CN110378366A - A kind of cross-domain image classification method based on coupling knowledge migration - Google Patents
A kind of cross-domain image classification method based on coupling knowledge migration Download PDFInfo
- Publication number
- CN110378366A CN110378366A CN201910482559.5A CN201910482559A CN110378366A CN 110378366 A CN110378366 A CN 110378366A CN 201910482559 A CN201910482559 A CN 201910482559A CN 110378366 A CN110378366 A CN 110378366A
- Authority
- CN
- China
- Prior art keywords
- data
- source domain
- domain
- field data
- label information
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of cross-domain image classification methods based on coupling knowledge migration to eliminate the difference of source domain and the distribution of aiming field data edges and the distribution of class condition by finding source domain and the common lower-dimensional subspace of aiming field based on Largest Mean difference criterion；According to the respective adjacent map of pseudo label information architecture of the label information of source domain data and target numeric field data, structural integrity of the data from luv space to lower-dimensional subspace is kept, while can dynamically adjust the structure of adjacent map, promotes the positive transfer of knowledge in domain；Using the source domain data training nearest neighbor classifier of tape label information in lower-dimensional subspace, continuous iteration optimization is carried out to the pseudo label information of target numeric field data and completes cross-domain image classification to obtain the final label information of target numeric field data；In addition the different confidence level of the pseudo label that the method for the present invention assigns target area image by design sample weight weighted strategy, is effectively reduced the negative transfer of knowledge in domain, improves the precision of cross-domain image classification.
Description
Technical field
The present invention relates to computer visual image sorting technique field more particularly to it is a kind of based on coupling knowledge migration across
Area image classification method.
Background technique
Traditional machine learning algorithm usually requires that the data sample of a large amount of tape label, and require training sample with
Test sample obeys independent and identical distribution.Direct method is by the data set migration of existing tape label to newly unknown
On data set, i.e., aiming field is moved to from source domain.But the difference of distribution is certainly existed between different data collection, " is obeyed independent
The hypothesis of same distribution " is often invalid.Therefore, traditional image classification method is difficult to obtain preferable classification performance.
In order to overcome the difference of sample distribution between source domain and aiming field, the generalization ability of sorting algorithm is promoted, at present master
It to adapt to be commonly divided into based on feature as the means for handling the knowledge migration problem in cross-cutting image classification by field
The method of transformation, the method for Case-based Reasoning and the method based on neural network migration.However current cross-cutting study remains
Many needs solve the problems, such as and perfect, for example, under source domain and the biggish situation of aiming field distributional difference, based on feature turn
The classification performance of the field adaptive method changed is far from reaching the initial expectation of people；The method of Case-based Reasoning needs to calculate additional
Sample weights, computation complexity is high, cannot carry out quickly cross-cutting study to large data sets；Method neural network based
It is required that the source domain sample of a large amount of tape labels carries out training for a long time, limited in actual application larger.
Summary of the invention
The present invention is to solve existing cross-domain image classification method to calculate the problems such as time is long, classification accuracy is insufficient, is mentioned
A kind of cross-domain image classification method based on coupling knowledge migration is supplied.
To realize the above goal of the invention, and the technological means used is:
A kind of cross-domain image classification method based on coupling knowledge migration, comprising the following steps:
S1. source domain image and target area image are obtained and carries out feature extraction, respectively obtains source domain eigenmatrix as source
Numeric field data obtains target domain characterization matrix as target numeric field data；
S2. the adjacent map of source domain data and the adjoining of target numeric field data are constructed respectively according to the label information of source domain data
Figure, the side right by calculating adjacent map respectively obtain the weight matrix of source domain and the weight matrix of aiming field again；
S3. data distribution is aligned between carrying out domain in the common lower-dimensional subspace of the source domain and aiming field, is utilized simultaneously
The weight matrix of source domain and the weight matrix of aiming field carry out the knowledge migration of data in domain；
S4. it in the lower-dimensional subspace, using the source domain data training nearest neighbor classifier of tape label information, and is used for
Predict the classification information of target numeric field data；
S5. the label information for the target numeric field data predicted is weighted again；
S6. judge whether to reach maximum number of iterations, if not return step S2, if then exporting predicted aiming field figure
The label information of picture.
In above scheme, by the common lower-dimensional subspace of searching source domain and target source, the difference of sample distribution is reduced, together
Shi Liyong adjacent map carries out knowledge positive transfer in domain, keeps leading to from luv space to the structural integrity of common lower-dimensional subspace
The difference for eliminating source domain and the distribution of aiming field data edges and the distribution of class condition is crossed, so that data distribution is aligned, thus in low-dimensional
The label information of target numeric field data is predicted in subspace using the source domain data training nearest neighbor classifier of tape label information, most
It is weighted again by the label information to the target numeric field data predicted afterwards and inhibits knowledge negative transfer, to realize cross-domain image
Accurate classification.
