CN109214441A - A kind of fine granularity model recognition system and method - Google Patents
A kind of fine granularity model recognition system and method Download PDFInfo
- Publication number
- CN109214441A CN109214441A CN201810968895.6A CN201810968895A CN109214441A CN 109214441 A CN109214441 A CN 109214441A CN 201810968895 A CN201810968895 A CN 201810968895A CN 109214441 A CN109214441 A CN 109214441A
- Authority
- CN
- China
- Prior art keywords
- vehicle
- fine granularity
- layer
- loss
- model
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of fine granularity model recognition system and method, method carries out target detection to the samples pictures in sample database using Fast-RCNN, intercepts out the picture of target vehicle the following steps are included: building vehicle sample database;Singular value decomposition convolutional neural networks are constructed, are lost using centre distance loss and the fusion of Classification Loss, the picture of target vehicle is imported into the singular value decomposition convolutional neural networks and is trained, training pattern is obtained;The weight matrix W for extracting full articulamentum in training pattern carries out after singular value decomposition assignment again to weight matrix W and adjusts, obtains fine granularity vehicle cab recognition model;Classification and Identification is carried out to the samples pictures in vehicle sample database using fine granularity vehicle cab recognition model, identifies the vehicle in vehicle sample database.Compared with the prior art, the present invention can remove the redundancy feature with correlation, and study can be improved the accuracy rate of fine granularity vehicle cab recognition to the differentiating characteristics of fine granularity rank.
Description
Technical field
The present invention relates to pattern-recognition and technical field of image processing, in particular to a kind of fine granularity model recognition system and
Method.
Background technique
Fine granularity vehicle cab recognition is the core in intelligent transportation system, has more widely answer in real life
With, such as: vehicle fake-license is checked, the vehicle of stolen vehicle and irregular driving carries out the identification etc. of vehicle, for mesh
Preceding traditional algorithm uses the deep learning of convolutional neural networks to the not high problem of the accuracy rate of fine granularity vehicle cab recognition, the present invention
Algorithm frame carries out singular value decomposition to the full articulamentum of model after training by constructing singular value decomposition convolutional neural networks
Retraining, and centre distance is combined to lose, the redundancy feature with correlation can be removed, the differentiation of fine granularity rank is arrived in study
Property feature, while the distance between making in class smaller, accuracy rate has biggish promotion compared to conventional method.
Summary of the invention
The object of the present invention is to provide a kind of fine granularity model recognition system and method, the technical problem to be solved is that:
Algorithm traditional at present is not high to the accuracy rate of fine granularity vehicle cab recognition.
The technical scheme to solve the above technical problems is that a kind of fine granularity model recognizing method, including it is following
Step:
Step S1: building vehicle sample database carries out mesh to the samples pictures in sample database using Fast-RCNN
Mark detection, when detecting target vehicle, then intercepts out the picture of target vehicle;
Step S2: building singular value decomposition convolutional neural networks are damaged using centre distance loss and the fusion of Classification Loss
It loses, the picture of target vehicle is imported into the singular value decomposition convolutional neural networks and is trained, training pattern is obtained;
Step S3: the weight matrix W of full articulamentum in training pattern is extracted, after carrying out singular value decomposition to weight matrix W
Again it assignment and adjusts, obtains fine granularity vehicle cab recognition model;
Step S4: classification knowledge is carried out to the samples pictures in vehicle sample database using fine granularity vehicle cab recognition model
Not, to identify the vehicle in vehicle sample database.
Further, the specific steps that Fast-RCNN is detected in the step 1 are as follows:
Multiple suggestion windows are extracted in samples pictures using Selective Search algorithm, while samples pictures being inputted
CNN carries out feature extraction;
Multiple suggestion windows are mapped on the last layer convolution feature map of CNN, pass through pooling layers of Rol
Each suggestion window is set to generate fixed-size feature map;
Finally joint training is returned to class probability and frame using Softmax Loss and Smooth L1 Loss to obtain
Target vehicle.
Further, the specific steps of step 2 are as follows: the singular value decomposition convolutional neural networks specifically include sequentially connected
Convolutional layer Conv1, pond layer Pool1, convolutional layer Conv2, pond layer Pool2, convolutional layer Conv3, pond layer Pool3, Quan Lian
Connect layer Fc1, drop_out1 layers, it is full articulamentum Fc2, drop_out2 layers and classification layer Softmax, further include centre distance loss
Layer, the centre distance loss layer connect full articulamentum Fc1 and data Layer;
The convolutional layer Conv1, convolutional layer Conv2 and convolutional layer Conv3 are used to carry out convolution behaviour to the image of input
Make, the output of convolution is calculated by activation primitive;
The pond layer Pool1, pond layer Pool2 and pond layer Pool3 are used to subtract image progress down-sampling operation
Few picture size size;
Described drop_out layers can prevent from training over-fitting, make e-learning to more compact and with more distinction
Feature;
Feature of the classification layer Softmax for the output of full articulamentum carries out tagsort.
The centre distance loss layer learns the feature with cluster property for the distance between reducing in class.
Further, the centre distance loss in the step S2 and the fusion of Classification Loss are lost, specific two losses letter
The objective function of number Weighted Fusion is as follows:
Wherein, L1For Classification Loss function;N is the sum of training data classification;M is sample attribute number;L2Centered on away from
From loss function;Wherein xiIndicate the characteristic pattern of the i-th picture;cyiIndicate class yiClass center;λ is L2Weighted value.
Feedback learning, Classification Loss function L (x are carried out by using BGD batch gradient descent algorithmi) to the anti-of input x
To calculating, derivation formula is as follows:
Wixi+biFor the output of neuron in neural network, feeds back in calculating and need to update xiWeight WiWith biasing bi, more
New formula is as follows:
Wherein, η is learning rate, and by updating neuron weight, characteristic layer can learn fusion loss.
Further, specific step is as follows by the step S3:
The weight matrix W connected entirely in training pattern is extracted, singular value decomposition, specific formula are carried out to weight matrix W
It is as follows:
W=USVT
Wherein W is the weight vectors of full articulamentum, and U is left unitary matrice, and S is singular value matrix, and V is right unitary matrice;
W is replaced with US, then full articulamentum uses WWTAll feature vectors as weight vectors, generated after replacement
Training pattern A;
The weight matrix of articulamentum complete in training pattern A is fixed, other layers are adjusted until convergence, are trained
Model B;Straightening is adjusted to convergence to all parameters on the basis of training pattern B, training is completed and obtains fine granularity vehicle
Identification model.
The beneficial effects of the present invention are: can be gone by carrying out singular value decomposition to the weight matrix in full articulamentum
Except the redundancy feature with correlation, the differentiating characteristics of fine granularity rank are arrived in study.It is compared with the traditional method, the present invention uses
The method of singular value decomposition convolutional neural networks can be improved the accuracy rate of fine granularity vehicle cab recognition.
Another technical solution that the present invention solves above-mentioned technical problem is as follows: a kind of fine granularity model recognition system, comprising:
Picture module is intercepted, for constructing vehicle sample database, using Fast-RCNN to the sample in sample database
Picture carries out target detection and then intercepts out the picture of target vehicle when detecting target vehicle;
Model module is constructed, for constructing singular value decomposition convolutional neural networks, is lost and is classified using centre distance and damaged
The fusion of mistake is lost, and the picture of target vehicle is imported into the singular value decomposition convolutional neural networks and is trained, is obtained
Training pattern;
Model training module carries out weight matrix W odd for extracting the weight matrix W of full articulamentum in training pattern
Different value assignment and adjusts again after decomposing, and obtains fine granularity vehicle cab recognition model;
Vehicle identification module, for using fine granularity vehicle cab recognition model to the samples pictures in vehicle sample database into
Row Classification and Identification, to identify the vehicle in vehicle sample database.
In above-described embodiment, the specific implementation of Fast-RCNN detection in the interception picture module are as follows: interception picture module
Multiple suggestion windows are extracted in samples pictures using Selective Search algorithm, while samples pictures input CNN being carried out
Feature extraction;Multiple suggestion windows are mapped on the last layer convolution feature map of CNN, pass through pooling layers of Rol
Each suggestion window is set to generate fixed-size feature map;Finally utilize Softmax Loss and Smooth L1 Loss
Joint training is returned to class probability and frame and obtains target vehicle.
In above-described embodiment, the singular value decomposition convolutional neural networks specifically include sequentially connected convolutional layer Conv1,
Pond layer Pool1, convolutional layer Conv2, pond layer Pool2, convolutional layer Conv3, pond layer Pool3, full articulamentum Fc1, drop_
Out1 layers, it is articulamentum Fc2, drop_out2 layers and classification layer Softmax complete, further include centre distance loss layer, the center away from
Full articulamentum Fc1 and data Layer are connected from loss layer.
In above-described embodiment, the centre distance loss and the fusion of Classification Loss of the building model module are lost, specifically
The objective function of two loss function Weighted Fusions is as follows:
Wherein, L1For Classification Loss function;N is the sum of training data classification;M is sample attribute number;L2Centered on away from
From loss function;Wherein xiIndicate the characteristic pattern of the i-th picture;cyiIndicate class yiClass center;λ is L2Weighted value.
Feedback learning, Classification Loss function L (x are carried out by using BGD batch gradient descent algorithmi) to the anti-of input x
To calculating, derivation formula is as follows:
Wixi+biFor the output of neuron in neural network, feeds back in calculating and need to update xiWeight WiWith biasing bi, more
New formula is as follows:
Wherein, η is learning rate, and by updating neuron weight, characteristic layer can learn fusion loss.
In above-described embodiment, the model training module extracts the weight matrix W connected entirely in training pattern, to weight
Matrix W carries out singular value decomposition, specific formula is as follows:
W=USVT
Wherein W is the weight vectors of full articulamentum, and U is left unitary matrice, and S is singular value matrix, and V is right unitary matrice;
W is replaced with US, then full articulamentum uses WWTAll feature vectors as weight vectors, generated after replacement
Training pattern A;
The weight matrix of articulamentum complete in training pattern A is fixed, other layers are adjusted until convergence, are trained
Model B;Straightening is adjusted to convergence to all parameters on the basis of training pattern B, training is completed and obtains fine granularity vehicle
Identification model.
The beneficial effects of the present invention are: can be gone by carrying out singular value decomposition to the weight matrix in full articulamentum
Except the redundancy feature with correlation, the differentiating characteristics of fine granularity rank are arrived in study.It is compared with the traditional method, the present invention uses
The method of singular value decomposition convolutional neural networks can be improved the accuracy rate of fine granularity vehicle cab recognition.
Detailed description of the invention
Fig. 1 is a kind of flow chart of fine granularity model recognizing method of the present invention;
Fig. 2 is a kind of module frame chart of fine granularity model recognition system of the present invention;
Fig. 3 is singular value decomposition convolutional neural networks block schematic illustration of the present invention.
In attached drawing, parts list represented by the reference numerals are as follows:
1, picture module, 2, building model module, 3, model training module, 4, vehicle identification module are intercepted.
Specific embodiment
The principle and features of the present invention will be described below with reference to the accompanying drawings, and the given examples are served only to explain the present invention, and
It is non-to be used to limit the scope of the invention.
As depicted in figs. 1 and 2, a kind of fine granularity model recognizing method, comprising the following steps:
Step S1: building vehicle sample database carries out mesh to the samples pictures in sample database using Fast-RCNN
Mark detection, when detecting target vehicle, then intercepts out the picture of target vehicle;
Step S2: building singular value decomposition convolutional neural networks are damaged using centre distance loss and the fusion of Classification Loss
It loses, the picture of target vehicle is imported into the singular value decomposition convolutional neural networks and is trained, training pattern is obtained;
Step S3: the weight matrix W of full articulamentum in training pattern is extracted, after carrying out singular value decomposition to weight matrix W
Again it assignment and adjusts, obtains fine granularity vehicle cab recognition model;
Step S4: classification knowledge is carried out to the samples pictures in vehicle sample database using fine granularity vehicle cab recognition model
Not, to identify the vehicle in vehicle sample database.
In above-described embodiment, the specific steps of Fast-RCNN detection in the step 1 are as follows:
Multiple suggestion windows are extracted in samples pictures using Selective Search algorithm, while samples pictures being inputted
CNN carries out feature extraction;
Multiple suggestion windows are mapped on the last layer convolution feature map of CNN, pass through pooling layers of Rol
Each suggestion window is set to generate fixed-size feature map;
Finally joint training is returned to class probability and frame using Softmax Loss and Smooth L1 Loss to obtain
Target vehicle.
In above-described embodiment, the specific steps of step 2 are as follows: the singular value decomposition convolutional neural networks specifically include successively
Convolutional layer Conv1, pond layer Pool1, convolutional layer Conv2, the pond layer Pool2, convolutional layer Conv3 of connection, pond layer
It is Pool3, full articulamentum Fc1, drop_out1 layers, articulamentum Fc2, drop_out2 layers and classification layer Softmax complete, further include
Heart range loss layer, the centre distance loss layer connect full articulamentum Fc1 and data Layer;
The convolutional layer Conv1, convolutional layer Conv2 and convolutional layer Conv3 are used to carry out convolution behaviour to the image of input
Make, the output of convolution is calculated by activation primitive;
The pond layer Pool 1, pond layer Pool2 and pond layer Pool3 are used to subtract image progress down-sampling operation
Few picture size size;
Described drop_out layers can prevent from training over-fitting, make e-learning to more compact and with more distinction
Feature;
Feature of the classification layer Softmax for the output of full articulamentum carries out tagsort.
The centre distance loss layer learns the feature with cluster property for the distance between reducing in class.
In above-described embodiment, centre distance loss and the fusion of Classification Loss in the step S2 are lost, and specific two
The objective function of loss function Weighted Fusion is as follows:
Wherein, L1For Classification Loss function;N is the sum of training data classification;M is sample attribute number;L2Centered on away from
From loss function;Wherein xiIndicate the characteristic pattern of the i-th picture;cyiIndicate class yiClass center;λ is L2Weighted value.
Feedback learning, classification are carried out by using BGD (Batch Gradient Descent) batch gradient descent algorithm
Loss function L (xi) it is as follows to the retrospectively calculate derivation formula of input x:
Wixi+biFor the output of neuron in neural network, feeds back in calculating and need to update xiWeight WiWith biasing bi, more
New formula is as follows:
Wherein, η is learning rate, and by updating neuron weight, characteristic layer can learn fusion loss.
