CN111797911A - Image data multi-label classification method - Google Patents

Image data multi-label classification method Download PDF

Info

Publication number
CN111797911A
CN111797911A CN202010573202.0A CN202010573202A CN111797911A CN 111797911 A CN111797911 A CN 111797911A CN 202010573202 A CN202010573202 A CN 202010573202A CN 111797911 A CN111797911 A CN 111797911A
Authority
CN
China
Prior art keywords
matrix
picture
label
pictures
correlation
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
Application number
CN202010573202.0A
Other languages
Chinese (zh)
Other versions
CN111797911B (en
Inventor
陈刚
胡天磊
陈珂
刘雨辰
李梦谨
王皓波
寿黎但
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202010573202.0A priority Critical patent/CN111797911B/en
Publication of CN111797911A publication Critical patent/CN111797911A/en
Application granted granted Critical
Publication of CN111797911B publication Critical patent/CN111797911B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a multi-label classification method for image data, which comprises the steps of firstly constructing a topological relation of pictures, and then solving the problem of semi-supervised multi-label learning (SSML) by using a multi-label propagation algorithm (CMLP) based on cooperation so as to obtain a confidence matrix of unmarked pictures

Description

Image data multi-label classification method
Technical Field
The invention relates to the field of machine learning, in particular to a multi-label classification method for image data.
Background
Image classification is one of the most widely used machine learning problems. In real-world image classification applications, an image instance is often associated with multiple labels, and there is a correlation between the labels. Therefore, classifying images with multiple associated labels is an important research task. In the conventional picture label classification problem, a common assumption is that each picture in the training data set is accurately labeled. Unfortunately, in many real world problems, although a vast number of images are readily available, the labeling of the images requires an expensive and time-consuming manual labeling process to obtain.
For this reason, a semi-supervised learning method is proposed, in which a large number of unlabeled images are added to a limited number of labeled images and trained together, so as to improve the performance of the classifier. The label propagation method is a semi-supervised classification learning method based on a graph. However, when an image instance has multiple correlated labels, simple label propagation does not make good use of the correlation, resulting in a degradation of classification performance.
Disclosure of Invention
In order to solve the problems in the background art, the invention provides a multi-label classification method for image data.
The purpose of the invention is realized by the following technical scheme: a multi-label classification method for image data comprises the following steps:
(1) acquiring a picture set D ═ { pic) to be classifiediI is more than or equal to 1 and less than or equal to n, wherein n is the total number of the pictures in the picture set D; in the picture set D, each picture has q ordered tags to be marked, and totally l pictures are marked whether to have the q tags or not, wherein the marked pictures are pici，1≤i≤l；
(2) Calculating a weight matrix W and constructing a graph relation between pictures, wherein the step comprises the following substeps:
(2.1) inputting the pictures in the picture set into the trained convolutional neural network VGG Net, and acquiring the feature vector set of the pictures from the outputV＝{xiI is more than or equal to 1 and less than or equal to n, wherein xiIs a picture piciThe output obtained by inputting the convolutional neural network is a p-dimensional column vector;
(2.2) selecting a hyper-parameter k, and calculating a feature vector x according to the picture feature vector set ViK neighbor set kNN (x)i)，kNN(xi) Is the picture feature vector set V and the feature vector xiA set of k picture feature vectors with the minimum Euclidean distance;
