CN108985217A - A kind of traffic sign recognition method and system based on deep space network - Google Patents
A kind of traffic sign recognition method and system based on deep space network Download PDFInfo
- Publication number
- CN108985217A CN108985217A CN201810751516.8A CN201810751516A CN108985217A CN 108985217 A CN108985217 A CN 108985217A CN 201810751516 A CN201810751516 A CN 201810751516A CN 108985217 A CN108985217 A CN 108985217A
- Authority
- CN
- China
- Prior art keywords
- network
- traffic sign
- image
- feature
- vgg
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- 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
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
- G06V20/582—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of traffic signs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to traffic information processing technology fields, disclose a kind of traffic sign recognition method and system based on deep space network, and the image information with space-invariance is extracted using spatial alternation network；By improving VGG network, the feature that converged network layer each stage extracts handles fused characteristics of image using spatial pyramid method, obtains the characteristic information of more Traffic Sign Images.Spatial alternation network is added in the present invention, and the deficiency of the image information with space-invariance cannot effectively be extracted by making up CNN；VGG network is improved, the feature that converged network layer each stage extracts handles fused characteristics of image using spatial pyramid method, obtains the characteristic information of more Traffic Sign Images.The experimental results showed that improving the accuracy of identification Traffic Sign Images on the basis of enhancing network characterization.
Description
Technical field
The invention belongs to traffic information processing technology field more particularly to a kind of traffic signs based on deep space network
Recognition methods and system.
Background technique
Currently, the prior art commonly used in the trade is such that
Sun Wei [] et al. proposes to extract multilayer feature using CNN network, and uses pyramid pond method, multiple dimensioned pond
These features are combined to obtain multiattribute traffic sign feature vector, and input in ELM classifier and carry out by each layer feature
Quickly identification traffic sign；
Traffic Sign Recognition (Traffic Sign Recognition, TSR)[1]It is intelligent transportation system
The important composition module of (intelligent transportation system, ITS).Using image, the information processing technology into
Row it is more accurate and timely traffic administration guarantee driving safety.Traffic Sign Recognition System mainly include road sign detection and
Identify two parts.Common lane marker detection mainly includes based on color segmentation[2], template matching is done based on road sign shape[3]With
And combine two methods[4]Target detection is carried out, and these technologies are all quite mature.The difficult point of TSR is mainly reflected in knowledge
On not, since Traffic Sign Recognition is carried out in outdoor complex environment, great challenge, main body are brought to researcher
Now[5]: the color of road sign becomes dim；Outdoor complicated light condition；Barrier blocks road sign；The road sign mould of high-speed motion
Paste etc..
In recent years, at convolutional neural networks (convolutional neural network, CNN) for the image of representative
Reason technology all achieves greater advance in target identification in semantic segmentation and target detection, also become intelligent transportation field
The emphasis of research.Convolutional neural networks directly input traffic image in model, automatically extract validity feature, improve image recognition
Efficiency.Based on this, many scholars explore the identification in terms of effective deep learning model is applied to traffic sign.Sermanet
Deng[6]Multi-scale image is put into training in network model, and the Feature Mapping of different convolutional layers is combined as final feature
Vector recently enters the identification of full articulamentum, and enhances data, increases sample diversity, meets model to image space not
Denaturation demand has fine recognition effect, but the acquisition of multi-scale image feature also increases the model training time.Jin etc.[7]
The inspiration of hinge loss function in supported vector machine (support vector machine, SVM) proposes to lose using hinge
Learn the convolution kernel in CNN, completion automatically extracts feature and identification image, but also results in heavy computation burden.Xie etc.[8]It will
Fish criterion is added in deep learning network, and obtains the feature for meeting chronic space-invariance, is identified using svm classifier.
Wang[9]Deng proposition to the independent pre-training of different convolutional layers in CNN network, retain network weight parameter, and comprehensive batch is added
It normalizes (Batch Normalization, BN), greatly shortens the training time, but Traffic Sign Images discrimination still has promotion
Space.Zhang[10]Et al. construct the CNN network that a variety of pond methods combine, not use GPU accelerate in the case where,
The network has good effect on training time and discrimination.Zeng etc.[11]Using CNN as feature extractor, with having added core
The extreme learning machine (Extreme Learning Machine) of function is used as classifier, when greatly shortening identification traffic sign
Between, but the feature lack of diversity that network obtains, especially to danger signal and other prohibitory sign effects of releasing far below it
His method.
In conclusion problem of the existing technology is:
(1) the existing traffic sign recognition method based on deep learning is not able to satisfy model for space-invariances such as rotations
Generalization ability and insufficient Small object image zooming-out characteristic quantity the problems such as, cause identification Traffic Sign Images poor accuracy.
(2) the existing traffic sign recognition method based on deep learning using data enhancing (translation including certain angle,
Rotation and scaling etc. map functions) method, increase sample it is multifarious simultaneously, greatly increase deep learning network training friendship
The time of logical mark is difficult to realize quick for traffic sign and has although can guarantee the accuracy of identification traffic sign
Effect identification
(3) network layer of neural network selected by the existing traffic sign recognition method based on deep learning is very deep, and
And there are many training parameter, cannot be efficiently applied to Traffic Sign Recognition.
Solve the difficulty and meaning of above-mentioned technical problem:
Traffic Sign Images are all Small object images, are difficult to obtain enough characteristic quantities using classical Let-5 network, and
And influence final discrimination.And VGG network model very good solution this problem is selected, this is because VGG Web vector graphic is more
Small convolution kernel obtains the feature of image, and more details feature can be expressed by network, and we change traditional VGG network
Into required parameter in reduction training process can quickly identify the friendship in image in this way while guaranteeing traffic identification
Logical mark.In addition, it is not direct picture that many traffic signs are concentrated in public data, certain transformation is had, therefore, it is necessary to
We correct original image, and traditional data enhancing technology is difficult to realize and is accurately suitably become to traffic sign
It changes, therefore, the present invention adaptively can correct these images, be conducive to image and exist by constructing spatial alternation network
Feature representation is carried out in depth network model, completes classification.
[1]Wu T,Ranganathan A.A practical system for road marking detection
and recognition[C]//Intelligent Vehicles Symposium.IEEE,2012:25-30.
[2]Luo L,Li X.A Method to Search for Color Segmentation Threshold in
Traffic Sign Detection.[C]//International Conference on Image and
Graphics.IEEE,2010:774-777.
[3]Landesa-Vázquez I,Parada-Loira F,Alba-Castro J L.Fast real-time
multiclass traffic sign detection based on novel shape and texture
descriptors[C]//International IEEE Conference on Intelligent Transportation
Systems.IEEE,2010:1388-1395.
[4]Bahlmann C,Zhu Y,Ramesh V,et al.A system for traffic sign
detection,tracking,and recognition using color,shape,and motion information
[C]//Intelligent Vehicles Symposium,2005.Proceedings.IEEE.IEEE,2005:255-260.
[5]Ellahyani A,Ansari M E,Jaafari I E.Traffic sign detection and
recognition based on random forests[M].Elsevier Science Publishers B.V.2016.
[6]Sermanet P,Lecun Y.Traffic sign recognition with multi-scale
Convolutional Networks[C]//International Joint Conference on Neural
Networks.IEEE,2012:2809-2813.
