CN114492620A - Credible multi-view classification method based on evidence deep learning - Google Patents

Credible multi-view classification method based on evidence deep learning Download PDF

Info

Publication number
CN114492620A
CN114492620A CN202210080384.7A CN202210080384A CN114492620A CN 114492620 A CN114492620 A CN 114492620A CN 202210080384 A CN202210080384 A CN 202210080384A CN 114492620 A CN114492620 A CN 114492620A
Authority
CN
China
Prior art keywords
evidence
view
deep learning
category
sample
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210080384.7A
Other languages
Chinese (zh)
Inventor
徐偲
赵京龙
赵伟
管子玉
詹涛
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.)
Xidian University
Original Assignee
Xidian 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 Xidian University filed Critical Xidian University
Priority to CN202210080384.7A priority Critical patent/CN114492620A/en
Publication of CN114492620A publication Critical patent/CN114492620A/en
Pending legal-status Critical Current

Links

Images

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
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2431—Multiple classes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a credible multi-view classification method based on evidence deep learning, which comprises the following steps: s1, sample definition, wherein N samples are set in a data set, and each sample has V visual angles; s2, estimating the classification uncertainty of the single-view data by using the single-view evidence; s3, fusing the multi-view evidences, and propagating the global information to each view by using a degenerated layer, so that each view can learn the evidence based on the global information; and S4, optimizing the target, and optimizing all parameters in the model by using a gradient descent algorithm. The method not only improves the accuracy of prediction, but also uses the degradation layer designed in the invention to mine complementary information between deep and easily ignored visual angles, thereby outputting uncertainty more conforming to human cognition in prediction.

