CN105335756B - A kind of image classification method and image classification system based on Robust Learning model - Google Patents

A kind of image classification method and image classification system based on Robust Learning model Download PDF

Info

Publication number
CN105335756B
CN105335756B CN201510726581.1A CN201510726581A CN105335756B CN 105335756 B CN105335756 B CN 105335756B CN 201510726581 A CN201510726581 A CN 201510726581A CN 105335756 B CN105335756 B CN 105335756B
Authority
CN
China
Prior art keywords
matrix
sample
training
label
samples
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.)
Active
Application number
CN201510726581.1A
Other languages
Chinese (zh)
Other versions
CN105335756A (en
Inventor
张召
江威明
李凡长
张莉
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.)
Suzhou University
Original Assignee
Suzhou 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 Suzhou University filed Critical Suzhou University
Priority to CN201510726581.1A priority Critical patent/CN105335756B/en
Publication of CN105335756A publication Critical patent/CN105335756A/en
Application granted granted Critical
Publication of CN105335756B publication Critical patent/CN105335756B/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
    • G06F18/2413—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

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of image classification method and image classification system based on Robust Learning model, it include: to be initialized to training set, initial category label matrix is obtained, the training sample in training set includes its known classification and demarcates the sample for having class label corresponding with its classification and its unknown classification and the sample for not demarcating class label；The building method that power is defined and reconstructed based on neighbour handles training sample, according to the similitude between sample, constructs reconstruction coefficients matrix, and carry out symmetrization, normalized；The soft label without calibration sample is determined using reconstruction coefficients matrix and initial category label matrix, and l is carried out to the soft label of training sample by the way of iteration2,1Norm regularization obtains projection matrix and soft label matrix；Sample to be tested is mapped using projection matrix, obtains its soft label；Sample to be tested is its unknown classification and the sample for not demarcating classification.Improve classification accuracy.