In above-described embodiment, specific step is as follows by the step S3:
The weight matrix W connected entirely in training pattern is extracted, singular value decomposition, specific formula are carried out to weight matrix W
It is as follows:
W=USVT
Wherein W is the weight vectors of full articulamentum, and U is left unitary matrice, and S is singular value matrix, and V is right unitary matrice;
W is replaced with US, then full articulamentum uses WWTAll feature vectors as weight vectors, generated after replacement
Training pattern A;
The weight matrix of articulamentum complete in training pattern A is fixed, other layers are adjusted until convergence, are trained
Model B;Straightening is adjusted to convergence to all parameters on the basis of training pattern B, training is completed and obtains fine granularity vehicle
Identification model.
The present embodiment can be removed by carrying out singular value decomposition to the weight matrix in full articulamentum with correlation
Redundancy feature, study arrive fine granularity rank differentiating characteristics.It is compared with the traditional method, the present invention is rolled up using singular value decomposition
The method of product neural network, can be improved the accuracy rate of fine granularity vehicle cab recognition.
Embodiment 2:
As shown in Figures 2 and 3, a kind of fine granularity model recognition system, comprising:
Picture module 1 is intercepted, for constructing vehicle sample database, using Fast-RCNN to the sample in sample database
This picture carries out target detection and then intercepts out the picture of target vehicle when detecting target vehicle;
Model module 2 is constructed, for constructing singular value decomposition convolutional neural networks, is lost and is classified using centre distance and damaged
The fusion of mistake is lost, and the picture of target vehicle is imported into the singular value decomposition convolutional neural networks and is trained, is obtained
Training pattern;
Model training module 3 carries out weight matrix W odd for extracting the weight matrix W of full articulamentum in training pattern
Different value assignment and adjusts again after decomposing, and obtains fine granularity vehicle cab recognition model;
Vehicle identification module 4, for using fine granularity vehicle cab recognition model to the samples pictures in vehicle sample database
Classification and Identification is carried out, to identify the vehicle in vehicle sample database.
In above-described embodiment, the specific implementation of Fast-RCNN detection in the interception picture module 1 are as follows: interception picture mould
Block 1 extracts multiple suggestion windows in samples pictures using Selective Search algorithm, at the same by samples pictures input CNN into
Row feature extraction;Multiple suggestion windows are mapped on the last layer convolution feature map of CNN, Rol pooling is passed through
Layer makes each suggestion window generate fixed-size feature map;Finally utilize Softmax Loss and Smooth L1
Loss returns joint training to class probability and frame and obtains target vehicle.
In above-described embodiment, the singular value decomposition convolutional neural networks specifically include sequentially connected convolutional layer Conv1,
Pond layer Pool1, convolutional layer Conv2, pond layer Pool2, convolutional layer Conv3, pond layer Pool3, full articulamentum Fc1, drop_
Out1 layers, it is articulamentum Fc2, drop_out2 layers and classification layer Softmax complete, further include centre distance loss layer, the center away from
Full articulamentum Fc1 and data Layer are connected from loss layer.
In above-described embodiment, the centre distance loss and the fusion of Classification Loss of the building model module 2 are lost, specifically
The objective function of two loss function Weighted Fusions is as follows:
Wherein, L1For Classification Loss function;N is the sum of training data classification;M is sample attribute number;L2Centered on away from
From loss function;Wherein xiIndicate the characteristic pattern of the i-th picture;cyiIndicate class yiClass center;λ is L2Weighted value.
Feedback learning, Classification Loss function L (x are carried out by using BGD batch gradient descent algorithmi) to the anti-of input x
To calculating, derivation formula is as follows:
Wixi+biFor the output of neuron in neural network, feeds back in calculating and need to update xiWeight WiWith biasing bi, more
New formula is as follows:
Wherein, η is learning rate, and by updating neuron weight, characteristic layer can learn fusion loss.
In above-described embodiment, the model training module 3 extracts the weight matrix W connected entirely in training pattern, to power
Weight matrix W carries out singular value decomposition, specific formula is as follows:
W=USVT
Wherein W is the weight vectors of full articulamentum, and U is left unitary matrice, and S is singular value matrix, and V is right unitary matrice;
W is replaced with US, then full articulamentum uses WWTAll feature vectors as weight vectors, generated after replacement
Training pattern A;
The weight matrix of articulamentum complete in training pattern A is fixed, other layers are adjusted until convergence, are trained
Model B;Straightening is adjusted to convergence to all parameters on the basis of training pattern B, training is completed and obtains fine granularity vehicle
Identification model.
The present embodiment can be removed by carrying out singular value decomposition to the weight matrix in full articulamentum with correlation
Redundancy feature, study arrive fine granularity rank differentiating characteristics.It is compared with the traditional method, the present invention is rolled up using singular value decomposition
The method of product neural network, can be improved the accuracy rate of fine granularity vehicle cab recognition.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (10)
1. a kind of fine granularity model recognizing method, which comprises the following steps:
Step S1: building vehicle sample database carries out target inspection to the samples pictures in sample database using Fast-RCNN
It surveys, when detecting target vehicle, then intercepts out the picture of target vehicle;
Step S2: building singular value decomposition convolutional neural networks are lost using centre distance loss and the fusion of Classification Loss, will
The picture of target vehicle imported into the singular value decomposition convolutional neural networks and is trained, and obtains training pattern;
Step S3: extracting the weight matrix W of full articulamentum in training pattern, carries out after singular value decomposition again to weight matrix W
Assignment simultaneously adjusts, and obtains fine granularity vehicle cab recognition model;
Step S4: carrying out Classification and Identification to the samples pictures in vehicle sample database using fine granularity vehicle cab recognition model, from
And identify the vehicle in vehicle sample database.
2. a kind of fine granularity model recognizing method according to claim 1, which is characterized in that Fast-RCNN in the step 1
The specific steps of detection are as follows:
Multiple suggestion windows are extracted in samples pictures using Selective Search algorithm, while samples pictures are inputted into CNN
Carry out feature extraction;
Multiple suggestion windows are mapped on the last layer convolution feature map of CNN, are made by pooling layers of RoI every
A suggestion window generates fixed-size feature map;
Joint training finally is returned to class probability and frame using Softmax Loss and Smooth L1Loss and obtains target carriage
?.
3. a kind of fine granularity model recognizing method according to claim 1, which is characterized in that the specific steps of step 2 are as follows: institute
It states singular value decomposition convolutional neural networks and specifically includes sequentially connected convolutional layer Conv1, pond layer Pool1, convolutional layer
Conv2, pond layer Pool2, convolutional layer Conv3, pond layer Pool3, articulamentum Fc1, drop_out1 layers complete, full articulamentum
Fc2, drop_out2 layers and classification layer Softmax, further include centre distance loss layer, and the centre distance loss layer connection is complete
Articulamentum Fc1 and data Layer.
4. a kind of fine granularity model recognizing method according to claim 2, which is characterized in that
Centre distance loss and the fusion of Classification Loss in the step S2 are lost, specific two loss function Weighted Fusions
Objective function is as follows:
Wherein, L1For Classification Loss function;N is the sum of training data classification;M is sample attribute number;L2Centered on distance damage
Lose function;Wherein xiIndicate the characteristic pattern of the i-th picture;cyiIndicate class yiClass center;λ is L2Weighted value;
Feedback learning, Classification Loss function L (x are carried out by using BGD batch gradient descent algorithmi) to the retrospectively calculate of input x
Derivation formula is as follows:
Wixi+biFor the output of neuron in neural network, feeds back in calculating and need to update xiWeight WiWith biasing bi, update public
Formula is as follows:
Wherein, η is learning rate, and by updating neuron weight, characteristic layer can learn fusion loss.
5. a kind of fine granularity model recognizing method according to claim 1, which is characterized in that the specific steps of the step S3
It is as follows:
The weight matrix W connected entirely in training pattern is extracted, singular value decomposition is carried out to weight matrix W, specific formula is as follows:
W=USVT
Wherein W is the weight vectors of full articulamentum, and U is left unitary matrice, and S is singular value matrix, and V is right unitary matrice;
W is replaced with US, then full articulamentum uses WWTAll feature vectors training mould is generated as weight vectors, after replacement
Type A;
The weight matrix of articulamentum complete in training pattern A is fixed, other layers are adjusted until convergence, training pattern is obtained
B;Straightening is adjusted to convergence to all parameters on the basis of training pattern B, training is completed and obtains fine granularity vehicle cab recognition
Model.
6. a kind of fine granularity model recognition system characterized by comprising
It intercepts picture module (1), for constructing vehicle sample database, using Fast-RCNN to the sample in sample database
Picture carries out target detection and then intercepts out the picture of target vehicle when detecting target vehicle;
It constructs model module (2), for constructing singular value decomposition convolutional neural networks, uses centre distance loss and Classification Loss
Fusion loss, the picture of target vehicle is imported into the singular value decomposition convolutional neural networks and is trained, is instructed
Practice model;
Model training module (3) carries out weight matrix W unusual for extracting the weight matrix W of full articulamentum in training pattern
Value assignment and adjusts again after decomposing, and obtains fine granularity vehicle cab recognition model;
Vehicle identification module (4), for using fine granularity vehicle cab recognition model to the samples pictures in vehicle sample database into
Row Classification and Identification, to identify the vehicle in vehicle sample database.
7. a kind of fine granularity model recognition system according to claim 6, which is characterized in that the interception picture module (1)
The specific implementation of middle Fast-RCNN detection are as follows: interception picture module (1) is using Selective Search algorithm in samples pictures
Multiple suggestion windows are extracted, while samples pictures input CNN is subjected to feature extraction;Multiple suggestion windows are mapped to CNN's
On the last layer convolution feature map, generate each suggestion window by pooling layers of RoI fixed-size
feature map;Finally joint training is returned to class probability and frame using Softmax Loss and Smooth L1Loss to obtain
To target vehicle.
8. a kind of fine granularity model recognition system according to claim 6, which is characterized in that the singular value decomposition convolution mind
Sequentially connected convolutional layer Conv1, pond layer Pool1, convolutional layer Conv2, pond layer Pool2, convolution are specifically included through network
Layer Conv3, pond layer Pool3, it is articulamentum Fc1, drop_out1 layers complete, it is full articulamentum Fc2, drop_out2 layers and classify layer
Softmax further includes centre distance loss layer, and the centre distance loss layer connects full articulamentum Fc1 and data Layer.
9. a kind of fine granularity model recognition system according to claim 8, which is characterized in that the building model module (2)
Centre distance loss and the fusion of Classification Loss lose, the objective function of specific two loss function Weighted Fusions is as follows:
Wherein, L1For Classification Loss function;N is the sum of training data classification;M is sample attribute number;L2Centered on distance damage
Lose function;Wherein xiIndicate the characteristic pattern of the i-th picture;cyiIndicate class yiClass center;λ is L2Weighted value;
Feedback learning, Classification Loss function L (x are carried out by using BGD batch gradient descent algorithmi) to the retrospectively calculate of input x
Derivation formula is as follows:
Wixi+biFor the output of neuron in neural network, feeds back in calculating and need to update xiWeight WiWith biasing bi, update public
Formula is as follows:
Wherein, η is learning rate, and by updating neuron weight, characteristic layer can learn fusion loss.
10. a kind of fine granularity model recognition system according to claim 6, which is characterized in that the model training module (3)
The weight matrix W connected entirely in training pattern is extracted, singular value decomposition is carried out to weight matrix W, specific formula is as follows:
W=USVT
Wherein W is the weight vectors of full articulamentum, and U is left unitary matrice, and S is singular value matrix, and V is right unitary matrice;
W is replaced with US, then full articulamentum uses WWTAll feature vectors training mould is generated as weight vectors, after replacement
Type A;
The weight matrix of articulamentum complete in training pattern A is fixed, other layers are adjusted until convergence, training pattern is obtained
B;Straightening is adjusted to convergence to all parameters on the basis of training pattern B, training is completed and obtains fine granularity vehicle cab recognition
Model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810968895.6A CN109214441A (en) | 2018-08-23 | 2018-08-23 | A kind of fine granularity model recognition system and method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810968895.6A CN109214441A (en) | 2018-08-23 | 2018-08-23 | A kind of fine granularity model recognition system and method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109214441A true CN109214441A (en) | 2019-01-15 |
Family
ID=64989742
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810968895.6A Pending CN109214441A (en) | 2018-08-23 | 2018-08-23 | A kind of fine granularity model recognition system and method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109214441A (en) |
Cited By (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109902732A (en) * | 2019-02-22 | 2019-06-18 | 哈尔滨工业大学(深圳) | Automobile automatic recognition method and relevant apparatus |
CN109919072A (en) * | 2019-02-28 | 2019-06-21 | 桂林电子科技大学 | Fine vehicle type recognition and flow statistics method based on deep learning and trajectory tracking |
CN110009722A (en) * | 2019-04-16 | 2019-07-12 | 成都四方伟业软件股份有限公司 | Three-dimensional rebuilding method and device |
CN110188692A (en) * | 2019-05-30 | 2019-08-30 | 南通大学 | A kind of reinforcing that effective target quickly identifies circulation Cascading Methods |
CN110334236A (en) * | 2019-07-12 | 2019-10-15 | 厦门大学 | A kind of fine granularity search method based on Schimidt orthogonalization constraint |
CN110503135A (en) * | 2019-07-31 | 2019-11-26 | 上海交通大学烟台信息技术研究院 | Deep learning model compression method and system for the identification of power equipment edge side |
CN110619304A (en) * | 2019-09-17 | 2019-12-27 | 中控智慧科技股份有限公司 | Vehicle type recognition method, system, device and computer readable medium |
CN110647912A (en) * | 2019-08-15 | 2020-01-03 | 深圳久凌软件技术有限公司 | Fine-grained image recognition method and device, computer equipment and storage medium |
CN110689043A (en) * | 2019-08-22 | 2020-01-14 | 长沙千视通智能科技有限公司 | Vehicle fine granularity identification method and device based on multiple attention mechanism |
CN110991374A (en) * | 2019-12-10 | 2020-04-10 | 电子科技大学 | Fingerprint singular point detection method based on RCNN |
CN111144372A (en) * | 2019-12-31 | 2020-05-12 | 上海眼控科技股份有限公司 | Vehicle detection method, device, computer equipment and storage medium |
CN111274893A (en) * | 2020-01-14 | 2020-06-12 | 中国人民解放军国防科技大学 | Aircraft image fine-grained identification method based on component segmentation and feature fusion |
CN111429215A (en) * | 2020-03-18 | 2020-07-17 | 北京互金新融科技有限公司 | Data processing method and device |
CN111797700A (en) * | 2020-06-10 | 2020-10-20 | 南昌大学 | Vehicle re-identification method based on fine-grained discrimination network and second-order reordering |
CN111914911A (en) * | 2020-07-16 | 2020-11-10 | 桂林电子科技大学 | Vehicle re-identification method based on improved depth relative distance learning model |
CN111931768A (en) * | 2020-08-14 | 2020-11-13 | 中国科学院重庆绿色智能技术研究院 | Vehicle identification method and system capable of self-adapting to sample distribution |
CN111967574A (en) * | 2020-07-20 | 2020-11-20 | 华南理工大学 | Convolutional neural network training method based on tensor singular value delimitation |
CN112070048A (en) * | 2020-09-16 | 2020-12-11 | 福州大学 | Vehicle attribute identification method based on RDSNet |
CN112149720A (en) * | 2020-09-09 | 2020-12-29 | 南京信息工程大学 | Fine-grained vehicle type identification method |
CN112418168A (en) * | 2020-12-10 | 2021-02-26 | 深圳云天励飞技术股份有限公司 | Vehicle identification method, device, system, electronic equipment and storage medium |
CN112949408A (en) * | 2021-02-02 | 2021-06-11 | 华电西藏能源有限公司 | Real-time identification method and system for target fish passing through fish channel |
CN113159115A (en) * | 2021-03-10 | 2021-07-23 | 中国人民解放军陆军工程大学 | Vehicle fine-grained identification method, system and device based on neural architecture search |
CN113343881A (en) * | 2021-06-21 | 2021-09-03 | 浪潮云信息技术股份公司 | Vehicle brand and model fine-grained classification system and method based on deep learning |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
CN107301383A (en) * | 2017-06-07 | 2017-10-27 | 华南理工大学 | A kind of pavement marking recognition methods based on Fast R CNN |
CN108052861A (en) * | 2017-11-08 | 2018-05-18 | 北京卓视智通科技有限责任公司 | A kind of nerve network system and the model recognizing method based on the nerve network system |
-
2018
- 2018-08-23 CN CN201810968895.6A patent/CN109214441A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
CN107301383A (en) * | 2017-06-07 | 2017-10-27 | 华南理工大学 | A kind of pavement marking recognition methods based on Fast R CNN |
CN108052861A (en) * | 2017-11-08 | 2018-05-18 | 北京卓视智通科技有限责任公司 | A kind of nerve network system and the model recognizing method based on the nerve network system |
Non-Patent Citations (1)
Title |
---|
袁公萍 等: "《基于深度卷积神经网络的车型识别方法》", 《浙江大学学报(工学版)》 * |
Cited By (31)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109902732A (en) * | 2019-02-22 | 2019-06-18 | 哈尔滨工业大学(深圳) | Automobile automatic recognition method and relevant apparatus |
CN109919072A (en) * | 2019-02-28 | 2019-06-21 | 桂林电子科技大学 | Fine vehicle type recognition and flow statistics method based on deep learning and trajectory tracking |
CN109919072B (en) * | 2019-02-28 | 2021-03-19 | 桂林电子科技大学 | Fine vehicle type recognition and flow statistics method based on deep learning and trajectory tracking |
CN110009722A (en) * | 2019-04-16 | 2019-07-12 | 成都四方伟业软件股份有限公司 | Three-dimensional rebuilding method and device |
CN110188692A (en) * | 2019-05-30 | 2019-08-30 | 南通大学 | A kind of reinforcing that effective target quickly identifies circulation Cascading Methods |
CN110188692B (en) * | 2019-05-30 | 2023-06-06 | 南通大学 | Enhanced cyclic cascading method for effective target rapid identification |
CN110334236A (en) * | 2019-07-12 | 2019-10-15 | 厦门大学 | A kind of fine granularity search method based on Schimidt orthogonalization constraint |
CN110503135A (en) * | 2019-07-31 | 2019-11-26 | 上海交通大学烟台信息技术研究院 | Deep learning model compression method and system for the identification of power equipment edge side |
CN110647912A (en) * | 2019-08-15 | 2020-01-03 | 深圳久凌软件技术有限公司 | Fine-grained image recognition method and device, computer equipment and storage medium |
CN110689043A (en) * | 2019-08-22 | 2020-01-14 | 长沙千视通智能科技有限公司 | Vehicle fine granularity identification method and device based on multiple attention mechanism |
CN110619304A (en) * | 2019-09-17 | 2019-12-27 | 中控智慧科技股份有限公司 | Vehicle type recognition method, system, device and computer readable medium |
CN110991374A (en) * | 2019-12-10 | 2020-04-10 | 电子科技大学 | Fingerprint singular point detection method based on RCNN |
CN111144372A (en) * | 2019-12-31 | 2020-05-12 | 上海眼控科技股份有限公司 | Vehicle detection method, device, computer equipment and storage medium |
CN111274893A (en) * | 2020-01-14 | 2020-06-12 | 中国人民解放军国防科技大学 | Aircraft image fine-grained identification method based on component segmentation and feature fusion |
CN111274893B (en) * | 2020-01-14 | 2022-11-08 | 中国人民解放军国防科技大学 | Aircraft image fine-grained identification method based on part segmentation and feature fusion |
CN111429215A (en) * | 2020-03-18 | 2020-07-17 | 北京互金新融科技有限公司 | Data processing method and device |
CN111429215B (en) * | 2020-03-18 | 2023-10-31 | 北京互金新融科技有限公司 | Data processing method and device |
CN111797700A (en) * | 2020-06-10 | 2020-10-20 | 南昌大学 | Vehicle re-identification method based on fine-grained discrimination network and second-order reordering |
CN111914911A (en) * | 2020-07-16 | 2020-11-10 | 桂林电子科技大学 | Vehicle re-identification method based on improved depth relative distance learning model |
CN111914911B (en) * | 2020-07-16 | 2022-04-08 | 桂林电子科技大学 | Vehicle re-identification method based on improved depth relative distance learning model |
CN111967574A (en) * | 2020-07-20 | 2020-11-20 | 华南理工大学 | Convolutional neural network training method based on tensor singular value delimitation |
CN111967574B (en) * | 2020-07-20 | 2024-01-23 | 华南理工大学 | Tensor singular value delimitation-based convolutional neural network training method |
CN111931768A (en) * | 2020-08-14 | 2020-11-13 | 中国科学院重庆绿色智能技术研究院 | Vehicle identification method and system capable of self-adapting to sample distribution |
CN112149720A (en) * | 2020-09-09 | 2020-12-29 | 南京信息工程大学 | Fine-grained vehicle type identification method |
CN112070048A (en) * | 2020-09-16 | 2020-12-11 | 福州大学 | Vehicle attribute identification method based on RDSNet |
CN112418168A (en) * | 2020-12-10 | 2021-02-26 | 深圳云天励飞技术股份有限公司 | Vehicle identification method, device, system, electronic equipment and storage medium |
CN112418168B (en) * | 2020-12-10 | 2024-04-02 | 深圳云天励飞技术股份有限公司 | Vehicle identification method, device, system, electronic equipment and storage medium |
CN112949408A (en) * | 2021-02-02 | 2021-06-11 | 华电西藏能源有限公司 | Real-time identification method and system for target fish passing through fish channel |
CN113159115A (en) * | 2021-03-10 | 2021-07-23 | 中国人民解放军陆军工程大学 | Vehicle fine-grained identification method, system and device based on neural architecture search |
CN113159115B (en) * | 2021-03-10 | 2023-09-19 | 中国人民解放军陆军工程大学 | Vehicle fine granularity identification method, system and device based on neural architecture search |
CN113343881A (en) * | 2021-06-21 | 2021-09-03 | 浪潮云信息技术股份公司 | Vehicle brand and model fine-grained classification system and method based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109214441A (en) | A kind of fine granularity model recognition system and method | |
CN108460356B (en) | Face image automatic processing system based on monitoring system | |
CN110598598A (en) | Double-current convolution neural network human behavior identification method based on finite sample set | |
CN105447473B (en) | A kind of any attitude facial expression recognizing method based on PCANet-CNN | |
CN107506722A (en) | One kind is based on depth sparse convolution neutral net face emotion identification method | |
CN106778604B (en) | Pedestrian re-identification method based on matching convolutional neural network | |
CN105825511B (en) | A kind of picture background clarity detection method based on deep learning | |
CN104182772B (en) | A kind of gesture identification method based on deep learning | |
CN106980854A (en) | Number-plate number recognition methods, device, storage medium and processor | |
CN108399435B (en) | Video classification method based on dynamic and static characteristics | |
CN106570477A (en) | Vehicle model recognition model construction method based on depth learning and vehicle model recognition method based on depth learning | |
CN109858390A (en) | The Activity recognition method of human skeleton based on end-to-end space-time diagram learning neural network | |
CN107766850A (en) | Based on the face identification method for combining face character information | |
CN111079674B (en) | Target detection method based on global and local information fusion | |
CN109063649B (en) | Pedestrian re-identification method based on twin pedestrian alignment residual error network | |
CN110473592B (en) | Multi-view human synthetic lethal gene prediction method | |
CN113239784B (en) | Pedestrian re-identification system and method based on space sequence feature learning | |
CN108090472B (en) | Pedestrian re-identification method and system based on multi-channel consistency characteristics | |
CN104992223A (en) | Intensive population estimation method based on deep learning | |
CN104281853A (en) | Behavior identification method based on 3D convolution neural network | |
CN111753207B (en) | Collaborative filtering method for neural map based on comments | |
CN107247952B (en) | Deep supervision-based visual saliency detection method for cyclic convolution neural network | |
CN108875819B (en) | Object and component joint detection method based on long-term and short-term memory network | |
CN107423747A (en) | A kind of conspicuousness object detection method based on depth convolutional network | |
CN110008819A (en) | A kind of facial expression recognizing method based on figure convolutional neural networks |
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: 20190115 |