Description

Credible multi-view classification method based on evidence deep learning
Technical Field
The invention relates to the technical field of deep learning, in particular to a credible multi-view classification method based on evidence deep learning.
Background
The multi-view classification refers to that each sample in the data set contains features of a plurality of different views, for example, the features of a plurality of different views are often considered by doctors when cancer diagnosis is performed. The multi-view classification task is realized by utilizing deep learning, so that the defect that deep information is difficult to mine from multi-view data in the traditional machine learning can be overcome. Most of the past research is focused on how to improve the prediction accuracy, but neglects the reliability of the decision. However, in many high-risk applications, not only the result of the decision obtained by the model, but also the confidence level of the decision obtained is required. For example, in medical diagnosis, the confidence level of the decision is crucial, and a decision with unknown confidence level may be unreliable, thereby misleading the doctor to make a wrong diagnosis and delaying the optimal treatment time of the patient.
In recent years, some methods have appeared to output uncertainty of prediction while predicting, but still cannot effectively mine complementary information between multiple views, and thus cannot output reasonable uncertainty. For example. In a certain medical diagnosis event, according to nuclear magnetic resonance imaging, three categories of health, left-side lesion and right-side lesion can be predicted according to a view angle 1, but the prediction precision is not high due to image blurring; according to the clinical test result, the view angle 2 can only predict two categories of health and lesion without distinguishing whether the lesion is on the left or on the right, but the accuracy is high. If, according to common human knowledge, a lesion is predicted at view 2, then a very large complementary information will be given at view 1. In fact, in the past, since it is impossible to distinguish whether a lesion is on the left or right side, that is, complete category information, when processing the view 2, only high uncertainty can be output, and thus deep complementary information of the two views cannot be mined. Ignoring information from any one perspective in a medical diagnosis increases the likelihood of causing a misdiagnosis.
In the multi-view classification problem, under the condition of ensuring that the accuracy is not reduced, it still has great challenges on how to mine complementary information among views and obtain uncertainty according with human cognition. Therefore, how to provide a credible multi-view classification method based on evidence deep learning is a problem that needs to be solved urgently by those skilled in the art.
Disclosure of Invention
The invention aims to provide a credible multi-view classification method based on evidence deep learning, and aims to solve the problems existing in the background technology by providing a deep learning method which can effectively mine complementarity information among views and obtain reasonable uncertainty under the condition of ensuring that the accuracy is not reduced aiming at the multi-view classification problem.
The credible multi-view classification method based on evidence deep learning comprises the following steps:
s1, sample definition, wherein N samples are set in a data set, and each sample has V visual angles;
s2, estimating the classification uncertainty of the single-view data by using the single-view evidence;
s3, fusing the multi-view evidences, and propagating the global information to each view by using a degenerated layer, so that each view can learn the evidence based on the global information;
and S4, optimizing the target, and optimizing all parameters in the model by using a gradient descent algorithm.
Preferably, in step S1:
characterization vector for nth sample vth view:
Figure BDA0003485879520000021
wherein D isvIs its dimension;
considering that the data set contains C categories, each category is represented as a one-hot vector as a true label of a sample, and the true label of the nth sample:
yn∈{0,1}C；
the goal is to give a sample, predict its label, and give the uncertainty u of the predictionn∈[0,1]。
Preferably, in step S2:
for the classification problem of C classes, the dirichlet distribution parameter of the nth sample at the v-th view angle is:
Figure BDA0003485879520000022
its probability density function is then expressed as:
Figure BDA0003485879520000023
wherein b (x) represents a probability density function of the polynomial distribution;
TCis C-dimensional simplex:
Figure BDA0003485879520000024
determining parameters of a dirichlet distribution using a deep neural network-based learner
Figure BDA0003485879520000025
Figure BDA0003485879520000026
Wherein f isv(x) A deep neural network-based learner is represented,
Figure BDA0003485879520000027
an evidence vector representing a v-th perspective of the nth sample;
dirichlet distribution is used to describe evidence distribution, let
Figure BDA0003485879520000031
Then the probability of prediction as class c for the nth sample and the prediction uncertainty are:
Figure BDA0003485879520000032
Figure BDA0003485879520000033
wherein the content of the first and second substances,
Figure BDA0003485879520000034
representing the intensity of the dirichlet distribution.
Preferably, the evidence value
Figure BDA0003485879520000035
The larger the likelihood that the corresponding category is a true category; uncertainty when the sum of the evidence values of C classes is larger
Figure BDA0003485879520000036
The smaller will be.
Preferably, in step S3:
the specific visual angle evidence and the fusion evidence satisfy a degradation relation:
Figure BDA0003485879520000037
wherein, calculating
Figure BDA0003485879520000038
Meaning that the corresponding elements of two vectors are multiplied, meaning that the data on both sides are not required to be exactly equal but are expected to be as close as possible, dv∈RCRepresenting an intensity degradation vector for scaling the intensity, U, between the perspective-specific evidence and the fused evidencev∈RC×CRepresenting a category degradation relationship matrix;
the objective function of the degradation layer can be formulated as:
Figure BDA0003485879520000039
wherein the content of the first and second substances,
Figure BDA00034858795200000310
the visual angle is a normalization weight coefficient of the v-th visual angle, the larger the weight is, the higher the reliability of the corresponding visual angle is, and when a sample is input into the model, the visual angle with high reliability plays a positive role in optimizing the model parameters; the view angle with low reliability may influence the updating direction of the parameters, so that the generalization capability of the model is poor, and therefore, the influence of the view angle with low reliability on the updating of the model parameters is weakened by adding the weight coefficient.
Preferably, the degradation loss has the meaning of expressing the difference between the fusion evidence and the characteristic perspective evidence.
Preferably, the degenerating layer includes two degenerating stages of intensity degenerating and category degenerating, the intensity degenerating is that the information intensity of the evidence at the specific view angle is inconsistent with that of the fused evidence, and the category degenerating is that the clustering definition at the specific view angle is inconsistent with that of the fused evidence.
Preferably, in step S4:
according to the dirichlet distribution in S2 and the evidence theory:
Figure BDA0003485879520000041
true label y of desired specimennThe corresponding category has more evidence than other categories, and only the log-likelihood function needs to be minimized:
Figure BDA0003485879520000042
however, it is not preferable that the larger the proof value is, and it is not desirable in the present invention that the proof values of the plurality of categories are scaled with each other. Therefore, the KL divergence is introduced in the present invention as a regularization term:
Figure BDA0003485879520000043
wherein the content of the first and second substances,
Figure BDA0003485879520000044
representing the dirichlet distribution parameters after removing the true category evidence; the KL divergence can on the one hand limit the growth of the overall evidence and on the other hand can force the evidence distribution towards a direction of less information entropy, in other words a greater difference in evidence values between different classes.
Giving the overall optimization objective of the model:
Figure BDA0003485879520000045
at the time of optimization, setting
Figure BDA0003485879520000046
t represents the number of iterations, it is expected that the model can learn more sufficient evidence at the beginning of training, and then gradually increase the penalty of the regularization term, setting δ to 0.1 represents the coefficient of the degenerated layer.
Preferably, the
Figure BDA0003485879520000047
The function stimulates the evidence values of the true category towards the total evidence value.
The invention has the beneficial effects that:
the method not only improves the accuracy of prediction, but also uses the degradation layer designed in the invention to mine complementary information between deep and easily ignored visual angles, thereby outputting uncertainty more conforming to human cognition in prediction.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
FIG. 1 is a flowchart of an overall depth network of a credible multi-view classification method based on evidence deep learning provided by the invention;
FIG. 2 is a visual angle 1 three-classification sample visualization example diagram of the credible multi-visual angle classification method based on evidence deep learning provided by the invention;
FIG. 3 is a visual angle 2 three-classification sample visualization example diagram of the credible multi-visual angle classification method based on evidence deep learning provided by the invention;
FIG. 4 is a bar graph of evidence value data for the TMC model;
fig. 5 is an evidence value data bar graph of the credible multi-view classification method based on evidence deep learning provided by the invention.
Detailed Description
The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic views, which illustrate the basic structure of the present invention only in a schematic manner, and thus show only the constitution related to the present invention, with reference to fig. 1 to 5:
example 1:
the embodiment provides a deep learning method for solving the multi-view classification problem containing uncertainty prediction by considering the consistency and complementarity of multi-view data, and verifies that the method obviously improves the classification accuracy and confidence.
The credible multi-view classification method based on evidence deep learning comprises the following steps:
s1, sample definition, wherein N samples are set in a data set, and each sample has V visual angles;
s2, estimating the classification uncertainty of the single-view data by using the single-view evidence;
s3, fusing the multi-view evidences, and propagating the global information to each view by using a degenerated layer, so that each view can learn the evidence based on the global information;
and S4, optimizing the target, and optimizing all parameters in the model by using a gradient descent algorithm.
In step S1:
characterization vector for nth sample vth view:
Figure BDA0003485879520000051
Dvis its dimension;
considering that the data set contains C categories, each category is represented as a one-hot vector as a true label of a sample, and the true label of the nth sample:
yn∈{0,1}C；
the goal is to give a sample, predict its label, and give the uncertainty u of the predictionn∈[0,1]。
In step S2:
for the classification problem of C classes, the dirichlet distribution parameter of the nth sample at the v-th view angle is:
Figure BDA0003485879520000061
its probability density function is expressed as:
Figure BDA0003485879520000062
wherein b (x) represents a probability density function of the polynomial distribution;
TCis C-dimensional simplex:
Figure BDA0003485879520000063
determining parameters of a dirichlet distribution using a deep neural network-based learner
Figure BDA0003485879520000064
Figure BDA0003485879520000065
Wherein f isv(x) A deep neural network-based learner is represented,
Figure BDA0003485879520000066
an evidence vector representing a v-th perspective of the nth sample;
dirichlet distribution is used to describe evidence distribution, let
Figure BDA0003485879520000067
Then the probability of prediction as class c for the nth sample and the prediction uncertainty are:
Figure BDA0003485879520000068
Figure BDA0003485879520000069
wherein the content of the first and second substances,
Figure BDA00034858795200000610
representing the intensity of the dirichlet distribution.
Preferably, the evidence value
Figure BDA00034858795200000611
The larger the likelihood that the corresponding category is a true category; uncertainty when the sum of the evidence values of C classes is larger
Figure BDA00034858795200000612
The smaller will be.
The model provided by the invention comprises two degradation stages of strength degradation and category degradation, wherein the strength degradation indicates that the information strength of the evidence at the specific visual angle is inconsistent with the information strength of the fused evidence, and the category degradation indicates that the clustering definition of the specific visual angle is inconsistent with the definition of the fused evidence.
Referring to fig. 2-3, in an embodiment, for a three-classification task involving two views, if category 2 and category 3 in view 1 do not have sufficient features to distinguish them, but are able to clearly distinguish category 1 from categories 2 and 3, then view 1 should have useful information, i.e., low uncertainty, to assist in a more clear classification of view 2.
In step S3:
the specific visual angle evidence and the fusion evidence satisfy a degradation relation:
Figure BDA0003485879520000071
wherein, calculating
Figure BDA0003485879520000072
Meaning that the corresponding elements of two vectors are multiplied, meaning that the data on both sides are not required to be exactly equal but are expected to be as close as possible, dv∈RCRepresenting an intensity degradation vector for scaling the intensity, U, between the perspective-specific evidence and the fused evidencev∈RC×CRepresenting a category degradation relationship matrix;
taking the sample distribution of view 1 as an example, the expected model can learn the matrix U1：
Figure BDA0003485879520000073
Since in View 1, Category 1 is clearly divisible, then
Figure BDA0003485879520000074
While classes 2 and 3 are completely indistinguishable, therefore class 2 or 3 fused evidence must be spread evenly among the evidence of classes 2 and 3 for perspective 1, and so
Figure BDA0003485879520000075
The objective function of the degradation layer can be formulated as:
Figure BDA0003485879520000076
wherein the content of the first and second substances,
Figure BDA0003485879520000077
the weight coefficient is a normalized weight coefficient of the v-th visual angle, and the higher the weight is, the higher the credibility of the corresponding visual angle is.
When a sample is input into the model, the view angle with high reliability plays a positive role in optimizing the model parameters; the view angle with low reliability may influence the updating direction of the parameters, so that the generalization capability of the model is poor, and therefore, the influence of the view angle with low reliability on the updating of the model parameters is weakened by adding the weight coefficient.
The significance of the degradation loss is to express the difference between the fusion evidence and the characteristic perspective evidence.
In step S4:
according to the dirichlet distribution in S2 and the evidence theory:
Figure BDA0003485879520000081
true label y of desired specimennThe corresponding category has more evidence than other categories, and only the log-likelihood function needs to be minimized:
Figure BDA0003485879520000082
however, it is not preferable that the evidence values are larger, and it is not desirable in the present invention that the evidence values of the plurality of categories are mutually scaled. Therefore, the KL divergence is introduced as the regularization term:
Figure BDA0003485879520000083
wherein the content of the first and second substances,
Figure BDA0003485879520000084
representing the dirichlet distribution parameters after removing the true category evidence; the KL divergence can limit the growth of the total evidence on one hand and can promote the evidence distribution to trend to a direction with smaller information entropy on the other hand, in other words, the evidence values of different classes are more different;
giving the overall optimization objective of the model:
Figure BDA0003485879520000085
at the time of optimization, setting
Figure BDA0003485879520000086
t represents the number of iterations, it is expected that the model can learn more sufficient evidence at the beginning of training, then gradually increasing the penalty of the regularization term, setting δ to 0.1 represents the coefficient of the degenerated layer
Figure BDA0003485879520000087
The function stimulates the evidence values of the true category towards the total evidence value.
After the implementation of the S1-S4, compared with the multi-view classification model provided in the prior art, the method provided by the invention not only can ensure that the accuracy is not lower than that of the conventional method, but also can calculate the uncertainty which is more in line with the practical logic by mining the complementary information among the multi-views.
Referring to fig. 4-5, in order to verify this, the present embodiment designs a test sample, which includes 2 views and 3 categories, the distribution of which is shown in fig. 2, and the best proposed method TMC (the paper "trained Multi-view Classification") and the method proposed in the present invention are used to perform experiments respectively, and the evidence values learned by the recording model are compared.
As can be seen from fig. 4 to 5, at view 1, when the categories 2 and 3 are not easily distinguished but are both clearly distinguishable from the category 1, TMC may have a condition of insufficient evidence mining, that is, information carried by view 1 cannot be fully mined, so that a higher uncertainty of 0.49 is output, which is not ideal, and the method provided by the present invention mines sufficient evidence at both categories 2 and 3 of view 1, which is attributed to mining of information of view 1 by the degraded layer, and further fuses view information, and predicts an accurate and reliable decision uncertainty of 0.22.
The invention has carried on the experiment on two open data sets, hand Written data set is the HandWritten digital picture, the extracted characteristic includes 6 visual angles, divide into 10 categories, Scene15 data set includes 4485 Scene pictures, divide into 15 categories, the invention uses GIST, HOG, LBP method to extract the characteristic of three visual angles to carry on the classification experiment of the multi-view separately, regard accuracy as the index, to each method, each data set carries on 10 times of experiments and gets the mean value, the experimental result is shown as table 1:
TABLE 1
Method HandWritten Scene15
TMC 98.04％ 63.85％
The invention 98.75％ 66.93％
The embodiment verifies that the method not only improves the accuracy of prediction, but also mines complementary information between deep and easily ignored visual angles by using the degradation layer designed in the invention, thereby outputting uncertainty more conforming to human cognition in prediction.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art should be considered to be within the technical scope of the present invention, and the technical solutions and the inventive concepts thereof according to the present invention should be equivalent or changed within the scope of the present invention.