for non-negative weight matrix W ═ Wij]n×nIt is required to satisfy:
wii＝0
wij≥0，j≠i，xj∈kNN(xi)
Figure BDA0002550140590000021
(2.3) feature vector x for each pictureiIs linearly reconstructed into sigma by other picture characteristic vectorsj≠iwijxjThe trade-off parameter θ is chosen, and the linear reconstruction error of the weight matrix W is defined as follows:
Figure BDA0002550140590000022
wherein | · | purple2A two-norm representing a matrix;
minimizing the linear reconstruction error (W) by using a constrained least square programming method, and establishing the following minimization model:
Figure BDA0002550140590000023
s.t.wii＝0
wij≥0，j≠i，xj∈kNN(xi)
Figure BDA0002550140590000024
j＝1，2，...，n
wherein, w·jIs the jth column vector of W, GjDenotes w·jCorresponding n rows and n columns of gray matrix, GjThe a-th row and the b-th column of the element (x)j-xa)′(xj-xb) V' denotes the transpose of a certain vector v;
solving the minimized model by an active set method of a convex quadratic programming problem to obtain a non-negative weight matrix W, thereby constructing a graph relation between pictures;
(3) solving semi-supervised multi-tag learning (SSML) problems with a collaboration-based multi-tag propagation algorithm (CMLP) to obtain a confidence matrix for unmarked pictures
Figure BDA0002550140590000028
The method comprises the following substeps:
(3.1) obtaining the propagation matrix P by normalizing the weight matrix W:
Figure BDA0002550140590000025
wherein D ═ diag { D ═ D1，d2，…，dnIs the diagonal matrix, the ith diagonal element of matrix D is
Figure BDA0002550140590000026
Figure BDA0002550140590000027
Thus, the propagation matrix P is obtained by normalizing the weight matrix W, and the labels of the pictures with similar characteristics are also similar;
constructing an object matrix Y of the marked picture as Yij]l×qThe following were used:
yij1, picture piciHaving the jth label
yij1, picture piciWithout jth tag
Selecting a cooperation degree parameter alpha and a regularization parameter gamma, and calculating a correlation matrix R ═ Rij]q×q：
Figure BDA0002550140590000031
Figure BDA0002550140590000032
Wherein, y·j，r·jDenotes Y, the jth column of R, I denotes a q × q identity matrix,
Figure BDA0002550140590000033
is that
Figure BDA0002550140590000034
Transposed matrix of (1), Ol×1Is a zero column vector of dimension l; a plurality of labels in the image often have certain correlation, for example, the labels "have the sun" and "are sunny days" have strong correlation; by calculating the correlation matrix, the invention extracts the correlation; when the classification task has a plurality of related labels, the related matrix provides a powerful tool for improving the accuracy of the prediction result;
(3.2) by iteratively updating F and Z alternately, the following loss function is minimized:
Figure BDA0002550140590000035
where F is the model output, F' is the transposed matrix of F, FlIs a matrix formed by the first row of the matrix F, represents the prediction result of the model on the marked image, Z is an intermediate variable of the model, P is a propagation matrix obtained in (2.1), the matrix Q is (1-alpha) I + alpha R, R is a correlation matrix obtained in (2.1), alpha is a cooperation degree parameter selected in (2.1), mu and lambda are balance parameters, tr is a trace function of the matrix, | | | | | I |, andFis the F-norm of the matrix;
initializing model output F by using the target matrix Y obtained in (2.1)0And intermediate variable Z0：
Figure BDA0002550140590000036
Z0＝Y
Wherein, O(n-l)×qA zero matrix that is (n-l) xq;
selecting a hyper-parameter learning rate beta and an iteration number T, and outputting F to the initialized model0And intermediate variable Z0Updating to obtain FtAnd ZtThe iterative update formula is as follows:
Figure BDA0002550140590000037
Figure BDA0002550140590000038
wherein,
Figure BDA0002550140590000039
is a matrix Ft+1The matrix formed by the first row, and Q' is a transposed matrix of the matrix Q;
by bringing the correlation matrix into the iteration process, the method makes full use of the correlation among the labels, considers the prediction of each label on other labels, and improves the accuracy of the prediction result;
