CN107563445A - A kind of method and apparatus of the extraction characteristics of image based on semi-supervised learning - Google Patents
A kind of method and apparatus of the extraction characteristics of image based on semi-supervised learning Download PDFInfo
- Publication number
- CN107563445A CN107563445A CN201710796349.4A CN201710796349A CN107563445A CN 107563445 A CN107563445 A CN 107563445A CN 201710796349 A CN201710796349 A CN 201710796349A CN 107563445 A CN107563445 A CN 107563445A
- Authority
- CN
- China
- Prior art keywords
- msub
- mrow
- training sample
- represent
- sample
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The embodiment of the invention discloses a kind of method and apparatus of the extraction characteristics of image based on semi-supervised learning, initialization model parameter, view data is pre-processed to obtain image pattern;And image pattern is divided into the training sample for including exemplar and unlabeled exemplars, the test sample for only including unlabeled exemplars;According to paired constraints, determine there is constraint set corresponding to exemplar in training sample;Using neighbor search algorithm, build neighbour corresponding to all training samples and scheme, and calculate weight matrix;By minimizing feature approximation mistake, low dimensional manifold characteristic processing is carried out to training sample, obtains low dimensional manifold feature and linear projection matrix;Training sample and the characteristics of image of test sample are extracted using linear projection matrix.Global and local structural information between sample data can be kept simultaneously by the technical scheme, improve the identifiability of feature.And realize and new test data is quickly mapped to low-dimensional, improve the performance of image characteristics extraction.
Description
Technical field
The present invention relates to computer vision and image identification technical field, more particularly to a kind of based on semi-supervised learning
Extract the method and apparatus of characteristics of image.
Background technology
In substantial amounts of practical application, the data in reality can be described with the attribute or feature of higher-dimension.It is but original
The dimension of feature may be very big, and sample is in one the very space of higher-dimension in other words, and passes through Feature Mapping or feature becomes
The method changed, high dimensional data can be transformed to a lower dimensional space.
Extraction obtains being computer vision and image recognition all the time to maximally effective feature of classifying from high dimensional feature
Etc. one of research topic extremely important and difficult in research field.Equidistant mapping algorithm (Isomap) is a kind of classical non-
Linear manifold learning method, its main thought are to find the subspace that can must preferably retain the geodesic distance between sample point.
For original high dimensional imageIt is its corresponding low-dimensional
The expression in space.Wherein N is sample size, and n, d are respectively the data dimension (d < < n) before and after dimensionality reduction.
Isomap's mainly comprises the following steps:1) Neighbor Points of each sample point are determined using k nearest neighbor or ε near neighbor methods;2) structure
A non-directed graph G (V, E) is made, wherein each node vi∈ V correspond to each sample point xi, dG(xi,xj) represent data point x in figure Gi
And xjBetween shortest path distance;3) pass throughTry to achieve low-dimensional insertion Y.
In recent years, much proposed in succession based on Isomap innovatory algorithm.Wherein, multiple manifold differentiates Isometric Maps algorithm
(MMD-Isomap) it is the multiple manifold supervised a full algorithm.It is using the thought constrained in pairs, according to the label information of sample
Sample point is divided into ML and CL, and makes the distance minimization between the data point under ML constraints, uses up the data point under CL constraints
It may possibly be separated, therefore the global structure of sample is maintained.But MMD-Isomap is a direct-push method, can not quickly be solved
The mapping problems of certainly external test data, in addition in actual applications, the sample data for having label is rare, and by artificial
The process of nominal data can expend a large amount of manpowers and time.
It is those skilled in the art's urgent problem to be solved it can be seen that how to improve the performance of image characteristics extraction.
The content of the invention
The purpose of the embodiment of the present invention is to provide a kind of method and apparatus of the extraction characteristics of image based on semi-supervised learning,
The performance of image characteristics extraction can be lifted.
In order to solve the above technical problems, the embodiment of the present invention provides a kind of extraction characteristics of image based on semi-supervised learning
Method, including:
Initialization model parameter, the view data of acquisition is pre-processed, obtains image pattern;And by described image sample
This division includes the training set of exemplar and unlabeled exemplars as training sample, a test set for including unlabeled exemplars
As test sample;
According to paired constraints, determine there is constraint set corresponding to exemplar in the training sample;
Using neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and weight matrix is calculated, be used for
Measure the Near-neighbor Structure information between test sample;
According to the constraint set and the weight matrix, by minimizing feature approximation mistake, the training sample is entered
Row low dimensional manifold characteristic processing, the low dimensional manifold feature of the training sample is obtained, and for obtaining low dimensional manifold feature
Linear projection matrix;
The training sample and the characteristics of image of the test sample are extracted using the linear projection matrix.
Optionally, the paired constraints of the foundation, determine to have in the training sample constraint corresponding to exemplar
Collection includes:
According to paired constraints, by the data point of the training sample to being divided into constraint set ML and constraint set CL,
ML={ (xi,xj)|i≠j,l(xi)=l (xj)};
CL={ (xi,xj)|i≠j,l(xi)≠l(xj)};
Wherein, l (xi) ∈ 1,2 ..., and c } represent image pattern xiThe class label of (i=1,2 ..., N);l(xj) table
Show image pattern xjThe class label of (j=1,2 ..., N);
According to minimum inter- object distance formulaTo it is described about
Point in constriction ML is to carrying out compact processing;
According to maximization between class distance formulaTo the constraint set
Point in CL is to carrying out decentralized processing;
Wherein, d (yi,yj) represent low-dimensional yiAnd yjBetween Euclidean distance, yiRepresent image pattern xiLow-dimensional represent, yjTable
Show image pattern xjLow-dimensional represent, dG(xi,xj) represent shortest path distance for approximate geodesic distance.
Optionally, it is described to utilize neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and power is calculated
Weight matrix includes:
Using K arest neighbors sorting algorithms, the training sample is handled, it is each right to obtain each training sample
Answer K neighbour, construction neighbour's figure;
According to formulaWeight square is calculated
Battle array W;
Wherein, xiFor i-th of data in constraint set, xjFor j-th of data in the constraint set, NN (xi) it is training
Sample xiNeighbour set, xj∈NN(xi) represent xjFor training sample xiNeighbour, | | | | represent the l of vector2Norm, WijRepresent
Summit i points to weight corresponding to summit j side.
Optionally, it is described according to the constraint set and the weight matrix, by minimizing feature approximation mistake, to described
Training sample carries out low dimensional manifold characteristic processing, obtains the low dimensional manifold feature of the training sample, and for obtaining low-dimensional
The linear projection matrix of manifold feature includes:
According to equation below, feature approximation error items are minimizedObtain low-dimensional corresponding to the training sample
Manifold characteristic Y, and for obtaining the linear projection matrix P of low dimensional manifold feature:
Wherein, α, β, γ are to weigh parameter, α ∈ (0,1), | ML | the element number in constraint set ML is represented, | CL | represent
Element number in constraint set CL.
The embodiment of the present invention additionally provides a kind of device of the extraction characteristics of image based on semi-supervised learning, including processing list
Member, determining unit, computing unit, obtain unit and extraction unit, the processing unit, for initialization model parameter, to obtaining
The view data taken is pre-processed, and obtains image pattern;And the division of described image sample is included into exemplar and without mark
The training set of signed-off sample sheet includes the test set of unlabeled exemplars as training sample, only as test sample;
The determining unit, for according to paired constraints, determining have exemplar corresponding in the training sample
Constraint set;
The computing unit, for utilizing neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and calculate
Weight matrix is obtained, for measuring the Near-neighbor Structure information between test sample;
It is described to obtain unit, for according to the constraint set and the weight matrix, by minimizing feature approximation mistake,
Low dimensional manifold characteristic processing is carried out to the training sample, obtains the low dimensional manifold feature of the training sample, and for obtaining
Take the linear projection matrix of low dimensional manifold feature;
The extraction unit, for extracting the training sample and the test sample using the linear projection matrix
Characteristics of image.
Optionally, the determining unit includes division subelement, compact subelement and scattered subelement, and division is single
Member, for according to paired constraints, by the data point of the training sample to being divided into constraint set ML and constraint set CL,
ML={ (xi,xj)|i≠j,l(xi)=l (xj)};
CL={ (xi,xj)|i≠j,l(xi)≠l(xj)};
Wherein, l (xi) ∈ 1,2 ..., and c } represent image pattern xiThe class label of (i=1,2 ..., N);l(xj) table
Show image pattern xjThe class label of (j=1,2 ..., N);
The compact subelement, for according to minimum inter- object distance formula
To the point in the constraint set ML to carrying out compact processing;
The scattered subelement, for according to maximization between class distance formula
To the point in the constraint set CL to carrying out decentralized processing;
Wherein, d (yi,yj) represent low-dimensional yiAnd yjBetween Euclidean distance, yiRepresent image pattern xiLow-dimensional represent, yjTable
Show image pattern xjLow-dimensional represent, dG(xi,xj) represent shortest path distance for approximate geodesic distance.
Optionally, the computing unit is specifically used for utilizing K arest neighbors sorting algorithms, at the training sample
Reason, obtain each training sample and each correspond to K neighbour, construction neighbour's figure;And according to formulaWeight matrix W is calculated;
Wherein, xiFor i-th of data in constraint set, xjFor j-th of data in the constraint set, NN (xi) it is training
Sample xiNeighbour set, xj∈NN(xi) represent xjFor training sample xiNeighbour, | | | | represent the l of vector2Norm, WijRepresent
Summit i points to weight corresponding to summit j side.
Optionally, the unit that obtains is specifically used for, according to equation below, minimizing feature approximation error items
Low dimensional manifold characteristic Y corresponding to the training sample is obtained, and for obtaining the linear projection matrix P of low dimensional manifold feature:
Wherein, α, β, γ are to weigh parameter, α ∈ (0,1), | ML | the element number in constraint set ML is represented, | CL | represent
Element number in constraint set CL.
The initialization model parameter it can be seen from above-mentioned technical proposal, the view data of acquisition is pre-processed, obtained
Image pattern;And using the division of described image sample include the training sets of exemplar and unlabeled exemplars as training sample,
Only the test set comprising unlabeled exemplars is as test sample;According to paired constraints, determine have in the training sample
Constraint set corresponding to exemplar;Using neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and be calculated
Weight matrix;According to the constraint set and the weight matrix, by minimizing feature approximation mistake, the training sample is entered
Row low dimensional manifold characteristic processing, the low dimensional manifold feature of the training sample is obtained, and for obtaining low dimensional manifold feature
Linear projection matrix;Can be to extract the characteristics of image of the test sample using the linear projection matrix.By in pairs about
Beam condition and minimum feature approximation mistake, realize while keep the global and local structural information between sample data, effectively
Improve the identifiability of feature.By minimizing feature approximation mistake, study obtains linear projection matrix, realizes quick incite somebody to action
New test data is mapped to low-dimensional, improves the performance of image characteristics extraction.
Brief description of the drawings
In order to illustrate the embodiments of the present invention more clearly, the required accompanying drawing used in embodiment will be done simply below
Introduce, it should be apparent that, drawings in the following description are only some embodiments of the present invention, for ordinary skill people
For member, on the premise of not paying creative work, other accompanying drawings can also be obtained according to these accompanying drawings.