Claims (9)

1. A credible multi-view classification method based on evidence deep learning is characterized by comprising the following method steps:
s1, sample definition, wherein N samples are set in a data set, and each sample has V visual angles;
s2, estimating the classification uncertainty of the single-view data by using the single-view evidence;
s3, fusing the multi-view evidences, and propagating the global information to each view by using a degenerated layer, so that each view can learn the evidence based on the global information;
and S4, optimizing the target, and optimizing all parameters in the model by using a gradient descent algorithm.
2. The credible multi-perspective classification method based on evidence deep learning of claim 1, wherein in step S1:
characterization vector for nth sample vth view:
Figure FDA0003485879510000011
wherein D isvIs its dimension;
considering that the data set contains C categories, each category is represented as a one-hot vector as a true label of a sample, and the true label of the nth sample:
yn∈{0,1}C；
the goal is to give a sample, predict its label, and give the uncertainty u of the predictionn∈[0,1]。
3. The credible multi-perspective classification method based on evidence deep learning of claim 1, wherein in step S2:
for the classification problem of C classes, the dirichlet distribution parameter of the nth sample at the v-th view angle is:
Figure FDA0003485879510000012
its probability density function is then expressed as:
Figure FDA0003485879510000013
wherein b (x) represents a probability density function of the polynomial distribution;
TCis C-dimensional simplex:
Figure FDA0003485879510000014
determining parameters of a dirichlet distribution using a deep neural network-based learner
Figure FDA0003485879510000017
Figure FDA0003485879510000015
Wherein f isv(x) A deep neural network-based learner is represented,
Figure FDA0003485879510000016
an evidence vector representing the nth sample vth view;
dirichlet distribution is used to describe evidence distribution, let
Figure FDA0003485879510000021
Then the probability of prediction as class c for the nth sample and the prediction uncertainty are:
Figure FDA0003485879510000022
Figure FDA0003485879510000023
wherein the content of the first and second substances,
Figure FDA0003485879510000024
representing the intensity of the dirichlet distribution.
4. The method of credible multi-view classification based on evidence deep learning of claim 3, characterized in that the evidence value
Figure FDA0003485879510000025
The larger the likelihood that the corresponding category is a true category; uncertainty when the sum of the evidence values of C classes is larger
Figure FDA0003485879510000026
The smaller will be.
5. The credible multi-perspective classification method based on evidence deep learning of claim 1, wherein in step S3:
the specific visual angle evidence and the fusion evidence satisfy a degradation relation:
Figure FDA0003485879510000027
where the term "operation" denotes the multiplication of corresponding elements of two vectors, the meaning of ≈ is that it does not require exact equality of the data on both sides but it is desirable that they are as close as possible, dv∈RCRepresenting an intensity degradation vector for scaling the intensity, U, between the perspective-specific evidence and the fused evidencev∈RC×CRepresenting a category degradation relationship matrix;
the objective function of the degradation layer can be formulated as:
Figure FDA0003485879510000028
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA0003485879510000029
the weighting coefficient is normalized at the v-th visual angle, and the greater the weighting, the higher the credibility of the corresponding visual angle is.
6. The credible multi-perspective classification method based on evidence deep learning according to claim 5, wherein the meaning of the degradation loss is to express the difference between the fusion evidence and the characteristic perspective evidence.
7. The credible multi-view classification method based on evidence deep learning of claim 5, wherein the degenerating layer comprises two degenerating stages of intensity degeneration and category degeneration, wherein the intensity degeneration is that the information intensity of evidence at a specific view is inconsistent with that of fused evidence, and the category degeneration is that the clustering definition at a specific view is inconsistent with that of fused evidence.
8. The credible multi-perspective classification method based on evidence deep learning of claim 1, wherein in step S4:
according to the dirichlet distribution in S2 and the evidence theory:
Figure FDA0003485879510000031
true label y of desired specimennThe corresponding category has more evidence than other categories, and only the log-likelihood function needs to be minimized:
Figure FDA0003485879510000032
introducing a KL divergence as a regularization term:
Figure FDA0003485879510000033
wherein the content of the first and second substances,
Figure FDA0003485879510000034
expressing Dirichlet distribution parameters after removing the real category evidence, the KL divergence can limit the increase of the total evidence on one hand, and can promote the evidence distribution to trend to the direction with smaller information entropy on the other hand;
giving the overall optimization objective of the model:
Figure FDA0003485879510000035
at the time of optimization, setting
Figure FDA0003485879510000036
t represents the number of iterations, it is expected that the model can learn more sufficient evidence at the beginning of training, and then gradually increase the penalty of the regularization term, setting δ to 0.1 represents the coefficient of the degenerated layer.
9. The evidence deep learning-based credible multi-view classification method according to claim 1, characterized in that, the method is used for credible multi-view classification based on evidence deep learning
Figure FDA0003485879510000037
The function stimulates the evidence values of the true category towards the total evidence value.
CN202210080384.7A 2022-01-24 2022-01-24 Credible multi-view classification method based on evidence deep learning Pending CN114492620A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210080384.7A CN114492620A (en) 2022-01-24 2022-01-24 Credible multi-view classification method based on evidence deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210080384.7A CN114492620A (en) 2022-01-24 2022-01-24 Credible multi-view classification method based on evidence deep learning

Publications (1)

Publication Number Publication Date
CN114492620A true CN114492620A (en) 2022-05-13

Family

ID=81473743

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210080384.7A Pending CN114492620A (en) 2022-01-24 2022-01-24 Credible multi-view classification method based on evidence deep learning

Country Status (1)

Country Link
CN (1) CN114492620A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116884094A (en) * 2023-09-07 2023-10-13 武汉理工大学 Multi-view behavior recognition method and system based on view and behavior decoupling
CN117974634A (en) * 2024-03-28 2024-05-03 南京邮电大学 Evidence deep learning-based reliable detection method for anchor-frame-free surface defects
CN117974634B (en) * 2024-03-28 2024-06-04 南京邮电大学 Evidence deep learning-based reliable detection method for anchor-frame-free surface defects

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116884094A (en) * 2023-09-07 2023-10-13 武汉理工大学 Multi-view behavior recognition method and system based on view and behavior decoupling
CN116884094B (en) * 2023-09-07 2023-12-12 武汉理工大学 Multi-view behavior recognition method and system based on view and behavior decoupling
CN117974634A (en) * 2024-03-28 2024-05-03 南京邮电大学 Evidence deep learning-based reliable detection method for anchor-frame-free surface defects
CN117974634B (en) * 2024-03-28 2024-06-04 南京邮电大学 Evidence deep learning-based reliable detection method for anchor-frame-free surface defects

Similar Documents

Publication Publication Date Title
Ozdemir et al. A 3D probabilistic deep learning system for detection and diagnosis of lung cancer using low-dose CT scans
CN108492297B (en) MRI brain tumor positioning and intratumoral segmentation method based on deep cascade convolution network
Shin et al. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning
Wan et al. Accurate segmentation of overlapping cells in cervical cytology with deep convolutional neural networks
Afshar et al. MIXCAPS: A capsule network-based mixture of experts for lung nodule malignancy prediction
Wang et al. Does non-COVID-19 lung lesion help? investigating transferability in COVID-19 CT image segmentation
CN112907555B (en) Survival prediction method and system based on image genomics
Durkee et al. Artificial intelligence and cellular segmentation in tissue microscopy images
CN109102498B (en) Method for segmenting cluster type cell nucleus in cervical smear image
CN111275118B (en) Chest film multi-label classification method based on self-correction type label generation network
Ferlaino et al. Towards deep cellular phenotyping in placental histology
Arjmand et al. Deep learning in liver biopsies using convolutional neural networks
CN114530222A (en) Cancer patient classification system based on multiomics and image data fusion
Bassi et al. COVID-19 detection using chest X-rays: Is lung segmentation important for generalization?
Dabass et al. An Atrous Convolved Hybrid Seg-Net Model with residual and attention mechanism for gland detection and segmentation in histopathological images
CN114492620A (en) Credible multi-view classification method based on evidence deep learning
CN114580501A (en) Bone marrow cell classification method, system, computer device and storage medium
Dumakude et al. Automated COVID-19 detection with convolutional neural networks
Bakır et al. Detection of pneumonia from x-ray images using deep learning techniques
Jiang et al. A semi-supervised learning approach with consistency regularization for tumor histopathological images analysis
Feng et al. Trusted multi-scale classification framework for whole slide image
Sameki et al. ICORD: Intelligent Collection of Redundant Data-A Dynamic System for Crowdsourcing Cell Segmentations Accurately and Efficiently.
Qiu et al. Scratch Each Other's Back: Incomplete Multi-Modal Brain Tumor Segmentation via Category Aware Group Self-Support Learning
CN112086174A (en) Three-dimensional knowledge diagnosis model construction method and system
Wang et al. 3D multi-scale DenseNet for malignancy grade classification of pulmonary nodules

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