Preferably, the specific steps of the step S1 are as follows:
S1.1. n is inputtedsOpen the source domain image and n of tape labeltOpen the target area image without label；Wherein ns、ntIt is positive whole
Number；
S1.2. feature extraction is carried out to the source domain image and target area image respectively, it will be to the source domain image zooming-out
Obtained feature vector obtains source domain data characteristics matrix by column arrangementWherein m is characterized
The dimension of vector, nsFor the number of feature vector；
The feature vector obtained to the aiming field image zooming-out is obtained into aiming field data characteristics matrix by column arrangementWherein m is the dimension of feature vector, ntFor the number of feature vector；
The feature vector of the source domain image indicates the source domain data of source domain image, the feature vector of the target area image
Indicate the target numeric field data of target area image.
Preferably, feature extraction described in step S1 extracts the source specifically by the neural network of pre-training respectively
The feature vector of area image and target area image.In this preferred embodiment, VGG16, GoogLeNet is can be used in neural network,
ResNet50 etc..
Preferably, feature extraction described in step S1 is specially and carries out respectively to the source domain image and target area image
After gray processing, the gray value of the source domain image is arranged in feature vector and arranges the gray value of the target area image
At feature vector.
Preferably, the S2 specific steps are as follows:
S2.1. first time iteration is judged whether it is, if otherwise directly carrying out step S2.2, if constructing arest neighbors classification
Device simultaneously trains the nearest neighbor classifier using the source domain data, for predicting the pseudo label information of target numeric field data
S2.2. according to the label information of the source domain data, the adjacent map of source domain data is constructed, i.e., in source domain data
Adjacent side is established between homogeneous data, and calculates the weight of adjacent side, obtains weight matrix
Wherein WsFor the weight matrix of source domain, it is calculate by the following formula:
WhereinFor the feature vector of i-th Zhang Yuan's area image,For the corresponding label information of i-th Zhang Yuan's area image, σ is
The bandwidth of Gaussian kernel, | | | | indicate two norm of vector；
According to the pseudo label information of the target numeric field data, the adjacent map of target numeric field data is constructed, in target numeric field data
Homogeneous data between establish adjacent side, and calculate the weight of adjacent side, obtain weight matrixWherein WtFor target
The weight matrix in domain is simultaneously calculate by the following formula:
WhereinFor the feature vector of i-th target area image,For the corresponding pseudo label of i-th target area image, βi
ForWeight, initial value is 1.
In this preferred embodiment, by respectively in source domain and aiming field construct adjacent map method, data can be kept
Structural integrity from luv space to the invariant subspace constructs adjacent map using the semantic information of source domain data tag information,
The structure of adjacent map can be dynamically adjusted, to excavate the more essential partial structurtes of data, promote in domain knowledge just
Migration.
Preferably, the S3 specific steps are as follows:
S3.1. it in lower-dimensional subspace common in the source domain and aiming field, is carried out using Largest Mean difference criterion
Data distribution is aligned between domain including edge distribution and conditional probability distribution, specifically by the following first object letter of minimum
Number:
Wherein C indicates the class number of label information, xs,cAnd xt,cRespectively indicate the classification category of source domain data tag information
Belong to the set of classification c in the set of classification c and the classification of aiming field data tag information,WithRespectively indicate source domain number
Belong to the number of the data of the classification c of label information according to the number and target numeric field data of the data for the classification c for belonging to label information,For projection matrix, wherein d < < m, projects to d n-dimensional subspace n for tieing up data from m；
S3.2. weight matrix W is utilizedsAnd WtThe knowledge migration for carrying out data in domain, specifically by below minimizing second
Objective function:
S3.3. the first object function and the second objective function are combined and a regularization is addedObtain with
Lower third objective function:
s.t.PTXHXTP=I
Wherein λ and α is balance factor, λ > 0 and α > 0, X=[Xs,Xt], change matrix centered on H:I is unit matrix,Complete 1 column vector for being 1 for value；
Projection matrix P is solved by generalized eigen decomposition, obtains source domain data and target numeric field data in low-dimensional
Expression in space is respectively as follows:With
In this preferred embodiment, by found based on Largest Mean difference criterion source domain and aiming field it is common low-dimensional son
The difference of source domain and the distribution of aiming field data edges and the distribution of class condition is eliminated in space.
Preferably, the specific steps of the step S4 are as follows:
According to expression of the source domain data in lower-dimensional subspace, the source of the tape label information in lower-dimensional subspace is obtained
Numeric field dataFor training the nearest neighbor classifier, and predict the target numeric field data in lower dimensional space's
Classification information
In this preferred embodiment, classified in lower-dimensional subspace using the source domain data training arest neighbors of tape label information
Device carries out continuous iteration optimization to the pseudo label information of target numeric field data, to obtain the final label information of target numeric field data.
Preferably, the specific steps of the step S5 include:
S5.1. to the target numeric field data in lower dimensional spaceA new adjacent map is constructed, i.e., will
Each aiming field data point in lower dimensional spaceThe k aiming field number nearest with its Euclidean distance respectively
Adjacent side is established at strong point；
S5.2. based on adjacent map described in step S5.1, for i-th of aiming field data point on authorized adjacent mapIf withK connected aiming field data point all withA classification is belonged to, then aiming field data pointLabel
Confidence level βi=1, otherwise βi=0, wherein k is positive integer.
In this preferred embodiment, based on sample weight weighted strategy, assign aiming field sample different confidence levels, thus effectively
Ground inhibits the negative transfer of knowledge in domain.
Compared with prior art, the beneficial effect of technical solution of the present invention is:
The method of the present invention by carried out using Largest Mean difference criterion include edge distribution and conditional probability distribution domain
Between data distribution be aligned, find source domain and the common lower-dimensional subspace of aiming field, eliminate source domain and aiming field data edges are distributed
The difference being distributed with the distribution of class condition, so that data distribution is aligned；
According to the respective adjacent map of pseudo label information architecture of the label information of source domain data and target numeric field data, number is kept
According to the structural integrity from luv space to lower-dimensional subspace, while the structure of adjacent map can be dynamically adjusted, promote to know in domain
The positive transfer of knowledge；Using the source domain data training nearest neighbor classifier of tape label information in lower-dimensional subspace, to aiming field number
According to pseudo label information carry out continuous iteration optimization and complete cross-domain figure to obtain the final label information of target numeric field data
As classification；
In addition the different confidence of the pseudo label that the method for the present invention assigns target area image by design sample weight weighted strategy
Degree, is effectively reduced the negative transfer of knowledge in domain, improves the precision of cross-domain image classification.
The present invention is based on the method for iteration, data distribution alignment can be effectively combined, promote in domain knowledge positive transfer and
Inhibit knowledge negative transfer in domain, to achieve the purpose that couple knowledge migration, greatly improves the performance of cross-domain image classification.
It solves existing cross-domain image classification method and calculates the problems such as time is long, classification accuracy is insufficient.
Detailed description of the invention
Fig. 1 is the flow chart of the method for the present invention.
Fig. 2 is the effect picture between carrying out domain in the present embodiment 2 after data distribution alignment.
Fig. 3 is the label information category distribution effect picture in the present embodiment 2 to aiming field in lower dimensional space.
Specific embodiment
The attached figures are only used for illustrative purposes and cannot be understood as limitating the patent；
In order to better illustrate this embodiment, the certain components of attached drawing have omission, zoom in or out, and do not represent actual product
Size；
To those skilled in the art, it is to be understood that certain known features and its explanation, which may be omitted, in attached drawing
's.
The following further describes the technical solution of the present invention with reference to the accompanying drawings and examples.
Embodiment 1
A kind of cross-domain image classification method based on coupling knowledge migration, as shown in Figure 1, comprising the following steps:
S1. source domain image and target area image are obtained and carries out feature extraction, respectively obtains source domain eigenmatrix as source
Numeric field data obtains target domain characterization matrix as target numeric field data:
S1.1. n is inputtedsOpen the source domain image and n of tape labeltOpen the target area image without label；Wherein ns、ntIt is positive whole
Number；
S1.2. feature extraction is carried out to the source domain image and target area image respectively, by the VGG16 of pre-training,
Any one neural network of GoogLeNe, ResNet50 extracts the feature vector of the source domain image and target area image respectively, will be right
The feature vector that the source domain image zooming-out obtains obtains source domain data characteristics matrix by column arrangement
Wherein m is the dimension of feature vector, nsFor the number of feature vector；
The feature vector obtained to the aiming field image zooming-out is obtained into aiming field data characteristics matrix by column arrangementWherein m is the dimension of feature vector, ntFor the number of feature vector；
The feature vector of the source domain image indicates the source domain data of source domain image, the feature vector of the target area image
Indicate the target numeric field data of target area image.
S2. the adjacent map of source domain data and the adjoining of target numeric field data are constructed respectively according to the label information of source domain data
Figure, the side right by calculating adjacent map respectively obtain the weight matrix of source domain and the weight matrix of aiming field again:
S2.1. first time iteration is judged whether it is, if otherwise directly carrying out step S2.2, if constructing arest neighbors classification
Device simultaneously trains the nearest neighbor classifier using the source domain data, for predicting the pseudo label information of target numeric field data
S2.2. according to the label information of the source domain data, the adjacent map of source domain data is constructed, i.e., in source domain data
Adjacent side is established between homogeneous data, and calculates the weight of adjacent side, obtains weight matrix
Wherein WsFor the weight matrix of source domain, it is calculate by the following formula:
WhereinFor the feature vector of i-th Zhang Yuan's area image,For the corresponding label information of i-th Zhang Yuan's area image, σ is
The bandwidth of Gaussian kernel, | | | | indicate two norm of vector；
According to the pseudo label information of the target numeric field data, the adjacent map of target numeric field data is constructed, in target numeric field data
Homogeneous data between establish adjacent side, and calculate the weight of adjacent side, obtain weight matrix
Wherein WtFor aiming field weight matrix and be calculate by the following formula:
WhereinFor the feature vector of i-th target area image,For the corresponding pseudo label of i-th target area image, βi
ForWeight, initial value is 1.