Fig. 1 is a kind of flow of the method for extraction characteristics of image based on semi-supervised learning provided in an embodiment of the present invention
Figure;
Fig. 2 is that a kind of structure of the device of the extraction characteristics of image based on semi-supervised learning provided in an embodiment of the present invention is shown
It is intended to.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Site preparation describes, it is clear that described embodiment is only part of the embodiment of the present invention, rather than whole embodiments.Based on this
Embodiment in invention, for those of ordinary skill in the art under the premise of creative work is not made, what is obtained is every other
Embodiment, belong to the scope of the present invention.
In order that those skilled in the art more fully understand the present invention program, with reference to the accompanying drawings and detailed description
The present invention is described in further detail.
Next, a kind of extraction characteristics of image based on semi-supervised learning that the embodiment of the present invention is provided is discussed in detail
Method.Fig. 1 is a kind of flow chart of the method for extraction characteristics of image based on semi-supervised learning provided in an embodiment of the present invention, should
Method includes:
S101:Initialization model parameter, the view data of acquisition is pre-processed, obtains image pattern;And by described in
Image pattern division include the training sets of exemplar and unlabeled exemplars as training sample, only include unlabeled exemplars
Test set is as test sample.
Model parameter can include neighbour's number K, control parameter α, β, γ.Original view data is due to by a variety of
Part limits and random disturbances, tends not to directly directly use in vision system, it is therefore desirable to enter original view data
Row data normalization etc. pre-processes.
For view dataWherein, n is the dimension of sample, and N is the quantity of sample, can
To be divided into the training sample for including class label (common c classification, c > 2)With include nothing
The test sample of labelAnd meet sample size l+u=N.
S102:According to paired constraints, determine there is constraint set corresponding to exemplar in the training sample.
In embodiments of the present invention, can be by the data point of the training sample to being divided into according to paired constraints
Constraint set ML and constraint set CL,
ML={ (xi,xj)|i≠j,l(xi)=l (xj)};
CL={ (xi,xj)|i≠j,l(xi)≠l(xj)};
Wherein, l (xi)∈{1,2 ..., c } represent image pattern xiThe class label of (i=1,2 ..., N);l(xj) table
Show image pattern xjThe class label of (j=1,2 ..., N);
According to minimum inter- object distance formulaTo the constraint
Collect the point in ML to carrying out compact processing;
According to maximization between class distance formulaTo the constraint set
Point in CL is to carrying out decentralized processing;
Wherein, d (yi,yj) represent low-dimensional yiAnd yjBetween Euclidean distance, yiRepresent image pattern xiLow-dimensional represent, yjTable
Show image pattern xjLow-dimensional represent, dG(xi,xj) represent shortest path distance for approximate geodesic distance.
By minimizing inter- object distance, between class distance is maximized, training sample global structure information can be kept, after being easy to
It is continuous that accurate classification processing is carried out to image pattern, reduce the error of image classification.
S103:Using neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and weight square is calculated
Battle array, for measuring the Near-neighbor Structure information between test sample.
In embodiments of the present invention, K arest neighbors sorting algorithms can be utilized, the training sample is handled, obtained
Each training sample each corresponds to K neighbour, construction neighbour's figure;
For original picture dataG represents a figure for having N number of summit, wherein, each
One data sample x of vertex correspondencei, represent that summit i points to summit j side with i~j.And then each edge is weighted, it is used in combination
W represents weight matrix, wherein WijWeight on representative edge i~j, Wij=0 two summits for representing to connect are not neighbor relationships.
It can be tried to achieve for weight matrix W by minimizing following optimization problem,
Wherein, xiFor i-th of data in constraint set, xjFor j-th of data in the constraint set, NN (xi) it is training
Sample xiNeighbour set, xj∈NN(xi) represent xjFor training sample xiNeighbour, | | | | represent the l of vector2Norm, WijRepresent
Summit i points to weight corresponding to summit j side.
S104:According to the constraint set and the weight matrix, by minimizing feature approximation mistake, to the training sample
This progress low dimensional manifold characteristic processing, the low dimensional manifold feature of the training sample is obtained, and for obtaining low dimensional manifold spy
The linear projection matrix of sign.
For the low-dimensional non-linearity manifold feature of training sample, the training data in training sample can be projected to low-dimensional sky
Between(wherein, d < < n), obtain the low-dimensional non-linearity manifold feature of training data
Established in embodiments of the present invention by introducing linear projection matrix P between manifold feature and training sample data
Relation, minimize feature approximation error itemsThe linear projection matrix P for making to obtain possesses directly to be obtained from sample extraction
To the ability of non-linearity manifold feature.
Specifically, feature approximation error items can be minimized according to equation belowObtain the training sample
Corresponding low dimensional manifold characteristic Y, and for obtaining the linear projection matrix P of low dimensional manifold feature:
Wherein, α, β, γ are to weigh parameter, α ∈ (0,1), | ML | the element number in constraint set ML is represented, | CL | represent
Element number in constraint set CL.
When solving low dimensional manifold characteristic Y and linear projection matrix P, above-mentioned formula can be converted into following feature decomposition and asked
Topic:
K=γ (I-XT(XXT)-1X);
Wherein, I is N × N unit matrix, | ML | represent the element number in constraint ML.
Make A=RML+VCL- M-K, feature decomposition is carried out to it, obtain before d maximum characteristic value corresponding to feature to
The set that amount is formed is the manifold characteristic Y of required d dimensions.
By step S101 to the step S103 training preparatory stage, can get training sample constraint set ML and
CL, and get the K neighbour (k nearest neighbor set) of each training sample.Point under being constrained for each ML is to (xi,xj)∈
ML, its shortest path distance d is calculated with Dijkstra or Floyd algorithmsG(xi,xj)。
Low dimensional manifold characteristic Y can be obtained with accordance with the following steps 1 to step 7 training process in embodiments of the present invention
With linear projection matrix P,
Step 1. initialization e is complete 1 vector, and I is N × N unit matrix.
Step 2. calculates H=I-eeT/N,
Step 3. passes through RML=- (1- α) (H (QML)2H)/2 | ML | calculate RML。
Step 4. calculates
Step 5. passes through M=β (I-WT) (I-W), K=γ (I-XT(XXT)-1X), M, K are calculated.
Step 6. calculates A=RML+VCL-M-K。
Step 7. is to matrix A feature decomposition, characteristic vector corresponding to d maximum characteristic value before obtainingLow dimensional manifold characteristic Y is formed, and by calculating P=YXT(XXT)-1Try to achieve linear projection square
Battle array P.
S105:The training sample and the characteristics of image of the test sample are extracted using the linear projection matrix.
By minimizing feature approximation mistake, study obtains linear projection matrix, can be fast using the linear projection matrix
New test data is mapped to low-dimensional by speed.
In embodiments of the present invention, can be tested by test sample, the performance of the linear projection matrix to obtaining.
To test sample xtest, linear projection space that test sample is embedded in obtain by useable linear projection matrix P, complete test sample
Feature extraction.Test sample xtestInsertion results expression it is as follows:WhereinFor the notable feature of test sample,
Nearest neighbor classifier is inputted, obtains soft class label to be measured, can according to position corresponding to maximum in soft class label
To determine the classification of test sample, and then classification accuracy is obtained as quantizating index, the feature extraction effect of evaluation model.
The initialization model parameter it can be seen from above-mentioned technical proposal, the view data of acquisition is pre-processed, obtained
Image pattern;And using the division of described image sample include the training sets of exemplar and unlabeled exemplars as training sample,
Only the test set comprising unlabeled exemplars is as test sample;According to paired constraints, determine have in the training sample
Constraint set corresponding to exemplar;Using neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and be calculated
Weight matrix;According to the constraint set and the weight matrix, by minimizing feature approximation mistake, the training sample is entered
Row low dimensional manifold characteristic processing, the low dimensional manifold feature of the training sample is obtained, and for obtaining low dimensional manifold feature
Linear projection matrix;Can be to extract the characteristics of image of the test sample using the linear projection matrix.By in pairs about
Beam condition and minimum feature approximation mistake, realize while keep the global and local structural information between sample data, effectively
Improve the identifiability of feature.By minimizing feature approximation mistake, study obtains linear projection matrix, realizes quick incite somebody to action
New test data is mapped to low-dimensional, improves the performance of image characteristics extraction.
Fig. 2 is that a kind of structure of the device of the extraction characteristics of image based on semi-supervised learning provided in an embodiment of the present invention is shown
It is intended to, including processing unit 21, determining unit 22, computing unit 23, obtain unit 24 and extraction unit 25, the processing unit
21, for initialization model parameter, the view data of acquisition is pre-processed, obtains image pattern;And by described image sample
This division includes the training set of exemplar and unlabeled exemplars as training sample, a test set for including unlabeled exemplars
As test sample;
The determining unit 22, for according to paired constraints, determining there is exemplar pair in the training sample
The constraint set answered;
The computing unit 23, for utilizing neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and count
Calculation obtains weight matrix, for measuring the Near-neighbor Structure information between test sample;
It is described to obtain unit 24, it is approximate wrong by minimizing feature for according to the constraint set and the weight matrix
By mistake, low dimensional manifold characteristic processing is carried out to the training sample, obtains the low dimensional manifold feature of the training sample, and be used for
Obtain the linear projection matrix of low dimensional manifold feature;
The extraction unit 25, for extracting the training sample and the test sample using the linear projection matrix
Characteristics of image.
Optionally, the determining unit includes division subelement, compact subelement and scattered subelement, and division is single
Member, for according to paired constraints, by the data point of the training sample to being divided into constraint set ML and constraint set CL,
ML={ (xi,xj)|i≠j,l(xi)=l (xj)};
CL={ (xi,xj)|i≠j,l(xi)≠l(xj)};
Wherein, l (xi)∈{1,2 ..., c } represent image pattern xiThe class label of (i=1,2 ..., N);l(xj) table
Show image pattern xjThe class label of (j=1,2 ..., N);
The compact subelement, for according to minimum inter- object distance formula
To the point in the constraint set ML to carrying out compact processing;
The scattered subelement, for according to maximization between class distance formula
To the point in the constraint set CL to carrying out decentralized processing;
Wherein, d (yi,yj) represent low-dimensional yiAnd yjBetween Euclidean distance, yiRepresent image pattern xiLow-dimensional represent, yjTable
Show image pattern xjLow-dimensional represent, dG(xi,xj) represent shortest path distance for approximate geodesic distance.
Optionally, the computing unit is specifically used for utilizing K arest neighbors sorting algorithms, at the training sample
Reason, obtain each training sample and each correspond to K neighbour, construction neighbour's figure;And according to formulaWeight matrix W is calculated;
Wherein, xiFor i-th of data in constraint set, xjFor j-th of data in the constraint set, NN (xi) it is training
Sample xiNeighbour set, xj∈NN(xi) represent xjFor training sample xiNeighbour, | | | | represent the l of vector2Norm, WijRepresent
Summit i points to weight corresponding to summit j side.
Optionally, the unit that obtains is specifically used for, according to equation below, minimizing feature approximation error items
Low dimensional manifold characteristic Y corresponding to the training sample is obtained, and for obtaining the linear projection matrix P of low dimensional manifold feature:
Wherein, α, β, γ are to weigh parameter, α ∈ (0,1), | ML | the element number in constraint set ML is represented, | CL | represent
Element number in constraint set CL.
The explanation of feature may refer to the related description of embodiment corresponding to Fig. 1 in embodiment corresponding to Fig. 2, here no longer
Repeat one by one.
The initialization model parameter it can be seen from above-mentioned technical proposal, the view data of acquisition is pre-processed, obtained
Image pattern;And using the division of described image sample include the training sets of exemplar and unlabeled exemplars as training sample,
Only the test set comprising unlabeled exemplars is as test sample;According to paired constraints, determine have in the training sample
Constraint set corresponding to exemplar;Using neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and be calculated
Weight matrix;According to the constraint set and the weight matrix, by minimizing feature approximation mistake, the training sample is entered
Row low dimensional manifold characteristic processing, the low dimensional manifold feature of the training sample is obtained, and for obtaining low dimensional manifold feature
Linear projection matrix;Can be to extract the characteristics of image of the test sample using the linear projection matrix.By in pairs about
Beam condition and minimum feature approximation mistake, realize while keep the global and local structural information between sample data, effectively
Improve the identifiability of feature.By minimizing feature approximation mistake, study obtains linear projection matrix, realizes quick incite somebody to action
New test data is mapped to low-dimensional, improves the performance of image characteristics extraction.
The method and dress of a kind of extraction characteristics of image based on semi-supervised learning provided above the embodiment of the present invention
Put and be described in detail.Each embodiment is described by the way of progressive in specification, and each embodiment stresses
The difference with other embodiment, between each embodiment identical similar portion mutually referring to.It is public for embodiment
For the device opened, because it is corresponded to the method disclosed in Example, so description is fairly simple, related part is referring to side
Method part illustrates.It should be pointed out that for those skilled in the art, the principle of the invention is not being departed from
Under the premise of, some improvement and modification can also be carried out to the present invention, these are improved and modification also falls into the claims in the present invention
In protection domain.
Professional further appreciates that, with reference to the unit of each example of the embodiments described herein description
And algorithm steps, can be realized with electronic hardware, computer software or the combination of the two, in order to clearly demonstrate hardware and
The interchangeability of software, the composition and step of each example are generally described according to function in the above description.These
Function is performed with hardware or software mode actually, application-specific and design constraint depending on technical scheme.Specialty
Technical staff can realize described function using distinct methods to each specific application, but this realization should not
Think beyond the scope of this invention.
Directly it can be held with reference to the step of method or algorithm that the embodiments described herein describes with hardware, processor
Capable software module, or the two combination are implemented.Software module can be placed in random access memory (RAM), internal memory, read-only deposit
Reservoir (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technology
In any other form of storage medium well known in field.
Claims (8)
- A kind of 1. method of the extraction characteristics of image based on semi-supervised learning, it is characterised in that including:Initialization model parameter, the view data of acquisition is pre-processed, obtains image pattern;And described image sample is drawn The test set conduct for dividing the training set for including exemplar and unlabeled exemplars to include unlabeled exemplars as training sample, only Test sample;According to paired constraints, determine there is constraint set corresponding to exemplar in the training sample;Using neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and weight matrix is calculated, for measuring Near-neighbor Structure information between test sample;According to the constraint set and the weight matrix, by minimizing feature approximation mistake, the training sample is carried out low Manifold characteristic processing is tieed up, obtains the low dimensional manifold feature of the training sample, and for obtaining the linear of low dimensional manifold feature Projection matrix;The training sample and the characteristics of image of the test sample are extracted using the linear projection matrix.
- 2. according to the method for claim 1, it is characterised in that the paired constraints of foundation, determine the training There is constraint set corresponding to exemplar to include in sample:According to paired constraints, by the data point of the training sample to being divided into constraint set ML and constraint set CL,ML={ (xi,xj)|i≠j,l(xi)=l (xj)};CL={ (xi,xj)|i≠j,l(xi)≠l(xj)};Wherein, l (xi) ∈ 1,2 ..., and c } represent image pattern xiThe class label of (i=1,2 ..., N);l(xj) represent figure Decent xjThe class label of (j=1,2 ..., N);According to minimum inter- object distance formulaTo the constraint set ML In point to carrying out compact processing;According to maximization between class distance formulaTo in the constraint set CL Point to carry out decentralized processing;Wherein, d (yi,yj) represent low-dimensional yiAnd yjBetween Euclidean distance, yiRepresent image pattern xiLow-dimensional represent, yjRepresent figure Decent xjLow-dimensional represent, dG(xi,xj) represent shortest path distance for approximate geodesic distance.
- 3. according to the method for claim 2, it is characterised in that it is described to utilize neighbor search algorithm, build the training sample Neighbour's figure corresponding to this, and weight matrix is calculated includes:Using K arest neighbors sorting algorithms, the training sample is handled, each training sample is obtained and each corresponds to K Individual neighbour, construction neighbour's figure;According to formulaWeight matrix W is calculated;Wherein, xiFor i-th of data in constraint set, xjFor j-th of data in the constraint set, NN (xi) it is training sample xi Neighbour set, xj∈NN(xi) represent xjFor training sample xiNeighbour, | | | | represent the l of vector2Norm, WijRepresent summit i Point to weight corresponding to summit j side.
- 4. according to the method for claim 3, it is characterised in that it is described according to the constraint set and the weight matrix, lead to Minimum feature approximation mistake is crossed, low dimensional manifold characteristic processing is carried out to the training sample, obtains the low of the training sample Manifold feature is tieed up, and the linear projection matrix for obtaining low dimensional manifold feature includes:According to equation below, feature approximation error items are minimizedObtain low dimensional manifold corresponding to the training sample Characteristic Y, and for obtaining the linear projection matrix P of low dimensional manifold feature,<mrow> <munder> <mi>min</mi> <mrow> <mi>Y</mi> <mo>,</mo> <mi>P</mi> </mrow> </munder> <mfenced open = "(" close = ")"> <mtable> <mtr> <mtd> <mrow> <mfrac> <mrow> <mn>1</mn> <mo>-</mo> <mi>&alpha;</mi> </mrow> <mrow> <mo>|</mo> <mi>M</mi> <mi>L</mi> <mo>|</mo> </mrow> </mfrac> <munder> <mo>&Sigma;</mo> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mi>j</mi> </msub> <mo>)</mo> <mo>&Element;</mo> <mi>M</mi> <mi>L</mi> </mrow> </munder> <msup> <mrow> <mo>(</mo> <mi>d</mi> <mo>(</mo> <mrow> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>y</mi> <mi>j</mi> </msub> </mrow> <mo>)</mo> <mo>-</mo> <msub> <mi>d</mi> <mi>G</mi> </msub> <mo>(</mo> <mrow> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mi>j</mi> </msub> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>-</mo> <mfrac> <mi>&alpha;</mi> <mrow> <mo>|</mo> <mi>C</mi> <mi>L</mi> <mo>|</mo> </mrow> </mfrac> <munder> <mo>&Sigma;</mo> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mi>j</mi> </msub> <mo>)</mo> <mo>&Element;</mo> <mi>C</mi> <mi>L</mi> </mrow> </munder> <msup> <mi>d</mi> <mn>2</mn> </msup> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>y</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <mo>+</mo> <msubsup> <mi>&beta;&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </msubsup> <mo>|</mo> <mo>|</mo> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>-</mo> <munder> <mo>&Sigma;</mo> <mrow> <mi>j</mi> <mo>:</mo> <msub> <mi>x</mi> <mi>j</mi> </msub> <mo>&Element;</mo> <mi>N</mi> <mi>N</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> </mrow> </munder> <msub> <mi>W</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <msub> <mi>y</mi> <mi>j</mi> </msub> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> <mo>+</mo> <msubsup> <mi>&gamma;&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </msubsup> <mo>|</mo> <mo>|</mo> <msub> <mi>Px</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>;</mo> </mrow>Wherein, α, β, γ are to weigh parameter, α ∈ (0,1), | ML | the element number in constraint set ML is represented, | CL | represent constraint Collect the element number in CL.
- 5. a kind of device of the extraction characteristics of image based on semi-supervised learning, it is characterised in that including processing unit, it is determined that single Member, computing unit, obtain unit and extraction unit, the processing unit, for initialization model parameter, to the picture number of acquisition According to being pre-processed, image pattern is obtained;And the division of described image sample is included to the instruction of exemplar and unlabeled exemplars Practice collection and the test set of unlabeled exemplars is included as training sample, only as test sample;The determining unit, for according to paired constraints, determining have in the training sample corresponding to exemplar about Constriction;The computing unit, for utilizing neighbor search algorithm, build neighbour corresponding to the training sample and scheme, and be calculated Weight matrix, for measuring the Near-neighbor Structure information between test sample;It is described to obtain unit, for according to the constraint set and the weight matrix, by minimizing feature approximation mistake, to institute State training sample and carry out low dimensional manifold characteristic processing, obtain the low dimensional manifold feature of the training sample, and it is low for obtaining Tie up the linear projection matrix of manifold feature;The extraction unit, for extracting the training sample and the image of the test sample using the linear projection matrix Feature.
- 6. device according to claim 5, it is characterised in that it is single that the determining unit includes division subelement, compact son First and scattered subelement, the division subelement, for according to paired constraints, by the data point of the training sample to drawing It is divided into constraint set ML and constraint set CL,ML={ (xi,xj)|i≠j,l(xi)=l (xj)};CL={ (xi,xj)|i≠j,l(xi)≠l(xj)};Wherein, l (xi) ∈ 1,2 ..., and c } represent image pattern xiThe class label of (i=1,2 ..., N);l(xj) represent figure Decent xjThe class label of (j=1,2 ..., N);The compact subelement, for according to minimum inter- object distance formula To the point in the constraint set ML to carrying out compact processing;The scattered subelement, for according to maximization between class distance formula To the point in the constraint set CL to carrying out decentralized processing;Wherein, d (yi,yj) represent low-dimensional yiAnd yjBetween Euclidean distance, yiRepresent image pattern xiLow-dimensional represent, yjRepresent figure Decent xjLow-dimensional represent, dG(xi,xj) represent shortest path distance for approximate geodesic distance.
- 7. device according to claim 6, it is characterised in that the computing unit is specifically used for classifying using K arest neighbors Algorithm, the training sample is handled, obtain each training sample and each correspond to K neighbour, construction neighbour's figure;And According to formulaWeight matrix W is calculated;Wherein, xiFor i-th of data in constraint set, xjFor j-th of data in the constraint set, NN (xi) it is training sample xi Neighbour set, xj∈NN(xi) represent xjFor training sample xiNeighbour, | | | | represent the l of vector2Norm, WijRepresent summit i Point to weight corresponding to summit j side.
- 8. device according to claim 7, it is characterised in that the unit that obtains is specifically used for according to equation below, most Smallization feature approximation error itemsLow dimensional manifold characteristic Y corresponding to the training sample is obtained, and it is low for obtaining Tie up the linear projection matrix P of manifold feature:<mrow> <munder> <mi>min</mi> <mrow> <mi>Y</mi> <mo>,</mo> <mi>P</mi> </mrow> </munder> <mfenced open = "(" close = ")"> <mtable> <mtr> <mtd> <mfrac> <mrow> <mn>1</mn> <mo>-</mo> <mi>&alpha;</mi> </mrow> <mrow> <mo>|</mo> <mi>M</mi> <mi>L</mi> <mo>|</mo> </mrow> </mfrac> <mstyle> <munder> <mo>&Sigma;</mo> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mi>j</mi> </msub> <mo>)</mo> <mo>&Element;</mo> <mi>M</mi> <mi>L</mi> </mrow> </munder> </mstyle> <msup> <mrow> <mo>(</mo> <mrow> <mi>d</mi> <mrow> <mo>(</mo> <mrow> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>y</mi> <mi>j</mi> </msub> </mrow> <mo>)</mo> </mrow> <mo>-</mo> <msub> <mi>d</mi> <mi>G</mi> </msub> <mrow> <mo>(</mo> <mrow> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mi>j</mi> </msub> </mrow> <mo>)</mo> </mrow> </mrow> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>-</mo> <mfrac> <mi>&alpha;</mi> <mrow> <mo>|</mo> <mi>C</mi> <mi>L</mi> <mo>|</mo> </mrow> </mfrac> <mstyle> <munder> <mo>&Sigma;</mo> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>x</mi> <mi>j</mi> </msub> <mo>)</mo> <mo>&Element;</mo> <mi>C</mi> <mi>L</mi> </mrow> </munder> </mstyle> <msup> <mi>d</mi> <mn>2</mn> </msup> <mo>(</mo> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>y</mi> <mi>j</mi> </msub> <mo>)</mo> </mtd> </mtr> <mtr> <mtd> <mo>+</mo> <mi>&beta;</mi> <msubsup> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </msubsup> <mo>|</mo> <mo>|</mo> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>-</mo> <munder> <mo>&Sigma;</mo> <mrow> <mi>j</mi> <mo>:</mo> <msub> <mi>x</mi> <mi>j</mi> </msub> <mo>&Element;</mo> <mi>N</mi> <mi>N</mi> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> </mrow> </munder> <msub> <mi>W</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <msub> <mi>y</mi> <mi>j</mi> </msub> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> <mo>+</mo> <mi>&gamma;</mi> <msubsup> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </msubsup> <mo>|</mo> <mo>|</mo> <mi>P</mi> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> </mtd> </mtr> </mtable> </mfenced> <mo>;</mo> </mrow>Wherein, α, β, γ are to weigh parameter, α ∈ (0,1), | ML | the element number in constraint set ML is represented, | CL | represent constraint Collect the element number in CL.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710796349.4A CN107563445A (en) | 2017-09-06 | 2017-09-06 | A kind of method and apparatus of the extraction characteristics of image based on semi-supervised learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710796349.4A CN107563445A (en) | 2017-09-06 | 2017-09-06 | A kind of method and apparatus of the extraction characteristics of image based on semi-supervised learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107563445A true CN107563445A (en) | 2018-01-09 |
Family
ID=60979330
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710796349.4A Pending CN107563445A (en) | 2017-09-06 | 2017-09-06 | A kind of method and apparatus of the extraction characteristics of image based on semi-supervised learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107563445A (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108491930A (en) * | 2018-03-23 | 2018-09-04 | 腾讯科技(深圳)有限公司 | A kind of processing method and data processing equipment of sample data |
CN108537269A (en) * | 2018-04-04 | 2018-09-14 | 中山大学 | A kind of the object detection deep learning method and its system of weak interactive mode |
CN108596027A (en) * | 2018-03-18 | 2018-09-28 | 西安电子科技大学 | The detection method of unknown sorting signal based on supervised learning disaggregated model |
CN109961088A (en) * | 2019-02-13 | 2019-07-02 | 苏州大学 | Unsupervised nonlinear adaptive manifold learning |
CN110210335A (en) * | 2019-05-16 | 2019-09-06 | 上海工程技术大学 | A kind of training method, system and the device of pedestrian's weight identification learning model |
CN110232690A (en) * | 2019-06-05 | 2019-09-13 | 广东工业大学 | A kind of method of image segmentation, system, equipment and computer readable storage medium |
CN110648276A (en) * | 2019-09-25 | 2020-01-03 | 重庆大学 | High-dimensional image data dimension reduction method based on manifold mapping and dictionary learning |
CN111353516A (en) * | 2018-12-21 | 2020-06-30 | 华为技术有限公司 | Sample classification method and model updating method for online learning |
CN111914906A (en) * | 2020-07-10 | 2020-11-10 | 云南电网有限责任公司带电作业分公司 | Semi-supervised data feature extraction method, device and system for semi-supervised neighbor reconstruction analysis and storage medium |
CN112966790A (en) * | 2021-04-30 | 2021-06-15 | 咪咕文化科技有限公司 | Image classification method and device, electronic equipment and storage medium |
CN113902950A (en) * | 2021-10-16 | 2022-01-07 | 江苏大学 | Image multi-label classification method based on geometric preservation double projection |
CN115861720A (en) * | 2023-02-28 | 2023-03-28 | 人工智能与数字经济广东省实验室(广州) | Small sample subclass image classification and identification method |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2013177586A1 (en) * | 2012-05-25 | 2013-11-28 | The Johns Hopkins University | An integrated real-time tracking system for normal and anomaly tracking and the methods therefor |
CN104751191A (en) * | 2015-04-23 | 2015-07-01 | 重庆大学 | Sparse self-adaptive semi-supervised manifold learning hyperspectral image classification method |
CN105608478A (en) * | 2016-03-30 | 2016-05-25 | 苏州大学 | Combined method and system for extracting and classifying features of images |
CN105608471A (en) * | 2015-12-28 | 2016-05-25 | 苏州大学 | Robust transductive label estimation and data classification method and system |
-
2017
- 2017-09-06 CN CN201710796349.4A patent/CN107563445A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2013177586A1 (en) * | 2012-05-25 | 2013-11-28 | The Johns Hopkins University | An integrated real-time tracking system for normal and anomaly tracking and the methods therefor |
CN104751191A (en) * | 2015-04-23 | 2015-07-01 | 重庆大学 | Sparse self-adaptive semi-supervised manifold learning hyperspectral image classification method |
CN105608471A (en) * | 2015-12-28 | 2016-05-25 | 苏州大学 | Robust transductive label estimation and data classification method and system |
CN105608478A (en) * | 2016-03-30 | 2016-05-25 | 苏州大学 | Combined method and system for extracting and classifying features of images |
Non-Patent Citations (1)
Title |
---|
BO YANG等: "Multi-manifold Discriminant Isomap for visualization and classification", 《PATTERN RECOGNITION》 * |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108596027A (en) * | 2018-03-18 | 2018-09-28 | 西安电子科技大学 | The detection method of unknown sorting signal based on supervised learning disaggregated model |
CN108491930A (en) * | 2018-03-23 | 2018-09-04 | 腾讯科技(深圳)有限公司 | A kind of processing method and data processing equipment of sample data |
CN108537269A (en) * | 2018-04-04 | 2018-09-14 | 中山大学 | A kind of the object detection deep learning method and its system of weak interactive mode |
CN111353516A (en) * | 2018-12-21 | 2020-06-30 | 华为技术有限公司 | Sample classification method and model updating method for online learning |
CN109961088A (en) * | 2019-02-13 | 2019-07-02 | 苏州大学 | Unsupervised nonlinear adaptive manifold learning |
CN109961088B (en) * | 2019-02-13 | 2020-09-29 | 苏州大学 | Unsupervised non-linear self-adaptive manifold learning method |
CN110210335A (en) * | 2019-05-16 | 2019-09-06 | 上海工程技术大学 | A kind of training method, system and the device of pedestrian's weight identification learning model |
CN110210335B (en) * | 2019-05-16 | 2021-05-11 | 上海工程技术大学 | Training method, system and device for pedestrian re-recognition learning model |
CN110232690A (en) * | 2019-06-05 | 2019-09-13 | 广东工业大学 | A kind of method of image segmentation, system, equipment and computer readable storage medium |
CN110648276A (en) * | 2019-09-25 | 2020-01-03 | 重庆大学 | High-dimensional image data dimension reduction method based on manifold mapping and dictionary learning |
CN111914906A (en) * | 2020-07-10 | 2020-11-10 | 云南电网有限责任公司带电作业分公司 | Semi-supervised data feature extraction method, device and system for semi-supervised neighbor reconstruction analysis and storage medium |
CN111914906B (en) * | 2020-07-10 | 2022-12-16 | 云南电网有限责任公司带电作业分公司 | Semi-supervised data feature extraction method, device and system for semi-supervised neighbor reconstruction analysis and storage medium |
CN112966790A (en) * | 2021-04-30 | 2021-06-15 | 咪咕文化科技有限公司 | Image classification method and device, electronic equipment and storage medium |
CN112966790B (en) * | 2021-04-30 | 2024-02-27 | 咪咕文化科技有限公司 | Image classification method, device, electronic equipment and storage medium |
CN113902950A (en) * | 2021-10-16 | 2022-01-07 | 江苏大学 | Image multi-label classification method based on geometric preservation double projection |
CN113902950B (en) * | 2021-10-16 | 2024-03-19 | 江苏大学 | Image multi-label classification method based on geometric retention double projection |
CN115861720A (en) * | 2023-02-28 | 2023-03-28 | 人工智能与数字经济广东省实验室(广州) | Small sample subclass image classification and identification method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107563445A (en) | A kind of method and apparatus of the extraction characteristics of image based on semi-supervised learning | |
CN104881689B (en) | A kind of multi-tag Active Learning sorting technique and system | |
CN105335756B (en) | A kind of image classification method and image classification system based on Robust Learning model | |
CN102982165B (en) | Large-scale human face image searching method | |
CN111738355B (en) | Image classification method and device with attention fused with mutual information and storage medium | |
CN107330100A (en) | Combine the two-way search method of image text of embedded space based on multi views | |
CN109242041A (en) | A kind of electric energy meter abnormal deviation data examination method, device, equipment and storage medium | |
CN107133213A (en) | A kind of text snippet extraction method and system based on algorithm | |
CN108446711A (en) | A kind of Software Defects Predict Methods based on transfer learning | |
CN108399414A (en) | Method of Sample Selection and device | |
CN104166731A (en) | Discovering system for social network overlapped community and method thereof | |
CN110135459A (en) | A kind of zero sample classification method based on double triple depth measure learning networks | |
CN105678261B (en) | Based on the direct-push Method of Data with Adding Windows for having supervision figure | |
CN104281835B (en) | Face recognition method based on local sensitive kernel sparse representation | |
CN106055573A (en) | Method and system for shoeprint image retrieval under multi-instance learning framework | |
CN105608478A (en) | Combined method and system for extracting and classifying features of images | |
CN107527065A (en) | A kind of flower variety identification model method for building up based on convolutional neural networks | |
CN113032613B (en) | Three-dimensional model retrieval method based on interactive attention convolution neural network | |
CN109284760A (en) | A kind of furniture detection method and device based on depth convolutional neural networks | |
CN103839074B (en) | Image classification method based on matching of sketch line segment information and space pyramid | |
CN108897750A (en) | Merge the personalized location recommendation method and equipment of polynary contextual information | |
CN104598889A (en) | Human action recognition method and device | |
CN110378872A (en) | A kind of multi-source adaptive equalization transfer learning method towards crack image detection | |
CN106971201A (en) | Multi-tag sorting technique based on integrated study | |
CN107392129A (en) | Face retrieval method and system based on Softmax |
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 |
Application publication date: 20180109 |
|
RJ01 | Rejection of invention patent application after publication |