Description

Image classification method and image classification system based on robust learning model
Technical Field
The invention relates to the technical field of pattern recognition and data mining, in particular to an image classification method and an image classification system based on a robust learning model.
Background
With the continuous development of computer technology and intelligence, image classification technology has developed into one of the most important research subjects in the fields of data mining, machine learning, and the like. The image classification technology is mainly used for judging the classification of unknown data, and has great significance in the fields of medical data analysis, text, webpage, credit card rating and the like, so that great social and economic benefits can be brought by putting the accurate classification technology into use. Multiple studies prove that the performance of the supervised learning method is obviously superior to that of the unsupervised learning method, but in the real world, the supervised data used for the supervised learning method is often difficult to obtain, and the class information of the supervised data obtained in a manual calibration mode consumes a large amount of time and manpower, so that the practicability of the supervised learning method is greatly reduced. Therefore, semi-supervised learning based on similarity graph construction has been developed as one of practical and popular classification tools due to its practicality and classification accuracy, and semi-supervised learning is mainly performed by calibrating the class of a small amount of data in each class of a large amount of data, and then spreading the class of data to unknown classes through similarity graphs, thereby predicting the class of data of the unknown classes.
In recent years, a label propagation method based on a label propagation theory in semi-supervised learning is one of typical representatives of semi-supervised learning due to the advantages of simplicity, effectiveness and rapidness. The label propagation method is used for propagating the category information of supervised labeled samples (i.e. samples with known categories) to unlabeled samples (i.e. samples with unknown categories) by learning the similarity between the samples, and then estimating the category information of the unlabeled samples. At present, most label propagation methods adopt soft labels as classification results of unlabeled samples, however, the inventor finds that in the label propagation method adopting the soft labels as the classification results of the unlabeled samples, original spaces corresponding to the labeled samples often contain mixed signals, and the mixed signals can cause adverse effects on the classification of the unlabeled samples, thereby causing the inaccurate classification results of the unlabeled samples.
In summary, there is a problem in the prior art that the accuracy of the classification result of the unlabeled sample is low due to the influence of the mixed signal of the original space.
Disclosure of Invention
The invention aims to provide an image classification method and an image classification system based on a robust learning model, and aims to solve the problem that the classification result accuracy of a label-free sample is low due to the influence of a mixed signal of an original space in the prior art.
In order to achieve the above purpose, the invention provides the following technical scheme:
an image classification method based on a robust learning model comprises the following steps:
initializing a pre-acquired training set to obtain an initial class label matrix, wherein the training set comprises a preset amount of training samples, and the training samples comprise samples of which the classes are known and are marked with class labels corresponding to the classes, and samples of which the classes are unknown and are not marked with class labels;
processing the training sample based on a construction method of neighbor definition and reconstruction weight to obtain a similarity measurement matrix corresponding to the training sample, and presetting the similarity measurement matrix to obtain a reconstruction coefficient matrix;
determining the soft label of the training sample without calibrating the class label through an effective balance manifold smooth item and a label fitting item based on the reconstruction coefficient matrix and the initial class label matrix, and carrying out l on the soft label of the training sample in an iteration mode2，1Normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample;
mapping a sample to be detected by using the projection matrix to obtain a soft label of the sample to be detected; the sample to be detected is a sample of which the category is unknown and the category label is not calibrated.
Preferably, the constructing method based on neighbor definition and reconstruction weight processes the training sample to obtain a similarity metric matrix corresponding to the training sample, and includes:
processing each training sample by using a K nearest neighbor algorithm to obtain K nearest neighbor samples of each training sample, wherein K is a positive integer;
and acquiring a similarity measurement matrix corresponding to the training sample by using the training sample and K nearest neighbor samples of each training sample by adopting an LLE-reconstruction weight construction method.
Preferably, the presetting the similarity metric matrix to obtain a reconstruction coefficient matrix includes:
and carrying out normalization processing and symmetry processing on the similarity measurement matrix to obtain the reconstruction coefficient matrix.
Preferably, the mapping the sample to be tested by using the projection matrix to obtain the soft label of the sample to be tested includes:
obtaining a target to be testedSample xnew；
By PTxnewThe sample x to be tested is processednewEmbedding the prediction vector into a projection matrix P to obtain a prediction vector;
determining that the soft label corresponding to the element with the maximum probability in the prediction vector is the sample x to be detectednewThe soft tag of (1).
Preferably, the initializing a pre-acquired training set to obtain an initial class label matrix includes:
obtaining an initialization matrix with Y0＝[y1,y2,…,yl+u]Representing the training setIt is shown that, among others,representing the matrix space of n times l + u, XLFor samples labeled with class labels, XL＝[x1,x2,...,xl]，XUFor samples not labeled with class labels, XU＝[xl+1,xl+2,...,xl+u]N, l and u are positive integers;
determining any y in the initialization matrixiIs a column vector, any column vector corresponds to the ith training sample x in the training seti，i＝1,2,...,l+u；
Training sample x with any calibration class labeljIf the training sample xjBelongs to the ith category, then y is determinedi，jIf the training sample x is 1jIf it does not belong to the i-th category, y is determinedi，j0; for any training sample x not marked with class labeljDetermining yi，jAn initial class label matrix Y is obtained, where j is 1, 2.
Preferably, the constructing method based on neighbor definition and reconstruction weight processes the training sample to obtain a similarity metric matrix corresponding to the training sample, and includes:
obtaining the similarity metric matrix using the following formula:
wherein,representing a training sample xiK nearest neighbor samples, q corresponds to K,for the current training sample xiThe reconstructed coefficient vector of (2) characterizing its neighborsOn reconstruction of training sample xiThe degree of contribution of the time-varying,representing the corresponding similarity reconstruction coefficient matrix.
Preferably, the soft label of the training sample is subjected to l in an iterative manner2，1And normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample, wherein the soft label matrix comprises the following steps:
the projection matrix and the soft label matrix are obtained by the following formulas:
subj fi≥0,eTfi＝1 for i＝1,2,...,l+u
wherein F represents a soft label matrix, P represents a projection matrix, W represents a reconstruction coefficient matrix, D represents a diagonal matrix, and D representsii＝∑jWi,j，fiTo representThe ith row vector of (2), XTP-FTRepresenting a regression residue term that is used to measure XTP and Soft tag FTThe degree of difference of (a); mu.siα, γ is the corresponding trade-off parameter;
or, the projection matrix and the soft label matrix are obtained by the following formulas:
subjF≥0,eTF＝eT
wherein, omega ═ I-WT-W+WTW and I are unit matrixes, and only diagonal elements of the unit matrixes are nonzero and are 1; u denotes a diagonal matrix with the ith diagonal element μiFor labeled training samples, μiSet to + ∞, for training samples not labeled with class labels, μiIs set to 0.
An image classification system comprising:
the system comprises a preprocessing module, a classification module and a classification module, wherein the preprocessing module is used for initializing a training set acquired in advance to obtain an initial classification label matrix, the training set comprises a preset amount of training samples, and the training samples comprise samples of which the classification is known and the classification labels corresponding to the classification are calibrated and samples of which the classification is unknown and the classification labels are not calibrated; the construction method is used for processing the training sample based on neighbor definition and reconstruction weight to obtain a similar measurement matrix corresponding to the training sample, and the similar measurement matrix is subjected to preset processing to obtain a reconstruction coefficient matrix;
a training module, configured to determine a soft label of a training sample without a class label by effectively balancing a manifold smoothing term and a label fitting term based on the reconstruction coefficient matrix and the initial class label matrix, and perform l on the soft label of the training sample in an iterative manner2，1Normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample;
the prediction module is used for mapping a sample to be detected by using the projection matrix to obtain a soft label of the sample to be detected; and the sample to be detected is a sample without a calibrated class label.
Preferably, the preprocessing module comprises:
the K nearest neighbor unit is used for processing each training sample by utilizing a K nearest neighbor algorithm to obtain K nearest neighbor samples of each training sample, and K is a positive integer;
and the LLE-reconstruction weight unit is used for acquiring a similarity measurement matrix corresponding to the training samples by using the K nearest neighbor samples of each training sample by adopting an LLE-reconstruction weight construction method.
Preferably, the preprocessing module comprises:
and the preset processing unit is used for carrying out normalization processing and symmetry processing on the similarity measurement matrix to obtain the reconstruction coefficient matrix.
The invention provides an image classification method and an image classification system based on a robust learning model, which comprise the following steps: initializing a pre-acquired training set to obtain an initial class label matrix, wherein the training set comprises a preset amount of training samples, and the training samples comprise samples of which the classes are known and are marked with class labels corresponding to the classes, and samples of which the classes are unknown and are not marked with class labels; processing the training sample based on a construction method of neighbor definition and reconstruction weight to obtain a similarity measurement matrix corresponding to the training sample, and presetting the similarity measurement matrix to obtain a reconstruction coefficient matrix; determining the soft label of the training sample without calibrating the class label through an effective balance manifold smooth item and a label fitting item based on the reconstruction coefficient matrix and the initial class label matrix, and carrying out l on the soft label of the training sample in an iteration mode2，1Normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample; mapping a sample to be detected by using the projection matrix to obtain a soft label of the sample to be detected; the sample to be detected is a sample of which the category is unknown and the category label is not calibrated.
Compared with the prior art, the method comprises the steps of firstly utilizing training samples marked with class labels and training samples not marked with class labels to jointly construct a similarity measurement matrix, further obtaining a reconstruction coefficient matrix, initializing to obtain an initial class label matrix, utilizing the initial class label matrix and the reconstruction coefficient matrix to determine soft labels of the training samples not marked with the class labels, and carrying out l on the soft labels of the training samples2,1Norm regularization effectively reduces the influence of mixed signals in an original space formed by the soft labels on the accuracy of the classification result; by being based on l2,1The robustness of regression residual measurement between the embedded features and the soft labels is effectively enhanced by the norm regularization measurement; by calculating on the basis of l2,1And the mapping of the metric of norm regularization enhances the descriptive property of the extracted features. In summary, by introducing l2,1Norm regularization technology effectively promotes the system to train samplesThe robustness of mixed signals such as medium noise, heterogeneous data and the like avoids the occurrence of the condition that the accuracy of the classification result of the label-free sample is lower due to the influence of the mixed signal of the original space in the background technology, namely, the influence of the mixed signal of the original space is reduced, the accuracy of the classification result of the sample to be detected is improved, and the classification performance is enhanced. In addition, this application is the model of promoting outside the sample, and the conclusion and the prediction of the outer data of sample need not to introduce extra reconstruction process, and the expansibility can be good for the high efficiency completion sample of embedding.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.
Fig. 1 is a flowchart of an image classification method based on a robust learning model according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of an image classification system according to an embodiment of the present invention;
fig. 3 is a schematic diagram of label prediction for obtaining classes of human face image test samples in an image classification method and an image classification system based on a robust learning model according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1, a flowchart of an image classification method based on a robust learning model according to an embodiment of the present invention is shown, which may include the following steps:
s11: initializing a pre-acquired training set to obtain an initial class label matrix, wherein the training set comprises a preset amount of training samples, and the training samples comprise samples of which the classes are known and are marked with class labels corresponding to the classes and samples of which the classes are unknown and are not marked with the class labels.
The number of training samples in the training set can be determined according to actual needs. The class label of the training sample corresponds to the class of the training sample, the class label of the training sample with unknown class is not calibrated, and the class label of the training sample with known class is calibrated by using the class label corresponding to the class.
S12: processing the training sample based on the construction method of neighbor definition and reconstruction weight to obtain a similar measurement matrix corresponding to the training sample, and presetting the similar measurement matrix to obtain a reconstruction coefficient matrix.
It should be noted that, step S11 and step S12 may be completed simultaneously, or S11 may be completed before S12 is completed, or S12 may be completed before S11 is completed, which may be determined according to actual needs.
S13: determining the soft label of the training sample without calibrating the class label through effectively balancing the manifold smooth item and the label fitting item based on the reconstruction coefficient matrix and the initial class label matrix, and carrying out l on the soft label of the training sample in an iterative mode2，1And normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample.