S3. data distribution is aligned between carrying out domain in the common lower-dimensional subspace of the source domain and aiming field, is utilized simultaneously
The weight matrix of source domain and the weight matrix of aiming field carry out the knowledge migration of data in domain:
S3.1. it in lower-dimensional subspace common in the source domain and aiming field, is carried out using Largest Mean difference criterion
Data distribution is aligned between domain including edge distribution and conditional probability distribution, specifically by the following first object letter of minimum
Number:
Wherein C indicates the class number of label information, xs,cAnd xt,cRespectively indicate the classification category of source domain data tag information
Belong to the set of classification c in the set of classification c and the classification of aiming field data tag information,WithRespectively indicate source domain number
Belong to the number of the data of the classification c of label information according to the number and target numeric field data of the data for the classification c for belonging to label information,For projection matrix, wherein d < < m, projects to d n-dimensional subspace n for tieing up data from m；
S3.2. weight matrix W is utilizedsAnd WtThe knowledge migration for carrying out data in domain, specifically by below minimizing second
Objective function:
S3.3. the first object function and the second objective function are combined and a regularization is addedObtain with
Lower third objective function:
s.t.PTXHXTP=I
Wherein λ and α is balance factor, λ > 0 and α > 0, X=[Xs,Xt], change matrix centered on H:I is unit matrix,Complete 1 column vector for being 1 for value；
Projection matrix P is solved by generalized eigen decomposition, obtains source domain data and target numeric field data in low-dimensional
Expression in space is respectively as follows:With
S4. it in the lower-dimensional subspace, using the source domain data training nearest neighbor classifier of tape label information, and is used for
It predicts the classification information of target numeric field data: according to expression of the source domain data in lower-dimensional subspace, obtaining empty in low-dimensional
Between middle tape label information source domain dataFor training the nearest neighbor classifier, and predict in lower dimensional space
Target numeric field dataClassification information
S5. the label information for the target numeric field data predicted is weighted again:
S5.1. to the target numeric field data in lower dimensional spaceA new adjacent map is constructed, i.e., will
Each aiming field data point in lower dimensional spaceThe k aiming field number nearest with its Euclidean distance respectively
Adjacent side is established at strong point；
S5.2. based on adjacent map described in step S5.1, for i-th of aiming field data point on authorized adjacent mapIf withK connected aiming field data point all withA classification is belonged to, then aiming field data pointLabel
Confidence level βi=1, otherwise βi=0, wherein k is positive integer.
S6. judge whether to reach maximum number of iterations, if not return step S2, if then exporting predicted aiming field figure
The label information of picture
Embodiment 2
In the present embodiment 2, the facial image composition of randomly selected 10 people from CMU PIE data set is selected
980 face images, every image are cut into 32X32 resolution ratio.Further, using C05 subset totally 490 images as
Source domain data set, using the C27 subset aiming field data set that 490 images are classified as needs in total.
S1. the target area image in the source domain image and aiming field data set in source domain data set is obtained, to the source domain
After image and target area image carry out gray processing respectively, the gray value of the source domain image is arranged in feature vector and by institute
The gray value for stating target area image is arranged in feature vector, in the present embodiment 2, the gray value of every image be arranged in feature to
Amount, i.e. every image is expressed as the column vector of one 1024 dimension, corresponding, and it is special to respectively obtain source domain data by column arrangement
Levy matrixWith target domain characterization matrix
S2. the adjacent map of source domain data and the adjoining of target numeric field data are constructed respectively according to the label information of source domain data
Figure, the side right by calculating adjacent map respectively obtain the weight matrix of source domain and the weight matrix of aiming field again:
S2.1. first time iteration is judged whether it is, if otherwise directly carrying out step S2.2, if constructing arest neighbors classification
Device and the source domain data for utilizing tape label informationTraining nearest neighbor classifier, for predicting target numeric field dataPseudo- classification information
S2.2. according to the label information y of source domain datasThe adjacent map for constructing source domain data, i.e., to same in source domain data
Adjacent side is established between class data, and calculates the weight of adjacent side, obtains weight matrix
Wherein WsFor the weight matrix of source domain, it is calculate by the following formula:
WhereinFor the feature vector of i-th Zhang Yuan's area image,For the corresponding label information of i-th Zhang Yuan's area image, σ is
The bandwidth of Gaussian kernel is set as 1 in the present embodiment 1, | | | | indicate two norm of vector；
According to the pseudo label information of the target numeric field dataThe adjacent map for constructing target numeric field data, to target numeric field data
In homogeneous data between establish adjacent side, and calculate the weight of adjacent side, obtain weight matrixWherein WtFor mesh
It marks the weight matrix in domain and is calculate by the following formula:
WhereinFor the feature vector of i-th target area image,For the corresponding pseudo label of i-th target area image, βi
ForWeight, initial value is 1.
S3. data distribution is aligned between carrying out domain in the common lower-dimensional subspace of the source domain and aiming field, is utilized simultaneously
The weight matrix of source domain and the weight matrix of aiming field carry out the knowledge migration of data in domain；
By minimizing following objective function:
s.t.PTXHXTP=I
Wherein λ=0.1 and α=0.01 are balance factor.X=[Xs,Xt], change matrix centered on H:
I980For unit matrix, 1980For complete 1 column vector.Above formula, which is solved, by generalized eigen decomposition obtains projection matrix
The dimension that lower-dimensional subspace is arranged in the present embodiment 2 is d=100.Source domain data and target numeric field data are finally obtained in low-dimensional
Expression in space:With
S4. the expression according to the source domain data in lower-dimensional subspace, obtains the tape label information in lower-dimensional subspace
Source domain dataFor training the nearest neighbor classifier, and predict the target numeric field data in lower dimensional spaceClassification information
S5. the label information for the target numeric field data predicted is weighted again；
S5.1. to the target numeric field data in lower dimensional spaceA new adjacent map is constructed, i.e., will
Each aiming field data point in lower dimensional spaceThe k=5 target nearest with its Euclidean distance respectively
Numeric field data point establishes adjacent side；
S5.2. based on adjacent map described in step S5.1, for i-th of aiming field data point on authorized adjacent mapIf with5 connected aiming field data points all withA classification is belonged to, then aiming field data pointLabel
Confidence level βi=1, otherwise βi=0.
S6. judge whether to reach maximum number of iterations 10, if not return step S2, if then exporting predicted aiming field
The label information of imageTo obtain classification results.
In the present embodiment 2, experiment porch is the MATLAB R2017a software in WIN10 system, the model of CPU
Intel i7-6700K@4.00GHz.The experimental result schematic diagram of the present embodiment 2 is as shown in Figure 2,3, wherein in Fig. 2, it is round
What is indicated is source domain data, and what five-pointed star indicated is target numeric field data, and the left subgraph of Fig. 2 is the original of source domain data and target numeric field data
Begin to be distributed, it can be clearly seen that the distribution of source domain data and target numeric field data has very big otherness；Fig. 2 right subgraph is this reality
It applies 2 methods and is carrying out the data point between domain after data distribution alignment in lower-dimensional subspace to source domain and data aiming field data
Cloth, it can be clearly seen that the otherness of source domain data and aiming field data distribution greatly reduces.It is of different shapes in Fig. 3
Mark indicates different label information classifications, a total of ten label information classifications.Wherein the left subgraph of Fig. 3 is original aiming field
The distribution of each label information of data, category distribution are more chaotic；Fig. 3 right subgraph is target numeric field data in 2 side of the present embodiment
Distribution in the obtained lower-dimensional subspace of method, it can be clearly seen that the category distribution of target numeric field data is relatively clear, Ge Geyou
The data of same label information are respectively at cluster.This illustrates the high efficiency of the method for the present invention.
The terms describing the positional relationship in the drawings are only for illustration, should not be understood as the limitation to this patent；
Obviously, the above embodiment of the present invention be only to clearly illustrate example of the present invention, and not be pair
The restriction of embodiments of the present invention.For those of ordinary skill in the art, may be used also on the basis of the above description
To make other variations or changes in different ways.There is no necessity and possibility to exhaust all the enbodiments.It is all this
Made any modifications, equivalent replacements, and improvements etc., should be included in the claims in the present invention within the spirit and principle of invention
Protection scope within.
Claims (8)
1. a kind of cross-domain image classification method based on coupling knowledge migration, which comprises the following steps:
S1. source domain image and target area image are obtained and carries out feature extraction, respectively obtains source domain eigenmatrix as source domain number
According to obtaining target domain characterization matrix as target numeric field data；
S2. the adjacent map of source domain data and the adjacent map of target numeric field data are constructed respectively according to the label information of source domain data, is led to
The side right for crossing calculating adjacent map respectively obtains the weight matrix of source domain and the weight matrix of aiming field again；
S3. data distribution is aligned between carrying out domain in the common lower-dimensional subspace of the source domain and aiming field, while utilizing source domain
Weight matrix and aiming field weight matrix carry out domain in data knowledge migration；
S4. in the lower-dimensional subspace, using the source domain data training nearest neighbor classifier of tape label information, and for predicting
The label information of target numeric field data；
S5. the label information for the target numeric field data predicted is weighted again；
S6. judge whether to reach maximum number of iterations, if not return step S2, if then exporting predicted target area image
Label information, to obtain classification results.
