CN110222213A - A kind of image classification method based on isomery tensor resolution - Google Patents

A kind of image classification method based on isomery tensor resolution Download PDF

Info

Publication number
CN110222213A
CN110222213A CN201910453844.4A CN201910453844A CN110222213A CN 110222213 A CN110222213 A CN 110222213A CN 201910453844 A CN201910453844 A CN 201910453844A CN 110222213 A CN110222213 A CN 110222213A
Authority
CN
China
Prior art keywords
tensor
matrix
rank
image classification
isomery
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
CN201910453844.4A
Other languages
Chinese (zh)
Other versions
CN110222213B (en
Inventor
井佩光
武宇廷
苏育挺
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.)
Tianjin University
Original Assignee
Tianjin University
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 Tianjin University filed Critical Tianjin University
Priority to CN201910453844.4A priority Critical patent/CN110222213B/en
Publication of CN110222213A publication Critical patent/CN110222213A/en
Application granted granted Critical
Publication of CN110222213B publication Critical patent/CN110222213B/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
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/55—Clustering; Classification
    • 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
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of image classification methods based on isomery tensor resolution, the described method comprises the following steps: constructing original tensor according to ORL data set；It is factor matrix corresponding with each rank of tensor and sample characteristics characterization matrix by original tensor resolution；Orthogonality constraint by column is added on factor matrix；Low-rank constraint is applied to sample characteristics characterization matrix, by identifying that the global low-rank structure of sample indicates to obtain lowest rank；Pass through l2,1Norm constraint loss function and regularization term realize the target for differentiating robust features selection；Obtain complete objective function；Iteration is optimized to objective function with alternating direction multiplier, optimal solution is obtained, further obtains classifier；Test set picture in ORL is inputted in above-mentioned trained classifier, image classification is completed.The present invention, which solves to be easy to cause in traditional vector or matrix method, to lose structural information and destroys the dependence between element and the isomerism during tensor resolution.