when T iterations are finished, obtaining model output F ═ FT；
(3.3) converting the model output F to a final prediction:
E＝FuQ＝[eij](n-l)×q
Figure BDA00025501405900000310
wherein, FuIs the matrix formed by the l +1 th row to the n th row of the matrix F, Q is the matrix used in the iteration in (2.2), Ψ is the post-processing operator:
Figure BDA0002550140590000041
wherein sgn is a sign function;
obtaining a predicted result when
Figure BDA0002550140590000042
Then, the (q) th label is not on the (i + l) th picture; when in use
Figure BDA0002550140590000043
Then, the (q) th label is arranged on the (i + l) th picture; when in use
Figure BDA0002550140590000044
In time, whether the q label is uncertain exists on the (i + l) th picture or not is determined.
The invention has the beneficial effects that: the invention solves the problem that the simple label propagation method can not utilize the correlation among a plurality of labels of the multi-label image data, introduces a correlation matrix in the iterative process of label propagation, only propagates the independent part of each label in the iterative process, considers the prediction of the label on the label when outputting the final label prediction result, also absorbs the prediction results of other labels on the label, fully utilizes the correlation among the labels and improves the accuracy of multi-label classification.
Detailed Description
Set of pictures to be classified D ═ { piciI is more than or equal to 1 and less than or equal to n, wherein n is the total number of the pictures in the picture set D. In the picture set D, each picture has q ordered tags to be marked, and totally l pictures are marked whether to have the q tags or not, wherein the marked pictures are pici，1≤i≤l。
(1) Calculating a weight matrix W and constructing a graph relation between pictures, wherein the step comprises the following substeps:
(1.1) Using the public data set ImageNet on the website http:// www.image-net.org/the convolutional neural network VGG Net proposed in the paper was trained according to the method provided by 3.1 in the paper VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION, co-published in 2015 by KarenSimnyan and Andrew Zisserman. VGG Net is a visual geometry group and valley of Oxford universityResearchers of the song deep mind company research and develop together, and feature vectors of pictures required by classification can be well extracted. Inputting the pictures in the picture set into a trained convolutional neural network, and acquiring a feature vector set V ═ x of the pictures from the outputiI is more than or equal to 1 and less than or equal to n, wherein xiIs a picture piciThe output obtained by inputting the convolutional neural network is a p-dimensional column vector;
(1.2) firstly selecting a hyper-parameter k, and calculating a feature vector x according to the picture feature vector set ViK neighbor set kNN (x)i)，kNN(xi) Is the picture feature vector set V and the feature vector xiA set of k picture feature vectors with the minimum Euclidean distance;
for the weight matrix W ═ Wij]n×nIt is required to satisfy:
wii＝0
wij≥0，j≠i，xj∈kNN(xi)
Figure BDA0002550140590000045
under such constraints, the weight matrix W is sparse, which may speed up training.
(1.3) feature vector x for each pictureiIs linearly reconstructed into sigma by other picture characteristic vectorsj≠iwijxjThe trade-off parameter θ is chosen, and the linear reconstruction error of the weight matrix W is defined as follows:
Figure BDA0002550140590000051
wherein | · | purple2A two-norm representing a matrix;
minimizing the linear reconstruction error (W) by using a constrained least square programming method, and establishing the following minimization model:
Figure BDA0002550140590000052
s.t.wii＝0
wij≥0，j≠i，xj∈kNN(xi)
Figure BDA0002550140590000053
j＝1，2，...，n
wherein, w·jIs the jth column vector of W, W'·jIs w·jTransposed vector of (1), GjDenotes w·jCorresponding n rows and n columns of gray matrix, GjThe a-th row and the b-th column of the element (x)j-xa)′(xj-xb)；
The non-negative weight matrix W can be obtained by solving the above minimization model by the active set method of the convex quadratic programming problem with reference to the Numerical Optimization (Second Edition) book 16.5, thereby constructing the graph relationship between the pictures.
(2) Solving semi-supervised multi-tag learning (SSML) problems with a collaboration-based multi-tag propagation algorithm (CMLP) to obtain a confidence matrix for unmarked pictures
Figure BDA0002550140590000054
The method comprises the following substeps:
(2.1) obtaining the propagation matrix P by normalizing the weight matrix W:
Figure BDA0002550140590000055
wherein D ═ diag { D ═ D1，d2，…，dnIs the diagonal matrix, the ith diagonal element of matrix D is
Figure BDA0002550140590000056
Figure BDA0002550140590000057
Thus, the propagation matrix P is obtained by normalizing the weight matrix W, and the labels of the pictures with similar characteristics are also similar;
constructing tagged picturesIs given as the target matrix Y ═ Yij]l×qThe following were used:
yij1, picture piciHaving the jth label
yij1, picture piciWithout jth tag
Selecting a cooperation degree parameter alpha and a regularization parameter gamma, and calculating a correlation matrix R ═ Rij]q×q：
Figure BDA0002550140590000061
Figure BDA0002550140590000062
Wherein, y·j，r·jDenotes Y, the jth column of R, I denotes a q × q identity matrix,
Figure BDA0002550140590000063
is that
Figure BDA0002550140590000064
Transposed matrix of (1), Ol×1Is a zero column vector of dimension l. The labels in the image often have a certain correlation, for example, the labels "have the sun" and "are sunny" have a strong correlation. The invention extracts this correlation by computing a correlation matrix. When the classification task has a plurality of related labels, the related matrix provides a powerful tool for improving the accuracy of the prediction result.
(2.2) by iteratively updating F and z alternately, the following loss function is minimized:
Figure BDA0002550140590000065
where F is the model output, F' is the transposed matrix of F, FlIs a matrix formed by the first row of the matrix F, represents the prediction result of the model on the marked image, Z is an intermediate variable of the model, P is a propagation matrix obtained in (2.1), the matrix Q is (1-alpha) I + alpha R, and R is in (2.1)The obtained correlation matrix, alpha is the cooperation degree parameter selected in (2.1), mu and lambda are balance parameters, tr (-) is the trace function of the matrix, | | |FIs the F-norm of the matrix.
Initializing model output F by using the target matrix Y obtained in (2.1)0And intermediate variable Z0：
Figure BDA0002550140590000066
Z0＝Y
Wherein, O(n-l)×qIs a zero matrix of (n-l) × q.
Selecting a hyper-parameter learning rate beta and an iteration number T, and outputting F to the initialized model0And intermediate variable Z0Updating to obtain FtAnd ZtThe iterative update formula is as follows:
Figure BDA0002550140590000067
Figure BDA0002550140590000068
wherein,
Figure BDA0002550140590000069
is a matrix Ft+1The first l rows form a matrix, and Q' is the transpose of matrix Q.
By incorporating the correlation matrix into the iterative process, the method makes full use of the correlation among the labels, considers the prediction of each label on other labels, and improves the accuracy of the prediction result.
When T iterations are finished, obtaining model output F ═ FT；
(2.3) converting the model output F to a final prediction:
E＝FuQ＝[eij](n-l)×q
Figure BDA00025501405900000610
wherein, FuIs a matrix formed by the l +1 th to n th rows of the matrix F, Q is the matrix used in (2.2), Ψ is the post-processing operator:
Figure BDA0002550140590000071
where sgn is a sign function.
Obtaining a predicted result when
Figure BDA0002550140590000072
Then, the (q) th label is not on the (i + l) th picture; when in use
Figure BDA0002550140590000073
Then, the (q) th label is arranged on the (i + l) th picture; when in use
Figure BDA0002550140590000074
In time, whether the q label is uncertain exists on the (i + l) th picture or not is determined.

Claims (1)

1. A multi-label classification method for image data is characterized by comprising the following steps: the method comprises the following steps:
(1) acquiring a picture set D ═ { pic) to be classifiediI is more than or equal to 1 and less than or equal to n, wherein n is the total number of the pictures in the picture set D. In the picture set D, each picture has q ordered tags to be marked, and totally l pictures are marked whether to have the q tags or not, wherein the marked pictures are pici，1≤i≤l。
(2) Calculating a weight matrix W and constructing a graph relation between pictures, wherein the step comprises the following substeps:
(2.1) inputting the pictures in the picture set into the trained convolutional neural network VGG Net, and acquiring the feature vector set V of the pictures from the output, wherein the feature vector set V is { x }iI is more than or equal to 1 and less than or equal to n, wherein xiIs a picture piciThe output obtained by inputting the convolutional neural network is a p-dimensional column vector;
(2.2) selecting a hyper-parameter k, and according to the picture feature vector set VComputing a feature vector xiK neighbor set kNN (x)i)，kNN(xi) Is the picture feature vector set V and the feature vector xiA set of k picture feature vectors with the minimum Euclidean distance;
for non-negative weight matrix W ═ Wij]n×nIt is required to satisfy:
wii＝0
wij≥0，j≠i，xj∈kNN(xi)
wij＝0，j≠i，
Figure FDA0002550140580000011
(2.3) feature vector x for each pictureiIs linearly reconstructed into sigma by other picture characteristic vectorsj≠iwijxjThe trade-off parameter θ is chosen, and the linear reconstruction error of the weight matrix W is defined as follows:
Figure FDA0002550140580000012
wherein | · | purple2A two-norm representing a matrix;
minimizing the linear reconstruction error (W) by using a constrained least square programming method, and establishing the following minimization model:
Figure FDA0002550140580000013
s.t.wii＝0
wij≥0，j≠i，xj∈kNN(xi)
wij＝0，j≠i，
Figure FDA0002550140580000014
j＝1，2，...，n
wherein, w·jIs the jth column vector of W, GjDenotes w·jCorresponding n rows and n columns of grayMatrix, GjThe a-th row and the b-th column of the element (x)j-xa)′(xj-xb) V' denotes the transpose of a certain vector v;
solving the minimized model by an active set method of a convex quadratic programming problem to obtain a non-negative weight matrix W, thereby constructing a graph relation between pictures.
(3) Solving semi-supervised multi-tag learning (SSML) problems with a collaboration-based multi-tag propagation algorithm (CMLP) to obtain a confidence matrix for unmarked pictures
Figure FDA00025501405800000210
The method comprises the following substeps:
(3.1) obtaining the propagation matrix P by normalizing the weight matrix W:
Figure FDA0002550140580000021
wherein D ═ diag { D ═ D1，d2，…，dnIs the diagonal matrix, the ith diagonal element of matrix D is
Figure FDA0002550140580000022
Figure FDA0002550140580000023
Thus, the propagation matrix P is obtained by normalizing the weight matrix W, and the labels of the pictures with similar characteristics are also similar;
constructing an object matrix Y of the marked picture as Yij]l×qThe following were used:
yij1, picture piciHaving the jth label
yij1, picture piciWithout jth tag
Selecting a cooperation degree parameter alpha and a regularization parameter gamma, and calculating a correlation matrix R ═ Rij]q×q：
Figure FDA0002550140580000024
Figure FDA0002550140580000025
Wherein, y·j，r·jDenotes Y, the jth column of R, I denotes a q × q identity matrix,
Figure FDA0002550140580000026
is that
Figure FDA0002550140580000027
Transposed matrix of (1), Ol×1Is a zero column vector of dimension l. The labels in the image often have a certain correlation, for example, the labels "have the sun" and "are sunny" have a strong correlation. The invention extracts this correlation by computing a correlation matrix. When the classification task has a plurality of related labels, the related matrix provides a powerful tool for improving the accuracy of the prediction result.
(3.2) by iteratively updating F and Z alternately, the following loss function is minimized:
Figure FDA0002550140580000028
where F is the model output, F' is the transposed matrix of F, FlIs a matrix formed by the first row of the matrix F, represents the prediction result of the model on the marked image, Z is an intermediate variable of the model, P is a propagation matrix obtained in (2.1), the matrix Q is (1-alpha) I + alpha R, R is a correlation matrix obtained in (2.1), alpha is a cooperation degree parameter selected in (2.1), mu and lambda are balance parameters, tr is a trace function of the matrix, | | | | | I |, andFis the F-norm of the matrix.
Initializing model output F by using the target matrix Y obtained in (2.1)0And intermediate variable Z0：
Figure FDA0002550140580000029
Z0＝Y
Wherein, O(n-l)×qIs a zero matrix of (n-l) × q.
Selecting a hyper-parameter learning rate beta and an iteration number T, and outputting F to the initialized model0And intermediate variable Z0Updating to obtain FtAnd ZtThe iterative update formula is as follows:
Figure FDA0002550140580000031
Figure FDA0002550140580000032
wherein,
Figure FDA0002550140580000033
is a matrix Ft+1The first l rows form a matrix, and Q' is the transpose of matrix Q.
By incorporating the correlation matrix into the iterative process, the method makes full use of the correlation among the labels, considers the prediction of each label on other labels, and improves the accuracy of the prediction result.
When T iterations are finished, obtaining model output F ═ FT；
(3.3) converting the model output F to a final prediction:
E＝FuQ＝[eij](n-l)×q
Figure FDA0002550140580000034
wherein, FuIs the matrix formed by the l +1 th row to the n th row of the matrix F, Q is the matrix used in the iteration in (2.2), Ψ is the post-processing operator:
Figure FDA0002550140580000035
where sgn is a sign function.
Obtaining a predicted result when
Figure FDA0002550140580000036
Then, the (q) th label is not on the (i + l) th picture; when in use
Figure FDA0002550140580000037
Then, the (q) th label is arranged on the (i + l) th picture; when in use
Figure FDA0002550140580000038
In time, whether the q label is uncertain exists on the (i + l) th picture or not is determined.
CN202010573202.0A 2020-06-22 2020-06-22 Multi-label classification method for image data Active CN111797911B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010573202.0A CN111797911B (en) 2020-06-22 2020-06-22 Multi-label classification method for image data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010573202.0A CN111797911B (en) 2020-06-22 2020-06-22 Multi-label classification method for image data

Publications (2)

Publication Number Publication Date
CN111797911A true CN111797911A (en) 2020-10-20
CN111797911B CN111797911B (en) 2024-02-09

Family

ID=72804571

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010573202.0A Active CN111797911B (en) 2020-06-22 2020-06-22 Multi-label classification method for image data

Country Status (1)

Country Link
CN (1) CN111797911B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113297382A (en) * 2021-06-21 2021-08-24 西南大学 Method for processing instrument and equipment function labeling
CN113537458A (en) * 2021-06-21 2021-10-22 中国科学技术大学 Rational function neural network construction method, system and readable storage medium
CN115100456A (en) * 2022-05-27 2022-09-23 福建天甫电子材料有限公司 Intelligent toxic and harmful gas alarm system for preparing electronic-grade hydrogen peroxide

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110210515A (en) * 2019-04-25 2019-09-06 浙江大学 A kind of image data multi-tag classification method
CN110647907A (en) * 2019-08-05 2020-01-03 广东工业大学 Multi-label image classification algorithm using multi-layer classification and dictionary learning
CN110717390A (en) * 2019-09-05 2020-01-21 杭州电子科技大学 Electroencephalogram signal classification method based on graph semi-supervised width learning
US20200160177A1 (en) * 2018-11-16 2020-05-21 Royal Bank Of Canada System and method for a convolutional neural network for multi-label classification with partial annotations
CN111259938A (en) * 2020-01-09 2020-06-09 浙江大学 Manifold learning and gradient lifting model-based image multi-label classification method

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200160177A1 (en) * 2018-11-16 2020-05-21 Royal Bank Of Canada System and method for a convolutional neural network for multi-label classification with partial annotations
CN110210515A (en) * 2019-04-25 2019-09-06 浙江大学 A kind of image data multi-tag classification method
CN110647907A (en) * 2019-08-05 2020-01-03 广东工业大学 Multi-label image classification algorithm using multi-layer classification and dictionary learning
CN110717390A (en) * 2019-09-05 2020-01-21 杭州电子科技大学 Electroencephalogram signal classification method based on graph semi-supervised width learning
CN111259938A (en) * 2020-01-09 2020-06-09 浙江大学 Manifold learning and gradient lifting model-based image multi-label classification method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
郭玉堂;李艳: "基于连续预测的半监督学习图像语义标注", 计算机工程与科学, vol. 37, no. 3, pages 553 - 558 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113297382A (en) * 2021-06-21 2021-08-24 西南大学 Method for processing instrument and equipment function labeling
CN113537458A (en) * 2021-06-21 2021-10-22 中国科学技术大学 Rational function neural network construction method, system and readable storage medium
CN113297382B (en) * 2021-06-21 2023-04-25 西南大学 Instrument and equipment function labeling processing method
CN115100456A (en) * 2022-05-27 2022-09-23 福建天甫电子材料有限公司 Intelligent toxic and harmful gas alarm system for preparing electronic-grade hydrogen peroxide
CN115100456B (en) * 2022-05-27 2023-02-03 福建天甫电子材料有限公司 Intelligent toxic and harmful gas alarm system for preparing electronic-grade hydrogen peroxide

Also Published As

Publication number Publication date
CN111797911B (en) 2024-02-09

Similar Documents

Publication Publication Date Title
CN111797911A (en) Image data multi-label classification method
CN111461157B (en) Self-learning-based cross-modal Hash retrieval method
CN110188827B (en) Scene recognition method based on convolutional neural network and recursive automatic encoder model
CN111476315B (en) Image multi-label identification method based on statistical correlation and graph convolution technology
CN113610173A (en) Knowledge distillation-based multi-span domain few-sample classification method
CN112199532B (en) Zero sample image retrieval method and device based on Hash coding and graph attention machine mechanism
Jia et al. ROML: A robust feature correspondence approach for matching objects in a set of images
CN109871454B (en) Robust discrete supervision cross-media hash retrieval method
CN110428424B (en) Radar echo image high-voltage line segmentation method based on deep learning
CN113010683B (en) Entity relationship identification method and system based on improved graph attention network
CN107346327A (en) The zero sample Hash picture retrieval method based on supervision transfer
CN110837736B (en) Named entity recognition method of Chinese medical record based on word structure
CN113947725B (en) Hyperspectral image classification method based on convolution width migration network
CN112465226B (en) User behavior prediction method based on feature interaction and graph neural network
CN111259938B (en) Manifold learning and gradient lifting model-based image multi-label classification method
CN115147607A (en) Anti-noise zero-sample image classification method based on convex optimization theory
CN114676769A (en) Visual transform-based small sample insect image identification method
CN112766400A (en) Semi-supervised classification integration method for high-dimensional data based on multiple data transformation spaces
CN114972904B (en) Zero sample knowledge distillation method and system based on fighting against triplet loss
CN114926691A (en) Insect pest intelligent identification method and system based on convolutional neural network
CN111461229B (en) Deep neural network optimization and image classification method based on target transfer and line search
CN109034387A (en) A kind of approximation method for quickly training self-encoding encoder based on pseudo- reversal learning
CN111259176A (en) Cross-modal Hash retrieval method based on matrix decomposition and integrated with supervision information
CN113158878B (en) Heterogeneous migration fault diagnosis method, system and model based on subspace
CN112801153B (en) Semi-supervised image classification method and system of image embedded with LBP (local binary pattern) features

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