2. the cross-domain image classification method according to claim 1 based on coupling knowledge migration, which is characterized in that the step
The specific steps of rapid S1 are as follows:
S1.1. n is inputtedsOpen the source domain image and n of tape labeltOpen the target area image without label；Wherein ns、ntFor positive integer；
S1.2. feature extraction is carried out to the source domain image and target area image respectively, the source domain image zooming-out will be obtained
Feature vector by column arrangement obtain source domain data characteristics matrixWherein m is feature vector
Dimension, nsFor the number of feature vector；
The feature vector obtained to the aiming field image zooming-out is obtained into aiming field data characteristics matrix by column arrangementWherein m is the dimension of feature vector, ntFor the number of feature vector；
The feature vector of the source domain image indicates that the source domain data of source domain image, the feature vector of the target area image indicate
The target numeric field data of target area image.
3. the cross-domain image classification method according to claim 2 based on coupling knowledge migration, which is characterized in that step S1
Described in feature extraction extract the source domain image and target area image respectively specifically by the neural network of pre-training
Feature vector.
4. the cross-domain image classification method according to claim 2 based on coupling knowledge migration, which is characterized in that step S1
Described in feature extraction be specially gray processing is carried out respectively to the source domain image and target area image after, by the source domain figure
The gray value of picture is arranged in feature vector and the gray value of the target area image is arranged in feature vector.
5. the cross-domain image classification method according to claim 3 or 4 based on coupling knowledge migration, which is characterized in that institute
The S2 specific steps stated are as follows:
S2.1. first time iteration is judged whether it is, if step S2.2 is otherwise directly carried out, if constructing nearest neighbor classifier simultaneously
Using the source domain data training nearest neighbor classifier, for predicting the pseudo label information of target numeric field data
S2.2. according to the label information of the source domain data, the adjacent map of source domain data is constructed, i.e., to similar in source domain data
Adjacent side is established between data, and calculates the weight of adjacent side, obtains weight matrix
Wherein WsFor the weight matrix of source domain, it is calculate by the following formula:
WhereinFor the feature vector of i-th Zhang Yuan's area image,For the corresponding label information of i-th Zhang Yuan's area image, σ is Gaussian kernel
Bandwidth, | | | | indicate two norm of vector；
According to the pseudo label information of the target numeric field data, the adjacent map of target numeric field data is constructed, to same in target numeric field data
Adjacent side is established between class data, and calculates the weight of adjacent side, obtains weight matrix
Wherein WtFor aiming field weight matrix and be calculate by the following formula:
WhereinFor the feature vector of i-th target area image,For the corresponding pseudo label of i-th target area image, βiFor
Weight, initial value is 1.
6. the cross-domain image classification method according to claim 5 based on coupling knowledge migration, which is characterized in that described
S3 specific steps are as follows:
S3.1. in lower-dimensional subspace common in the source domain and aiming field, included using Largest Mean difference criterion
Data distribution is aligned between edge distribution and the domain of conditional probability distribution, specifically by the following first object function of minimum:
Wherein C indicates the class number of label information, xs,cAnd xt,cThe classification for respectively indicating source domain data tag information belongs to class
The set of other c and the classification of aiming field data tag information belong to the set of classification c,WithSource domain data are respectively indicated to belong to
The number and target numeric field data of the data of the classification c of label information belong to the number of the data of the classification c of label information,For projection matrix, wherein d < < m, projects to d n-dimensional subspace n for tieing up data from m；
S3.2. weight matrix W is utilizedsAnd WtThe knowledge migration for carrying out data in domain, specifically by the second target below minimizing
Function:
S3.3. the first object function and the second objective function are combined and a regularization is addedObtain following
Three objective functions:
s.t.PTXHXTP=I
Wherein λ and α is balance factor, λ > 0 and α > 0, X=[Xs,Xt], change matrix centered on H:
I is unit matrix,Complete 1 column vector for being 1 for value；
Projection matrix P is solved by generalized eigen decomposition, obtains source domain data and target numeric field data in the lower-dimensional subspace
In expression be respectively as follows:With
7. the cross-domain image classification method according to claim 6 based on coupling knowledge migration, which is characterized in that the step
The specific steps of rapid S4 are as follows:
According to expression of the source domain data in lower-dimensional subspace, the source domain number of the tape label information in lower-dimensional subspace is obtained
According toFor training the nearest neighbor classifier, and predict the target numeric field data in lower dimensional spaceClassification