[7]Jin J,Fu K,Zhang C.Traffic Sign Recognition With Hinge Loss
Trained Convolutional Neural Networks[J].IEEE Transactions on Intelligent
Transportation Systems,2014,15(5):1991-2000.
[8]Xie Jin,Cai Zixing,Deng Haitao,et al.Classification of Traffic
Signs Based on Deep Learning of Image Invariant Features[J].Journal of
Computer-Aided Design&Computer Graphics,2017,29(4):632-640.(in Chinese)
(Xie Jin, Cai Zixing, Deng Haitao wait to calculate based on traffic sign classification [J] of image invariant features deep learning
Machine Computer Aided Design and graphics journal, 2017,29 (4): 632-640.)
[9]Wang Xiaobin,Huang Jinjie,Liu Wenju.Traffic Sign Recognition Based
on Optimized Convolutional Neural Network Structure[J].Journal of Computer
Applications,2017,37(2):530-534.(in Chinese)
(Wang Xiaobin, gold is outstanding, and Liu Wen is lifted and calculated based on Traffic Sign Recognition [J] of optimization convolutional neural networks structure
Machine application, 2017,37 (2): 530-534.)
[10]Zhang J,Huang Q,Wu H,et al.A Shallow Network with Combined
Pooling for Fast Traffic Sign Recognition[J].Information,2017,8(2):45.
[11]Huang Z,Yu Y,Gu J.A novel method for traffic sign recognition
based on extreme learning machine[C]//Intelligent Control and
Automation.IEEE,2015:1451-1456.
[12]Jaderberg M,Simonyan K,Zisserman A,et al.Spatial Transformer
Networks[J].2015:2017-2025.
[13]Goodfellow I J,Warde-Farley D,Mirza M,et al.Maxout Networks[J]
.Computer Science,2013:1319-1327.
[14]Simonyan K,Zisserman A.Very Deep Convolutional Networks for
Large-Scale Image Recognition[J].Computer Science,2014.
[15]He K,Zhang X,Ren S,et al.Spatial Pyramid Pooling in Deep
Convolutional Networks for Visual Recognition[J].IEEE Transactions on Pattern
Analysis&Machine Intelligence,2015,37(9):1904-1916.
[16]Srivastava N,Hinton G,Krizhevsky A,et al.Dropout:a simple way to
prevent neural networks from overfitting[J].Journal of Machine Learning
Research,2014,15(1):1929-1958.
[17]Dan C,Meier U,Masci J,et al.A committee of neural networks for
traffic sign classification[C]//International Joint Conference on Neural
Networks.IEEE,2011:1918-1921.
[18]Stallkamp J,Schlipsing M,Salmen J,et al.Man vs.computer:
benchmarking machine learning algorithms for traffic sign recognition.[J]
.Neural Networks,2012,32(2):323-332.
[19]Berkaya S K,Gunduz H,Ozsen O,et al.On circular traffic sign
detection and recognition[J].Expert Systems with Applications,2016,48:67-75.
Zaklouta F,Stanciulescu B,Hamdoun O.Traffic sign classification using
K-d trees and Random Forests[C]//International Joint Conference on Neural
Networks.IEEE,2011:2151-2155.
Summary of the invention
In view of the problems of the existing technology, the present invention provides a kind of Traffic Sign Recognitions based on deep space network
Method and system.
The invention is realized in this way a kind of traffic sign recognition method based on deep space network, comprising:
Firstly, Traffic Sign Images conveying space converting network is obtained the affine transformation square of image using localized network
Battle array, and using the backpropagation of neural network, optimal affine transformation matrix is obtained, more contains space-invariance to obtain
Image；
Then changing image is extracted in the feature of network layer different phase, to different phase using improved VGG network
Feature is merged, and the diversity of Traffic Sign Images feature is increased, and improved VGG network selects smaller size of convolution
Core obtains the minutia of more images, is conducive to feature and preferably expresses in deep learning network layer；Fusion feature passes through
Spatial pyramid pond method can obtain specified size characteristic vector, and being conducive to network acquisition feature can be in different point
Traffic sign is identified in class device；Meanwhile characteristics of image is further enhanced, and export as the feature of traffic sign；
Finally classified with sotfmax classifier to traffic sign.
Further, before the training of image input network model, spatial alternation network is added, enhances the more of image space structure
Sample, and sufficient spatial signature information is provided for CNN Network Recognition traffic sign；
Spatial alternation network is that affine transformation is carried out to image, passes through the change of 2 × 3 dimensions during the affine transformation of image
It changes matrix to translate Traffic Sign Images, scale, rotation；
Transformation for mula is expressed as
Wherein, θ indicates the parameter in transformation matrix, and (x, y) indicates the pixel of image, and α indicates image after rotation
State, image successively by translation, scaling and rotation spatial alternation；
Traffic Sign Images are after removing the preprocess method of illumination, by the office operated containing multilayer convolution sum pondization
In portion's network, the transformation matrix A1 of initial 2 × 3 dimension is obtained；Grid Sampling is established using transformation matrix A1, is obtained desired defeated
Image A2 out；Using input and output image corresponding relationship, bilinear interpolation is carried out to input picture, completes transformed image
A3 is the input of identification module；
Spatial network includes localized network, 3 parts of mesh generator and sampler；
In localized network, input is Traffic Sign ImagesWherein W, H are the width and height of input picture,
C indicates port number；Spatial alternation layer exports θ=floc(U), local network structure are as follows: input picture passes through flocFunction obtains office
The output of portion's network includes 6 neurons, transformation matrix θ needed for forming image.
Further, mesh generator is the sampling network of the parameter creation obtained according to localized network, is one group of point set, defeated
Enter mapping and generates desired conversion output through over-sampling；Using the method for inverse transformation, the respective value of output pixel point is obtained, is converted
Formula is as follows:
Wherein,Image input and output coordinate points are respectively indicated, θ indicates the output of localized network；To seat
Scale value is normalized, so that xi, yi∈ [- 1,1]；
Output is established by sampler to contact with upper all pixels point gray value is inputted, and is made using sampling and input picture
Bilinearity sampling is carried out, sampling formula indicates are as follows:
Wherein,Indicate input picture,Image after indicating output transform, H ', w ' expression
The height and width of sampling grid, c indicate that port number is identical as original input Traffic Sign Images；It samples in formulaWithIn
Partial derivative can be asked, and constantly update transformation matrix using improved VGG network backpropagation principle, and it is empty to obtain more images
Between characteristic, for convolutional neural networks CNN identify image Biodiversity Characteristics are provided.
Further, include: based on the feature extracting method for improving VGG
The convolution kernel of multiple 3 × 3 sizes is selected to replace 5 × 5 convolution kernel, to CNN network image Edge Gradient Feature；Net
The input of network is the gray level image after spatial alternation network processes, and 3 layers of convolutional layer and 1 layer of pond layer constitute VGG module, always
3 VGG modules are shared, the number of network layer convolution kernel is 32,32,32,32,32,32,64,64,64, VGG module different phase
Obtaining feature sizes is 16 × 16 × 64,8 × 8 × 64 and 4 × 4 × 64；
The Fusion Features that different VGG modules are obtained, since the dimension of feature is different, selection obtains feature to VGG1 and carries out
The operation of 2 × 2 pondizations, feature down-sampling are 8 × 8 × 32 dimension sizes, obtain feature 1, and obtain feature with VGG2 and merged,
Fusion feature 8 × 8 × (32+32)=8 × 8 × 64 dimension size is obtained, feature 2 is obtained；2 down-sampling of feature is operated, with
VGG3 feature, which blends, obtains feature 3, and dimension is after 4 × 4 × (64+64)=4 × 4 × 128.VGG modular character merges, more
Minutia by e-learning；It carries out enhancing fusion feature using pyramid pond method again.
Further, using pyramid pond method, (so-called pyramid pond method is exactly the pond using different scale
The characteristic pattern that core obtains convolution is converted into the feature vector of fixed size, is obtained not using pyramid pond method here
With the characteristic information of scale, Enhanced feature improves Traffic Sign Recognition effect) feature after fusion is further enhanced, it obtains
More sufficient target signature, the identification to Traffic Sign Images.
Another object of the present invention is to provide the Traffic Sign Recognition sides based on deep space network described in a kind of realize
The computer program of method.
Another object of the present invention is to provide the Traffic Sign Recognition sides based on deep space network described in a kind of realize
The information data processing terminal of method.
Another object of the present invention is to provide a kind of computer readable storage mediums, including instruction, when it is in computer
When upper operation, so that computer executes the traffic sign recognition method based on deep space network.
Another object of the present invention is to provide the Traffic Sign Recognition sides based on deep space network described in a kind of realize
The Traffic Sign Recognition System based on deep space network of method, the Traffic Sign Recognition System based on deep space network
Include:
The image collection module of space-invariance, by Traffic Sign Images conveying space converting network, acquisition more contains
The image of space-invariance；
Characteristics of image processing module extracts changing image in the spy of network layer different phase using improved VGG network
Sign, merges different phase feature；Fusion feature passes through spatial pyramid pond method, further enhances characteristics of image,
And it is exported as the feature of traffic sign；
Traffic sign categorization module classifies to traffic sign with sotfmax classifier.
Another object of the present invention is to provide a kind of equipped with the Traffic Sign Recognition based on deep space network
System, Traffic Sign Recognition equipment.
In conclusion advantages of the present invention and good effect are as follows:
The invention proposes a kind of deep space network models to be applied to Traffic Sign Recognition.Spatial alternation network is added,
The deficiency of the image information with space-invariance cannot effectively be extracted by making up CNN；Improve VGG network, each rank of converged network layer
The feature that section is extracted handles fused characteristics of image using spatial pyramid method, obtains the spy of more Traffic Sign Images
Reference breath.The experimental results showed that improving the accuracy of identification Traffic Sign Images on the basis of enhancing network characterization.
Detailed description of the invention
Fig. 1 is the traffic sign recognition method flow chart provided in an embodiment of the present invention based on deep space network.
Fig. 2 is the structural schematic diagram of STN network provided in an embodiment of the present invention.
Fig. 3 is improved VGG network model figure provided in an embodiment of the present invention.
Fig. 4 is SPP illustraton of model provided in an embodiment of the present invention.
Fig. 5 is input picture provided in an embodiment of the present invention and the image graph after STN is exported.
Fig. 6 is the cost function comparison diagram before and after addition STN provided in an embodiment of the present invention.
Fig. 7 is training set discrimination comparison diagram before and after addition STN network provided in an embodiment of the present invention.
Fig. 8 is the Traffic Sign Recognition System schematic diagram provided in an embodiment of the present invention based on deep space network.
In figure: 1, the image collection module of space-invariance；2, characteristics of image processing module；3, traffic sign classification mould
Block.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
It is not able to satisfy model for space invariances such as rotations for the existing traffic sign recognition method based on deep learning
Property generalization ability and insufficient Small object image zooming-out characteristic quantity the problems such as, propose and a kind of net converted based on deep space
The traffic sign recognition method of network.Firstly, obtaining and more containing space invariance Traffic Sign Images conveying space converting network
The image of property；Then changing image is extracted in the feature of network layer different phase, to different phase using improved VGG network
Feature is merged；Fusion feature passes through spatial pyramid pond method, further enhances characteristics of image, and as traffic sign
Feature output；Finally classified with sotfmax classifier to traffic sign.It is tested, is tied on disclosed data set
Fruit shows that this method can effectively improve Traffic Sign Recognition effect.
Below with reference to concrete analysis, the invention will be further described.
1, the Traffic Sign Recognition based on spatial alternation network
A kind of traffic sign model based on deep space network proposed, by spatial alternation network (Spatial
Transformer Networks,STN)[12]It is introduced into CNN model, spatial alternation adaptively is carried out to Traffic Sign Images, is obtained
Take more space-invariance characteristic informations.Input of the image as convolutional neural networks after transformation, by multilayer convolution sum pond
The feature extraction of Traffic Sign Images is completed in operation, and network layer acquisition feature is merged, and obtains Biodiversity Characteristics, and make
With pyramid pond model, strengthens characteristics of image, convey full articulamentum, complete the identification of image, overall framework is as shown in Figure 1.
The 1.1 image diversity enhancings based on spatial alternation network
The influence that Traffic Sign Images are shot by illumination condition, motion blur and different angle, the space knot of image
Structure changes, and the feature of image also becomes diversification, and traditional CNN model can identify these image sections, but cannot
Image space invariance is solved, as part speed limit 60km/h traffic sign rotates by a certain angle and cannot effectively identify.The present invention mentions
Out before the training of image input network model, spatial alternation network is added, enhances the diversity of image space structure, and is CNN
Network Recognition traffic sign provides sufficient spatial signature information.The essence of spatial alternation network is to carry out affine change to image
It changes, the affine transformation process of image is to be translated by the transformation matrix of 2 × 3 dimensions to Traffic Sign Images, is scaled, rotation
Deng operation.
Transformation for mula can be expressed as
Wherein, θ indicates the parameter in transformation matrix, and (x, y) indicates the pixel of image, and α indicates image after rotation
State, image successively by translation, scaling and rotation spatial alternation.
Traffic Sign Images are after removing the preprocess method of illumination, by the office operated containing multilayer convolution sum pondization
In portion's network, the transformation matrix A1 of initial 2 × 3 dimension is obtained；Grid Sampling is established using transformation matrix A1, is obtained desired defeated
Image A2 out；Using input and output image corresponding relationship, bilinear interpolation is carried out to input picture, completes transformed image
A3 is the input of identification module.Spatial network mainly includes localized network, 3 parts of mesh generator and sampler, Fig. 2 table
Show the structural schematic diagram of STN network.
In localized network, input is Traffic Sign ImagesWherein W, H are the width and height of input picture,
C indicates port number.So spatial alternation layer exports α=floc(U), local network structure is similar to convolutional neural networks, contains volume
Lamination, pond layer, full articulamentum etc..Different places is input picture by flocFunction obtains the output of localized network
Comprising 6 neurons, transformation matrix θ needed for forming image.
Mesh generator is to create sampling network according to the parameter of localized network acquisition, this is one group of point set, that is, is inputted
Mapping generates desired conversion through over-sampling and exports.Using the method for inverse transformation, the respective value of output pixel point is obtained, transformation is public
Formula is as follows:
Wherein,Image input and output coordinate points are respectively indicated, θ indicates the output of localized network.In order to
The interpolation calculation for facilitating sampler needs that coordinate value is normalized, so that xi, yi∈ [- 1,1].
The effect of sampler is exactly to establish output to contact with upper all pixels point gray value is inputted, and use sampling and input
Image is made to carry out bilinearity sampling, and sampling formula indicates are as follows:
Wherein,Indicate input picture,Image after indicating output transform, H ', W ' expression
The height and width of sampling grid, C indicate that port number is identical as original input Traffic Sign Images.Due in sampling formulaWithMiddle partial derivative can be asked, and transformation matrix is constantly updated using the backpropagation principle of identification module part, obtain
More image space characteristics identify that image provides Biodiversity Characteristics for convolutional neural networks, improve Traffic Sign Images discrimination.
1.2 based on the feature extraction for improving VGG
Convolutional layer (Convolutional Layer) and pond layer (Pooling Layer) are alternately schemed in CNN model
The feature extraction of picture, in classification of the full articulamentum to Traffic Sign Images.Convolutional layer is to extract characteristics of image using convolution operation,
Relationship between the character representation pixel and adjacent pixel of image, characteristics of image carry out convolution operation by weighting adjacent pixel and obtain
?.Traffic Sign Images are filtered by convolution kernel, and be activated Function Mapping, generate output characteristic pattern.Convolution operation
It can be expressed as
Wherein,Indicate the i-th input of l network layer,It is the i-th output of l network layer, g indicates activation primitive,
MiIndicate the i-th selected input, * essence is that convolution kernel w is allowed to make convolution algorithm on l-1 layers of associated characteristic pattern, so
After sum, add an offset parameter b.
Pond layer is generally connected to behind convolutional layer, is replaced with convolutional layer, and the feature extractor of image is constituted.Due to convolution
It obtains characteristic pattern to be trained directly as feature vector, will cause huge computation complexity.Therefore, it is necessary to pass through Chi Huacao
Make to carry out dimensionality reduction to characteristic pattern.Goodfellow etc.[13]It is proposed the activation primitive for using maximum pond function as pond layer,
The result that the maximum value in receptive field region is operated as pondization is directly chosen during pond.This pond has CNN model
There is preferably anti-noise ability, while also reducing the dimension of feature.
Since traffic indication map seems Small object image, the convolutional layer of CNN model selects big convolution kernel to lose many weights
The minutia wanted, influences target identification.The present invention proposes improved VGG[14]Network model selects multiple 3 in VGG network
The convolution kernel of × 3 sizes replaces 5 × 5 convolution kernel, makes up CNN network and extracts incomplete defect to picture edge characteristic.Choosing
It is with the advantage of small convolution kernel: on the one hand can reduces the study of parameter in neural network, is on the other hand carried out more
Nonlinear Mapping can increase the fitting and ability to express of network.Improved VGG model as shown in figure 3, the input of network be through
Gray level image after crossing spatial alternation network processes, 3 layers of convolutional layer and 1 layer of pond layer constitute VGG module, a total of 3 VGG moulds
Block, the number of network layer convolution kernel is 32,32,32,32,32,32,64,64,64, VGG module different phase obtain feature sizes
It is 16 × 16 × 64,8 × 8 × 64 and 4 × 4 × 64.
Document[6]By convolutional neural networks heterogeneous networks layer Fusion Features, characteristics of image is easier to express, and is effectively promoted and is handed over
Logical sign image discrimination.The Fusion Features that different VGG modules are obtained, since the dimension of feature is different, selection obtains VGG1
It obtains feature and carries out the operation of 2 × 2 pondizations, feature down-sampling is 8 × 8 × 32 dimension sizes, obtains feature 1, and obtain spy with VGG2
Sign is merged, and is obtained fusion feature 8 × 8 × (32+32)=8 × 8 × 64 dimension size, is obtained feature 2；To being adopted under feature 2
Sample operation blends with VGG3 feature and obtains feature 3, and dimension is that 4 × 4 × (64+64)=4 × 4 × 128.VGG modular character melts
After conjunction, more minutias are by e-learning.In next step, fusion feature is enhanced using pyramid pond method, improves figure
As the discrimination in depth network model.
1.3 traffic sign characteristic strengthenings
Spatial pyramid pond method (Spatial Pyramid Pooling, SPP)[15]Be proposed to solve it is defeated
The problem of entering graphical rule variability, network with full articulamentum during connecting, using pyramid pond method, so that appointing
The characteristic pattern of meaning size can be converted into the feature vector of specified dimension.Assuming that CNN network model, which obtains characteristic pattern, to be drawn
It is divided into 22×22Then a characteristic area carries out characteristic area using three kinds of different size of scales (4 × 4,2 × 2,1 × 1)
It divides, available 16+4+1=21 block, the maximum value in each region is acquired using maximum pond method, thus may be used in total
To obtain the feature vector of 21 fixed dimensions.However select that fixed scale size divides characteristic pattern is nonoverlapping piece of area
The marginal information of image is lost in domain, is unfavorable for the identification of overall profile information, leads to the decline of recognition accuracy.
The present invention further enhances feature after fusion using pyramid pond method, and it is special to obtain more sufficient target
Sign completes the identification to Traffic Sign Images.Fig. 4 indicates the mistake of spatial pyramid pond fusion feature used in experimentation
Journey, merging the feature in VGG network layer is 4 × 4 × 128, wherein 128 indicate port number, fusion feature is by 3 kinds of different scales
Pondization operation, the pond method choice maximum pond pond Hua Zhi, obtaining character pair dimension vector is Isosorbide-5-Nitrae respectively, 16, always it is obtained
The feature vector of 21 dimensions, then feature vector length is 21 × 128=2688, and original fusion feature length is 2048,
More characteristic informations are obtained to express in network model.Target object ruler can be effectively adapted to using SPP method
Variability, the more flexible marginal information for obtaining image are spent, and uses the phase on the pondization reinforcement boundary and adjacent area that have overlapping
Guan Xing, the boundary between blurred picture block, edge pixel point can also provide characteristic information, be conducive to the whole of Traffic Sign Images
The extraction and identification of body profile information.On the basis of fusion feature, using pyramid pond method, more traffic can be obtained
The characteristic information of sign image, and it is transported to full articulamentum, final identification classification is completed by softmax classifier, is being connected entirely
Dropout module is added in socket part point[16], prevent the generation of over-fitting.
Below with reference to experimental result and analysis, the invention will be further described.
2 experimental results and analysis
Experiment is run on Hewlett-Packard pavilion notebook, 64 bit manipulation system of win10, CPU 2.6GHz, interior
8G is saved as, video card NVIDIA GTX 960M uses Tensorflow deep learning frame training data.
2.1 data set
Disclosed German traffic data collection (Germany Traffic Sign Recognition Benchmark,
GTSRB it is tested on), wherein training set has 39209, and test set has 12630, shares 43 class traffic signs.Original coloured silk
All include a traffic sign in chromatic graph picture, 10% boundary is had around mark, these image pixels are from 15 × 15 to 250
× 250 differ in size, and are not necessarily rectangular.Network training is gray level image in experiment, because gray level image identifies
Rate can faster complete image training, improve recognition efficiency in the case where guaranteeing and color image is very nearly the same.Deep learning
Network Recognition target image takes the image of identical size, and Traffic Sign Images are scaled 32 × 32 sizes by selection.
The design of 2.2 network models
The given Traffic Sign Recognition algorithm based on deep space network first has to determine the net for extracting characteristics of image
Network model.When extracting feature to Traffic Sign Images, the convolutional neural networks model that VGG module is added, referred to as VCNN are selected
Model, so-called VGG module be exactly utilize the convolution kernels of 33 × 3 sizes instead of single 5 × 5 convolution kernel method, make up
CNN network extracts incomplete defect to picture edge characteristic, and local acknowledgement normalization layer (Local Response is added
Normalization, LRN), prevent the generation of over-fitting.Used network model such as table 2, wherein deep learning network packet
Containing 9 convolutional layers, 3 pond layers, 3 LRN layers and 2 full articulamentums.
The network model that 1 present invention of table uses
Image inputs before network model training, needs to remove influence of the different light conditions to Traffic Sign Recognition.Choosing
Use document[6]The enhancing contrast method of middle proposition mainly includes greyscale transformation, histogram equalization and limits contrast certainly
3 kinds of preprocess methods such as adaptive equalization (CLAHE), reducing different illumination conditions influences Traffic Sign Recognition.Initial data
Collection Data1 obtains data set Data2, Data3, Data4 after greyscale transformation, histogram equalization, CLAHE image procossing,
4 kinds of different data sets are respectively delivered to be trained on CNN model and VCNN model.Table 3 provides different data collection two
Recognition result compares under kind network model, wherein it is 0.001 that initial learning rate, which is arranged, and dropout is added in full articulamentum
2 CNN model of table and VCNN model compare
Module reduces model over-fitting.The experimental results showed that 4 kinds of different data collection are equal in the recognition effect of VCNN model
Better than CNN, wherein on Data2 data set, the accuracy of Traffic Sign Recognition can at most be promoted to 2.2%, illustrate to select
The VCNN model of small convolution kernel obtains more image detail features, improves the validity of model training and the accuracy of identification.?
Highest discrimination 96.87% is obtained on Data4 data set, illustrates that traffic sign after CLAHE processing, effectively removes in image
Redundancy.
2.3STN increases image diversity
It proposes that STN network is added before VCNN network, adaptively carry out spatial alternation to image, obtaining has space not
The traffic sign of Vertic features.Local network model is designed in STN, as shown in table 3, localized network is by convolutional layer, pond layer
And full articulamentum composition, the full articulamentum of the last layer export 6 neurons, the parameter of corresponding affine transformation matrix.Utilize mind
Transformation matrix is constantly updated in backpropagation through network, is increased the diversity of traffic sign feature representation, is promoted the accurate of identification
Degree.Fig. 5 indicates that test image enters the Contrast on effect before and after STN network in 4 data set of Data, and first row is test image,
3 localized network model of table
Second row is the transformed image of STN, it can be seen that gradually eliminates image after spatial alternation e-learning
Geometry noise and background information, only retain the area-of-interest of input picture, and are input in VCNN network, complete traffic sign
Identification.
Fig. 6 and Fig. 7 indicates that VCNN network is added before and after STN network, the cost function and discrimination situation of traffic sign
Comparison, explanation in Fig. 6 (a), after STN network is added, noise and complicated background information due to removal image, network model
Convergence rate faster, optimize network model, be obtained with higher Traffic Sign Recognition rate using lesser cost value；Figure
It is provided in 6 (b), the final cost value of VCNN+STN network is 0.256, and the cost value of VCNN network is 0.275, furtherly
Bright, VCNN+STN network needs less cost value to complete Traffic Sign Recognition.Explanation in Fig. 7 (a), after STN network is added, net
Network obtains feature more with space-invariance, is conducive to traffic sign expression characteristic in a model, improves network model pair
The accuracy of Traffic Sign Recognition；It is provided in Fig. 7 (b), VCNN+STN network discrimination on training set is 99.5%, and VCNN
Network Recognition rate is 99.1%, and discrimination improves 0.4%；On test set, the model after STN network is added is in Data4 number
According to the discrimination for obtaining 97.45% on collection, and the discrimination of VCNN network is then 96.87%, and discrimination improves 0.58%.Institute
With spatial alternation network enhances the diversity of image space structure, effectively promotes the discrimination of traffic sign.Such as following table.
Traffic Sign Recognition effect before and after STN network is added in table
Table 4 Traffic sign recognition effect before and after STN network
STN network is added as can be seen from the above table in comparison before and after deep learning network layer carries out that STN network is added
Afterwards, there is good promotion on the training set core test set of traffic sign, wherein training set improves 0.4%, is promoted on test set
0.58%, illustrate that STN network, which is added, has higher discrimination to traffic sign.In addition, STN network is added in penalty values
Afterwards, it needs smaller penalty values to be obtained with traffic discrimination, illustrates that improved network can quickly identify traffic sign,
Promote recognition effect.
2.4 SPP strengthen feature
The present invention is directed to the characteristics of traffic indication map seems Small object image, and it is strong to carry out feature to traffic sign using SPP
Change.Document[6]Network layer different phase characteristic binding is chosen, Traffic Sign Recognition rate is effectively promoted.Therefore, by heterogeneous networks layer
Characteristic binding, obtain 4 × 4 × 128 dimensions feature.Select 3 kinds of different characteristic regions such as 4 × 4,2 × 2,1 × 1 to joint
Feature carries out pyramid pond, obtains the feature vector of 16+4+1=21 dimension, then the length of feature vector is 21 × 128=
Feature after reinforcing is inputted full articulamentum by 2688., completes Traffic Sign Recognition using softmax classifier.To prevent network
Over-fitting occurs, dropout module is added.As dropout=0.1, inhibit more multi-neuron, network cannot effectively identify friendship
Logical mark；As dropout=1, neuron is completely in state of activation, is easy to happen over-fitting, influences recognition effect.In order to
Estimate different dropout values, iteration 20 times on VCNN+STN+SPP network, more different dropout values are to traffic sign
The influence of identification, as shown in table 4, as dropout=0.4, discrimination highest can achieve 98.42%, than VCNN+STN's
Discrimination improves 0.97%；As dropout=0.1, discrimination minimum 97.54%, also the discrimination than VCNN+STN is promoted
0.09%, so pyramid pondization effectively enhancing characteristics of image, further promotes the accuracy of Traffic Sign Recognition.
Influence of the 4 difference dropout value of table to feature is strengthened
Table 5 and table 6 list method proposed by the present invention compared with the arithmetic result of current mainstream respectively.Table 5 provides Germany
Open traffic data collection contest first place Committee of CNNS method[17]、Multi-scale CNN[6]Method, Human
average[18]Method, GABOR+LBP+HOG[19]4 kinds of algorithm comparisons such as method, the present invention classify correctly on GTSRB data set
Rate is higher than the methods of multiple dimensioned CNN method and GABOR+LBP+HOG multiple features fusion, the slightly below side CNNS Committee of
Method and human eye Direct Recognition, wherein Committee of CNNS method utilizes multiple Network Recognition traffic signs, training time
37h is needed, especially time-consuming, real-time recognition efficiency substantially reduces in assorting process；The present invention proposes to improve CNN network model,
Completing the training required time is 102.1s, and whole test images are completed the time required to identification to be 6.2s, can be effectively right in real time
Traffic Sign Recognition.
Table 5 and classic algorithm compare
The performance of this algorithm is studied for deep layer, table 6 provides inventive algorithm and other algorithms to traffic sign Sub Data Set
Recognition result comparison, by traffic sign data set be divided into speed(-)limit sign dataset1, other prohibitory signs dataset2, release
Prohibitory sign dataset3, Warning Mark dataset4, danger signal dataset5 and other marks dataset6 etc. 6
Sub Data Set.Algorithm proposed by the present invention is to speed(-)limit sign, other prohibitory signs, Warning Mark, other marks 4 Sub Data Set
Discrimination is higher, but to lift a ban mark and danger signal discrimination lower than other classic algorithms, this is because lifting a ban
Indicate that original image is gray image, after gray processing is handled, loses many marginal informations；Danger signal original image is by contracting
Dimension normalization is discharged into, deformation occurs for traffic indication map case, and this kind of traffic sign shape is much like, Network Recognition traffic mark
Will obscures these images, classification error.And method proposed by the present invention is to ignore part edge loss of learning and image generation
Deformation bring influences, and using spatial alternation network, obtains multifarious space characteristics, SPP network is added, it is special to strengthen network layer
Sign, further increases Traffic Sign Recognition rate.
The recognition result of the different traffic sign Sub Data Sets of table 6
Below with reference to effect, the invention will be further described.
The traffic sign recognition method based on deep space network is analyzed, spatial alternation network is introduced, it is more to increase image
The space characteristics of sample remove geometry noise and extra background information, only retain the area-of-interest of image.STN net is being added
On the basis of network, by the Fusion Features of heterogeneous networks layer, SPP method is introduced, is chosen under various sizes of pondization verification fusion feature
Characteristics of image is strengthened in sampling, obtains Network Recognition traffic sign abundance feature, promotes Traffic Sign Recognition rate.
For the enhancing and acquisition of characteristics of image after dimension normalization, optimize network model, it can be with using lesser cost
It obtains to higher Traffic Sign Recognition rate.But image passes through gray processing, dimension normalization method, lost part marginal information.
Further work, it is the methods of unrelated using scale by image transfer zone others color space, more marginal informations are obtained, are mentioned
Rise the accuracy of Traffic Sign Recognition.
As shown in figure 8, the embodiment of the present invention provides a kind of Traffic Sign Recognition System based on deep space network includes:
The image collection module 1 of space-invariance, by Traffic Sign Images conveying space converting network, acquisition more contains
The image of space-invariance；
Characteristics of image processing module 2 extracts changing image in the spy of network layer different phase using improved VGG network
Sign, merges different phase feature；Fusion feature passes through spatial pyramid pond method, further enhances characteristics of image,
And it is exported as the feature of traffic sign；
Traffic sign categorization module 3 classifies to traffic sign with sotfmax classifier.
In the above-described embodiments, can come wholly or partly by software, hardware, firmware or any combination thereof real
It is existing.When using entirely or partly realizing in the form of a computer program product, the computer program product include one or
Multiple computer instructions.When loading on computers or executing the computer program instructions, entirely or partly generate according to
Process described in the embodiment of the present invention or function.The computer can be general purpose computer, special purpose computer, computer network
Network or other programmable devices.The computer instruction may be stored in a computer readable storage medium, or from one
Computer readable storage medium is transmitted to another computer readable storage medium, for example, the computer instruction can be from one
A web-site, computer, server or data center pass through wired (such as coaxial cable, optical fiber, Digital Subscriber Line (DSL)
Or wireless (such as infrared, wireless, microwave etc.) mode is carried out to another web-site, computer, server or data center
Transmission).The computer-readable storage medium can be any usable medium or include one that computer can access
The data storage devices such as a or multiple usable mediums integrated server, data center.The usable medium can be magnetic Jie
Matter, (for example, floppy disk, hard disk, tape), optical medium (for example, DVD) or semiconductor medium (such as solid state hard disk Solid
State Disk (SSD)) etc..
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (10)
1. a kind of traffic sign recognition method based on deep space network, which is characterized in that described to be based on deep space network
Traffic sign recognition method include:
Firstly, Traffic Sign Images conveying space converting network is obtained into the affine transformation matrix of image using localized network, and
Using the backpropagation of neural network, optimal affine transformation matrix is obtained, to obtain more containing the figure of space-invariance
Picture；
Then changing image is extracted in the feature of network layer different phase, to different phase feature using improved VGG network
It is merged, increases the diversity of Traffic Sign Images feature, improved VGG network is selected smaller size of convolution kernel, obtained
The minutia for obtaining more images, for being expressed in deep learning network layer；Fusion feature is by spatial pyramid pondization side
Method obtains specified size characteristic vector, identifies traffic sign in different classifiers；Meanwhile characteristics of image is further enhanced,
And it is exported as the feature of traffic sign；
Finally classified with sotfmax classifier to traffic sign.
2. the traffic sign recognition method as described in claim 1 based on deep space network, which is characterized in that image input
Before network model training, spatial alternation network is added, enhances the diversity of image space structure, and hand over for CNN Network Recognition
Logical mark provides sufficient spatial signature information；
Spatial alternation network is that affine transformation is carried out to image, passes through the transformation square of 2 × 3 dimensions during the affine transformation of image
Battle array translates Traffic Sign Images, scales, rotation；
Transformation for mula is expressed as
Wherein, θ indicates the parameter in transformation matrix, and (x, y) indicates the pixel of image, and α indicates shape of the image after rotation
State, image is successively by translation, the spatial alternation of scaling and rotation；
Traffic Sign Images are after removing the preprocess method of illumination, by the local area network operated containing multilayer convolution sum pondization
In network, the transformation matrix A1 of initial 2 × 3 dimension is obtained；Grid Sampling is established using transformation matrix A1, obtains desired output figure
As A2；Using input and output image corresponding relationship, bilinear interpolation is carried out to input picture, completing transformed image A3 is
The input of identification module；
Spatial network includes localized network, 3 parts of mesh generator and sampler；
In localized network, input is Traffic Sign ImagesWherein W, H are the width and height of input picture, C table
Show port number；Spatial alternation layer exports θ=floc(U), local network structure are as follows: input picture passes through flocFunction obtains local area network
The output of network includes 6 neurons, transformation matrix θ needed for forming image.
3. the traffic sign recognition method as claimed in claim 2 based on deep space network, which is characterized in that
Mesh generator is the sampling network of the parameter creation obtained according to localized network, is one group of point set, and input mapping is passed through
Sampling generates desired conversion output；Using the method for inverse transformation, the respective value of output pixel point is obtained, transformation for mula is as follows:
Wherein,Image input and output coordinate points are respectively indicated, θ indicates the output of localized network；To coordinate value
It is normalized, so that xi, yi∈ [- 1,1]；
Output is established by sampler to contact with upper all pixels point gray value is inputted, and is carried out using sampling and input picture
Bilinearity sampling, sampling formula indicate are as follows:
Wherein,Indicate input picture,Image after indicating output transform, H ', W ' expression sampling
The height and width of grid, C indicate that port number is identical as original input Traffic Sign Images.
4. the traffic sign recognition method as described in claim 1 based on deep space network, which is characterized in that based on improvement
The feature extracting method of VGG includes:
The convolution kernel of multiple 3 × 3 sizes is selected to replace 5 × 5 convolution kernel, to CNN network image Edge Gradient Feature；Network
Input is the gray level image after spatial alternation network processes, and 3 layers of convolutional layer and 1 layer of pond layer constitute VGG module, a total of
3 VGG modules, the number of network layer convolution kernel is 32,32,32,32,32,32,64,64,64, VGG module different phase obtains
Feature sizes are 16 × 16 × 64,8 × 8 × 64 and 4 × 4 × 64；
The Fusion Features that different VGG modules are obtained, since the dimension of feature is different, selection obtains feature to VGG1 and carries out 2 × 2
Pondization operation, feature down-sampling are 8 × 8 × 32 dimension sizes, obtain feature 1, and obtain feature with VGG2 and merged, obtain
Fusion feature 8 × 8 × (32+32)=8 × 8 × 64 dimension size, obtains feature 2；2 down-sampling of feature is operated, with VGG3 spy
Sign, which blends, obtains feature 3, and dimension is minutia quilt after the fusion of 4 × 4 × (64+64)=4 × 4 × 128.VGG modular character
E-learning；It carries out enhancing fusion feature using pyramid pond method again.
5. the traffic sign recognition method as described in claim 1 based on deep space network, which is characterized in that use golden word
Tower basin method further enhances feature after fusion, obtains more sufficient target signature, the identification to Traffic Sign Images.
6. a kind of realize described in Claims 1 to 5 any one based on the traffic sign recognition method of deep space network
Calculation machine program.
7. a kind of letter for realizing the traffic sign recognition method based on deep space network described in Claims 1 to 5 any one
Cease data processing terminal.
8. a kind of computer readable storage medium, including instruction, when run on a computer, so that computer is executed as weighed
Benefit requires the traffic sign recognition method based on deep space network described in 1-5 any one.
9. a kind of traffic sign recognition method realized based on deep space network described in claim 8 based on deep space net
The Traffic Sign Recognition System of network, which is characterized in that the Traffic Sign Recognition System based on deep space network includes:
Traffic Sign Images conveying space converting network is obtained and more contains space by the image collection module of space-invariance
The image of invariance；
Characteristics of image processing module, using improved VGG network extract changing image network layer different phase feature, it is right
Different phase feature is merged；Fusion feature passes through spatial pyramid pond method, further enhances characteristics of image, and conduct
The feature of traffic sign exports；
Traffic sign categorization module classifies to traffic sign with sotfmax classifier.
10. a kind of Traffic Sign Recognition System equipped with based on deep space network described in claim 9, traffic sign is known
Other equipment.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810751516.8A CN108985217A (en) | 2018-07-10 | 2018-07-10 | A kind of traffic sign recognition method and system based on deep space network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810751516.8A CN108985217A (en) | 2018-07-10 | 2018-07-10 | A kind of traffic sign recognition method and system based on deep space network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108985217A true CN108985217A (en) | 2018-12-11 |
Family
ID=64537579
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810751516.8A Pending CN108985217A (en) | 2018-07-10 | 2018-07-10 | A kind of traffic sign recognition method and system based on deep space network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108985217A (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635784A (en) * | 2019-01-10 | 2019-04-16 | 重庆邮电大学 | Traffic sign recognition method based on improved convolutional neural networks |
| CN109815875A (en) * | 2019-01-17 | 2019-05-28 | 柳州康云互联科技有限公司 | One kind is for versicolor system in internet detection |
| CN109829848A (en) * | 2019-01-17 | 2019-05-31 | 柳州康云互联科技有限公司 | A kind of system and method for Image space transformation neural network based in internet detection |
| CN109902609A (en) * | 2019-02-22 | 2019-06-18 | 淮阴工学院 | A kind of road traffic sign detection and recognition methods based on YOLOv3 |
| CN109902584A (en) * | 2019-01-28 | 2019-06-18 | 深圳大学 | A kind of recognition methods, device, equipment and the storage medium of mask defect |
| CN110473255A (en) * | 2019-07-17 | 2019-11-19 | 南京畅淼科技有限责任公司 | A kind of ship bollard localization method divided based on multi grid |
| CN111274971A (en) * | 2020-01-21 | 2020-06-12 | 南京航空航天大学 | Traffic identification method based on color space fusion network and space transformation network |
| CN111310837A (en) * | 2020-02-21 | 2020-06-19 | 广州华工邦元信息技术有限公司 | Vehicle refitting recognition method, device, system, medium and equipment |
| CN111476230A (en) * | 2020-03-05 | 2020-07-31 | 重庆邮电大学 | License plate positioning method for improving combination of MSER and multi-feature support vector machine |
| CN111806702A (en) * | 2020-06-30 | 2020-10-23 | 鑫喆喆 | Parachute jumping mechanism pop-up platform and method based on signal detection |
| CN112115777A (en) * | 2020-08-10 | 2020-12-22 | 杭州优行科技有限公司 | Method, device and equipment for detecting and identifying traffic sign categories |
| CN112528058A (en) * | 2020-11-23 | 2021-03-19 | 西北工业大学 | Fine-grained image classification method based on image attribute active learning |
| CN112699898A (en) * | 2020-12-29 | 2021-04-23 | 山西大学 | Image direction identification method based on multi-layer feature fusion |
| CN113052057A (en) * | 2021-03-19 | 2021-06-29 | 北京工业大学 | Traffic sign identification method based on improved convolutional neural network |
| CN113313118A (en) * | 2021-06-25 | 2021-08-27 | 哈尔滨工程大学 | Self-adaptive variable-proportion target detection method based on multi-scale feature fusion |
| CN113409393A (en) * | 2019-05-17 | 2021-09-17 | 百度在线网络技术（北京）有限公司 | Method and device for identifying traffic sign |
| CN113537131A (en) * | 2021-07-30 | 2021-10-22 | 中国科学院地理科学与资源研究所 | Land resource analysis model training method and analysis method based on image recognition |
| CN114202733A (en) * | 2022-02-18 | 2022-03-18 | 青岛海信网络科技股份有限公司 | Video-based traffic fault detection method and device |
| WO2023016290A1 (en) * | 2021-08-12 | 2023-02-16 | 北京有竹居网络技术有限公司 | Video classification method and apparatus, readable medium and electronic device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140180986A1 (en) * | 2012-12-24 | 2014-06-26 | Google Inc. | System and method for addressing overfitting in a neural network |
| CN105956532A (en) * | 2016-04-25 | 2016-09-21 | 大连理工大学 | Traffic scene classification method based on multi-scale convolution neural network |
| CN107292875A (en) * | 2017-06-29 | 2017-10-24 | 西安建筑科技大学 | A kind of conspicuousness detection method based on global Local Feature Fusion |
-
2018
- 2018-07-10 CN CN201810751516.8A patent/CN108985217A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140180986A1 (en) * | 2012-12-24 | 2014-06-26 | Google Inc. | System and method for addressing overfitting in a neural network |
| CN105956532A (en) * | 2016-04-25 | 2016-09-21 | 大连理工大学 | Traffic scene classification method based on multi-scale convolution neural network |
| CN107292875A (en) * | 2017-06-29 | 2017-10-24 | 西安建筑科技大学 | A kind of conspicuousness detection method based on global Local Feature Fusion |
Non-Patent Citations (3)
| Title |
|---|
| 乔堃等: "基于深度学习的交通标志识别", 《信息技术与信息化》 * |
| 方向: "纯干货：深度学习实现之空间变换网络-part1", 《HTTPS://YQ.ALIYUN.COM/ARTICLES/69259》 * |
| 方向: "纯干货：深度学习实现之空间变换网络-part2", 《HTTPS://YQ.ALIYUN.COM/ARTICLES/580026?SPM=A2C4E.11153940.0.0.6F3C7046YJCEBO》 * |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635784A (en) * | 2019-01-10 | 2019-04-16 | 重庆邮电大学 | Traffic sign recognition method based on improved convolutional neural networks |
| CN109815875A (en) * | 2019-01-17 | 2019-05-28 | 柳州康云互联科技有限公司 | One kind is for versicolor system in internet detection |
| CN109829848A (en) * | 2019-01-17 | 2019-05-31 | 柳州康云互联科技有限公司 | A kind of system and method for Image space transformation neural network based in internet detection |
| CN109815875B (en) * | 2019-01-17 | 2023-04-07 | 柳州康云互联科技有限公司 | System for color transformation in Internet detection |
| CN109902584A (en) * | 2019-01-28 | 2019-06-18 | 深圳大学 | A kind of recognition methods, device, equipment and the storage medium of mask defect |
| CN109902584B (en) * | 2019-01-28 | 2022-02-22 | 深圳大学 | Mask defect identification method, device, equipment and storage medium |
| CN109902609A (en) * | 2019-02-22 | 2019-06-18 | 淮阴工学院 | A kind of road traffic sign detection and recognition methods based on YOLOv3 |
| CN113409393B (en) * | 2019-05-17 | 2023-10-03 | 百度在线网络技术（北京）有限公司 | Method and device for identifying traffic sign |
| CN113409393A (en) * | 2019-05-17 | 2021-09-17 | 百度在线网络技术（北京）有限公司 | Method and device for identifying traffic sign |
| CN110473255A (en) * | 2019-07-17 | 2019-11-19 | 南京畅淼科技有限责任公司 | A kind of ship bollard localization method divided based on multi grid |
| CN111274971A (en) * | 2020-01-21 | 2020-06-12 | 南京航空航天大学 | Traffic identification method based on color space fusion network and space transformation network |
| CN111310837A (en) * | 2020-02-21 | 2020-06-19 | 广州华工邦元信息技术有限公司 | Vehicle refitting recognition method, device, system, medium and equipment |
| CN111476230B (en) * | 2020-03-05 | 2023-04-18 | 重庆邮电大学 | License plate positioning method for improving combination of MSER and multi-feature support vector machine |
| CN111476230A (en) * | 2020-03-05 | 2020-07-31 | 重庆邮电大学 | License plate positioning method for improving combination of MSER and multi-feature support vector machine |
| CN111806702A (en) * | 2020-06-30 | 2020-10-23 | 鑫喆喆 | Parachute jumping mechanism pop-up platform and method based on signal detection |
| CN112115777A (en) * | 2020-08-10 | 2020-12-22 | 杭州优行科技有限公司 | Method, device and equipment for detecting and identifying traffic sign categories |
| CN112528058B (en) * | 2020-11-23 | 2022-09-02 | 西北工业大学 | Fine-grained image classification method based on image attribute active learning |
| CN112528058A (en) * | 2020-11-23 | 2021-03-19 | 西北工业大学 | Fine-grained image classification method based on image attribute active learning |
| CN112699898A (en) * | 2020-12-29 | 2021-04-23 | 山西大学 | Image direction identification method based on multi-layer feature fusion |
| CN113052057A (en) * | 2021-03-19 | 2021-06-29 | 北京工业大学 | Traffic sign identification method based on improved convolutional neural network |
| CN113313118A (en) * | 2021-06-25 | 2021-08-27 | 哈尔滨工程大学 | Self-adaptive variable-proportion target detection method based on multi-scale feature fusion |
| CN113537131A (en) * | 2021-07-30 | 2021-10-22 | 中国科学院地理科学与资源研究所 | Land resource analysis model training method and analysis method based on image recognition |
| WO2023016290A1 (en) * | 2021-08-12 | 2023-02-16 | 北京有竹居网络技术有限公司 | Video classification method and apparatus, readable medium and electronic device |
| CN114202733A (en) * | 2022-02-18 | 2022-03-18 | 青岛海信网络科技股份有限公司 | Video-based traffic fault detection method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108985217A (en) | A kind of traffic sign recognition method and system based on deep space network | |
| Roy et al. | Handwritten isolated Bangla compound character recognition: A new benchmark using a novel deep learning approach | |
| Mo et al. | Multi-level deep supervised networks for retinal vessel segmentation | |
| Sarkhel et al. | A multi-scale deep quad tree based feature extraction method for the recognition of isolated handwritten characters of popular indic scripts | |
| CN110428428B (en) | Image semantic segmentation method, electronic equipment and readable storage medium | |
| CN107506761B (en) | Brain image segmentation method and system based on significance learning convolutional neural network | |
| Zhang et al. | DeepPap: deep convolutional networks for cervical cell classification | |
| Sun et al. | Automatic multi-fault recognition in TFDS based on convolutional neural network | |
| Zeng et al. | Multi-stage contextual deep learning for pedestrian detection | |
| CN103605972B (en) | Non-restricted environment face verification method based on block depth neural network | |
| Islalm et al. | Recognition bangla sign language using convolutional neural network | |
| CN106682569A (en) | Fast traffic signboard recognition method based on convolution neural network | |
| Li et al. | Pedestrian detection based on deep learning model | |
| CN107704859A (en) | A kind of character recognition method based on deep learning training framework | |
| Yamashita et al. | Hand posture recognition based on bottom-up structured deep convolutional neural network with curriculum learning | |
| CN110163069A (en) | Method for detecting lane lines for assisting driving | |
| CN108230330A (en) | A kind of quick express highway pavement segmentation and the method for Camera Positioning | |
| Khanday et al. | Taxonomy, state-of-the-art, challenges and applications of visual understanding: A review | |
| CN106650798A (en) | Indoor scene recognition method combining deep learning and sparse representation | |
| Ma et al. | Retinal vessel segmentation by deep residual learning with wide activation | |
| Fardous et al. | Handwritten isolated Bangla compound character recognition | |
| Zeng et al. | Convolutional Neural Network Based Multi-feature Fusion for Non-rigid 3D Model Retrieval. | |
| Wang et al. | Pedestrian recognition in multi-camera networks based on deep transfer learning and feature visualization | |
| CN116311387B (en) | Cross-modal pedestrian re-identification method based on feature intersection | |
| Li et al. | HEp-2 specimen classification with fully convolutional network |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181211 |