Specifically, the manifold smoothing item is determined based on the reconstruction coefficient matrix, the label fitting item is determined based on the initial category label matrix, and then the soft label corresponding to the manifold smoothing item and the label fitting item is determined.
The soft label of each training sample is also associated with its class, and the soft label matrix is determined based on the soft labels of each training sample. The classes of all training samples can be obtained through the obtained soft label matrix.
And l2，1Norm regularization is an important means in machine learning, and is actually an optimal process for solving a cost function in the learning process of a support vector machine, and the norm is added into the cost function, so that the learning result meets sparseness (sparsity), and the characteristics can be conveniently extracted by a human.
S14: mapping a sample to be detected by using a projection matrix to obtain a soft label of the sample to be detected; the sample to be detected is a sample of which the category is unknown and the category label is not calibrated.
And mapping the sample to be detected by using the projection matrix to obtain a soft label of the sample to be detected, and further determining the category of the sample to be detected.
The method comprises the steps of firstly, utilizing training samples marked with class labels and training samples not marked with class labels to jointly construct a similarity measurement matrix, further obtaining a reconstruction coefficient matrix, initializing to obtain an initial class label matrix, utilizing the initial class label matrix and the reconstruction coefficient matrix to determine soft labels of the training samples not marked with the class labels, and carrying out l on the soft labels of the training samples2,1Norm regularization effectively reduces the influence of mixed signals in an original space formed by the soft labels on the accuracy of the classification result; by being based on l2,1The robustness of regression residual measurement between the embedded features and the soft labels is effectively enhanced by the norm regularization measurement; by calculating on the basis of l2,1And the mapping of the metric of norm regularization enhances the descriptive property of the extracted features. In summary, by introducing l2,1Norm regularization technology effectively promotes system pairThe robustness of mixed signals such as noise, heterogeneous data and the like in the training samples avoids the occurrence of the condition that the accuracy of the classification result of the label-free sample is lower due to the influence of the mixed signal of the original space in the background technology, namely, the influence of the mixed signal of the original space is reduced, the accuracy of the classification result of the sample to be detected is improved, and the classification performance is enhanced. In addition, this application is the model of promoting outside the sample, and the conclusion and the prediction of the outer data of sample need not to introduce extra reconstruction process, and the expansibility can be good for the high efficiency completion sample of embedding.
In the image classification method based on the robust learning model provided in the above embodiment, initializing a training set acquired in advance to obtain an initial class label matrix may include:
obtaining an initialization matrix with Y0＝[y1,y2,…,yl+u]For presentation and training setIt is shown that, among others,representing the matrix space of n times l + u, XLFor samples labeled with class labels, XL＝[x1,x2,...,xl]，XUFor samples not labeled with class labels, XU＝[xl+1,xl+2,...,xl+u]N, l and u are positive integers;
determining any y in the initialization matrixiIs a column vector, any column vector corresponds to the ith training sample x in the training seti，i＝1,2,...,l+u；
Training sample x with any calibration class labeljIf the training sample xjBelongs to the ith category, then y is determinedi，jIf the training sample x is 1jIf it does not belong to the i-th category, y is determinedi，j0; for any training sample not marked with class labelxjDetermining yi，jAn initial class label matrix Y is obtained, where j is 1, 2.
Therefore, through the steps, the training samples are converted into the initial class label matrix capable of expressing the class of each training sample based on the class labels of each training sample, so that smooth implementation of the subsequent steps is guaranteed.
In the image classification method based on the robust learning model provided in the foregoing embodiment, the processing of the training sample based on the neighbor definition and the reconstruction weight construction method to obtain the similarity metric matrix corresponding to the training sample may include:
processing each training sample by using a K nearest neighbor algorithm to a training set containing c training samples to obtain K nearest neighbor samples of each training sample, wherein K is a positive integer;
and acquiring a similarity measurement matrix corresponding to the training samples by using the training samples and the K nearest neighbor samples of each training sample by adopting an LLE-reconstruction weight construction method.
The idea of the K nearest neighbor algorithm is as follows: if a sample belongs to a certain class in the K most similar samples in the feature space (i.e., the nearest neighbors in the feature space), then the sample also belongs to this class.
The construction method of LLE-reconstruction weight is an optimization method aiming at the nonlinear signal feature vector dimension, and the dimension optimization is not only simple reduction in quantity, but also mapping the signal of high-dimensional space to low-dimensional space under the condition of keeping the original data property unchanged, namely secondary extraction of feature value.
Specifically, processing the training sample based on the neighbor definition and reconstruction weight construction method to obtain a similarity metric matrix corresponding to the training sample may include:
the similarity metric matrix is obtained using the following formula:
wherein,representing a training sample xiK nearest neighbor samples, q corresponds to K,for the current training sample xiThe reconstructed coefficient vector of (2) characterizing its neighborsOn reconstruction of training sample xiThe degree of contribution of the time-varying,a similarity metric matrix is represented. And a similarity metric matrixObtained by repeating the above steps related to the K-nearest neighbor algorithm and the construction method of LLE-reconstruction weights for all training samples.
It should be noted that, the construction method of LLE-reconstruction weights is used to obtain the similarity metric matrix corresponding to the training samples by using the training samples and the K nearest neighbor samples of each training sample, specifically, the construction method of LLE-reconstruction weights is used to calculate and measure the similarity between vertices to construct the similarity metric matrix of the similar neighbor graph.
In the image classification method based on the robust learning model provided in the above embodiment, the pre-setting processing is performed on the similarity metric matrix to obtain a reconstruction coefficient matrix, which may include:
and carrying out normalization processing and symmetry processing on the similarity measurement matrix to obtain a reconstruction coefficient matrix.
And mapping the sample to be detected by using the projection matrix to obtain the soft label of the sample to be detected, which may include:
obtaining a sample x to be testednew；
By PTxnewTo-be-detected sample xnewEmbedding the prediction vector into a projection matrix P to obtain a prediction vector;
determining the soft label corresponding to the element with the maximum probability in the prediction vector as the sample x to be detectednewThe soft tag of (1).
The projection matrix is a mapping classifier, the sample to be detected is embedded into the mapping classifier, the prediction vector can be obtained, the prediction vector is a soft label matrix corresponding to the sample to be detected, the soft label of the position corresponding to the element with the maximum probability in the soft label matrix is determined to be the soft label of the sample to be detected, and the hard class label of the sample to be detected can be classified as argmaxi≤c(fnew)iWhereina prediction vector f representing a predictionnewThe ith element position.
Through the steps, all the possible classes to which the sample to be detected belongs can be obtained according to the projection matrix, and the class with the highest probability is used as the class of the sample to be detected, so that the accuracy of the classification result is ensured.
Assume a predicted soft label matrix, i.e., an initialization matrix of F ═ F1,f2,…fl+u]Wherein the training sample xjClass and each column fjMiddle maximum termfi，jIs associated with the location of (a); in the image classification method based on the robust learning model provided in the above embodiment, an iteration mode is adopted to perform l on the soft label2，1Normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample, which may include:
the projection matrix and the soft label matrix are obtained by the following formulas:
subj fi≥0,eTfi＝1for i＝1,2,...,l+u
wherein F represents a soft label matrix, P represents a projection matrix, W represents a reconstruction coefficient matrix, D represents a diagonal matrix, and D representsii＝∑jWi,j，fiTo representThe ith row vector of (2), XTP-FTRepresenting regression residue, which is used to measure XTP and Soft tag FTThe degree of difference of (a); mu.siα, gamma is a trade-off parameter corresponding to each of the above data, and a constraint of 1 for column sum, e, can be addedTfi1 and a non-negative constraint, i.e. fiAnd the distribution probability of each training sample in the output soft label category matrix F is ensured to meet the logic that the probability sum is 1 and the distribution probability is a positive value (e is a column vector with elements of 1). And the projection matrix P, i.e. the mapping classifier, can be expressed as:
in the actual iterative solution process, the projection matrix and the soft label matrix are usually obtained by using the following formulas:
subjF≥0,eTF＝eT
wherein, omega ═ I-WT-W+WTW and I are unit matrixes, and only diagonal elements of the unit matrixes are nonzero and are 1; u denotes a diagonal matrix with the ith diagonal element μiFor labeled training samples, μiSet to + ∞, for training samples not labeled with class labels, μiIs set to 0. While projecting l of the matrix P2,1Norm, i.e. P2,1Many rows of the projection matrix P can be changed into zero, thereby reducing the influence of mixed signals in the original space on the soft label and ensuring that the projection matrix P is sparse.
It should be noted that, in the actual iterative solution process, on the basis of the above formula, the formula may be further rewritten as:
wherein G, H, V are diagonal matrices, and:
wherein r isiIs R ═ XTP-FTThe ith row vector of (1).
By passingFor a partial derivative of P set to 0, the following equation can be obtained:
wherein,
by passingFor a partial derivative of F set to 0, the following equation can be obtained:
wherein,
after each iteration update, the initial label matrix Y, the diagonal matrix D, the soft label matrix F, and the diagonal matrix H, V are updated as:
wherein,due to the fact thatAnd YlRespectively a prediction soft label and a label matrix of an original training sample marked with a class label,and YuFor the predicted soft label and label matrix of the original training sample not labeled with class label, the above formula can be rewritten as:
wherein, U can be divided into 4 parts according to the original training samples with the class labels and without the class labels in the initial training samples. The above formula can be converted into the following formula:
after the (k + 1) th iteration is estimated by the above two equations, the soft labels of the training samples with class labels and without class labels in the original space are updated as follows:
setting upBy substituting the above second equation into the first equation:
suppose thatThe soft label of the original training sample in the iteration process can be obtained by the following formula:
f can be updated by the above iteration stepsl k+1Andtherefore, when the original input signal contains noise or even contains error marks, the image classification method based on the robust learning model provided by the invention can update F through iterative stepsl k+1Thereby making the discrimination stronger. By F after updatel k+1Andcan train to obtain Pk+1And Gk+1，Andthe updating is as follows:
wherein,is (F)l k+1)TOrJth row vector, (r)j)k+1J is 1,2, …, l isThe jth row vector, and (r)j)k+1J is l +1, …, l + u isThe row vector of (1).
In the actual iteration process, in order to facilitate solving, the image classification method based on the robust learning model provided by the invention sets F in each iteration processl k+1＝YlTo simplify the optimization, it is therefore possible to solve the following problems
Wherein,let the marked data x in the original input signaljIs normalized by the parameter psijIs set to psilUnmarked data xjIs normalized by the parameter psijIs set to psiuWhen psilWhen 0, it indicates that the label prediction result for the labeled data is the same as the original label. I isψ＝(1+U)-1，Iξ＝I-IψRespectively mixing IψAnd IξThe method is divided into four parts as follows:
wherein,andrespectively diagonal element is psilAnd 1-psilThe diagonal matrix of (a) is,andrespectively diagonal element is psiuAnd 1-psiuThe diagonal matrix of (a). Is provided withThe k +1 th iteration yields:
the specific algorithm is as follows:
inputting:
matrix of raw dataInitial label matrix Y, control parameters ψ, α, γ, and K.
And (3) outputting:
1) initialization G0，H0，V0Is an identity matrix, F0＝Y，P0＝0，k＝0；
2) Computing a weight matrix W, constructing a diagonal matrix D such that Dii＝∑jWi,j；
3) Define matrix Ω ═ I-WT-W+WTW＝[Ωll,Ωlu,Ωul,Ωuu]。
When not converging:
the soft label matrix F is updated by fixing other values to solve the following problem:
updating the projection matrix P by fixing other values solves the following problem:
update matrix G is:
computingUpdating by the method for updating H and VAnd
if | | | Fk+1-Fk||F≤ε,||Pk+1-Pk||FStopping circulation when the epsilon is less than or equal to epsilon; otherwise, K is K +1, and the circulation is continued.
And (3) outputting:
optimal solutionAnd an optimal projection matrix P ═ Pk+1。
Corresponding to the foregoing method embodiment, an embodiment of the present invention further provides an image classification system, as shown in fig. 2, which may include:
the preprocessing module 21 is configured to initialize a training set obtained in advance to obtain an initial class label matrix, where the training set includes a preset number of training samples, and the training samples include samples of which the classes are known and are labeled with class labels corresponding to the classes, and samples of which the classes are unknown and are not labeled with class labels; the construction method is used for processing the training samples based on the neighbor definition and the reconstruction weight to obtain a similarity measurement matrix corresponding to the training samples, and the similarity measurement matrix is subjected to preset processing to obtain a reconstruction coefficient matrix;
a training module 22, configured to determine a soft label of a training sample without a class label by effectively balancing a manifold smoothing term and a label fitting term based on the reconstruction coefficient matrix and the initial class label matrix, and perform l on the soft label of the training sample in an iterative manner2，1Normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample;
the prediction module 23 is configured to map the sample to be detected by using the projection matrix to obtain a soft label of the sample to be detected; wherein, the sample to be detected is a sample without a calibrated class label.
The method comprises the steps of firstly, utilizing training samples marked with class labels and training samples not marked with class labels to jointly construct a similarity measurement matrix, further obtaining a reconstruction coefficient matrix, initializing to obtain an initial class label matrix, utilizing the initial class label matrix and the reconstruction coefficient matrix to determine soft labels of the training samples not marked with the class labels, and carrying out l on the soft labels2,1Norm of normThe influence of the mixed signal in the original space formed by the soft label on the accuracy of the classification result is effectively reduced; by being based on l2,1The robustness of regression residual measurement between the embedded features and the soft labels is effectively enhanced by the norm regularization measurement; by calculating on the basis of l2,1And the mapping of the metric of norm regularization enhances the descriptive property of the extracted features. In summary, by introducing l2,1The norm regularization technology effectively improves the robustness of the system to mixed signals such as noise, heterogeneous data and the like in training samples, avoids the occurrence of the condition that the classification result accuracy of a label-free sample is lower due to the influence of the mixed signals of an original space in the background technology, namely, reduces the influence of the mixed signals of the original space, improves the accuracy of the classification result of a sample to be detected, and enhances the classification performance. In addition, this application is the model of promoting outside the sample, and the conclusion and the prediction of the outer data of sample need not to introduce extra reconstruction process, and the expansibility can be good for the high efficiency completion sample of embedding.
In the image classification system provided by the above embodiment, the preprocessing module may include:
and the K neighbor unit is used for processing each training sample by utilizing a K neighbor algorithm to obtain K nearest neighbor samples of each training sample, wherein K is a positive integer.
And the LLE-reconstruction weight unit is used for acquiring a similarity measurement matrix corresponding to the training samples by using the K nearest neighbor samples of each training sample by adopting a LLE-reconstruction weight construction method.
And the preset processing unit is used for carrying out normalization processing and symmetry processing on the similarity measurement matrix to obtain a reconstruction coefficient matrix.
An initialization unit to: obtaining an initialization matrix with Y0＝[y1,y2,…,yl+u]For presentation and training setIt is shown that, among others,representing the matrix space of n times l + u, XLFor samples labeled with class labels, XL＝[x1,x2,...,xl]，XUFor samples not labeled with class labels, XU＝[xl+1,xl+2,...,xl+u]N, l and u are positive integers; determining any y in the initialization matrixiIs a column vector, any column vector corresponds to the ith training sample x in the training seti1, 2., l + u; training sample x with any calibration class labeljIf the training sample xjBelongs to the ith category, then y is determinedi，jIf the training sample x is 1jIf it does not belong to the i-th category, y is determinedi，j0; for any training sample x not marked with class labeljDetermining yi，jAn initial class label matrix Y is obtained, where j is 1, 2.
The LLE-reconstructing weight unit may include:
an LLE-reconfiguration weight section for: the similarity metric matrix is obtained using the following formula:
wherein,representing a training sample xiK nearest neighbor samples ofIn this case, q corresponds to K,for the current training sample xiThe reconstructed coefficient vector of (2) characterizing its neighborsOn reconstruction of training sample xiThe degree of contribution of the time-varying,representing the corresponding similarity reconstruction coefficient matrix.
In the image classification system provided by the above embodiment, the prediction module may include:
a prediction unit to: obtaining a sample x to be testednew(ii) a By PTxnewTo-be-detected sample xnewEmbedding the prediction vector into a projection matrix P to obtain a prediction vector; determining the soft label corresponding to the element with the maximum probability in the prediction vector as the sample x to be detectednewThe soft tag of (1).
In the image classification system provided by the above embodiment, the training module may include:
a training unit to: the projection matrix and the soft label matrix are obtained by the following formulas:
subj fi≥0,eTfi＝1for i＝1,2,...,l+u
wherein F represents a soft label matrix, P represents a projection matrix, W represents a reconstruction coefficient matrix, D represents a diagonal matrix, and D representsii＝∑jWi,j，fiTo representThe ith row vector of (2), XTP-FTRepresenting regression residue, which is used to measure XTP and Soft tag FTThe degree of difference of (a); mu.siα, γ is the corresponding trade-off parameter;
or, the projection matrix and the soft label matrix are obtained by the following formulas:
subjF≥0,eTF＝eT
wherein, omega ═ I-WT-W+WTW and I are unit matrixes, and only diagonal elements of the unit matrixes are nonzero and are 1; u denotes a diagonal matrix with the ith diagonal element μiFor labeled training samples, μiSet to + ∞, for training samples not labeled with class labels, μiIs set to 0.
For an image classification system disclosed in the embodiment of the present invention, since it corresponds to an image classification method based on a robust learning model disclosed in the embodiment of the present invention, for a specific description thereof, reference may be made to relevant contents of the above method embodiment.
In addition, in the embodiment of the present invention, the image classification method based on the robust learning model provided in the embodiment of the present invention is tested on 4 real data sets, including Extended Yale-B, AR face, CMU peer face, and CMU place face. Based on computational efficiency considerations, all real images are compressed to 32x32 in size, so in the experiment one 1024-dimensional vector is for each picture. In order to test the performance of the method, a certain number of samples are selected from each class to form training samples with labeled class labels in a training set to serve as labeled training sample subsets, and then the same number of training samples without labeled class labels are selected from each class to serve as unlabeled training sample subsets. These data sets are collected from multiple aspects, and thus the test results are generally illustrative.
Please refer to table 1, which shows the average recognition rate and standard deviation of each method experiment for the image classification method based on the robust learning model and the comparison table of the testing recognition results of the GFHF, LLGC, SLP, LNP, SDA, Lap-LDA, FME and ELP methods in the extended Yale-B, AR face, CMU PIE and CMU pos face data set. In the experiment, a certain amount of samples are selected from each type of samples to form a training set, and in order to increase the experiment fairness, experiment parameters participating in a comparison method are also carefully selected, wherein the number of training samples marked with class labels for each type in the training set is in parentheses in a table, and the units of identification results are percentages.
TABLE 1 comparison table of test identification results
In addition, when the image is a face image, the face image test sample is equivalent to a sample to be tested in the present application, and the face image training sample is equivalent to a training sample in the present application, and a label prediction schematic diagram for obtaining the category of the face image test sample is shown in fig. 3.
The experimental result shows that the image classification method based on the robust learning model provided by the invention has an effect obviously superior to that of the traditional label propagation algorithm and has higher applicability and robustness. Therefore, the image classification system provided by the invention has the same effect.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. An image classification method based on a robust learning model is characterized by comprising the following steps:
initializing a pre-acquired training set to obtain an initial class label matrix, wherein the training set comprises a preset amount of training samples, and the training samples comprise samples of which the classes are known and are marked with class labels corresponding to the classes, and samples of which the classes are unknown and are not marked with class labels;
processing the training sample based on a construction method of neighbor definition and reconstruction weight to obtain a similarity measurement matrix corresponding to the training sample, and presetting the similarity measurement matrix to obtain a reconstruction coefficient matrix;
determining the soft label of the training sample without calibrating the class label through an effective balance manifold smooth item and a label fitting item based on the reconstruction coefficient matrix and the initial class label matrix, and carrying out l on the soft label of the training sample in an iteration mode2，1Normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample;
mapping a sample to be detected by using the projection matrix to obtain a soft label of the sample to be detected; the sample to be detected is a sample of which the category is unknown and the category label is not calibrated;
the training sample and the sample to be detected are both face images.
2. The method according to claim 1, wherein the constructing method based on neighbor definition and reconstruction weight processes the training samples to obtain a similarity metric matrix corresponding to the training samples, and comprises:
processing each training sample by using a K nearest neighbor algorithm to obtain K nearest neighbor samples of each training sample, wherein K is a positive integer;
and acquiring a similarity measurement matrix corresponding to the training sample by using the training sample and K nearest neighbor samples of each training sample by adopting an LLE-reconstruction weight construction method.
3. The method according to claim 2, wherein the pre-processing the similarity metric matrix to obtain a reconstruction coefficient matrix comprises:
and carrying out normalization processing and symmetry processing on the similarity measurement matrix to obtain the reconstruction coefficient matrix.
4. The method of claim 3, wherein mapping a sample to be tested with the projection matrix to obtain a soft label of the sample to be tested comprises:
obtaining a sample x to be testednew；
By PTxnewThe sample x to be tested is processednewEmbedding the prediction vector into a projection matrix P to obtain a prediction vector;
determining that the soft label corresponding to the element with the maximum probability in the prediction vector is the sample x to be detectednewThe soft tag of (1).
5. The method of claim 2, wherein initializing a pre-acquired training set to obtain an initial class label matrix comprises:
obtaining an initialization matrix with Y0＝[y1,y2,…,yl+u]Representing the training setIt is shown that, among others,representing the matrix space of n times l + u, XLFor samples labeled with class labels, XL＝[x1,x2,...,xl]，XUFor samples not labeled with class labels, XU＝[xl+1,xl+2,...,xl+u]N, l and u are positive integers;
determining any y in the initialization matrixiIs a column vector, any column vector corresponds to the ith training sample x in the training seti，i＝1,2,...,l+u；
Training sample x with any calibration class labeljIf the training sample xjBelongs to the ith category, then y is determinedi，jIf the training sample x is 1jIf it does not belong to the i-th category, y is determinedi，j0; for any training sample x not marked with class labeljDetermining yi，jThe initial class label matrix Y is obtained, where j is 1,2,...,l+u。
6. The method according to claim 5, wherein the constructing method based on neighbor definition and reconstruction weight processes the training samples to obtain a similarity metric matrix corresponding to the training samples, and comprises:
obtaining the similarity metric matrix using the following formula:
wherein,representing a training sample xiK nearest neighbor samples, q corresponds to K,for the current training sample xiThe reconstructed coefficient vector of (2) characterizing its neighborsOn reconstruction of training sample xiThe degree of contribution of the time-varying,representing the corresponding similarity reconstruction coefficient matrix.
7. The method of claim 6, wherein said iteratively applying is performed on said sampleTraining sample soft label2，1And normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample, wherein the soft label matrix comprises the following steps:
the projection matrix and the soft label matrix are obtained by the following formulas:
subj fi≥0,eTfi＝1for i＝1,2,...,l+u
wherein F represents a soft label matrix, P represents a projection matrix, W represents a reconstruction coefficient matrix, D represents a diagonal matrix, and D representsii＝∑jWi,j，fiTo representThe ith row vector of (2), XTP-FTRepresenting a regression residue term that is used to measure XTP and Soft tag FTThe degree of difference of (a); mu.siα, γ is the corresponding trade-off parameter;
or, the projection matrix and the soft label matrix are obtained by the following formulas:
subjF≥0,eTF＝eT
wherein, omega ═ I-WT-W+WTW and I are unit matrixes, and only diagonal elements of the unit matrixes are nonzero and are 1; u denotes a diagonal matrix with the ith diagonal element μiFor labeled training samples, μiSet to + ∞, for training samples not labeled with class labels, μiIs set to 0.
8. An image classification system, comprising:
the system comprises a preprocessing module, a classification module and a classification module, wherein the preprocessing module is used for initializing a training set acquired in advance to obtain an initial classification label matrix, the training set comprises a preset amount of training samples, and the training samples comprise samples of which the classification is known and the classification labels corresponding to the classification are calibrated and samples of which the classification is unknown and the classification labels are not calibrated; the construction method is used for processing the training sample based on neighbor definition and reconstruction weight to obtain a similar measurement matrix corresponding to the training sample, and the similar measurement matrix is subjected to preset processing to obtain a reconstruction coefficient matrix;
a training module, configured to determine a soft label of a training sample without a class label by effectively balancing a manifold smoothing term and a label fitting term based on the reconstruction coefficient matrix and the initial class label matrix, and perform l on the soft label of the training sample in an iterative manner2，1Normalizing the norm to obtain a projection matrix and a soft label matrix corresponding to the soft label of the training sample;
the prediction module is used for mapping a sample to be detected by using the projection matrix to obtain a soft label of the sample to be detected; the sample to be detected is a sample without a calibrated class label;
the training sample and the sample to be detected are both face images.
9. The system of claim 8, wherein the pre-processing module comprises:
the K nearest neighbor unit is used for processing each training sample by utilizing a K nearest neighbor algorithm to obtain K nearest neighbor samples of each training sample, and K is a positive integer;
and the LLE-reconstruction weight unit is used for acquiring a similarity measurement matrix corresponding to the training samples by using the K nearest neighbor samples of each training sample by adopting an LLE-reconstruction weight construction method.
10. The system of claim 9, wherein the pre-processing module comprises:
and the preset processing unit is used for carrying out normalization processing and symmetry processing on the similarity measurement matrix to obtain the reconstruction coefficient matrix.
CN201510726581.1A 2015-10-30 2015-10-30 A kind of image classification method and image classification system based on Robust Learning model Active CN105335756B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201510726581.1A CN105335756B (en) 2015-10-30 2015-10-30 A kind of image classification method and image classification system based on Robust Learning model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201510726581.1A CN105335756B (en) 2015-10-30 2015-10-30 A kind of image classification method and image classification system based on Robust Learning model

Publications (2)

Publication Number Publication Date
CN105335756A CN105335756A (en) 2016-02-17
CN105335756B true CN105335756B (en) 2019-06-11

Family

ID=55286271

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201510726581.1A Active CN105335756B (en) 2015-10-30 2015-10-30 A kind of image classification method and image classification system based on Robust Learning model

Country Status (1)

Country Link
CN (1) CN105335756B (en)

Families Citing this family (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105608478B (en) * 2016-03-30 2019-12-06 苏州大学 image feature extraction and classification combined method and system
CN106529604B (en) * 2016-11-24 2019-09-27 苏州大学 A kind of adaptive image tag Robust Prediction method and system
CN106845358B (en) * 2016-12-26 2020-11-10 苏州大学 Method and system for recognizing image features of handwritten characters
CN108108769B (en) * 2017-12-29 2020-08-25 咪咕文化科技有限公司 Data classification method and device and storage medium
CN109993191B (en) * 2018-01-02 2021-07-06 ***通信有限公司研究院 Information processing method and device, electronic device and storage medium
CN108848065B (en) * 2018-05-24 2020-12-11 中电运行(北京)信息技术有限公司 Network intrusion detection method, system, medium and equipment
CN108932539B (en) * 2018-05-31 2021-10-08 重庆微标科技股份有限公司 Medical examination sample and medical test tube batch identification device and method
CN111008637A (en) * 2018-10-08 2020-04-14 北京京东尚科信息技术有限公司 Image classification method and system
CN110365583B (en) * 2019-07-17 2020-05-22 南京航空航天大学 Symbol prediction method and system based on bridge domain transfer learning
CN110781926B (en) * 2019-09-29 2023-09-19 武汉大学 Multi-spectral band image analysis method of support vector machine based on robust auxiliary information reconstruction
CN111046933B (en) * 2019-12-03 2024-03-05 东软集团股份有限公司 Image classification method, device, storage medium and electronic equipment
CN111461345B (en) * 2020-03-31 2023-08-11 北京百度网讯科技有限公司 Deep learning model training method and device
CN113537389B (en) * 2021-08-05 2023-11-07 京东科技信息技术有限公司 Robust image classification method and device based on model embedding
CN114049567B (en) * 2021-11-22 2024-02-23 齐鲁工业大学 Adaptive soft label generation method and application in hyperspectral image classification
CN114998640A (en) * 2022-04-20 2022-09-02 安徽文达信息工程学院 Hyperspectral semi-supervised classification method for self-adaptive neighbor graph optimization
CN115511012B (en) * 2022-11-22 2023-04-07 南京码极客科技有限公司 Class soft label identification training method with maximum entropy constraint

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103714536A (en) * 2013-12-17 2014-04-09 深圳先进技术研究院 Sparse-representation-based multi-mode magnetic resonance image segmentation method and device
CN104794489A (en) * 2015-04-23 2015-07-22 苏州大学 Deep label prediction based inducing type image classification method and system
CN104933428A (en) * 2015-07-23 2015-09-23 苏州大学 Human face recognition method and device based on tensor description

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103714536A (en) * 2013-12-17 2014-04-09 深圳先进技术研究院 Sparse-representation-based multi-mode magnetic resonance image segmentation method and device
CN104794489A (en) * 2015-04-23 2015-07-22 苏州大学 Deep label prediction based inducing type image classification method and system
CN104933428A (en) * 2015-07-23 2015-09-23 苏州大学 Human face recognition method and device based on tensor description

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
"Robust Bilinear Matrix Recovery by Tensor Low Rank Representation";zhang zhao 等;《2014 International Conference on Neural Networks》;20140711;第2945-2951页

Also Published As

Publication number Publication date
CN105335756A (en) 2016-02-17

Similar Documents

Publication Publication Date Title
CN105335756B (en) A kind of image classification method and image classification system based on Robust Learning model
CN105354595B (en) A kind of robust visual pattern classification method and system
CN105608471B (en) Robust direct-push label estimation and data classification method and system
Zhao et al. Learning from normalized local and global discriminative information for semi-supervised regression and dimensionality reduction
CN104123560B (en) Fuzzy facial image verification method based on phase code feature and more metric learnings
CN107679466B (en) Information output method and device
CN103064941B (en) Image search method and device
CN110210625B (en) Modeling method and device based on transfer learning, computer equipment and storage medium
CN112232407B (en) Neural network model training method and device for pathological image samples
CN102422324B (en) Age estimation device and method
CN108021930B (en) Self-adaptive multi-view image classification method and system
CN109492750B (en) Zero sample image classification method based on convolutional neural network and factor space
CN111814804B (en) Human body three-dimensional size information prediction method and device based on GA-BP-MC neural network
WO2019205729A1 (en) Method used for identifying object, device and computer readable storage medium
CN113128671B (en) Service demand dynamic prediction method and system based on multi-mode machine learning
CN108664986B (en) Based on lpNorm regularized multi-task learning image classification method and system
CN111090764A (en) Image classification method and device based on multitask learning and graph convolution neural network
CN109766942A (en) A kind of small-sample learning image-recognizing method based on attention neural network
Wang et al. Energy based competitive learning
CN108985161B (en) Low-rank sparse representation image feature learning method based on Laplace regularization
Zhao et al. Soft label based linear discriminant analysis for image recognition and retrieval
CN115147607A (en) Anti-noise zero-sample image classification method based on convex optimization theory
Yang et al. Unsupervised images segmentation via incremental dictionary learning based sparse representation
Konukoglu et al. Neighbourhood approximation forests
CN110263808B (en) Image emotion classification method based on LSTM network and attention mechanism

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant