 10 based on integrating channel feature and convolutional neural networks, it is special
Levy and be, the feature pyramid generation unit specifically for：
Step 1：The multi-channel feature of described image to be detected under different scale is calculated by below equation：
<mrow>
<msub>
<mi>F</mi>
<mi>s</mi>
</msub>
<mo>=</mo>
<mi>R</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>F</mi>
<mo>,</mo>
<mi>s</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>&times;</mo>
<msup>
<mi>s</mi>
<mrow>
<mo>-</mo>
<msub>
<mi>&lambda;</mi>
<mi>&Omega;</mi>
</msub>
</mrow>
</msup>
<mo>,</mo>
</mrow>
Wherein, FsThe corresponding features of yardstick s are represented, R represents to use image yardstick s resamplings, and F=Ω (I) represent image correspondence
The feature of passage, Ω corresponds to different feature passages；
Step 2：According to the multi-channel feature, the spy of described image to be detected under generation different scale is calculated by below equation
Levy figure：
<mrow>
<msub>
<mi>F</mi>
<mi>s</mi>
</msub>
<mo>&ap;</mo>
<mi>R</mi>
<mrow>
<mo>(</mo>
<mrow>
<msub>
<mi>F</mi>
<msup>
<mi>s</mi>
<mo>&prime;</mo>
</msup>
</msub>
<mo>,</mo>
<mi>s</mi>
<mo>/</mo>
<msup>
<mi>s</mi>
<mo>&prime;</mo>
</msup>
</mrow>
<mo>)</mo>
</mrow>
<mo>&times;</mo>
<msup>
<mrow>
<mo>(</mo>
<mrow>
<mi>s</mi>
<mo>/</mo>
<msup>
<mi>s</mi>
<mo>&prime;</mo>
</msup>
</mrow>
<mo>)</mo>
</mrow>
<mrow>
<mo>-</mo>
<msub>
<mi>&lambda;</mi>
<mi>&Omega;</mi>
</msub>
</mrow>
</msup>
<mo>,</mo>
</mrow>
Wherein, Fs′It is characteristic pattern of the image to be detected in yardstick s ', λΩIt is the corresponding scale factor of Ω passages, other s pairs of yardsticks
The feature F answeredsAccording to the dimension scale relation between different scale images and characteristic pattern under obtained a certain yardstick is calculated
Approximately obtain the characteristic pattern under another yardstick；
Step 3：Feature pyramid is generated according to characteristic pattern of the described image to be detected under each yardstick.
12. the Vehicle License Plate Recognition System according to claim 11 based on integrating channel feature and convolutional neural networks, it is special
Levy and be, the corresponding λ of the different passage Ω of calculatingΩProcess include：
Statistics collection global feature is with the mean μ of change of scales：
<mrow>
<msub>
<mi>&mu;</mi>
<mi>s</mi>
</msub>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>N</mi>
</mfrac>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</msubsup>
<msub>
<mi>f</mi>
<mi>&Omega;</mi>
</msub>
<mrow>
<mo>(</mo>
<msubsup>
<mi>I</mi>
<mi>s</mi>
<mi>i</mi>
</msubsup>
<mo>)</mo>
</mrow>
<mo>/</mo>
<msub>
<mi>f</mi>
<mi>&Omega;</mi>
</msub>
<mrow>
<mo>(</mo>
<msup>
<mi>I</mi>
<mi>i</mi>
</msup>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
By formulaμ can be obtainedsWith λΩRelational expression it is as follows：
<mrow>
<msub>
<mi>&mu;</mi>
<mi>s</mi>
</msub>
<mo>=</mo>
<msup>
<mi>s</mi>
<mrow>
<mo>-</mo>
<msub>
<mi>&lambda;</mi>
<mi>&Omega;</mi>
</msub>
</mrow>
</msup>
<mo>+</mo>
<mi>E</mi>
<mrow>
<mo>&lsqb;</mo>
<mi>&epsiv;</mi>
<mo>&rsqb;</mo>
</mrow>
<mo>,</mo>
</mrow>
Wherein, E [ε] represents the desired value of error, fΩ(Is) be all passages weighted sum, i.e. fΩ(I)=∑i,j,kωi,j,kF
(i, j, k), ω features are the weights of respective channel, and k represents the sequence of passage.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710416776.5A CN107273832B (en) | 2017-06-06 | 2017-06-06 | License plate recognition method and system based on integral channel characteristics and convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710416776.5A CN107273832B (en) | 2017-06-06 | 2017-06-06 | License plate recognition method and system based on integral channel characteristics and convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107273832A true CN107273832A (en) | 2017-10-20 |
| CN107273832B CN107273832B (en) | 2020-09-22 |
Family
ID=60064600
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710416776.5A Active CN107273832B (en) | 2017-06-06 | 2017-06-06 | License plate recognition method and system based on integral channel characteristics and convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107273832B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108154149A (en) * | 2017-12-08 | 2018-06-12 | 济南中维世纪科技有限公司 | Licence plate recognition method based on deep learning network share |
| CN108520278A (en) * | 2018-04-10 | 2018-09-11 | 陕西师范大学 | A kind of road surface crack detection method and its evaluation method based on random forest |
| CN108960026A (en) * | 2018-03-10 | 2018-12-07 | 王洁 | Unmanned plane during flying Orientation system |
| CN109325449A (en) * | 2018-01-04 | 2019-02-12 | 苏州中科天启遥感科技有限公司 | Convolutional neural networks target detection frame based on Sample Refreshment |
| CN110020651A (en) * | 2019-04-19 | 2019-07-16 | 福州大学 | Car plate detection localization method based on deep learning network |
| CN110069651A (en) * | 2017-10-23 | 2019-07-30 | 腾讯科技（北京）有限公司 | Picture screening technique and device, storage medium |
| CN111127327A (en) * | 2019-11-14 | 2020-05-08 | 贝壳技术有限公司 | Picture inclination detection method and device |
| CN111801689A (en) * | 2018-04-17 | 2020-10-20 | 赫尔实验室有限公司 | System for real-time object detection and recognition using image and size features |
| CN112381150A (en) * | 2020-11-17 | 2021-02-19 | 上海科技大学 | Confrontation sample detection method based on sample robustness difference |
| US20210117762A1 (en) * | 2018-06-25 | 2021-04-22 | Olympus Corporation | Arithmetic processing device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104239867A (en) * | 2014-09-17 | 2014-12-24 | 深圳市捷顺科技实业股份有限公司 | License plate locating method and system |
| CN104809443A (en) * | 2015-05-05 | 2015-07-29 | 上海交通大学 | Convolutional neural network-based license plate detection method and system |
| CN106295507A (en) * | 2016-07-25 | 2017-01-04 | 华南理工大学 | A kind of gender identification method based on integrated convolutional neural networks |
| CN106529522A (en) * | 2016-11-07 | 2017-03-22 | 湖南源信光电科技有限公司 | License plate location method based on integral channel features |
| EP3182334A1 (en) * | 2015-12-17 | 2017-06-21 | Xerox Corporation | License plate recognition using coarse-to-fine cascade adaptations of convolutional neural networks |
-
2017
- 2017-06-06 CN CN201710416776.5A patent/CN107273832B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104239867A (en) * | 2014-09-17 | 2014-12-24 | 深圳市捷顺科技实业股份有限公司 | License plate locating method and system |
| CN104809443A (en) * | 2015-05-05 | 2015-07-29 | 上海交通大学 | Convolutional neural network-based license plate detection method and system |
| EP3182334A1 (en) * | 2015-12-17 | 2017-06-21 | Xerox Corporation | License plate recognition using coarse-to-fine cascade adaptations of convolutional neural networks |
| CN106295507A (en) * | 2016-07-25 | 2017-01-04 | 华南理工大学 | A kind of gender identification method based on integrated convolutional neural networks |
| CN106529522A (en) * | 2016-11-07 | 2017-03-22 | 湖南源信光电科技有限公司 | License plate location method based on integral channel features |
Non-Patent Citations (1)
| Title |
|---|
| PIOTR DOLL´AR等: "Fast Feature Pyramids for Object Detection", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110069651B (en) * | 2017-10-23 | 2023-04-07 | 腾讯科技（北京）有限公司 | Picture screening method and device and storage medium |
| CN110069651A (en) * | 2017-10-23 | 2019-07-30 | 腾讯科技（北京）有限公司 | Picture screening technique and device, storage medium |
| CN108154149A (en) * | 2017-12-08 | 2018-06-12 | 济南中维世纪科技有限公司 | Licence plate recognition method based on deep learning network share |
| CN108154149B (en) * | 2017-12-08 | 2021-12-10 | 济南中维世纪科技有限公司 | License plate recognition method based on deep learning network sharing |
| CN109325449A (en) * | 2018-01-04 | 2019-02-12 | 苏州中科天启遥感科技有限公司 | Convolutional neural networks target detection frame based on Sample Refreshment |
| CN108960026A (en) * | 2018-03-10 | 2018-12-07 | 王洁 | Unmanned plane during flying Orientation system |
| CN108520278A (en) * | 2018-04-10 | 2018-09-11 | 陕西师范大学 | A kind of road surface crack detection method and its evaluation method based on random forest |
| CN111801689A (en) * | 2018-04-17 | 2020-10-20 | 赫尔实验室有限公司 | System for real-time object detection and recognition using image and size features |
| US20210117762A1 (en) * | 2018-06-25 | 2021-04-22 | Olympus Corporation | Arithmetic processing device |
| CN110020651B (en) * | 2019-04-19 | 2022-07-08 | 福州大学 | License plate detection and positioning method based on deep learning network |
| CN110020651A (en) * | 2019-04-19 | 2019-07-16 | 福州大学 | Car plate detection localization method based on deep learning network |
| CN111127327A (en) * | 2019-11-14 | 2020-05-08 | 贝壳技术有限公司 | Picture inclination detection method and device |
| CN111127327B (en) * | 2019-11-14 | 2024-04-12 | 贝壳技术有限公司 | Picture inclination detection method and device |
| CN112381150A (en) * | 2020-11-17 | 2021-02-19 | 上海科技大学 | Confrontation sample detection method based on sample robustness difference |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107273832B (en) | 2020-09-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107273832A (en) | Licence plate recognition method and system based on integrating channel feature and convolutional neural networks | |
| CN109977812B (en) | Vehicle-mounted video target detection method based on deep learning | |
| CN112200161B (en) | Face recognition detection method based on mixed attention mechanism | |
| CN108830188B (en) | Vehicle detection method based on deep learning | |
| US8509478B2 (en) | Detection of objects in digital images | |
| CN103366602B (en) | Method of determining parking lot occupancy from digital camera images | |
| CN104268528B (en) | A kind of crowd massing method for detecting area and device | |
| CN109816024A (en) | A kind of real-time automobile logo detection method based on multi-scale feature fusion and DCNN | |
| CN109033950A (en) | Vehicle based on multiple features fusion cascade deep model, which is disobeyed, stops detection method | |
| CN106295636A (en) | Passageway for fire apparatus based on multiple features fusion cascade classifier vehicle checking method | |
| CN107122776A (en) | A kind of road traffic sign detection and recognition methods based on convolutional neural networks | |
| CN107886080A (en) | One kind is parked position detecting method | |
| CN108734189A (en) | Vehicle License Plate Recognition System based on atmospherical scattering model and deep learning under thick fog weather | |
| CN105975929A (en) | Fast pedestrian detection method based on aggregated channel features | |
| CN107085696A (en) | A kind of vehicle location and type identifier method based on bayonet socket image | |
| CN105809121A (en) | Multi-characteristic synergic traffic sign detection and identification method | |
| JPWO2020181685A5 (en) | ||
| CN108960266A (en) | Image object detection method and device | |
| CN105574550A (en) | Vehicle identification method and device | |
| CN106650731A (en) | Robust license plate and logo recognition method | |
| CN105844257A (en) | Early warning system based on machine vision driving-in-fog road denoter missing and early warning method | |
| CN106295605A (en) | Traffic lights detection and recognition methods | |
| CN112084869A (en) | Compact quadrilateral representation-based building target detection method | |
| CN102855500A (en) | Haar and HoG characteristic based preceding car detection method | |
| CN109886147A (en) | A kind of more attribute detection methods of vehicle based on the study of single network multiple-task |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |