CN108549926A - A kind of deep neural network and training method for refining identification vehicle attribute - Google Patents
A kind of deep neural network and training method for refining identification vehicle attribute Download PDFInfo
- Publication number
- CN108549926A CN108549926A CN201810194741.6A CN201810194741A CN108549926A CN 108549926 A CN108549926 A CN 108549926A CN 201810194741 A CN201810194741 A CN 201810194741A CN 108549926 A CN108549926 A CN 108549926A
- Authority
- CN
- China
- Prior art keywords
- feature
- identification
- attribute
- neural network
- branch
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
Abstract
The invention discloses a kind of deep neural network and training method for refining identification vehicle attribute, which includes：Depth residual error network carries out feature extraction for the picture to input, obtains characteristic pattern；Feature migrating layer, including multiple feature migration units, the feature for being shared for each Attribute Recognition task immigration, are allowed to be adapted to specific task；Multiple full articulamentums, corresponding each Attribute Recognition task branch, are connected to feature migrating layer, to obtain corresponding to the feature vector of each Attribute Recognition task；Multiple costing bio disturbance units, corresponding each Attribute Recognition task branch, are separately connected each full articulamentum, for using loss function counting loss of the cross entropy as multi-categorizer；Multiple parameters updating unit, corresponding to each Attribute Recognition task, each costing bio disturbance unit is connected, is returned and is lost by stochastic gradient descent optimization algorithm, undated parameter, the present invention realize the purpose for the vehicle attribute that can only identify a variety of finings simultaneously with a neural network.
Description
Technical field
The present invention relates to computer visions and mode identification technology, more particularly to one kind for refining identification vehicle
The deep neural network and its training method of attribute.
Background technology
Vehicle fining Attribute Recognition technology is a basic fundamental of intelligent transportation safety-security area, and identification vehicle attribute can
To improve understanding of the computer to target vehicle, help to solve the problems, such as the traffic safety-security area that some are more difficult to, for example vehicle is certainly
Dynamic retrieval, vehicle identify again.
Vehicle attribute identification is computer vision and a classical problem in pattern-recognition, and the vehicle attribute of identification is general
There is the brand manufacturers etc. of the color of vehicle, the type of vehicle, vehicle, the scene of technology application is generally in road camera shooting, identification
The fining attribute of vehicular traffic, to vehicle classification labeling.The key technology of such issues that solution is that image recognition classification is calculated
Method, difficult point are illumination, scale, block influence to accuracy of identification.
The vehicle attribute identification technology that early stage uses is added with manual feature based on Machine learning classifiers, and effect is relatively good
Have using scale invariant feature (SIFT feature), histograms of oriented gradients feature (HOG) is as manual feature, with visual angle bag of words
Frame and support vector machines (SVM) solve vehicle attribute classification problem, are modeled respectively to different attributes respectively.Later everybody
Widely used deep learning solves the problems, such as image classification, by big data and powerful calculating power, with stochastic gradient descent
Optimization algorithm optimizes the up to ten million parameter of network, allows network to extract the feature of vehicle well, be different from conventional method is
The feature of depth network output can be learned, and to be much better than manual feature eventually by the feature of study out, use grader again in this way
Classify, effect is better plus the way of grader than traditional manual feature.The method of deep learning has very strong
Habit ability, as long as training data is related to all kinds of complex scenes (illumination, is blocked scale), that model just can be in these complicated fields
Scape shows excellent performance.
The Chinese patent application of Publication No. 105678275A discloses a kind of vehicle identification method, first to image data
Pretreatment, calculates the average vector of whole pictures, then calculate each training sample picture constitute vector with it is described averagely to
The difference of amount；Then the covariance matrix for calculating the training sample picture finds out feature vector according to the covariance matrix
With characteristic value and arranged in sequence, feature vector is extracted with Principal Component Analysis (PCA), finally use a BP neural network as dividing
Class device.The Chinese patent application of Publication No. 105787466A discloses fine recognition methods and the system of a kind of type of vehicle,
It carries out gray processing and standardization to the original vehicle image of acquisition；The gradient of each pixel of normalized image and
Direction；Histograms of oriented gradients feature extraction and local linear constraint are carried out to standardized images according to the gradient of calculating and direction
Coding, obtains the coding vector of standardized images；According to obtained coding vector using weights spatial pyramid to local linear
Standardized images after constraint coding are handled, and the final expression vector of vehicle image are obtained, finally with a SVM to this
Feature vector into driving classification of type.
The Chinese patent application of Publication No. 102737221B provides recognition methods and the device of a kind of vehicle color,
According to the reference zone that the texture of image and structural information positioning vehicle color identify, main identification region and auxiliary are carried out again later
The identification of identification region, main identification region and the result Nearest Neighbor with Weighted Voting for assisting in identifying region obtain the knot of final color identification
Fruit.The Chinese patent application of Publication No. 103544480A provides a kind of vehicle color identification method, by reference area
Vehicle is divided into colored vehicle and black and white silver gray vehicle by the statistics of the characteristic parameters such as domain saturation degree first；If it is determined that result is colour
Vehicle isolates colored region, and carries out color identification to this region；If it is determined that result is the silver grey vehicle of black and white, then by reference zone
Segmentation, the color of vehicle is determined by the method for ballot.The Chinese patent application of Publication No. 106203420A provides one kind
Bayonet vehicle color identification method, with a depth convolutional neural networks, but simultaneously one complete auto graph of non-input, it is defeated
It is color identification area-of-interest to enter, and output is a feature vector, then is classified to the vector with SVM or softmax.
The Chinese patent application of Publication No. 103500327A provides a kind of same brand based on spatial positional information
The characteristic area of N number of vehicle to be measured is extracted in the position of the model recognizing method of vehicle, first positioning licence plate position, opposite car plate,
The feature vector for calculating each characteristic area is used as classification, and this way can dexterously be kept away due to there is the help of car plate position
Some redundancies are opened, classification accuracy is improved.Publication No. 105930812A Chinese patent applications provide one kind and are based on melting
The vehicle brand kind identification method for closing feature sparse coding model, thinks traditional feature, such as Curvelet transformation, HOG
Feature, PHOG features, Harr features, EOH features, Gabor wavelet etc. have respective limitation, it is proposed that a kind of fusion feature
The method of sparse coding carrys out the judgement index of Enhanced feature.The Chinese patent application of Publication No. 105488517A provides a kind of base
Prepare more than 2,000 ten thousand using the high definition bayonet point collection of different cities in the vehicle brand type identifier method of deep learning
Capture picture is accurately divided into 3,000 multiclass by artificial screening according to the brand and model of vehicle, establish it is abundant cover it is various
The vehicle brand model picture database of condition.Then license plate recognition technology is recycled to be accurately positioned the car plate position in bayonet picture
It sets, vehicle face region, normalization vehicle face region is accurately estimated according to car plate location information.They devise 4 convolutional layers two
The convolutional neural networks of a full articulamentum extract the feature of this batch data, then make grader with SVM.
However, the disclosed above number China for 105678275A, 105787466A, 103500327A, 105930812A is specially
Profit application is main to characterize the type of vehicle with some statistics with the low-level feature in gradient, does not use high-rise semanteme
Feature；The Chinese patent application of Publication No. 105678275A, 105787466A, 102737221B, 103544480A have number
Data preprocess or result post-processing step, but its flow is cumbersome；The Chinese patent application of Publication No. 105488517A is one
A big data vehicle brand knows method for distinguishing, has applied to depth learning technology, but its depth network is too simple, there are four
Convolutional layer and two full articulamentums, and identifiable vehicle visual angle is single, is only limitted to the product of the monitoring vehicle pictures of front face
Board identifies.
Invention content
In order to overcome the deficiencies of the above existing technologies, purpose of the present invention is to provide one kind for refining identification vehicle
The deep neural network and its training method of attribute only can identify a variety of finings simultaneously to realize with a neural network
Vehicle attribute purpose, save calculate, improve recognition time.
In order to achieve the above object, the present invention proposes a kind of deep neural network for refining identification vehicle attribute, including：
Depth residual error network carries out feature extraction for the picture to input, obtains characteristic pattern, the feature extracted is for more
A Attribute Recognition task sharing；
Feature migrating layer, including multiple feature migration units, the feature for being shared for each Attribute Recognition task immigration,
It is allowed to be more suitable for specific task；
Multiple full articulamentums, corresponding each Attribute Recognition task branch, are connected to the feature migrating layer, corresponding each to obtain
The feature vector of Attribute Recognition task；
Multiple costing bio disturbance units, corresponding each Attribute Recognition task branch, are separately connected each full articulamentum, for using friendship
Loss function counting loss of the entropy as multi-categorizer is pitched, and after costing bio disturbance, is returned by stochastic gradient descent optimization algorithm
Pass loss, undated parameter.
Preferably, the deep neural network further includes multiple bilinearity ponds unit, connects the feature migrating layer, is used
Full articulamentum is accessed in obtaining the stronger feature vector of identification using the operation of bilinearity pondization.
Preferably for vehicle type identification, visual angle identifies, the bigger identification mission of gap, uses between color identification class
General depth network image sorting algorithm accesses a full articulamentum, exports the probability vector of a 1-D, and length is classification
Class number.
Preferably for brand/sub-brand name identification mission, is operated using the bilinearity pondization of compression, input a feature
Figure exports the stronger one-dimensional feature vector of an identification, this feature vector is then switched to length with a full articulamentum
For the probability vector of class categories number.
Preferably, each feature migration units include multiple convolutional layers, and each convolutional layer uses convolution kernel, shared to migrate
Feature, be allowed to be more suitable for specific task.
Preferably, each feature migration units include two convolutional layers, and each convolutional layer uses 1*1 convolution kernels.
In order to achieve the above objectives, the present invention also provides a kind of for refining the deep neural network of identification vehicle attribute
Training method includes the following steps：
Step 1 is trained the classification branch of a certain Attribute Recognition task and shared convolutional layer using a data set, is obtained
To the preferable pre-training model of identification；
Step 2 fixes shared convolutional layer using the pre-training model, utilizes larger of different data sets
Habit rate trains each branch；
All inconsistent data collection are integrated into and become one large-scale mixed data set together by step 3, and utilization is smaller
All identification mission of learning rate joint training, optimize whole parameters of the neural network.
Preferably, it in step 1, using data set training sub-brand name classification branch and shared convolutional layer, obtains
The preferable pre-training model of identification.
Preferably, in step 2, other inconsistent data collection training color classifications branch, vehicle classification of type point are utilized
Branch, viewpoint classification branch and brand branch.
Preferably, in step 3, the label for being not involved in counting loss with one for default attribute replaces.
Compared with prior art, a kind of deep neural network and its training for refining identification vehicle attribute of the present invention
Method realizes the purpose for the vehicle attribute that can only identify a variety of finings simultaneously with a neural network, and is not limited to list
Attribute Recognition field；Same feature is shared by more Attribute Recognition branches of the present invention, need not be each identification mission with one
A neural network extracts feature, saves and calculates, and improves recognition time；Present invention introduces feature migrating layers to make each Attribute Recognition
Branch can make full use of sharing feature, improve the accuracy rate of single branch and do not influence the recognition effect of other branches；The present invention
Support end-to-end training, frame succinctly graceful；The present invention supports the training on inconsistent data collection, supports to there is default attribute
Data are trained, strong applicability；A variety of Attribute Recognition tasks of the present invention play synergistic work when multi-task coordination is trained
With improving the accuracy rate of each branch.
Description of the drawings
Fig. 1 is a kind of system architecture diagram for refining the deep neural network of identification vehicle attribute of the present invention；
The step of Fig. 2 is a kind of training method of deep neural network being used to refine identification vehicle attribute of the present invention is flowed
Cheng Tu；
Fig. 3 is the data set A picture examples of the specific embodiment of the invention；
Fig. 4 is the data set BCD picture examples of the specific embodiment of the invention.
Specific implementation mode
Below by way of specific specific example and embodiments of the present invention are described with reference to the drawings, those skilled in the art can
Understand the further advantage and effect of the present invention easily by content disclosed in the present specification.The present invention can also pass through other differences
Specific example implemented or applied, details in this specification can also be based on different perspectives and applications, without departing substantially from
Various modifications and change are carried out under the spirit of the present invention.
Fig. 1 is a kind of system architecture diagram for refining the deep neural network of identification vehicle attribute of the present invention.Such as Fig. 1
It is shown, a kind of deep neural network for refining identification vehicle attribute of the present invention, including：
Depth residual error network 101 carries out feature extraction for the picture to input, obtains characteristic pattern, the feature extracted
For multiple Attribute Recognition task sharings.
It is 512 by input picture holding length-width ratio scaling to most short side in the specific embodiment of the invention, it is random to overturn,
The random subgraph for taking out 448*448, is input into 101 layers of depth residual error network 101, which will scheme input
32 times of piece down-sampling obtains a 14*14, the characteristic pattern that port number is 2048.
Feature migrating layer 102, including multiple feature migration units, the spy for being shared for each Attribute Recognition task immigration
Sign, is allowed to be more suitable for specific task.Specifically, each feature migration units correspond to an Attribute Recognition task, each
Feature migration units include multiple convolutional layers (in the specific embodiment of the invention, including two convolutional layers), and each convolutional layer is adopted
It is allowed to be more suitable for specific task to migrate shared feature with 1*1 convolution kernels.It is multiple in the specific embodiment of the invention
Attribute Recognition task includes brand recognition, sub-brand name identification, vehicle type identification, visual angle identification and color identification etc., but this hair
It is bright to be not limited.
That is, the characteristic pattern that depth residual error network 101 obtains is the input of five different attribute identification missions respectively.
But since the focal point that five required by task are wanted is different, for each Attribute Recognition task, the present invention is added
Two convolutional layers being made of 1*1 convolution kernels do feature migration to shared feature, generate the specific features for particular task,
Here the convolution kernel of 1*1 only weights the value in the different channels of each position again, does not influence each other between different location, also not
Original 14*14 resolution ratio can be maintained to former characteristic pattern down-sampling.
Multiple bilinearity ponds unit 103, connection features migrating layer 102, for being sentenced using the operation of bilinearity pondization
The other stronger feature vector of property, to adapt to for for brand and the such fine granularity identification mission of sub-brand name identification.
Multiple full articulamentums 104, corresponding each Attribute Recognition task branch, are connected to feature migrating layer 102 or bilinearity pond
Change unit 103, to obtain corresponding to the feature vector of each Attribute Recognition task.Specifically, for vehicle type, visual angle, color etc.
Identification mission obtains feature vector using full articulamentum 104, and such fine granularity identification mission is identified for brand and sub-brand name,
It then first passes through bilinearity pond unit 103 and feature vector is obtained by full articulamentum 104 again.
Multiple costing bio disturbance units 105, corresponding each Attribute Recognition task branch, are separately connected each full articulamentum 104, are used for
Using softmax cross entropies as loss function, counting loss, and after costing bio disturbance, is optimized by stochastic gradient descent and calculated
Method passback loss, undated parameter.
Specifically, the calculating of Softmax intersections entropy loss is as follows：Softmax intersects entropy loss and calculates in two steps, counts first
The softmax functions for calculating probability vector, then calculate intersection entropy loss again.It is assumed that z be input counting loss probability to
Amount, f (z) is the output of softmax, then：
The true classification of sample is corresponded to assuming that y is z, loss function is defined as：
L (y, z)=- log f (zy)
Above formula is substituted into obtain：
To above formula derivation, then have：
In the specific embodiment of the invention, the gradient calculated using Back Propagation Algorithm passback above formula, echelon update is each
The weight and bias of neuron.The principle of Back Propagation Algorithm is using the chain rule differentiated, step by step from damage
Function derivation forward is lost, until obtaining the gradient of the weight of each neuron and biasing about loss function in neural network, i.e.,
Derivative.Then can training parameter (weight and biasing) along gradient direction using learning rate as step-length updated value, formula is as follows：
Wherein α refers to learning rate, and w is neuron weight, and b is neuron biasing.
That is, for vehicle type identification, visual angle identification, color identifies that the identification that gap is bigger between these three classes is appointed
Business, the present invention use general depth network image sorting algorithm, access a full articulamentum 104, export a 1-D (one-dimensional)
Probability vector, length be classification class number (class categories number here is determined by task.For example it needs to be known with the model
Other 10 kinds of colors, 12 kinds of vehicle types, 5 visual angles, 100 brands, 300 sub-brand names.So classification class of color classification branch
Shuo not be 10, the class categories number of vehicle type is 12, and the class categories number at visual angle is 5, and the class categories number of brand is 100, son
The class categories number of brand is 300), softmax then to be used to intersect entropy function as loss function, calculate prediction classification and to
The supervision message given compares, and is returned and is lost by stochastic gradient descent optimization algorithm, undated parameter；For brand/sub-brand name
The characteristics of identification mission, this generic task is that gap is smaller than gap in class between class, uses the identification of general image classification algorithms accurate
True rate is undesirable, therefore present invention uses the operations of the bilinearity pondization of compression, input the characteristic pattern of a 2048*14*14, defeated
Go out the extremely strong 1-D feature vectors of an identification, then length 32768 is switched to this feature vector using a full articulamentum
Length is the probability vector of class categories number, uses softmax cross entropies as loss function, counting loss, and pass through boarding steps
Degree declines undated parameter.
The step of Fig. 2 is a kind of training method of deep neural network being used to refine identification vehicle attribute of the present invention is flowed
Cheng Tu.As shown in Fig. 2, a kind of training method for refining the deep neural network of identification vehicle attribute of the present invention, including
Following steps：
Step S201 trains the classification branch of a certain Attribute Recognition task and shared convolutional layer using a data set,
Obtain the relatively good pre-training model of identification.
The present invention supports the training on inconsistent data collection, often will appear the feelings of inconsistent data in practical applications
Condition, i.e., certain data only have the label of certain classifications, but lack the label of some classifications, for example data set A only has color and vehicle
Type label, data set B have the label at visual angle and brand sub-brand name.The method of the present invention can train on these data sets,
And each Branch Tasks can obtain good effect, for example, utilizing data set B training sub-brand name classification branch and shared
Convolutional layer obtains the relatively good pre-training model of identification, extensive to other tasks because the class categories of sub-brand name are most wide
Ability is best, uses the model as subsequent pre-training model.Specific method is：Shield color, vehicle type, visual angle, brand point
Class branch individually trains the parameter of the parameter and shared network of sub-brand name classification branch, until being reached in sub-brand name classification task
To best effects.
Step S202 fixes shared convolutional layer using the pre-training model, using different data sets with larger
Learning rate trains each branch.For example, using data set A training color classifications branch, training vehicle classification of type branch utilizes number
According to training viewpoint classification branch is collected, training brand classification branch trains sub-brand name classification branch.In the specific embodiment of the invention
In, larger learning rate refers between 0.01 to 0.001, and training process is fixed inclusion layer and other irrelevant branches layer
Parameter, only train component parameter.
All inconsistent data collection are integrated into and become one large-scale mixed data set together by step S203, and using compared with
All identification missions of small learning rate joint training, optimize whole parameters of the neural network, for default attribute with one
A label for being not involved in counting loss replaces, such as 255.Here smaller learning rate refers between 0.001 to 0.0001,
Whole parameters of training whole network.
By the following specific examples further illustrate the invention：
Embodiment one：The embodiment of a variety of vehicles fining attribute recognition approach based on VGG16 includes data standard
Standby, environment configurations, model training and model measurement four-stage.
1.1 data preparation stage：
In the present embodiment, inconsistent data training is carried out on several different data sets：Data set A has 160,000 figures
The attribute of piece, mark has vehicle type, brand, sub-brand name；Data set B has the auto graph at 25000 monitoring visual angles, the category of mark
Property has vehicle type, brand, sub-brand name；Data set C has 15000 monitoring visual angle auto graphs, the attribute of mark to have color；Data
Collection D has 26000 monitoring multi-perspective pictures, the attribute of mark to have visual angle.Data set A is high definition auto graph, as shown in figure 3, number
It is monitoring visual angle auto graph according to collection BCD, as shown in Figure 4.
1.2 environmental preparation stages：
The present embodiment is developed based on deep learning frame Caffe, needs to have configured experiment according to the study course of the official websites Caffe
Environment.Key relies on version：CUDA 8.0, CUDNN V5, OpenCV 2.4.9, Python 2.7.2.
1.3. model training:
Specifically, model training step is divided into two parts, individually trains each branch first, then carries out joint fine tuning,
It has used multiple inconsistent data collection and has participated in training.
1.3.1 input data pre-processes
Random Level overturning is carried out to the data of input, it is 512 then to keep length-width ratio scaling, scaling to most short side, most
Random crop goes out 448*448 subgraphs afterwards, subtracts the gray average of each Color Channel, is input into network.Specific practice is：Assuming that figure
Piece a height of h at this time, width w, every time [0, w-448) and [0, h-448) random round numbers x and y in range, on this figure
It is (x, y) to take out upper left corner starting point, and the coordinate of lower right corner terminal is the subgraph of (x+448, y+448).
1.3.2 model structure
In the present embodiment, model structure is removed as shown in Figure 1, wherein replace depth residual error network with VGG16 networks
The last one pond layer of VGG16, the feature for taking out conv5_3 are input into each Attribute Recognition branch.
There are two feature migrating layers before each Attribute Recognition branch to filter shared feature, the accomplished task
Specific features, feature migrating layer is made of the convolutional layer of the convolution kernel of 1*1, the port number of output and the port number phase of input
Together.
For vehicle type identification, visual angle identification, color identifies these three Attribute Recognition tasks, accesses a full articulamentum,
The probability vector of a 1-D is exported, length is the class number of classification.Then softmax is used to intersect entropy function as loss letter
Number.For brand/sub-brand name identification mission, the bilinearity pondization operation of compression is used, the feature of a 512*28*28 is inputted
Figure exports the extremely strong 1-D feature vectors of a discriminant line, length 8192.Then this feature vector is turned with a full articulamentum
For the probability vector that length is class categories number, use softmax cross entropies as loss function.
1.3.3 each branch is individually trained
First with the big large data sets A training sub-brand name classification branches of data volume and shared convolutional layer, sentenced
The relatively good pre-training model of other property.
Then the pre-training model is utilized, shared convolutional layer is fixed, each branch is trained with different data sets.Tool
Body, with data set B training brand classification branch, training sub-brand name classification branch, training vehicle classification of type branch.Use data set
C trains color classification branch.With data set D training viewpoint classifications branch.
1.3.4 joint fine tuning
All data sets are integrated together and become one large-scale mixed data set, all there was only the attribute of part per pictures
It marks, not whole attribute labelings.Here the method handled be by default attribute labeling be one counting loss when
Wait the label ignored, for example assume to ignore 255 this label in cross entropy costing bio disturbance, i.e., to label be 255 classification not
Counting loss does not return gradient.
Optimize the whole of this network on this mixed data set using a smaller learning rate (such as 0.0001)
Parameter finds that the recognition accuracy of each task has promotion, it was demonstrated that it is special to share depth convolution between each Attribute Recognition branch
Sign, which can not only play, improves computational efficiency (not having to calculate a depth characteristic figure for each task), can also play complementary raising
Effect.
1.4 model measurement.
Model measurement process includes data input processing, as a result model prediction exports three parts.
Input processing improves the accuracy rate of model prediction using the method for multiple crop (cutting).To picture to be tested into
It is 512 that row, which keeps the ratio of width to height scaling, scaling to most short side,.The upper left corner on the picture of this 512*x, the lower left corner, the upper right corner,
The lower right corner, the intermediate subgraph for all distinguishing mono- 448*448 of crop.Flip horizontal is carried out to the picture of 512*x, similarly on its left side
Upper angle, the lower left corner, the upper right corner, the lower right corner, the intermediate subgraph for all distinguishing mono- 448*448 of crop.Here 10 are generated in total
They are formed a batch, are input into network by the subgraph of 448*448.
With trained model before, the batch (batch processing) that a batch_size is 10 is inputted, each attribute is exported
Classification results.Model is that each categorical attribute of every subgraph predicts a probability vector, by the knot of this 10 subgraphs
Fruit is averaged, and obtains final classification results.
Embodiment two：The embodiment of a variety of vehicles fining attribute recognition approach based on ResNet-101 includes number
According to preparation, environment configurations, model training and model measurement four-stage：
2.1 data preparation stage.
The present invention carries out inconsistent data training on several different data sets.Data set A has 160,000 pictures, mark
Attribute have vehicle type, brand, sub-brand name；Data set B has the auto graph at 25000 monitoring visual angles, the attribute of mark to have vehicle
Type, brand, sub-brand name；Data set C has 15000 monitoring visual angle auto graphs, the attribute of mark to have color；Data set D has
26000 monitoring multi-perspective pictures, the attribute of mark have visual angle.Data set A is high definition auto graph, as shown in figure 3, data set
BCD is monitoring visual angle auto graph, as shown in Figure 4.
2.2 environmental preparation stages.
The present embodiment is developed based on deep learning frame Caffe, needs to have configured experiment according to the study course of the official websites Caffe
Environment.Key relies on version：CUDA 8.0, CUDNN V5, OpenCV 2.4.9, Python 2.7.2.
2.3 model training.
The model training step of the present embodiment is divided into two parts, individually trains each branch first, then carries out combining micro-
It adjusts, has used multiple inconsistent data collection and participated in training.
2.3.1 input data pre-processes
Random Level overturning is carried out to the data of input, it is 512 then to keep length-width ratio scaling, scaling to most short side, most
Random crop goes out 448*448 subgraphs afterwards, subtracts the gray average of each Color Channel, is input into network.Specific practice is：Assuming that figure
Piece a height of h at this time, width w, every time [0, w-448) and [0, h-448) random round numbers x and y in range, on this figure
It is (x, y) to take out upper left corner starting point, and the coordinate of lower right corner terminal is the subgraph of (x+448, y+448).
2.3.2 model structure
In the present embodiment, model structure is as shown in Figure 1, shared 101 layers of the depth residual error of all properties classification branch
Network, the feature for taking out res5c are input into each Attribute Recognition branch.
There are two feature migrating layers before each Attribute Recognition branch to filter shared feature, obtain doing this task
Specific features, feature migrating layer is made of the convolutional layer of the convolution kernel of 1*1, the port number of output and the port number phase of input
Together.
For vehicle type identification, visual angle identification, color identifies these three Attribute Recognition tasks, accesses a full articulamentum,
The probability vector of a 1-D is exported, length is the class number of classification.Then softmax is used to intersect entropy function as loss letter
Number.For brand/sub-brand name identification mission, the bilinearity pondization operation of compression is used, the spy of a 2048*14*14 is inputted
Sign figure, exports the extremely strong 1-D feature vectors of a discriminant line, length 32768.Then with a full articulamentum by this feature vector
Switch to the probability vector that length is class categories number, uses softmax cross entropies as loss function.
2.3.3 each branch is individually trained
First with the big large data sets A training sub-brand name classification branches of data volume and shared convolutional layer, sentenced
The relatively good pre-training model of other property.
Then this pre-training model is utilized, shared convolutional layer is fixed, each branch is trained with different data sets.
With data set B training brand classification branch, training sub-brand name classification branch, training vehicle classification of type branch.It is trained with data set C
Color classification branch.With data set D training viewpoint classifications branch.
2.3.4 joint fine tuning
All data sets are integrated together and become one large-scale mixed data set, all there was only the attribute of part per pictures
It marks, not whole attribute labelings.Here the method that we are handled is for one by default attribute labeling in counting loss
When the label ignored, for example we assume that ignore 255 this label in cross entropy costing bio disturbance, i.e., be 255 to label
Classification not counting loss, do not return gradient.
Optimize whole parameters of this network on this mixed data set with a smaller learning rate (0.0001), sends out
The recognition accuracy of existing each task has promotion, it was demonstrated that sharing depth convolution feature between each Attribute Recognition branch can not only
Play the role of improving computational efficiency (not having to calculate a depth characteristic figure for each task), can also play complementary raising.
2.4 model measurement.
Model measurement process includes data input processing, as a result model prediction exports three parts.
Input processing improves the accuracy rate of model prediction using the method for multiple crop.Picture to be tested is kept
The ratio of width to height scaling, scaling to most short side are 512.The upper left corner on the picture of this 512*x, the lower left corner, the upper right corner, bottom right
Angle, the intermediate subgraph for all distinguishing mono- 448*448 of crop.Flip horizontal is carried out to the picture of 512*x, similarly in its upper left
Angle, the lower left corner, the upper right corner, the lower right corner, the intermediate subgraph for all distinguishing mono- 448*448 of crop.Here 10 448* are generated in total
They are formed a batch, are input into network by 448 subgraph.
With trained model before, the batch that a batch_size is 10 is inputted, the classification knot of each attribute is exported
Fruit.Model is that each categorical attribute of every subgraph predicts a probability vector, we do the result of this 10 subgraphs
It is average, obtain final classification results.
Compared with prior art, the invention has the advantages that：
(1) vehicle attribute that can only identify a variety of finings simultaneously with a neural network, is not limited to single Attribute Recognition
Field；
(2) same feature is shared by more Attribute Recognition branches, need not be one neural network of each identification mission
Feature is extracted, saves and calculates, improves recognition time；
(3) introduced feature migrating layer enables each Attribute Recognition branch to make full use of sharing feature, improves single branch
Accuracy rate do not influence the recognition effects of other branches again；
(4) support end-to-end training, frame succinctly graceful；
(5) it supports the training on inconsistent data collection, supports to having the training of the data of default attribute, strong applicability；
(6) a variety of Attribute Recognition tasks play the role of synergistic when multi-task coordination is trained, and improve each branch
Accuracy rate.
The above-described embodiments merely illustrate the principles and effects of the present invention, and is not intended to limit the present invention.Any
Field technology personnel can without violating the spirit and scope of the present invention, and modifications and changes are made to the above embodiments.Therefore,
The scope of the present invention, should be as listed in the claims.
Claims (10)
1. a kind of deep neural network for refining identification vehicle attribute, including：
Depth residual error network carries out feature extraction for the picture to input, obtains characteristic pattern, the feature extracted supplies multiple categories
Property identification mission it is shared；
Feature migrating layer, including multiple feature migration units, the feature for being shared for each Attribute Recognition task immigration, are allowed to
It is more suitable for specific task；
Multiple full articulamentums, corresponding each Attribute Recognition task branch, are connected to the feature migrating layer, to obtain corresponding to each attribute
The feature vector of identification mission；
Multiple costing bio disturbance units, corresponding each Attribute Recognition task branch, are separately connected each full articulamentum, for using cross entropy
As the loss function counting loss of multi-categorizer, and after costing bio disturbance, is returned and damaged by stochastic gradient descent optimization algorithm
It loses, undated parameter.
2. a kind of deep neural network for refining identification vehicle attribute as described in claim 1, it is characterised in that：Institute
It further includes multiple bilinearity ponds unit to state deep neural network, the feature migrating layer is connected, for using bilinearity pond
Operation obtains the stronger feature vector of identification and accesses full articulamentum.
3. a kind of deep neural network for refining identification vehicle attribute as described in claim 1, it is characterised in that：It is right
In vehicle type identification, visual angle identifies, the bigger identification mission of gap, uses general depth network image between color identification class
Sorting algorithm accesses a full articulamentum, exports an one-dimensional probability vector, and length is the class number of classification.
4. a kind of deep neural network for refining identification vehicle attribute as claimed in claim 2, it is characterised in that：It is right
In brand/sub-brand name identification mission, is operated using the bilinearity pondization of compression, input a characteristic pattern, export an identification
Extremely strong one-dimensional feature vector, then with a full articulamentum by this feature vector switch to probability that length is class categories number to
Amount.
5. a kind of deep neural network for refining identification vehicle attribute as described in claim 1, it is characterised in that：Often
A feature migration units include multiple convolutional layers, and each convolutional layer uses convolution kernel, to migrate shared feature, is allowed to more adapt to
In specific task.
6. a kind of deep neural network for refining identification vehicle attribute as claimed in claim 5, it is characterised in that：Often
A feature migration units include two convolutional layers, and each convolutional layer uses 1*1 convolution kernels.
7. a kind of training method for refining the deep neural network of identification vehicle attribute includes the following steps：
Step 1 is trained the classification branch of a certain Attribute Recognition task and shared convolutional layer using a data set, is sentenced
The other preferable pre-training model of property；
Step 2 fixes shared convolutional layer using the pre-training model, utilizes the larger learning rate of different data sets
The each branch of training；
All inconsistent data collection are integrated into and become one large-scale mixed data set together, and utilize smaller by step 3
All identification missions of habit rate joint training optimize whole parameters of the neural network.
8. a kind of training method for refining the deep neural network of identification vehicle attribute as claimed in claim 7,
It is characterized in that：In step 1, using data set training sub-brand name classification branch and shared convolutional layer, identification is obtained
Preferable pre-training model.
9. a kind of training method for refining the deep neural network of identification vehicle attribute as claimed in claim 7,
It is characterized in that：In step 2, other inconsistent data collection training color classifications branch, vehicle classification of type branch, visual angle are utilized
Classification branch and brand branch.
10. a kind of training method for refining the deep neural network of identification vehicle attribute as claimed in claim 7,
It is characterized in that：In step 3, the label for being not involved in counting loss with one for default attribute replaces.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810194741.6A CN108549926A (en) | 2018-03-09 | 2018-03-09 | A kind of deep neural network and training method for refining identification vehicle attribute |
| PCT/CN2018/095060 WO2019169816A1 (en) | 2018-03-09 | 2018-07-10 | Deep neural network for fine recognition of vehicle attributes, and training method thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810194741.6A CN108549926A (en) | 2018-03-09 | 2018-03-09 | A kind of deep neural network and training method for refining identification vehicle attribute |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108549926A true CN108549926A (en) | 2018-09-18 |
Family
ID=63516097
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810194741.6A Pending CN108549926A (en) | 2018-03-09 | 2018-03-09 | A kind of deep neural network and training method for refining identification vehicle attribute |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN108549926A (en) |
| WO (1) | WO2019169816A1 (en) |
Cited By (38)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109472360A (en) * | 2018-10-30 | 2019-03-15 | 北京地平线机器人技术研发有限公司 | Update method, updating device and the electronic equipment of neural network |
| CN109685110A (en) * | 2018-11-28 | 2019-04-26 | 北京陌上花科技有限公司 | Training method, image classification method and device, the server of image classification network |
| CN109753999A (en) * | 2018-12-21 | 2019-05-14 | 西北工业大学 | For the fine granularity model recognizing method of visual angle auto graph |
| CN109816001A (en) * | 2019-01-10 | 2019-05-28 | 高新兴科技集团股份有限公司 | A kind of more attribute recognition approaches of vehicle based on deep learning, device and equipment |
| CN109978132A (en) * | 2018-12-24 | 2019-07-05 | 中国科学院深圳先进技术研究院 | A kind of neural network method and system refining vehicle identification |
| CN110070082A (en) * | 2019-04-22 | 2019-07-30 | 苏州科达科技股份有限公司 | Licence plate recognition method, device, equipment and storage medium |
| CN110135437A (en) * | 2019-05-06 | 2019-08-16 | 北京百度网讯科技有限公司 | For the damage identification method of vehicle, device, electronic equipment and computer storage medium |
| CN110147707A (en) * | 2018-10-25 | 2019-08-20 | 初速度（苏州）科技有限公司 | A kind of high-precision vehicle identification method and system |
| CN110209857A (en) * | 2019-05-23 | 2019-09-06 | 厦门市美亚柏科信息股份有限公司 | The more attribute recognition approaches of vehicle, device, medium based on neural network structure search |
| CN110210401A (en) * | 2019-06-03 | 2019-09-06 | 多维协同人工智能技术研究院(重庆)有限公司 | A kind of Intelligent target detection method under dim light |
| CN110321970A (en) * | 2019-07-11 | 2019-10-11 | 山东领能电子科技有限公司 | A kind of fine-grained objective classification method of multiple features based on branch neural network |
| CN110334572A (en) * | 2019-04-04 | 2019-10-15 | 南京航空航天大学 | The fine recognition methods of vehicle under a kind of multi-angle |
| CN110378237A (en) * | 2019-06-21 | 2019-10-25 | 浙江工商大学 | Human facial expression recognition method based on depth measure converged network |
| CN110399828A (en) * | 2019-07-23 | 2019-11-01 | 吉林大学 | A kind of vehicle recognition methods again of the depth convolutional neural networks based on multi-angle |
| CN110472664A (en) * | 2019-07-17 | 2019-11-19 | 杭州有盾网络科技有限公司 | A kind of certificate image identification method, device and equipment based on deep learning |
| CN110533024A (en) * | 2019-07-10 | 2019-12-03 | 杭州电子科技大学 | Biquadratic pond fine granularity image classification method based on multiple dimensioned ROI feature |
| CN110598749A (en) * | 2019-08-13 | 2019-12-20 | 广东数鼎科技有限公司 | Image vehicle recognition implementation system based on convolutional neural network AI algorithm |
| CN110598621A (en) * | 2019-09-06 | 2019-12-20 | 高新兴科技集团股份有限公司 | Vehicle brand identification method, device, equipment and computer readable storage medium |
| CN110991349A (en) * | 2019-12-05 | 2020-04-10 | 中国科学院重庆绿色智能技术研究院 | Lightweight vehicle attribute identification method based on metric learning |
| CN111091165A (en) * | 2020-03-25 | 2020-05-01 | 北京精诊医疗科技有限公司 | Fine-grained identification method of image |
| CN111126224A (en) * | 2019-12-17 | 2020-05-08 | 成都通甲优博科技有限责任公司 | Vehicle detection method and classification recognition model training method |
| CN111222409A (en) * | 2019-11-26 | 2020-06-02 | 北京迈格威科技有限公司 | Vehicle brand labeling method, device and system |
| CN111275061A (en) * | 2018-12-04 | 2020-06-12 | 北京嘀嘀无限科技发展有限公司 | Vehicle attribute identification method and model training method and device thereof, and electronic equipment |
| CN111401251A (en) * | 2020-03-17 | 2020-07-10 | 北京百度网讯科技有限公司 | Lane line extraction method and device, electronic equipment and computer-readable storage medium |
| CN111598000A (en) * | 2020-05-18 | 2020-08-28 | 中移（杭州）信息技术有限公司 | Face recognition method, device, server and readable storage medium based on multiple tasks |
| CN111612855A (en) * | 2020-04-09 | 2020-09-01 | 北京旷视科技有限公司 | Object color identification method and device and electronic equipment |
| CN111754784A (en) * | 2020-06-23 | 2020-10-09 | 高新兴科技集团股份有限公司 | Attention mechanism-based vehicle main and sub brand identification method of multilayer network |
| CN111814857A (en) * | 2020-06-29 | 2020-10-23 | 浙江大华技术股份有限公司 | Target re-identification method, network training method thereof and related device |
| CN111860068A (en) * | 2019-04-30 | 2020-10-30 | 四川大学 | Fine-grained bird identification method based on cross-layer simplified bilinear network |
| CN112016617A (en) * | 2020-08-27 | 2020-12-01 | 中国平安财产保险股份有限公司 | Fine-grained classification method and device and computer-readable storage medium |
| CN111325739B (en) * | 2020-02-28 | 2020-12-29 | 推想医疗科技股份有限公司 | Method and device for detecting lung focus and training method of image detection model |
| CN112200193A (en) * | 2020-12-03 | 2021-01-08 | 中国科学院自动化研究所 | Distributed license plate recognition method, system and device based on multi-attribute fusion |
| CN112541374A (en) * | 2019-09-20 | 2021-03-23 | 南京行者易智能交通科技有限公司 | Passenger attribute acquisition method and device based on deep learning and model training method |
| CN112766349A (en) * | 2021-01-12 | 2021-05-07 | 齐鲁工业大学 | Object description generation method based on machine vision and tactile perception |
| CN112861724A (en) * | 2021-02-06 | 2021-05-28 | 安徽理工大学 | Traditional Chinese medicine identification system |
| CN113221982A (en) * | 2021-04-28 | 2021-08-06 | 深圳市捷顺科技实业股份有限公司 | Vehicle identification method, vehicle identification model creation method and related components |
| CN114882448A (en) * | 2022-04-01 | 2022-08-09 | 北京卓视智通科技有限责任公司 | Vehicle monitoring method and electronic equipment |
| CN115294644A (en) * | 2022-06-24 | 2022-11-04 | 北京昭衍新药研究中心股份有限公司 | Rapid monkey behavior identification method based on 3D convolution parameter reconstruction |
Families Citing this family (43)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110674884A (en) * | 2019-09-30 | 2020-01-10 | 山东浪潮人工智能研究院有限公司 | Image identification method based on feature fusion |
| CN110852417B (en) * | 2019-10-10 | 2023-04-07 | 华东师范大学 | Single-depth neural network model robustness improving method for application of Internet of things |
| CN111860087A (en) * | 2019-10-16 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | Information detection method, information detection device, computer device, and storage medium |
| CN110929610B (en) * | 2019-11-12 | 2023-05-23 | 上海五零盛同信息科技有限公司 | Plant disease identification method and system based on CNN model and transfer learning |
| CN110909742A (en) * | 2019-11-19 | 2020-03-24 | 上海眼控科技股份有限公司 | License plate detection method, system, platform and storage medium |
| CN111008694B (en) * | 2019-12-02 | 2023-10-27 | 许昌北邮万联网络技术有限公司 | Depth convolution countermeasure generation network-based data model quantization compression method |
| CN110956214B (en) * | 2019-12-03 | 2023-10-13 | 北京车和家信息技术有限公司 | Training method and device for automatic driving vision positioning model |
| CN111079584A (en) * | 2019-12-03 | 2020-04-28 | 东华大学 | Rapid vehicle detection method based on improved YOLOv3 |
| CN111401122B (en) * | 2019-12-27 | 2023-09-26 | 航天信息股份有限公司 | Knowledge classification-based complex target asymptotic identification method and device |
| TWI755669B (en) * | 2020-01-02 | 2022-02-21 | 中華電信股份有限公司 | Homogeneous vehicle retrieval system and method |
| CN111314934B (en) * | 2020-02-14 | 2021-08-10 | 西北工业大学 | Network cooperative detection method for unified optimal decision |
| CN111429465B (en) * | 2020-03-18 | 2023-05-23 | 浙江科技学院 | Parallax-cleaning-based binary residual binocular significant object image segmentation method |
| CN111310862B (en) * | 2020-03-27 | 2024-02-09 | 西安电子科技大学 | Image enhancement-based deep neural network license plate positioning method in complex environment |
| CN111695587B (en) * | 2020-04-01 | 2024-04-16 | 新华三技术有限公司 | Vehicle brand and model identification method and device |
| CN111598948B (en) * | 2020-04-03 | 2024-02-20 | 上海嘉奥信息科技发展有限公司 | CT image vertebral arch root nail implantation channel planning method and system based on deep learning |
| CN111476307B (en) * | 2020-04-13 | 2023-06-20 | 河北工业大学 | Lithium battery surface defect detection method based on depth field adaptation |
| CN111738302B (en) * | 2020-05-28 | 2023-06-20 | 华南理工大学 | System for classifying and diagnosing Alzheimer's disease based on multi-modal data |
| CN111696101A (en) * | 2020-06-18 | 2020-09-22 | 中国农业大学 | Light-weight solanaceae disease identification method based on SE-Inception |
| CN111881739A (en) * | 2020-06-19 | 2020-11-03 | 安徽清新互联信息科技有限公司 | Automobile tail lamp state identification method |
| CN111967492A (en) * | 2020-06-29 | 2020-11-20 | 北京百度网讯科技有限公司 | Method and device for training classification model, electronic equipment and storage medium |
| CN111881987A (en) * | 2020-07-31 | 2020-11-03 | 西安工业大学 | Apple virus identification method based on deep learning |
| CN112149500B (en) * | 2020-08-14 | 2023-11-17 | 浙江大学 | Face recognition small sample learning method with partial shielding |
| CN111967528B (en) * | 2020-08-27 | 2023-12-26 | 北京大学 | Image recognition method for deep learning network structure search based on sparse coding |
| CN112149720A (en) * | 2020-09-09 | 2020-12-29 | 南京信息工程大学 | Fine-grained vehicle type identification method |
| CN112508078A (en) * | 2020-12-02 | 2021-03-16 | 携程旅游信息技术（上海）有限公司 | Image multitask multi-label identification method, system, equipment and medium |
| CN112464846B (en) * | 2020-12-03 | 2024-04-02 | 武汉理工大学 | Automatic recognition method for abnormal faults of freight train carriage of station |
| CN112613536A (en) * | 2020-12-08 | 2021-04-06 | 燕山大学 | Near infrared spectrum diesel grade identification method based on SMOTE and deep learning |
| CN112818736A (en) * | 2020-12-10 | 2021-05-18 | 西南交通大学 | Emergency command big data supporting platform |
| CN112766493B (en) * | 2021-01-19 | 2023-04-07 | 北京市商汤科技开发有限公司 | Training method and device for multitask neural network, electronic equipment and storage medium |
| CN113011436A (en) * | 2021-02-26 | 2021-06-22 | 北京工业大学 | Traditional Chinese medicine tongue color and fur color collaborative classification method based on convolutional neural network |
| CN112884653A (en) * | 2021-03-01 | 2021-06-01 | 西北大学 | Broken terracotta warriors fragment splicing method and system based on fracture surface information |
| CN113094808A (en) * | 2021-03-22 | 2021-07-09 | 大连理工大学 | Simulation data and artificial intelligence based automobile collision damage grade real-time prediction method |
| CN113468978B (en) * | 2021-05-26 | 2023-11-24 | 北京邮电大学 | Fine granularity car body color classification method, device and equipment based on deep learning |
| CN113780535A (en) * | 2021-09-27 | 2021-12-10 | 华中科技大学 | Model training method and system applied to edge equipment |
| CN113963231A (en) * | 2021-10-15 | 2022-01-21 | 中国石油大学(华东) | Pedestrian attribute identification method based on image enhancement and sample balance optimization |
| CN114155496B (en) * | 2021-11-29 | 2024-04-26 | 西安烽火软件科技有限公司 | Vehicle attribute multitasking collaborative recognition method based on self-attention |
| CN114237046B (en) * | 2021-12-03 | 2023-09-26 | 国网山东省电力公司枣庄供电公司 | Partial discharge pattern recognition method based on SIFT data feature extraction algorithm and BP neural network model |
| CN114694171B (en) * | 2022-02-22 | 2023-10-10 | 电子科技大学 | Pedestrian attribute identification method based on self-supervision mode feature enhancement |
| CN114881940A (en) * | 2022-04-21 | 2022-08-09 | 北京航空航天大学 | Method for identifying head defects of high-temperature alloy bolt after hot heading |
| CN115083169B (en) * | 2022-06-14 | 2023-07-18 | 公安部交通管理科学研究所 | Method for discovering suspected vehicles of counterfeit ambulances |
| CN115731436B (en) * | 2022-09-21 | 2023-09-26 | 东南大学 | Highway vehicle image retrieval method based on deep learning fusion model |
| CN117315430B (en) * | 2023-11-28 | 2024-03-12 | 华侨大学 | Incomplete modal feature fusion method for large-range vehicle re-identification |
| CN117409298B (en) * | 2023-12-15 | 2024-04-02 | 西安航空学院 | Multi-size target accurate identification method and equipment for road surface vehicle identification |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106599869A (en) * | 2016-12-22 | 2017-04-26 | 安徽大学 | Vehicle attribute identification method based on multi-task convolutional neural network |
| CN107145857A (en) * | 2017-04-29 | 2017-09-08 | 深圳市深网视界科技有限公司 | Face character recognition methods, device and method for establishing model |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106326893A (en) * | 2016-08-25 | 2017-01-11 | 安徽水滴科技有限责任公司 | Vehicle color recognition method based on area discrimination |
| CN106650813B (en) * | 2016-12-27 | 2019-11-15 | 华南理工大学 | A kind of image understanding method based on depth residual error network and LSTM |
| CN106845371B (en) * | 2016-12-31 | 2019-10-25 | 中国科学技术大学 | A kind of city road network automotive emission remote sensing monitoring system |
| CN106934396A (en) * | 2017-03-09 | 2017-07-07 | 深圳市捷顺科技实业股份有限公司 | A kind of license number search method and system |
| CN107273828A (en) * | 2017-05-29 | 2017-10-20 | 浙江师范大学 | A kind of guideboard detection method of the full convolutional neural networks based on region |
-
2018
- 2018-03-09 CN CN201810194741.6A patent/CN108549926A/en active Pending
- 2018-07-10 WO PCT/CN2018/095060 patent/WO2019169816A1/en active Application Filing
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106599869A (en) * | 2016-12-22 | 2017-04-26 | 安徽大学 | Vehicle attribute identification method based on multi-task convolutional neural network |
| CN107145857A (en) * | 2017-04-29 | 2017-09-08 | 深圳市深网视界科技有限公司 | Face character recognition methods, device and method for establishing model |
Cited By (56)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110147707B (en) * | 2018-10-25 | 2021-07-20 | 初速度（苏州）科技有限公司 | High-precision vehicle identification method and system |
| CN110147707A (en) * | 2018-10-25 | 2019-08-20 | 初速度（苏州）科技有限公司 | A kind of high-precision vehicle identification method and system |
| US11328180B2 (en) | 2018-10-30 | 2022-05-10 | Beijing Horizon Robotics Technology Research And Development Co., Ltd. | Method for updating neural network and electronic device |
| CN109472360B (en) * | 2018-10-30 | 2020-09-04 | 北京地平线机器人技术研发有限公司 | Neural network updating method and updating device and electronic equipment |
| CN109472360A (en) * | 2018-10-30 | 2019-03-15 | 北京地平线机器人技术研发有限公司 | Update method, updating device and the electronic equipment of neural network |
| CN109685110A (en) * | 2018-11-28 | 2019-04-26 | 北京陌上花科技有限公司 | Training method, image classification method and device, the server of image classification network |
| CN111275061A (en) * | 2018-12-04 | 2020-06-12 | 北京嘀嘀无限科技发展有限公司 | Vehicle attribute identification method and model training method and device thereof, and electronic equipment |
| CN109753999A (en) * | 2018-12-21 | 2019-05-14 | 西北工业大学 | For the fine granularity model recognizing method of visual angle auto graph |
| CN109978132A (en) * | 2018-12-24 | 2019-07-05 | 中国科学院深圳先进技术研究院 | A kind of neural network method and system refining vehicle identification |
| CN109816001A (en) * | 2019-01-10 | 2019-05-28 | 高新兴科技集团股份有限公司 | A kind of more attribute recognition approaches of vehicle based on deep learning, device and equipment |
| CN110334572B (en) * | 2019-04-04 | 2021-07-13 | 南京航空航天大学 | Multi-angle vehicle model fine recognition method |
| CN110334572A (en) * | 2019-04-04 | 2019-10-15 | 南京航空航天大学 | The fine recognition methods of vehicle under a kind of multi-angle |
| CN110070082A (en) * | 2019-04-22 | 2019-07-30 | 苏州科达科技股份有限公司 | Licence plate recognition method, device, equipment and storage medium |
| CN111860068A (en) * | 2019-04-30 | 2020-10-30 | 四川大学 | Fine-grained bird identification method based on cross-layer simplified bilinear network |
| CN110135437B (en) * | 2019-05-06 | 2022-04-05 | 北京百度网讯科技有限公司 | Loss assessment method and device for vehicle, electronic equipment and computer storage medium |
| CN110135437A (en) * | 2019-05-06 | 2019-08-16 | 北京百度网讯科技有限公司 | For the damage identification method of vehicle, device, electronic equipment and computer storage medium |
| CN110209857B (en) * | 2019-05-23 | 2022-04-22 | 厦门市美亚柏科信息股份有限公司 | Vehicle multi-attribute identification method, device and medium based on neural network structure search |
| CN110209857A (en) * | 2019-05-23 | 2019-09-06 | 厦门市美亚柏科信息股份有限公司 | The more attribute recognition approaches of vehicle, device, medium based on neural network structure search |
| WO2020232840A1 (en) * | 2019-05-23 | 2020-11-26 | 厦门市美亚柏科信息股份有限公司 | Vehicle multi-attribute identification method and device employing neural network structure search, and medium |
| CN110210401B (en) * | 2019-06-03 | 2022-10-21 | 多维协同人工智能技术研究院(重庆)有限公司 | Intelligent target detection method under weak light |
| CN110210401A (en) * | 2019-06-03 | 2019-09-06 | 多维协同人工智能技术研究院(重庆)有限公司 | A kind of Intelligent target detection method under dim light |
| CN110378237A (en) * | 2019-06-21 | 2019-10-25 | 浙江工商大学 | Human facial expression recognition method based on depth measure converged network |
| CN110378237B (en) * | 2019-06-21 | 2021-06-11 | 浙江工商大学 | Facial expression recognition method based on depth measurement fusion network |
| CN110533024A (en) * | 2019-07-10 | 2019-12-03 | 杭州电子科技大学 | Biquadratic pond fine granularity image classification method based on multiple dimensioned ROI feature |
| CN110533024B (en) * | 2019-07-10 | 2021-11-23 | 杭州电子科技大学 | Double-quadratic pooling fine-grained image classification method based on multi-scale ROI (region of interest) features |
| CN110321970A (en) * | 2019-07-11 | 2019-10-11 | 山东领能电子科技有限公司 | A kind of fine-grained objective classification method of multiple features based on branch neural network |
| CN110472664A (en) * | 2019-07-17 | 2019-11-19 | 杭州有盾网络科技有限公司 | A kind of certificate image identification method, device and equipment based on deep learning |
| CN110399828A (en) * | 2019-07-23 | 2019-11-01 | 吉林大学 | A kind of vehicle recognition methods again of the depth convolutional neural networks based on multi-angle |
| CN110399828B (en) * | 2019-07-23 | 2022-10-28 | 吉林大学 | Vehicle re-identification method based on multi-angle deep convolutional neural network |
| CN110598749A (en) * | 2019-08-13 | 2019-12-20 | 广东数鼎科技有限公司 | Image vehicle recognition implementation system based on convolutional neural network AI algorithm |
| CN110598621A (en) * | 2019-09-06 | 2019-12-20 | 高新兴科技集团股份有限公司 | Vehicle brand identification method, device, equipment and computer readable storage medium |
| CN112541374B (en) * | 2019-09-20 | 2024-04-30 | 南京行者易智能交通科技有限公司 | Deep learning-based passenger attribute acquisition method, device and model training method |
| CN112541374A (en) * | 2019-09-20 | 2021-03-23 | 南京行者易智能交通科技有限公司 | Passenger attribute acquisition method and device based on deep learning and model training method |
| CN111222409A (en) * | 2019-11-26 | 2020-06-02 | 北京迈格威科技有限公司 | Vehicle brand labeling method, device and system |
| CN110991349A (en) * | 2019-12-05 | 2020-04-10 | 中国科学院重庆绿色智能技术研究院 | Lightweight vehicle attribute identification method based on metric learning |
| CN110991349B (en) * | 2019-12-05 | 2023-02-10 | 中国科学院重庆绿色智能技术研究院 | Lightweight vehicle attribute identification method based on metric learning |
| CN111126224A (en) * | 2019-12-17 | 2020-05-08 | 成都通甲优博科技有限责任公司 | Vehicle detection method and classification recognition model training method |
| CN111325739B (en) * | 2020-02-28 | 2020-12-29 | 推想医疗科技股份有限公司 | Method and device for detecting lung focus and training method of image detection model |
| CN111401251B (en) * | 2020-03-17 | 2023-12-26 | 北京百度网讯科技有限公司 | Lane line extraction method, lane line extraction device, electronic equipment and computer readable storage medium |
| CN111401251A (en) * | 2020-03-17 | 2020-07-10 | 北京百度网讯科技有限公司 | Lane line extraction method and device, electronic equipment and computer-readable storage medium |
| CN111091165A (en) * | 2020-03-25 | 2020-05-01 | 北京精诊医疗科技有限公司 | Fine-grained identification method of image |
| CN111612855A (en) * | 2020-04-09 | 2020-09-01 | 北京旷视科技有限公司 | Object color identification method and device and electronic equipment |
| CN111598000A (en) * | 2020-05-18 | 2020-08-28 | 中移（杭州）信息技术有限公司 | Face recognition method, device, server and readable storage medium based on multiple tasks |
| CN111754784A (en) * | 2020-06-23 | 2020-10-09 | 高新兴科技集团股份有限公司 | Attention mechanism-based vehicle main and sub brand identification method of multilayer network |
| CN111754784B (en) * | 2020-06-23 | 2022-05-24 | 高新兴科技集团股份有限公司 | Method for identifying main and sub brands of vehicle based on multi-layer network of attention mechanism |
| CN111814857A (en) * | 2020-06-29 | 2020-10-23 | 浙江大华技术股份有限公司 | Target re-identification method, network training method thereof and related device |
| CN112016617B (en) * | 2020-08-27 | 2023-12-01 | 中国平安财产保险股份有限公司 | Fine granularity classification method, apparatus and computer readable storage medium |
| CN112016617A (en) * | 2020-08-27 | 2020-12-01 | 中国平安财产保险股份有限公司 | Fine-grained classification method and device and computer-readable storage medium |
| CN112200193B (en) * | 2020-12-03 | 2021-03-16 | 中国科学院自动化研究所 | Distributed license plate recognition method, system and device based on multi-attribute fusion |
| CN112200193A (en) * | 2020-12-03 | 2021-01-08 | 中国科学院自动化研究所 | Distributed license plate recognition method, system and device based on multi-attribute fusion |
| CN112766349A (en) * | 2021-01-12 | 2021-05-07 | 齐鲁工业大学 | Object description generation method based on machine vision and tactile perception |
| CN112861724A (en) * | 2021-02-06 | 2021-05-28 | 安徽理工大学 | Traditional Chinese medicine identification system |
| CN113221982A (en) * | 2021-04-28 | 2021-08-06 | 深圳市捷顺科技实业股份有限公司 | Vehicle identification method, vehicle identification model creation method and related components |
| CN114882448A (en) * | 2022-04-01 | 2022-08-09 | 北京卓视智通科技有限责任公司 | Vehicle monitoring method and electronic equipment |
| CN114882448B (en) * | 2022-04-01 | 2023-10-31 | 北京卓视智通科技有限责任公司 | Vehicle monitoring method and electronic equipment |
| CN115294644A (en) * | 2022-06-24 | 2022-11-04 | 北京昭衍新药研究中心股份有限公司 | Rapid monkey behavior identification method based on 3D convolution parameter reconstruction |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2019169816A1 (en) | 2019-09-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108549926A (en) | A kind of deep neural network and training method for refining identification vehicle attribute | |
| CN106650806B (en) | A kind of cooperating type depth net model methodology for pedestrian detection | |
| CN106096531B (en) | A kind of traffic image polymorphic type vehicle checking method based on deep learning | |
| CN104281853B (en) | A kind of Activity recognition method based on 3D convolutional neural networks | |
| CN103514456B (en) | Image classification method and device based on compressed sensing multi-core learning | |
| CN104268583B (en) | Pedestrian re-recognition method and system based on color area features | |
| CN103714181B (en) | A kind of hierarchical particular persons search method | |
| CN106845487A (en) | A kind of licence plate recognition method end to end | |
| CN109902806A (en) | Method is determined based on the noise image object boundary frame of convolutional neural networks | |
| CN110232316A (en) | A kind of vehicle detection and recognition method based on improved DSOD model | |
| CN108491797A (en) | A kind of vehicle image precise search method based on big data | |
| CN104134068B (en) | Monitoring vehicle characteristics based on sparse coding represent and sorting technique | |
| CN106485214A (en) | A kind of eyes based on convolutional neural networks and mouth state identification method | |
| CN104992142A (en) | Pedestrian recognition method based on combination of depth learning and property learning | |
| CN107133569A (en) | The many granularity mask methods of monitor video based on extensive Multi-label learning | |
| CN108734719A (en) | Background automatic division method before a kind of lepidopterous insects image based on full convolutional neural networks | |
| Li et al. | Pedestrian detection based on deep learning model | |
| CN104063719A (en) | Method and device for pedestrian detection based on depth convolutional network | |
| CN109299644A (en) | A kind of vehicle target detection method based on the full convolutional network in region | |
| CN110163077A (en) | A kind of lane recognition method based on full convolutional neural networks | |
| CN104504395A (en) | Method and system for achieving classification of pedestrians and vehicles based on neural network | |
| CN108647700A (en) | Multitask vehicle part identification model based on deep learning, method and system | |
| CN109033175A (en) | A kind of method and system to scheme to search vehicle | |
| CN107578091A (en) | A kind of Pedestrians and vehicles real-time detection method based on lightweight depth network | |
| CN108681696A (en) | Merge the infrared target robustness recognition methods of multiple features dimensionality reduction and transfer learning |
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: 20180918 |