Description

A kind of image classification method based on isomery tensor resolution
Technical field
The present invention relates to image classification field more particularly to a kind of image classification methods based on isomery tensor resolution.
Background technique
Nowadays, with the fast development of present information acquisition and storage technology, the never homologous or channel in multiple views Mass data is generated, while by noisy damage, lacks entry and high-dimensional.For example, multidimensional data is big in many fields Amount production, comprising: remote sensing, medical science, biology, computer vision and image procossing.
In this case, the complicated composed structure in data description highlights the limitation of conventional vector or matrix method Property, therefore such scheme makes data become flattening and partly destroys the complete structure letter established in multiple dimensions Breath.
Traditional small sample image classification and generally two steps the step of detection: feature extraction and training classifier.
In feature extraction phases, designer can attempt various generic features or self-designed feature and carry out spy to image Sign is extracted, and by taking Face datection as an example, general feature is exactly histograms of oriented gradients (HOG), accelerates robust features (SURF), office The features such as portion's binary pattern (LBP), and the feature relatively good for face effect has haar feature.Classical small sample image point Class method has support vector machines (SVM), principal component analysis (PCA) etc..
After deep learning in 2015 captures each image procossing match umber one, existing image classification largely makes Method is all deep learning, that is, neural network.Neural network is built into layer by many neurons Network, by active coating come so that model has the ability of very strong nonlinear fitting.Designer need to only input image, then accuse Tell that model needs the result is that, model will be extracted automatically and result mapping learning characteristic.
By deep learning, the work of designer is concentrated mainly in image characteristics extraction.Furthermore designer only needs to set Network structure is counted, so that the feature that network automatically extracts is better, effect will be more preferable, and accuracy is higher.Deep learning is for figure It is highly useful as handling, but there are also drawbacks simultaneously, such as: it needs mass data, adjust parameter by experience, the calculating needed Ability is very high, the real life scenarios for being suitble to processing very complicated.
Summary of the invention
The present invention provides a kind of image classification method based on isomery tensor resolution, the present invention solve traditional vector or It is easy to cause in matrix method and loses structural information and destroy the dependence between element and the isomery during tensor resolution Property, it is described below:
A kind of image classification method based on isomery tensor resolution, the described method comprises the following steps:
Original tensor is constructed according to ORL data set；By original tensor resolution be factor matrix corresponding with each rank of tensor and Sample characteristics characterization matrix；
Orthogonality constraint by column is added on factor matrix；Low-rank constraint is applied to sample characteristics characterization matrix, passes through identification The global low-rank structure of sample indicates to obtain lowest rank；
Pass through l2,1Norm constraint loss function and regularization term realize the target for differentiating robust features selection；
Obtain complete objective function；Iteration is optimized to objective function with alternating direction multiplier, obtains optimal solution, into One step obtains classifier；
Test set picture in ORL is inputted in above-mentioned trained classifier, image classification is completed.
Wherein, described to have original tensor resolution for factor matrix corresponding with each rank of tensor and sample characteristics characterization matrix Body are as follows:
1) to the preceding M rank of tensor, learn the low-dimensional insertion that M orthogonal factor matrixs realize original tensor, to tensor Last first-order learning characteristic present matrix corresponding with sample；
2) learn linear projection matrix, the mapping between building low-dimensional characterization and corresponding label space.
Further, described to pass through l2,1Norm constraint loss function and regularization term specifically:
Wherein, α is tradeoff parameter；A1 ... Am is factor matrix；Z is characteristic present matrix；||·||FIt indicates Frobenius norm；Indicate core tensor；I1×I2×…×IMIt is the size of sample space.
Wherein, the complete objective function specifically:
Wherein, χ is M rank target tensor；β, λ, η are that tradeoff parameter is tradeoff parameter；F is binary label matrix；H is Slack variable matrix.
The beneficial effect of the technical scheme provided by the present invention is that:
1, the present invention proposes joint isomery tensor resolution and differentiates robust features selection, and institute's climbing form type passes through isomery tensor point The intrinsic character representation of solution study more low-rank is obtained the feature for having more identification by differentiation robust features selection, improves and divide Class accuracy；
2, in order to better solve the heterogeneity being embedded in tensor data, object of the present invention is to minimize nuclear norm, and Apply low-rank constraint, and the factor matrix application orthogonalization behaviour to generate for the matrix of the last one mode of tensor resolution generation Make, and then the classifier optimized；
3, in order to preferably instruct decomposable process and improve the discriminating power of algorithm, the present invention strengthen about loss function and The joint " 2,1- norm minimum " of regularization；
4, the present invention is based on isomery tensor resolutions to propose a kind of efficient image classification, merges low-rank and orthogonal thought, Image classification accuracy is significantly improved while reducing calculation amount.
Detailed description of the invention
Fig. 1 is a kind of flow chart of image classification method based on isomery tensor resolution.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, embodiment of the present invention is made below further Ground detailed description.
Tensor is a multidimensional variable array, can be for multichannel array provides the simple of structure attribute and naturally indicates. For example, grey video can be modeled as three one point of the rank tensor sums indexed by two spaces variable in computer vision Not Dui Yingyu width, highly with the time variable of time；In electroencephalogram, each test can there are two types of channel and times with apparatus " channel × time × test " matrix of mode characterizes, and test of many times forms three rank tensors, i.e.,；In social tagging system In, user, the relationship triple between theme and project can also be by three rank tensor representations, i.e., " user × theme × project ". In fact, tensor be vector sum matrix high-order it is extensive, for example, scalar is zeroth order tensor, vector is single order tensor, and matrix is two Rank tensor.
In the modeling and analysis of multidimensional data, due to the strong coherence in each possible hyperspace with α, in structure It inevitably include bulk redundancy in the tensor data made.Therefore, a critical issue is to reduce this redundancy to obtain more More intrinsic representation, with the application for real world.Tensor resolution occurs as promising data mining technology, wide It is general to be used to reduce the inherent low dimensional structures being embedded in this redundancy and heuristic data.The most important motivation of tensor resolution is from a large amount of Compact expression is generated in redundant data.In various tensor resolution models, Tucker decomposes [1] and CANDECOMP/ It is two most popular frames that PARAFAC (CP), which decomposes [2], they are formed naturally two different definitions of tensor grade, i.e., Polyteny grade and CP order.It is low-dimensional core tensor multiplied by the group factor square along different mode that Tucker, which is decomposed tensor resolution, Battle array.CP decomposes the summation by tensor resolution for -1 tensor of order.In some cases, CP decomposition is considered as having super diagonal core The Tucker of the specific type of amount is decomposed.
Embodiment 1
The embodiment of the invention provides a kind of image classification methods based on isomery tensor resolution, referring to Fig. 1, this method packet Include following steps:
101: original tensor is constructed according to ORL data set；
Wherein, ORL data set be univ cambridge uk establish face image database, by those skilled in the art public affairs Know, the embodiment of the present invention does not repeat them here this.
102: being factor matrix corresponding with each rank of tensor and sample characteristics characterization matrix by original tensor resolution；
103: orthogonality constraint by column is added on factor matrix；
104: low-rank constraint being applied to sample characteristics characterization matrix Z, is obtained by identifying the global low-rank structure of sample Lowest rank indicates；
105: passing through l2,1Norm constraint loss function and regularization term realize the target for differentiating robust features selection；
106: obtaining complete object function；
107: iteration being optimized to objective function with alternating direction multiplier, optimal solution is obtained, further obtains classifier；
108: the test set picture in ORL being inputted in above-mentioned trained classifier, image classification is completed.
In conclusion the embodiment of the present invention proposes a kind of image classification method based on isomery tensor resolution, with higher Accuracy solve the problems, such as image classification.
Embodiment 2
The scheme in embodiment 1 is further introduced below with reference to specific calculation formula, example, it is as detailed below Description
1, problem definition
Assuming that providing a M rank target tensorTraining set, wherein N is sample number Amount, I1×I2×…×IMIt is the size of sample space, R is real number sample space.N number of target tensor is constituted jointly into (a M + 1) tensor of rankMeanwhile by F=[f1,f2,…,fN]T∈{0,1}N×CIt is expressed as binary label square Battle array, wherein C is the quantity of class, fiIt is i-th of sample χiLabel vector.And if only if χiBelong to jth class (j=1, 2, C) when, element fi,jIt is 1, remaining situation is 0.
The target of mentioned algorithm is to learn jointly:
1) to the preceding M rank of tensor, learn M orthogonal factor matrixsRealize original The low-dimensional of amount is embedded in, characteristic present matrix Z ∈ R corresponding with sample to the last first-order learning of tensorN×D；
Wherein, AiFor orthogonal factor matrix, D is characterized dimension.
2) learn linear projection matrix W ∈ RD×C, the mapping between low-dimensional characterization and corresponding label space is constructed, wherein Ji< IiWith D < N.
2, the isomery tensor resolution of low-rank regularization:
Tucker decomposes the one group factor matrix of relatively low-rank core tensor sum by original tensor resolution for each mode.Connection The Tucker decomposition of tensor χ can be expressed as following form:
Wherein, ×iIndicate the modular multiplication along the i-th rank of tensor,Indicate core tensor.I.e. formula (1) is above-mentioned reality Apply the original tensor and label matrix corresponding with sample in example 1 in step 101.
Latent factor matrix A1,A2,…,AMIt can be obtained by minimizing reconstruction loss function as follows with Z, i.e., it is above-mentioned Factor matrix and sample characteristics characterization matrix in embodiment 1 in step 102:
Wherein, factor matrix is A1 ... Am, and characteristic present matrix is Z, | | | |FIndicate Frobenius norm in general, nothing It is not unique for constraining the solution that Tucker is decomposed.A kind of direct strategy for overcoming the problems, such as this is to apply volume to factor matrix Outer constraint allows to obtain unique and physically significant explanation.
By adding orthogonality constraint by column on factor matrix, following formula is derived:
Orthogonality is applied to factor matrix to attempt to avoid any scaling of intermediate representation and retain as much as possible to believe Breath.
Therefore tensor can be rewritten and rebuild loss function
Since the first item in equation (4) is constant, so minimization problem is converted into consequent minimum:
Wherein,For objective function.
In fact,It is considered potential projection tensorIt is compressed into the humble dictionary tensor for indicating Z. The inspiration of exceptional value [3] or missing content [4] is solved by the low-rank method being most recently successfully, Z is applied in consideration of the embodiment of the present invention Low-rank is added to constrain, so that by identifying that the global low-rank structure of sample indicates to obtain lowest rank.As order minimization problem Common practice, the embodiment of the present invention are replaced low-rank constraint function, i.e. the sum of singular value of Z with nuclear norm ‖ Z ‖ *, are obtained following Convex optimization problem.Therefore, the objective function comprising the constraint of orthogonal and low-rank specifically:
Wherein, α is tradeoff parameter.
Wherein, above-mentioned formula (6) is that the lowest rank in embodiment 1 in step 104 indicates.
3、l2,1Norm regularization differentiates robust features selection:
Wherein, l2,1Norm regularization is widely used in due to its efficiency and robustness when selection differentiates feature Various classification tasks.Least square regression is one of the most popular method connected between modeling input element and output label.For Simplified model (6), solve Tag Estimation using least square regression, as follows:
Wherein, W ∈ RD×CLinear projection matrix, β and λ are tradeoff parameters, for control loss function and regularization term it Between balance.
In order to enhance the robustness of model, robust loss function is used:
Therefore, above-mentioned equation can convert are as follows:
Wherein,wijFor the element that the i-th row jth in matrix W arranges, above-mentioned formula (9) is step Loss function and regularization term in 105.
It is expected that the stringent binary constraint in H can be loosened in soft-constraint, so that the model proposed has oneself bigger Label is adapted to by spending and avoids overfitting problem.For this purpose, replacing original tag matrix F using slack variable matrix H.Cause This, improved discriminate learning model is restated as follows:
Wherein, η is tradeoff parameter.
By combining target function type (10) and formula (6) in equation, the complete objective function (HTDRC) got It is as follows:
44, Optimization Solution
Objective function optimization in equation (11) can be solved by alternating direction multipliers method (ADMM), and this method will answer Miscellaneous problem is divided into subproblem, wherein each subproblem is easier to be handled with iterative process.It is firstly introduced into two auxiliary variable P And E, keep objective function (11) relaxation as follows:
Then, Lagrange's multiplier Y is introduced1And Y2, and it is fitted Augmented Lagrangian Functions againIt is as follows:
Wherein, μ indicates positve term punishment parameter, and formula (13) is the classifier further obtained by optimal solution.
In order to preferably explain the process, introducing variable k simultaneously willZk,Ek,Wk,Pk,FkAnd μkIt is defined as in kth The variable updated in secondary iteration.
Wherein, alternating direction multipliers method (ADMM)[5]For technology well known in the art, specific iterative process is this field Well known to technical staff, the embodiment of the present invention is not repeated them here.
Embodiment 3
Feasibility verifying is carried out to the scheme in Examples 1 and 2 below with reference to specific implementation process, result of implementation, in detail See below description:
This method is compared with state-of-the-art sorting algorithm, to verify conjunctive use core tensor sum l2,1Norm it is different Structure tensor resolution carries out the validity of image classification, the data set used are as follows: ORL.
In embodiment 1, the image based on isomery tensor resolution of proposition is had evaluated on widely used data set ORL Classification.This data set is widely used by various image classification methods.Classification performance is assessed using nearest neighbour classification device.
ORL data set includes 40 different themes, and each theme has 10 images.Image in same subject be It is shot under very big difference, changes illumination, facial expression (opening/closing eye, smile/unhappy) and face detail (wear glasses/ It does not wear glasses).In an experiment, facial image is tailored to 32 × 32 pixels.Gray value is used as input element.Use nearest-neighbors It is repeated this five times as classifier, and reports average result.For each class, { 2,3,4,5 } a image therein is randomly choosed As training sample, remaining image is as test sample.
Aspect has evaluated this hair in convergence, parametric sensitivity, characteristic dimension and compared with several most advanced methods The benchmark test algorithm that bright embodiment proposes.Mentioned algorithm is used for initial factor matrix A1,A2,A3,…,AMAnd Z.Rule of thumb α=10 and λ=0.1 are set by the auto-adaptive parameter of four data sets.By trellis search method setting proposed based on Tradeoff parameter beta and η in the image classification method of isomery tensor resolution.
In embodiments of the present invention, it adjusts and records the optimal classification accuracy of every kind of method to carry out fair comparison.Table 1 Show the recognition accuracy obtained by the distinct methods of the training sample of all size on ORL data set.It can from table To find out, the embodiment of the present invention proposes the small sample image classification method (HTDRC) based on isomery tensor resolution in most of feelings Other image classification algorithms are better than under condition.
Classification accuracy of 1 distinct methods of table on ORL data set
The method of comparison includes: linear regression classification linear regression classification (LRC), sparse Presentation class sparse representation classification (SRC) is classified based on constant canonical regression is turned round Rotational invariant norm-based regression for classification (RRC), optimization feature choosing Select classification optimal feature selection classification (OFSC), Higher-order Singular value decomposition high order Singular value decomposition (HOSVD), synchronous tensor resolution and completion simultaneous tensor Decomposition and completion (STDC), exceptional value robust tensor principal component analysis outlier-robust tensor PCA(OR-TPCA)。
The embodiment of the present invention proposes one and nuclear norm and l is used in combination2,1The image of specification regularization isomery tensor resolution Sorting algorithm.In order to better solve the isomerism being embedded in tensor data, tensor expansion is minimised as most based on nuclear norm The lowest rank representing matrix of latter mode finds one group of orthogonal factor matrix.
In order to preferably instruct tensor resolution process and improve the discriminating power of algorithm, the differentiation for introducing a robust is special Selected section is levied, emphasizes the joint " l about loss function and regularization2,1Norm minimum ".Meanwhile potential label matrix This part is further relaxed, keeps frame more stable and flexible.Result of implementation shows this method not only in a small amount of iteration Reach convergence, and also achieves more accurate image classification result compared with state-of-the-art method.
Bibliography
[1]K.T.G and B.B.W,“Tensor decompositions and applications,”SIAM review,vol.51,no.3, pp.455–500,2009.
[2]K.H.AL,“Towards a standardized notation and terminology in multiway analysis,”Journal of chemometrics,vol.14,no.3,pp.105–122,2000.
[3]C.JiaandY.Fu,“Low-rank tensor subspace learning for rgb- dactionrecognition,”IEEE Transactions on Image Processing,vol.25,no.10, pp.4641–4652,2016.
[4]P.Rai,Y.Wang,S.Guo,G.Chen,D.Dunson,and L.Carin,“Scalable bayesian low-rank decomposition of incomplete multiway tensors,”in Proceedinds of International Conference on Machine Learning,2014,pp.1800–1808.
[5]W.Shi,Q.Ling,K.Yuan,G.Wu and W.Yin,"On the Linear Convergence of the ADMM in Decentralized Consensus Optimization,"in IEEE Transactions on Signal Processing,vol.62, no.7,pp.1750-1761,April1,2014.
The embodiment of the present invention to the model of each device in addition to doing specified otherwise, the model of other devices with no restrictions, As long as the device of above-mentioned function can be completed.
It will be appreciated by those skilled in the art that attached drawing is the schematic diagram of a preferred embodiment, the embodiments of the present invention Serial number is for illustration only, does not represent the advantages or disadvantages of the embodiments.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.

Claims (4)

1. a kind of image classification method based on isomery tensor resolution, which is characterized in that the described method comprises the following steps:
Original tensor is constructed according to ORL data set；It is factor matrix corresponding with each rank of tensor and sample by original tensor resolution Characteristic present matrix；
Orthogonality constraint by column is added on factor matrix；Low-rank constraint is applied to sample characteristics characterization matrix, by identifying sample Global low-rank structure come obtain lowest rank indicate；
Pass through l2,1Norm constraint loss function and regularization term realize the target for differentiating robust features selection；
Obtain complete objective function；Iteration is optimized to objective function with alternating direction multiplier, obtains optimal solution, further Obtain classifier；
Test set picture in ORL is inputted in above-mentioned trained classifier, image classification is completed.
2. a kind of image classification method based on isomery tensor resolution according to claim 1, which is characterized in that described to incite somebody to action Original tensor resolution is factor matrix corresponding with each rank of tensor and sample characteristics characterization matrix specifically:
1) to the preceding M rank of tensor, learn the low-dimensional insertion that M orthogonal factor matrixs realize original tensor, to the last of tensor First-order learning characteristic present matrix corresponding with sample；
2) learn linear projection matrix, the mapping between building low-dimensional characterization and corresponding label space.
3. a kind of image classification method based on isomery tensor resolution according to claim 1, which is characterized in that described logical Cross l2,1Norm constraint loss function and regularization term specifically:
Wherein, α is tradeoff parameter；A1 ... Am is factor matrix；Z is characteristic present matrix；||·||FIndicate Frobenius model Number；Indicate core tensor；I1×I2×…×IMIt is the size of sample space.
4. a kind of image classification method based on isomery tensor resolution according to claim 1, which is characterized in that described complete Whole objective function specifically:
Wherein,For M rank target tensor；β, λ, η are that tradeoff parameter is tradeoff parameter；F is binary label matrix；H is relaxation Matrix of variables.
CN201910453844.4A 2019-05-28 2019-05-28 Image classification method based on heterogeneous tensor decomposition Active CN110222213B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910453844.4A CN110222213B (en) 2019-05-28 2019-05-28 Image classification method based on heterogeneous tensor decomposition

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910453844.4A CN110222213B (en) 2019-05-28 2019-05-28 Image classification method based on heterogeneous tensor decomposition

Publications (2)

Publication Number Publication Date
CN110222213A true CN110222213A (en) 2019-09-10
CN110222213B CN110222213B (en) 2021-07-16

Family

ID=67818419

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910453844.4A Active CN110222213B (en) 2019-05-28 2019-05-28 Image classification method based on heterogeneous tensor decomposition

Country Status (1)

Country Link
CN (1) CN110222213B (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110717538A (en) * 2019-10-08 2020-01-21 广东工业大学 Color picture clustering method based on non-negative tensor ring
CN110766066A (en) * 2019-10-18 2020-02-07 天津理工大学 FNN-based tensor heterogeneous integrated internet of vehicles missing data estimation method
CN111274525A (en) * 2020-01-19 2020-06-12 东南大学 Tensor data recovery method based on multi-linear augmented Lagrange multiplier method
CN112085704A (en) * 2020-08-07 2020-12-15 深圳先进技术研究院 Medical image classification method and device, terminal equipment and storage medium
CN112329604A (en) * 2020-11-03 2021-02-05 浙江大学 Multi-modal emotion analysis method based on multi-dimensional low-rank decomposition
CN112600697A (en) * 2020-12-07 2021-04-02 中山大学 QoS prediction method and system based on federal learning, client and server
CN113349795A (en) * 2021-06-15 2021-09-07 杭州电子科技大学 Depression electroencephalogram analysis method based on sparse low-rank tensor decomposition
WO2023065525A1 (en) * 2021-10-22 2023-04-27 西安闻泰信息技术有限公司 Object feature matrix determination method and apparatus, device, and storage medium
CN117784940A (en) * 2024-02-23 2024-03-29 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) High-dimensional tensor analysis method and device aiming at Chinese character stroke writing imagination

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105608478A (en) * 2016-03-30 2016-05-25 苏州大学 Combined method and system for extracting and classifying features of images
CN107392107A (en) * 2017-06-24 2017-11-24 天津大学 A kind of face feature extraction method based on isomery tensor resolution
CN107563444A (en) * 2017-09-05 2018-01-09 浙江大学 A kind of zero sample image sorting technique and system
CN109522956A (en) * 2018-11-16 2019-03-26 哈尔滨理工大学 A kind of low-rank differentiation proper subspace learning method

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105608478A (en) * 2016-03-30 2016-05-25 苏州大学 Combined method and system for extracting and classifying features of images
CN107392107A (en) * 2017-06-24 2017-11-24 天津大学 A kind of face feature extraction method based on isomery tensor resolution
CN107563444A (en) * 2017-09-05 2018-01-09 浙江大学 A kind of zero sample image sorting technique and system
CN109522956A (en) * 2018-11-16 2019-03-26 哈尔滨理工大学 A kind of low-rank differentiation proper subspace learning method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
JING ZHANG等: "Low-Rank Regularized Heterogeneous Tensor Decomposition for Subspace Clustering", 《IEEE》 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110717538A (en) * 2019-10-08 2020-01-21 广东工业大学 Color picture clustering method based on non-negative tensor ring
CN110717538B (en) * 2019-10-08 2022-06-24 广东工业大学 Color picture clustering method based on non-negative tensor ring
CN110766066A (en) * 2019-10-18 2020-02-07 天津理工大学 FNN-based tensor heterogeneous integrated internet of vehicles missing data estimation method
CN111274525A (en) * 2020-01-19 2020-06-12 东南大学 Tensor data recovery method based on multi-linear augmented Lagrange multiplier method
CN112085704A (en) * 2020-08-07 2020-12-15 深圳先进技术研究院 Medical image classification method and device, terminal equipment and storage medium
CN112329604A (en) * 2020-11-03 2021-02-05 浙江大学 Multi-modal emotion analysis method based on multi-dimensional low-rank decomposition
CN112600697A (en) * 2020-12-07 2021-04-02 中山大学 QoS prediction method and system based on federal learning, client and server
CN113349795A (en) * 2021-06-15 2021-09-07 杭州电子科技大学 Depression electroencephalogram analysis method based on sparse low-rank tensor decomposition
CN113349795B (en) * 2021-06-15 2022-04-08 杭州电子科技大学 Depression electroencephalogram analysis method based on sparse low-rank tensor decomposition
WO2023065525A1 (en) * 2021-10-22 2023-04-27 西安闻泰信息技术有限公司 Object feature matrix determination method and apparatus, device, and storage medium
CN117784940A (en) * 2024-02-23 2024-03-29 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) High-dimensional tensor analysis method and device aiming at Chinese character stroke writing imagination

Also Published As

Publication number Publication date
CN110222213B (en) 2021-07-16

Similar Documents

Publication Publication Date Title
CN110222213A (en) A kind of image classification method based on isomery tensor resolution
Xie et al. Hyper-Laplacian regularized multilinear multiview self-representations for clustering and semisupervised learning
Zhang et al. Joint low-rank and sparse principal feature coding for enhanced robust representation and visual classification
Nie et al. Towards Robust Discriminative Projections Learning via Non-Greedy $\ell _ {2, 1} $ ℓ 2, 1-Norm MinMax
Zhuang et al. Label information guided graph construction for semi-supervised learning
Yang et al. Supervised translation-invariant sparse coding
Li et al. Linestofacephoto: Face photo generation from lines with conditional self-attention generative adversarial networks
Yan et al. Graph embedding and extensions: A general framework for dimensionality reduction
Kapoor et al. Active learning with gaussian processes for object categorization
Chen et al. Ambiguously labeled learning using dictionaries
Lin et al. Construction of dependent Dirichlet processes based on Poisson processes
Xing et al. Towards robust and accurate multi-view and partially-occluded face alignment
Yang et al. Incomplete-data oriented multiview dimension reduction via sparse low-rank representation
Fang et al. Graph-based learning via auto-grouped sparse regularization and kernelized extension
Wen et al. Discriminative dictionary learning with two-level low rank and group sparse decomposition for image classification
Du et al. Low-rank graph preserving discriminative dictionary learning for image recognition
Jia et al. Semi-supervised cross-modality action recognition by latent tensor transfer learning
Jia et al. Adaptive neighborhood propagation by joint L2, 1-norm regularized sparse coding for representation and classification
Nguyen et al. Discriminative low-rank dictionary learning for face recognition
CN109063555B (en) Multi-pose face recognition method based on low-rank decomposition and sparse representation residual error comparison
Yuan et al. Generative modeling of infinite occluded objects for compositional scene representation
Wang et al. Generative partial multi-view clustering
Du et al. Discriminative low-rank graph preserving dictionary learning with Schatten-p quasi-norm regularization for image recognition
Zhang et al. Multi-view dimensionality reduction via canonical random correlation analysis
Wang et al. Locality-constrained discriminative learning and coding

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
CB03 Change of inventor or designer information
CB03 Change of inventor or designer information

Inventor after: Jing Peiguang

Inventor after: Cui Kai

Inventor after: Wu Yuting

Inventor after: Su Yuting

Inventor before: Jing Peiguang

Inventor before: Wu Yuting

Inventor before: Su Yuting

GR01 Patent grant
GR01 Patent grant