Information
8. the cross-domain image classification method according to claim 7 based on coupling knowledge migration, which is characterized in that the step
Suddenly the specific steps of S5 include:
S5.1. to the target numeric field data in lower dimensional spaceConstruct a new adjacent map, i.e., it will be each
Aiming field data point in lower dimensional spaceThe k aiming field data point nearest with its Euclidean distance respectively
Establish adjacent side；
S5.2. based on adjacent map described in step S5.1, for i-th of aiming field data point on authorized adjacent mapIf
WithK connected aiming field data point all withA classification is belonged to, then aiming field data pointLabelConfidence level
βi=1, otherwise βi=0, wherein k is positive integer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910482559.5A CN110378366B (en) | 2019-06-04 | 2019-06-04 | Cross-domain image classification method based on coupling knowledge migration |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910482559.5A CN110378366B (en) | 2019-06-04 | 2019-06-04 | Cross-domain image classification method based on coupling knowledge migration |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110378366A true CN110378366A (en) | 2019-10-25 |
| CN110378366B CN110378366B (en) | 2023-01-17 |
Family
ID=68249801
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910482559.5A Active CN110378366B (en) | 2019-06-04 | 2019-06-04 | Cross-domain image classification method based on coupling knowledge migration |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110378366B (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110781970A (en) * | 2019-10-30 | 2020-02-11 | 腾讯科技（深圳）有限公司 | Method, device and equipment for generating classifier and storage medium |
| CN111312403A (en) * | 2020-01-21 | 2020-06-19 | 山东师范大学 | Disease prediction system, device and medium based on instance and feature sharing cascade |
| CN111814871A (en) * | 2020-06-13 | 2020-10-23 | 浙江大学 | Image classification method based on reliable weight optimal transmission |
| CN111929053A (en) * | 2020-07-07 | 2020-11-13 | 中国矿业大学 | DA-RVFLN-based pneumatic control valve field adaptive fault diagnosis method |
| CN112085085A (en) * | 2020-09-03 | 2020-12-15 | 哈尔滨工业大学 | Multi-source transfer learning method based on graph structure |
| CN112733970A (en) * | 2021-03-31 | 2021-04-30 | 腾讯科技（深圳）有限公司 | Image classification model processing method, image classification method and device |
| CN113420775A (en) * | 2021-03-31 | 2021-09-21 | 中国矿业大学 | Image classification method under extremely small quantity of training samples based on adaptive subdomain field adaptation of non-linearity |
| CN113469405A (en) * | 2021-05-18 | 2021-10-01 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Cross-city tail gas migration prediction method, system and medium based on subgraph segmentation |
| CN113657651A (en) * | 2021-07-27 | 2021-11-16 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Diesel vehicle emission prediction method, medium and equipment based on deep migration learning |
| CN113836996A (en) * | 2021-08-10 | 2021-12-24 | 中国地质大学（武汉） | Hyperspectral remote sensing image part migration method and system |
| CN114743043A (en) * | 2022-03-15 | 2022-07-12 | 北京迈格威科技有限公司 | Image classification method, electronic device, storage medium and program product |
| CN115393362A (en) * | 2022-10-28 | 2022-11-25 | 中南大学 | Method, equipment and medium for selecting automatic glaucoma recognition model |
| CN117216507A (en) * | 2023-10-08 | 2023-12-12 | 深圳大学 | Deep neural network model mobility measurement method based on geographic partition |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103971129A (en) * | 2014-05-27 | 2014-08-06 | 浙江大学 | Classification method and device based on learning image content recognition in cross-data field subspace |
| CN107895177A (en) * | 2017-11-17 | 2018-04-10 | 南京邮电大学 | A kind of migration classification learning method for keeping image classification sparsity structure |
| CN109389174A (en) * | 2018-10-23 | 2019-02-26 | 四川大学 | A kind of crowd massing Sensitive Image Detection Method |
-
2019
- 2019-06-04 CN CN201910482559.5A patent/CN110378366B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103971129A (en) * | 2014-05-27 | 2014-08-06 | 浙江大学 | Classification method and device based on learning image content recognition in cross-data field subspace |
| CN107895177A (en) * | 2017-11-17 | 2018-04-10 | 南京邮电大学 | A kind of migration classification learning method for keeping image classification sparsity structure |
| CN109389174A (en) * | 2018-10-23 | 2019-02-26 | 四川大学 | A kind of crowd massing Sensitive Image Detection Method |
Non-Patent Citations (1)
| Title |
|---|
| 覃姜维: "迁移学习方法研究及其在跨领域数据分类中的应用", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110781970A (en) * | 2019-10-30 | 2020-02-11 | 腾讯科技（深圳）有限公司 | Method, device and equipment for generating classifier and storage medium |
| CN110781970B (en) * | 2019-10-30 | 2024-04-26 | 腾讯科技（深圳）有限公司 | Classifier generation method, device, equipment and storage medium |
| CN111312403A (en) * | 2020-01-21 | 2020-06-19 | 山东师范大学 | Disease prediction system, device and medium based on instance and feature sharing cascade |
| CN111814871A (en) * | 2020-06-13 | 2020-10-23 | 浙江大学 | Image classification method based on reliable weight optimal transmission |
| CN111814871B (en) * | 2020-06-13 | 2024-02-09 | 浙江大学 | Image classification method based on reliable weight optimal transmission |
| CN111929053A (en) * | 2020-07-07 | 2020-11-13 | 中国矿业大学 | DA-RVFLN-based pneumatic control valve field adaptive fault diagnosis method |
| CN112085085A (en) * | 2020-09-03 | 2020-12-15 | 哈尔滨工业大学 | Multi-source transfer learning method based on graph structure |
| CN112085085B (en) * | 2020-09-03 | 2024-02-23 | 哈尔滨工业大学 | Multi-source migration learning method based on graph structure |
| CN112733970B (en) * | 2021-03-31 | 2021-06-18 | 腾讯科技（深圳）有限公司 | Image classification model processing method, image classification method and device |
| CN113420775A (en) * | 2021-03-31 | 2021-09-21 | 中国矿业大学 | Image classification method under extremely small quantity of training samples based on adaptive subdomain field adaptation of non-linearity |
| CN113420775B (en) * | 2021-03-31 | 2024-03-29 | 中国矿业大学 | Picture classification method based on nonlinearity self-adaptive subdomain field adaptation under extremely small quantity of training samples |
| CN112733970A (en) * | 2021-03-31 | 2021-04-30 | 腾讯科技（深圳）有限公司 | Image classification model processing method, image classification method and device |
| CN113469405A (en) * | 2021-05-18 | 2021-10-01 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Cross-city tail gas migration prediction method, system and medium based on subgraph segmentation |
| CN113469405B (en) * | 2021-05-18 | 2024-04-02 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Subimage segmentation-based cross-city tail gas migration prediction method, system and medium |
| CN113657651A (en) * | 2021-07-27 | 2021-11-16 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Diesel vehicle emission prediction method, medium and equipment based on deep migration learning |
| CN113836996A (en) * | 2021-08-10 | 2021-12-24 | 中国地质大学（武汉） | Hyperspectral remote sensing image part migration method and system |
| CN113836996B (en) * | 2021-08-10 | 2024-02-02 | 中国地质大学（武汉） | Hyperspectral remote sensing image part migration method and hyperspectral remote sensing image part migration system |
| CN114743043A (en) * | 2022-03-15 | 2022-07-12 | 北京迈格威科技有限公司 | Image classification method, electronic device, storage medium and program product |
| CN114743043B (en) * | 2022-03-15 | 2024-04-26 | 北京迈格威科技有限公司 | Image classification method, electronic device, storage medium and program product |
| CN115393362B (en) * | 2022-10-28 | 2023-02-03 | 中南大学 | Method, equipment and medium for selecting automatic glaucoma recognition model |
| CN115393362A (en) * | 2022-10-28 | 2022-11-25 | 中南大学 | Method, equipment and medium for selecting automatic glaucoma recognition model |
| CN117216507A (en) * | 2023-10-08 | 2023-12-12 | 深圳大学 | Deep neural network model mobility measurement method based on geographic partition |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110378366B (en) | 2023-01-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110378366A (en) | A kind of cross-domain image classification method based on coupling knowledge migration | |
| CN109344736B (en) | Static image crowd counting method based on joint learning | |
| Hu et al. | Recognition of pornographic web pages by classifying texts and images | |
| CN103514456B (en) | Image classification method and device based on compressed sensing multi-core learning | |
| CN108564029A (en) | Face character recognition methods based on cascade multi-task learning deep neural network | |
| CN111079674B (en) | Target detection method based on global and local information fusion | |
| CN110163258A (en) | A kind of zero sample learning method and system reassigning mechanism based on semantic attribute attention | |
| CN107341447A (en) | A kind of face verification mechanism based on depth convolutional neural networks and evidence k nearest neighbor | |
| Zhang et al. | Unsupervised difference representation learning for detecting multiple types of changes in multitemporal remote sensing images | |
| CN108345850A (en) | The scene text detection method of the territorial classification of stroke feature transformation and deep learning based on super-pixel | |
| CN106408039A (en) | Off-line handwritten Chinese character recognition method carrying out data expansion based on deformation method | |
| CN107506786A (en) | A kind of attributive classification recognition methods based on deep learning | |
| CN108960260B (en) | Classification model generation method, medical image classification method and medical image classification device | |
| CN110163239A (en) | A kind of Weakly supervised image, semantic dividing method based on super-pixel and condition random field | |
| CN110751027B (en) | Pedestrian re-identification method based on deep multi-instance learning | |
| Blanchart et al. | A semi-supervised algorithm for auto-annotation and unknown structures discovery in satellite image databases | |
| Xiao et al. | Signature verification using a modified Bayesian network | |
| CN113808166B (en) | Single-target tracking method based on clustering difference and depth twin convolutional neural network | |
| CN106874862A (en) | People counting method based on submodule technology and semi-supervised learning | |
| Jiao et al. | Exploring temporal community structure and constant evolutionary pattern hiding in dynamic networks | |
| CN110473195A (en) | It is a kind of can automatic customization medicine lesion detection framework and method | |
| CN110084136A (en) | Context based on super-pixel CRF model optimizes indoor scene semanteme marking method | |
| Lyzinski et al. | On the consistency of the likelihood maximization vertex nomination scheme: Bridging the gap between maximum likelihood estimation and graph matching | |
| CN105809200B (en) | Method and device for autonomously extracting image semantic information in bioauthentication mode | |
| CN104598898A (en) | Aerially photographed image quick recognizing system and aerially photographed image quick recognizing method based on multi-task topology 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |