CN110348283A - Fine granularity model recognizing method based on the fusion of bilinear model multilayer feature - Google Patents
Fine granularity model recognizing method based on the fusion of bilinear model multilayer feature Download PDFInfo
- Publication number
- CN110348283A CN110348283A CN201910391455.3A CN201910391455A CN110348283A CN 110348283 A CN110348283 A CN 110348283A CN 201910391455 A CN201910391455 A CN 201910391455A CN 110348283 A CN110348283 A CN 110348283A
- Authority
- CN
- China
- Prior art keywords
- layer
- feature
- bilinearity
- correlation information
- auto
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention discloses a kind of fine granularity model recognizing method based on the fusion of bilinear model multilayer feature, include the following steps: (10) first interior auto-correlation information extractions layer by layer: according to convolution feature in each layer of bilinear model convolutional layer, parameter matrix is decomposed into two single order vectors by auto-correlation information bilinearity feature in extract layer;(20) second layer by layer in auto-correlation information extractions: according to convolution feature in each layer of bilinear model convolutional layer, parameter matrix is decomposed into two single order vectors by auto-correlation information bilinearity feature in extract layer;(30) interlayer cross-correlation information extraction: according to each interlayer convolution feature of bilinear model convolutional layer, the interlayer cross-correlation information bilinearity feature of different convolution interlayers is extracted, parameter matrix is decomposed into two single order vectors;(40) multilayer feature merges: fusion multilayer feature obtains vehicle fine granularity feature.Fine granularity model recognizing method of the invention, computation complexity is low, accuracy is high.
Description
Technical field
The invention belongs to intelligent image identification technology field, especially a kind of tool is merged based on bilinear model multilayer feature
Fine granularity model recognizing method.
Background technique
Fine granularity vehicle cab recognition is the classification more refined to subclass vehicles different under same major class.Its purpose
It is that the specifying informations such as Vehicle manufacturers, vehicle model are judged to the vehicle image under any angle and scene.Utilize particulate
Degree vehicle cab recognition technology can preferably analyze road conditions in automatic Pilot field, while traffic police can also cross and pass through fine granularity
Vehicle cab recognition more easily checks violation vehicle.In addition to this, fine granularity vehicle cab recognition also has in other many fields
Have a wide range of applications, such as road section traffic volume flow monitoring analysis, 0vehicle marketing assistant analysis and road video monitoring etc..
In terms of fine granularity model recognizing method is concentrated mainly on following two: the target detection of vehicle and to detection target
Classification.The components such as the preceding face of vehicle, headlight are positioned by algorithm of target detection first, then each component is mentioned respectively
Feature is taken, finally the characteristic synthetic of all components gets up and classifies to target vehicle.Wherein basis uses sets by hand
Fine granularity model recognizing method can be divided into based on manual feature and based on deep by the feature or convolutional neural networks feature of meter
Spend convolution feature two major classes.
Since the target of fine granularity image studies is visually closely similar, the information with distinction is frequently found in carefully
Small regional area, therefore early stage many methods first carry out target detection to input vehicle pictures using the feature artificially designed,
Feature is extracted again later to classify.But in general, the feature of hand-designed is not necessarily optimal for final classification,
And be difficult accurately to detect vehicle under complex scene, thus generalization ability is not strong.
The fast development of promotion and depth learning technology recently as hardware computing capability, some domestic and foreign scholars open
The research begun by the excellent ability in feature extraction development fine granularity vehicle cab recognition of convolutional neural networks.Based on depth convolution feature
Method achieve significant achievement in fine granularity vehicle cab recognition task, but convolutional neural networks for component detection very
The computation complexity of time-consuming, model is very high, this is fatal for large-scale image analysis task.
Therefore, problem of the existing technology is: fine granularity vehicle cab recognition computation complexity is high, accuracy is low.
Summary of the invention
The purpose of the present invention is to provide a kind of fine granularity vehicle cab recognition sides based on the fusion of bilinear model multilayer feature
Method, computation complexity is low, accuracy is high.
Realize the technical solution of the object of the invention are as follows:
A kind of fine granularity model recognizing method based on the fusion of bilinear model multilayer feature, includes the following steps:
(10) first interior auto-correlation information extractions layer by layer: it according to convolution feature in each layer of bilinear model convolutional layer, extracts
Auto-correlation information bilinearity feature in layer, while parameter matrix is decomposed into two single order vectors using bilinearity pond is decomposed;
(20) second interior auto-correlation information extractions layer by layer: it according to convolution feature in each layer of bilinear model convolutional layer, extracts
Auto-correlation information bilinearity feature in layer, while parameter matrix is decomposed into two single order vectors using bilinearity pond is decomposed;
(30) according to each interlayer convolution feature of bilinear model convolutional layer, different volumes interlayer cross-correlation information extraction: are extracted
Interlayer cross-correlation information bilinearity feature between lamination, while parameter matrix is decomposed into two one using bilinearity pond is decomposed
Rank vector;
(40) multilayer feature merges: using the bilinearity pondization of pairs of cross-layer interaction and the bilinearity pond of single cross-layer interaction
Change, merges multilayer feature, obtain vehicle fine granularity feature.
Compared with prior art, remarkable advantage of the invention are as follows:
1, computation complexity is low: the present invention is classified end to end using the completion of B-CNN model, avoids localized region
Detection, simplifies network model, reduces computation complexity;
2, accuracy is high: the present invention is based on B-CNN to extract local detail feature, is based on cross-layer bilinearity pond method, will
Auto-correlation information is merged in the interlayer interactive information of each convolutional layer, layer, and intermediate convolutional layer is made full use of to activate, and is obtained more
The feature representation of robust improves the accuracy rate of fine granularity vehicle cab recognition.
Detailed description of the invention
Fig. 1 is the main flow chart of the fine granularity model recognizing method merged the present invention is based on bilinear model multilayer feature.
Fig. 2 be in Fig. 1 first layer by layer in auto-correlation information extracting step flow chart.
Fig. 3 be in Fig. 1 second layer by layer in auto-correlation information extracting step flow chart.
The flow chart of Fig. 4 cross-correlation information extracting step between the middle layer Fig. 1.
Fig. 5 is the flow chart of multilayer feature fusion steps in Fig. 1.
Fig. 6 is the network structure of bilinearity fusion on VGG-16.
Fig. 7 is influence comparison diagram of the different dimensions to precision.
Specific embodiment
As shown in Figure 1, the present invention is based on the fine granularity model recognizing methods of bilinear model multilayer feature fusion, including such as
Lower step:
(10) first interior auto-correlation information extractions layer by layer: it according to convolution feature in each layer of bilinear model convolutional layer, extracts
Auto-correlation information bilinearity feature in layer, while parameter matrix is decomposed into two single order vectors using bilinearity pond is decomposed;
As shown in Fig. 2, auto-correlation information extracting step includes: in the layer of (10) first layer
(11) first layer by layer in bilinearity character representation: the bilinearity feature that auto-correlation information C × C is tieed up in the layer of first layer
It is expressed as,
bilinear(l,I,fA,fA)=fA(l,I)TfA(l, I)=XTX (1)
In formula, X is the feature note that first layer convolutional layer extracts, and dimension is C × M, and x indicates X in the feature point of a certain position
Amount, the i.e. feature vector in a certain channel of convolutional neural networks;
(12) first interior bilinearity character representation outputs layer by layer: complete bilinear model are as follows:
zi=xTWix (2)
Wherein Wi∈Rc×cFor projection matrix, ziIt is exported for the bilinearity character representation of position B-CNN;
(13) projection matrix decomposes: by projection matrix WiIt is decomposed into two single order vectors:
Wherein Ai∈Rc;
(14) auto-correlation information exports in layer: study tensor representation W=[W1,W2,...,Wo]∈Rc×c×o, obtain first layer
Auto-correlation information in the layer of convolutional layer, o dimension output z ∈ Ro:
Wherein A ∈ Rc×dFor projection matrix, P ∈ Rd×oIt is classification matrix,It is Hadamard product, d is then to determine that joint is embedding
Enter the hyper parameter of dimension.
(20) second interior auto-correlation information extractions layer by layer: it according to convolution feature in each layer of bilinear model convolutional layer, extracts
Auto-correlation information bilinearity feature in layer, while parameter matrix is decomposed into two single order vectors using bilinearity pond is decomposed;
As shown in figure 3, described (20) second layer by layer in auto-correlation information extracting step include:
(21) second layer by layer in bilinearity character representation: the bilinearity feature that auto-correlation information C × C is tieed up in the layer of the second layer
It is expressed as,
bilinear(l,I,fA,fA)=fA(l,I)TfA(l, I)=YTY (5)
In formula, the feature that second layer convolutional layer extracts is denoted as Y, and dimension is C × N, and y indicates Y in the feature point of a certain position
Amount;
(22) second interior bilinearity character representation outputs layer by layer: complete bilinear model are as follows:
zi=yTWiy (6)
Wherein Wi∈Rc×cFor projection matrix, ziIt is exported for the bilinearity character representation of position B-CNN.
(23) second layer projection matrix decomposes: by projection matrix WiTwo single order vectors are decomposed into,
Wherein Bi∈Rc。
(24) second interior auto-correlation information outputs layer by layer: study tensor representation W=[W1,W2,...,Wo]∈Rc×c×o, obtain
Auto-correlation information in the layer of second layer convolutional layer, o dimension output z ∈ Ro:
Wherein B ∈ Rc×dFor projection matrix, P ∈ Rd×oIt is classification matrix,It is Hadamard product, d is then to determine that joint is embedding
Enter the hyper parameter of dimension.
(30) according to each interlayer convolution feature of bilinear model convolutional layer, different volumes interlayer cross-correlation information extraction: are extracted
Interlayer cross-correlation information bilinearity feature between lamination, while parameter matrix is decomposed into two one using bilinearity pond is decomposed
Rank vector;
As shown in figure 4, (30) the interlayer cross-correlation information extracting step includes:
(31) interlayer cross-correlation acquisition of information: interlayer cross-correlation information XTY、YTX is two extracted to different convolutional layers
Feature seeks apposition,
bilinear(l,I,fA,fB)=fA(l,I)TfB(l, I)=XTY (9)
bilinear(l,I,fA,fB)=fB(l,I)TfA(l, I)=YTX (10)
Wherein, fA、fBThe feature of the two adjacent convolutional layers extracted is denoted as X and Y, and dimension is respectively C × M and C × N, x
The characteristic component of X and Y at same position, the i.e. feature vector in the same channel of convolutional neural networks are respectively indicated with y;
(32) interlayer bilinear model exports: complete bilinear model is respectively as follows:
zi=xTWiy (11)
zi=yTWix (12)
Wherein Wi∈Rc×cFor projection matrix, ziIt is exported for the bilinearity character representation of position B-CNN;
(33) interlayer projection matrix decomposes: by projection matrix WiIt is separately disassembled into two single order vectors:
Wherein Ai∈Rc、Bi∈Rc、
(34) interlayer cross-correlation information exports: study tensor representation W=[W1,W2,...,Wo]∈Rc×c×o, it is mutual to obtain interlayer
Relevant information XTY、YTX o dimension output z ∈ Ro:
WhereinFor projection matrix, P ∈ Rd×oIt is classification matrix,It is Hadamard product, d then determines
Surely combine the hyper parameter of Embedded dimensions.
(40) multilayer feature merges: using the bilinearity pondization of pairs of cross-layer interaction and the bilinearity pond of single cross-layer interaction
Change, merges multilayer feature, obtain vehicle fine granularity feature.
As shown in figure 5, (40) the multilayer feature fusion steps include:
(41) cross-layer bilinearity the bilinearity pond based on the interaction of pairs of cross-layer: is indicated into YTX is handed over based on single cross-layer
Mutual bilinearity pondization is superimposed, forms another fusion feature vector Z after splicing,
Z=XTX+YTY+XTY+YTX (17)
The calculation formula of its backpropagation indicates are as follows:
(42) bilinearity of two convolutional layers the bilinearity pond based on the interaction of single cross-layer: is indicated into XTX and YTY and across
Layer bilinearity indicates XTY is superimposed, and a fusion feature vector Z is formed after splicing:
Z=XTX+YTY+XTY (20)
The calculation formula of its backpropagation indicates are as follows:
For the validity for verifying the method for the present invention, inventor has carried out following confirmatory experiment.
Confirmatory experiment of the invention is using data set: Stanford Cars-196 and Stanford BMW-10.
Stanford Cars-196 and Stanford BMW-10 are to be currently being widely used due to having certain scale and complexity
Two fine granularity vehicle cab recognition data sets.Stanford Cars-196 includes 196 kinds of class of vehicle, totally 16185 automobile figures
Piece is divided according to the production firm of vehicle, model and productive year, such as: Audi A4 in 2011.Each subclass difference
The image not waited comprising 48~136, wherein 24~68 are divided into training image, remaining image is as test set.
Stanford BMW-10 data set then includes the different angle picture of 10 BMW car systems, and every class vehicle packet 2 is containing the training of about 25 width
Picture.Use disclosed VGG-16 that water is used only in order to compare under same standard with other methods as basic network model
Flat overturning is to expand training samples number.The classification results of original image and its flipped image are averaged during the test
Value, the result as final classification.
Experimental Hardware environment: Ubuntu 16.04, GTX1080ti, video memory 12G, Core (TM) i7 processor, dominant frequency are
3.4G inside saves as 16G.
Code running environment: source code library (MatConvNet), python2.7, Matlab 2014a.
1, model validation experimental analysis
Feasibility analysis is carried out to two different Fusion Features schemes.Because in VGG-16, compared with shallow-layer feature,
Relu5_1, relu5_2 and relu5_3 include more part of semantic information, therefore using VGG-D as basic model, are used
Tri- kinds of integration programs of relu5_3+relu5_2, relu5_3+relu5_1 and relu5_3+relu5_2 are in Stanford Cars-
196 compare experiment.Experimental result (black overstriking data indicate precision highest) as shown in table 3-1.
Although DCL-Fusion introduces additional interlayer interactive information, Stanford compared to SCL-Fusion
It is on Cars-196 the experimental results showed that DCL-Fusion performance there is no promoted, increase calculating cost instead, cause to instruct
It is slack-off to practice the time.Infer by careful analysis, it may be possible to XTY and YTX transposition and include similar information each other, is added YTX it
Cause the ratio of X and Y cross-correlation information excessive afterwards, SCL-Fusion Fusion Features are not helped instead.Therefore later
Experiment in, do not add YTX only selects network structure on the first Fusion Features scheme SCL-Fusion, VGG-16 such as
Shown in Fig. 6.
Model efficiency analysis on table 1 relu5_3+relu5_2, relu5_3+relu5_1 and relu5_3+relu5_2
2, the experimental analysis of parameter d adjustment
D is the hyper parameter for determining joint Embedded dimensions in decomposing bilinearity pond model, in order to study SCL-Fusion
Influence of the middle parameter d to experimental result, is tested on Stanford Cars-196 data set, by the reality of SCL-Fusion
It tests result to compare with general decomposition bilinearity pond, as a result as shown in Figure 7.
As it can be seen from table 1 in the integration program of three kinds of different layers, relu5_3 and two layers of relu5_1 of Fusion Features
Better nicety of grading can be obtained, therefore in an experiment, be carried out using the feature of two convolutional layers of relu5_1 and relu5_3
Fusion, captures this two layers auto-correlation and cross-correlation information.What general decomposition bilinearity pondization was extracted is the last one convolution
The auto-correlation information of layer relu5_3.By testing it can be found that the result of SCL-Fusion is significantly better than logical at identical d
Decomposition bilinearity pond shows that richer information can be extracted by the interlayer interaction of feature, enhances sentencing for model
Other ability.
Observing Fig. 5 simultaneously can see, as hyper parameter d is gradually varied to 8192 from 512, general decomposition bilinearity pond
Change and the nicety of grading of two models of SCL-Fusion can gradually rise.As d=8192, SCL-Fusion is in Stanford
93.05% optimum efficiency is obtained on Cars-196.Therefore hyper parameter d=8192 in next experiment.
3, Fusion Features effect analysis
Experiment before shows that convolutional layer Fusion Features are obviously improved nicety of grading, therefore in Stanford
Quantitative experiment is carried out on Cars-196 data set, the Fusion Features for analyzing which convolutional layer in SCL-Fusion can obtain highest
Nicety of grading.According to Fig.5, as a result, setting insertion dimension d=8192, combines different convolutional layers and carry out Fusion Features
Classifying quality is studied, and the results are shown in Table 2 (black overstriking data indicate precision highest).
It can be found that experiment can obtain when the Fusion Features of tri- convolutional layers of relu5_3, relu5_2 and relu5_1
Best effect.Compared with the scheme that relu5_1+relu5_3 fusion is used only, nicety of grading improves 0.4%.Show centre
Convolutional layer activation is effective to fine grit classification task.Because in the forward and reverse communication process of CNN, there are information to lose
Lose, when propagating among vital distinguishing characteristics, which may have lost, to be identified for fine granularity in convolutional layer.With general point
Solution bilinear model is compared, and SCL-Fusion considers the interaction feature of intermediate convolutional layer, the auto-correlation information including convolutional layer
It is therefore more steady with cross-correlation information.The auto-correlation of relu5_3, relu5_2 and relu5_1 are used in subsequent experiment
It is merged with cross-correlation information.
The Comparative result of the different convolutional layer fusions of table 2
4, with the Comparision of forefathers
In order to further verify the fine granularity image classification proposed by the present invention based on the fusion of bilinear model multilayer feature
The validity of algorithm is compared it with performance of the mainstream algorithm on Stanford Cars-196 data set, experiment knot
Fruit is as shown in table 3 (black overstriking data indicate precision highest).
Comparison of experiment results on 3 Stanford Cars-196 of table
Method proposed by the present invention achieves 93.45% recognition accuracy under weak supervision condition, is higher by than BoT algorithm
0.95%, while lower than HSnet algorithm 0.45%.But two algorithms of BoT and HSnet all employ additionally in the training stage
Markup information belongs to strong supervision algorithm.Algorithm proposed by the present invention and unsupervised method RA-CNN, MA-CNN based on component
It compares, nicety of grading improves 1.05% and 0.95% respectively.In addition, the algorithm ratio Improved B-CNN and LRBP effect are all
It is better, but effect is lower than the bilinearity pond algorithm HBP of layering.Because HBP is not only supported compared with method proposed by the present invention
The feature interaction of interlayer, while learning fine granularity in a manner of mutually enhancing and indicate.Network knot of the DLA algorithm to current main-stream
Structure (VGG, ResNet, ResNeXt, DenseNet etc.) summarize abstract, proposes a kind of significantly more efficient mode to take
Establishing network structure achieves highest accuracy of identification on current Stanford Cars-196 data set.
The recognition effect comparison of method proposed by the present invention and typical fine granularity recognition methods on BMW-10 data set is such as
Shown in table 4.Because the difference between each subclass vehicle of fine granularity vehicle classification is very subtle, the knowledge in the past based on manual feature
Other classification effect is bad, and SPM algorithm only achieves 66.1% classification accuracy on BMW-10 data set.BB algorithm is
The ability in feature extraction of enhancing localized region, artificial selection have the image-region of discrimination, achieve 69.3% knowledge
Other effect shows that the effective position for regional area can be obviously improved the accuracy of fine granularity vehicle cab recognition.BB-3D-G exists
It is improved on the basis of this, BB method is promoted to 3d space to the influence for eliminating visual angle, so that recognition accuracy improves
6.7%.The fine granularity vehicle cab recognition that BoxCars uses 3D rectangle frame relevant information to achieve 77.2% as CNN input is accurate
Rate.And method proposed by the present invention is not under the premise of by additional callout box, by relu5_3, relu5_2 and relu5_1 tri-
The convolutional layer feature of layer finds out the cross-correlation information (relu5_3*relu5_ of its autocorrelative bilinearity feature and interlayer respectively
2+relu5_2*relu5_1+relu5_3*relu5_1), using decompose bilinearity pond by projection matrix be decomposed into two it is one-dimensional
Vector achieves 79.27% vehicle cab recognition accuracy to reduce calculation amount, improves 2.95% compared to BB-3D-G method, table
Bright fine granularity model recognizing method proposed by the present invention can effectively improve the accuracy of fine granularity vehicle cab recognition.
Comparison of experiment results on 4 Stanford BMW-10 of table
Claims (5)
1. a kind of fine granularity model recognizing method based on the fusion of bilinear model multilayer feature, which is characterized in that including as follows
Step:
(10) first layer by layer in auto-correlation information extractions: according to convolution feature in each layer of bilinear model convolutional layer, in extract layer
Auto-correlation information bilinearity feature, while parameter matrix is decomposed into two single order vectors using bilinearity pond is decomposed;
(20) second layer by layer in auto-correlation information extractions: according to convolution feature in each layer of bilinear model convolutional layer, in extract layer
Auto-correlation information bilinearity feature, while parameter matrix is decomposed into two single order vectors using bilinearity pond is decomposed;
(30) according to each interlayer convolution feature of bilinear model convolutional layer, different convolutional layers interlayer cross-correlation information extraction: are extracted
Between interlayer cross-correlation information bilinearity feature, while using decompose bilinearity pond by parameter matrix be decomposed into two single orders to
Amount;
(40) multilayer feature merges: bilinearity pondization and the bilinearity pond of single cross-layer interaction using the interaction of pairs of cross-layer,
Multilayer feature is merged, vehicle fine granularity feature is obtained.
2. fine granularity model recognizing method according to claim 1, which is characterized in that in the layer of (10) first layer certainly
Related information extraction step includes:
(11) first layer by layer in bilinearity character representation: the bilinearity character representation that auto-correlation information C × C is tieed up in the layer of first layer
For,
bilinear(l,I,fA,fA)=fA(l,I)TfA(l, I)=XTX (1)
In formula, X be first layer convolutional layer extract feature note, dimension be C × M, x indicate X a certain position characteristic component, i.e.,
The feature vector in a certain channel of convolutional neural networks;
(12) first interior bilinearity character representation outputs layer by layer: complete bilinear model are as follows:
zi=xTWix (2)
Wherein Wi∈Rc×cFor projection matrix, ziIt is exported for the bilinearity character representation of position B-CNN;
(13) projection matrix decomposes: by projection matrix WiIt is decomposed into two single order vectors:
Wherein Ai∈Rc;
(14) auto-correlation information exports in layer: study tensor representation W=[W1,W2,...,Wo]∈Rc×c×o, obtain first layer convolution
Auto-correlation information in the layer of layer, o dimension output z ∈ Ro:
Wherein A ∈ Rc×dFor projection matrix, P ∈ Rd×oIt is classification matrix,It is Hadamard product, d is then to determine joint insertion dimension
Several hyper parameters.
3. fine granularity model recognizing method according to claim 1, which is characterized in that described (20) second is interior from phase layer by layer
Closing information extracting step includes:
(21) second layer by layer in bilinearity character representation: the bilinearity character representation that auto-correlation information C × C is tieed up in the layer of the second layer
For,
bilinear(l,I,fA,fA)=fA(l,I)TfA(l, I)=YTY (5)
In formula, the feature that second layer convolutional layer extracts is denoted as Y, and dimension is C × N, and y indicates Y in the characteristic component of a certain position;
(22) second interior bilinearity character representation outputs layer by layer: complete bilinear model are as follows:
zi=yTWiy (6)
Wherein Wi∈Rc×cFor projection matrix, ziIt is exported for the bilinearity character representation of position B-CNN.
(23) second layer projection matrix decomposes: by projection matrix WiTwo single order vectors are decomposed into,
Wherein Bi∈Rc。
(24) second interior auto-correlation information outputs layer by layer: study tensor representation W=[W1,W2,...,Wo]∈Rc×c×o, obtain second
Auto-correlation information in the layer of layer convolutional layer, o dimension output z ∈ Ro:
Wherein B ∈ Rc×dFor projection matrix, P ∈ Rd×oIt is classification matrix,It is Hadamard product, d is then to determine joint Embedded dimensions
Hyper parameter.
4. fine granularity model recognizing method according to claim 1, which is characterized in that (30) the interlayer cross-correlation information
Extraction step includes:
(31) interlayer cross-correlation acquisition of information: interlayer cross-correlation information XTY、YTX is two features extracted to different convolutional layers
Apposition is sought,
bilinear(l,I,fA,fB)=fA(l,I)TfB(l, I)=XTY (9)
bilinear(l,I,fA,fB)=fB(l,I)TfA(l, I)=YTX (10)
Wherein, fA、fBThe feature of the two adjacent convolutional layers extracted is denoted as X and Y, and dimension is respectively C × M and C × N, and x and y divide
It Biao Shi not characteristic component of the X and Y at same position, the i.e. feature vector in the same channel of convolutional neural networks;
(32) interlayer bilinear model exports: complete bilinear model is respectively as follows:
zi=xTWiy (11)
zi=yTWix (12)
Wherein Wi∈Rc×cFor projection matrix, ziIt is exported for the bilinearity character representation of position B-CNN;
(33) interlayer projection matrix decomposes: by projection matrix WiIt is separately disassembled into two single order vectors:
Wherein Ai∈Rc、Bi∈Rc、
(34) interlayer cross-correlation information exports: study tensor representation W=[W1,W2,...,Wo]∈Rc×c×o, obtain interlayer cross-correlation
Information XTY、YTX o dimension output z ∈ Ro:
Z=XTY=PT(ATxoBTy) (15)
WhereinFor projection matrix, P ∈ Rd×oIt is classification matrix, o is Hadamard product, and d is then to determine connection
Close the hyper parameter of Embedded dimensions.
5. fine granularity model recognizing method according to claim 1, which is characterized in that (40) the multilayer feature fusion step
Suddenly include:
(41) cross-layer bilinearity the bilinearity pond based on the interaction of pairs of cross-layer: is indicated into YTX interacts double with based on single cross-layer
Linear pondization is superimposed, forms another fusion feature vector Z after splicing,
Z=XTX+YTY+XTY+YTX (17)
The calculation formula of its backpropagation indicates are as follows:
(42) bilinearity of two convolutional layers the bilinearity pond based on the interaction of single cross-layer: is indicated into XTX and YTY and cross-layer are double
Linear expression XTY is superimposed, and a fusion feature vector Z is formed after splicing:
Z=XTX+YTY+XTY (20)
The calculation formula of its backpropagation indicates are as follows:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910391455.3A CN110348283A (en) | 2019-05-13 | 2019-05-13 | Fine granularity model recognizing method based on the fusion of bilinear model multilayer feature |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910391455.3A CN110348283A (en) | 2019-05-13 | 2019-05-13 | Fine granularity model recognizing method based on the fusion of bilinear model multilayer feature |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110348283A true CN110348283A (en) | 2019-10-18 |
Family
ID=68174251
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910391455.3A Pending CN110348283A (en) | 2019-05-13 | 2019-05-13 | Fine granularity model recognizing method based on the fusion of bilinear model multilayer feature |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110348283A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112183602A (en) * | 2020-09-22 | 2021-01-05 | 天津大学 | Multi-layer feature fusion fine-grained image classification method with parallel rolling blocks |
CN113011362A (en) * | 2021-03-29 | 2021-06-22 | 吉林大学 | Fine-grained fundus image grading algorithm based on bilinear pooling and attention mechanism |
CN114998964A (en) * | 2022-06-02 | 2022-09-02 | 天津道简智创信息科技有限公司 | Novel license quality detection method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108280490A (en) * | 2018-02-28 | 2018-07-13 | 北京邮电大学 | A kind of fine granularity model recognizing method based on convolutional neural networks |
CN109359684A (en) * | 2018-10-17 | 2019-02-19 | 苏州大学 | Fine granularity model recognizing method based on Weakly supervised positioning and subclass similarity measurement |
CN109685115A (en) * | 2018-11-30 | 2019-04-26 | 西北大学 | A kind of the fine granularity conceptual model and learning method of bilinearity Fusion Features |
-
2019
- 2019-05-13 CN CN201910391455.3A patent/CN110348283A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108280490A (en) * | 2018-02-28 | 2018-07-13 | 北京邮电大学 | A kind of fine granularity model recognizing method based on convolutional neural networks |
CN109359684A (en) * | 2018-10-17 | 2019-02-19 | 苏州大学 | Fine granularity model recognizing method based on Weakly supervised positioning and subclass similarity measurement |
CN109685115A (en) * | 2018-11-30 | 2019-04-26 | 西北大学 | A kind of the fine granularity conceptual model and learning method of bilinearity Fusion Features |
Non-Patent Citations (3)
Title |
---|
CHAOJIAN YU等: "Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition", 《ARXIV》 * |
SIJIA CAI 等: "Higher-order Integration of Hierarchical Convolutional Activations for Fine-grained Visual Categorization", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER》 * |
王昊然: "基于多层卷积特征高阶融合的多任务目标检测***研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112183602A (en) * | 2020-09-22 | 2021-01-05 | 天津大学 | Multi-layer feature fusion fine-grained image classification method with parallel rolling blocks |
CN112183602B (en) * | 2020-09-22 | 2022-08-26 | 天津大学 | Multi-layer feature fusion fine-grained image classification method with parallel rolling blocks |
CN113011362A (en) * | 2021-03-29 | 2021-06-22 | 吉林大学 | Fine-grained fundus image grading algorithm based on bilinear pooling and attention mechanism |
CN114998964A (en) * | 2022-06-02 | 2022-09-02 | 天津道简智创信息科技有限公司 | Novel license quality detection method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110852316B (en) | Image tampering detection and positioning method adopting convolution network with dense structure | |
Chen et al. | A serial image copy-move forgery localization scheme with source/target distinguishment | |
CN106897714B (en) | Video motion detection method based on convolutional neural network | |
Kong et al. | Detect and locate: Exposing face manipulation by semantic-and noise-level telltales | |
CN105550701B (en) | Realtime graphic extracts recognition methods and device | |
CN111311563A (en) | Image tampering detection method based on multi-domain feature fusion | |
CN110348283A (en) | Fine granularity model recognizing method based on the fusion of bilinear model multilayer feature | |
CN104616664B (en) | A kind of audio identification methods detected based on sonograph conspicuousness | |
CN111080629A (en) | Method for detecting image splicing tampering | |
CN107368787A (en) | A kind of Traffic Sign Recognition algorithm that application is driven towards depth intelligence | |
CN107767405A (en) | A kind of nuclear phase for merging convolutional neural networks closes filtered target tracking | |
CN109815867A (en) | A kind of crowd density estimation and people flow rate statistical method | |
CN109657551B (en) | Face detection method based on context information enhancement | |
CN112150450B (en) | Image tampering detection method and device based on dual-channel U-Net model | |
CN113762138B (en) | Identification method, device, computer equipment and storage medium for fake face pictures | |
CN107909081A (en) | The quick obtaining and quick calibrating method of image data set in a kind of deep learning | |
CN101364263A (en) | Method and system for detecting skin texture to image | |
CN107305691A (en) | Foreground segmentation method and device based on images match | |
Chen et al. | ASF-Net: Adaptive screening feature network for building footprint extraction from remote-sensing images | |
CN112990282B (en) | Classification method and device for fine-granularity small sample images | |
Li et al. | Image manipulation localization using attentional cross-domain CNN features | |
CN113221770B (en) | Cross-domain pedestrian re-recognition method and system based on multi-feature hybrid learning | |
CN111160356A (en) | Image segmentation and classification method and device | |
CN104978569A (en) | Sparse representation based incremental face recognition method | |
CN104200202B (en) | A kind of upper half of human body detection method based on cumulative perceptron |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191018 |
|
WD01 | Invention patent application deemed withdrawn after publication |