CN106250924A - A kind of newly-increased category detection method based on multi-instance learning - Google Patents
A kind of newly-increased category detection method based on multi-instance learning Download PDFInfo
- Publication number
- CN106250924A CN106250924A CN201610600041.3A CN201610600041A CN106250924A CN 106250924 A CN106250924 A CN 106250924A CN 201610600041 A CN201610600041 A CN 201610600041A CN 106250924 A CN106250924 A CN 106250924A
- Authority
- CN
- China
- Prior art keywords
- bag
- classification
- super
- learning
- newly
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 21
- 238000012549 training Methods 0.000 claims description 40
- 239000000203 mixture Substances 0.000 claims description 16
- 238000000605 extraction Methods 0.000 claims description 11
- 238000012360 testing method Methods 0.000 claims description 11
- 238000013145 classification model Methods 0.000 claims description 8
- 238000000034 method Methods 0.000 claims description 6
- 241001269238 Data Species 0.000 claims description 4
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 230000000452 restraining effect Effects 0.000 claims description 3
- 239000000284 extract Substances 0.000 abstract description 2
- 238000002372 labelling Methods 0.000 description 15
- 238000010801 machine learning Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 241000282376 Panthera tigris Species 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 241000406668 Loxodonta cyclotis Species 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
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)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention discloses a kind of newly-increased category detection method based on multi-instance learning, and the crucial example in many examples " being wrapped " first with crucial example detection algorithm more ripe in multi-instance learning extracts；Afterwards, for each known class, the crucial example of its correspondence being combined into one " the super bag of classification ", all examples not simultaneously being identified as crucial example form one " the super bag of unit "；Subsequently, the distance wrapped between super bag just can be determined by follow-up metric learning.In the practical stage, for the bag of known class, surpass bag according to its closest classification and judge its concept classification；And for the bag of newly-increased classification, owing to there is not the super bag of the classification of its correspondence, should be the super bag of unit apart from its nearest super Bao Ze, so can determine that it is newly-increased classification according to this situation.
Description
Technical field
The present invention relates to machine learning, application technology, particularly to multi-instance learning, newly-increased classification monitoring technology, tolerance
Study, is that one both can carry out automatic concept classification prediction/classification to existing classification, can detect newly-increased classification simultaneously
Robust multi-instance learning algorithm.
Background technology
It is considered as most promising machine learning approach from sample learning.If using the ambiguousness of training sample as
The criteria for classifying, then at present the research in this field is substantially set up under three kinds of learning frameworks, i.e. supervised learning, unsupervised learning and strong
Chemistry is practised.
Supervised learning is by learning the training example with concept labelling, outside as correctly as possible to training set
The concept labelling of example be predicted.The most all of training sample is all markd, and therefore its ambiguousness is minimum.Non-prison
Educational inspector practises by learning the training example not having concept labelling, to find the structure hidden in data.The most all of instruction
Practicing sample is all not have markd, and therefore its ambiguousness is the highest.Intensified learning by there is no concept labelling but with a delay
The training example that award or effectiveness (can be considered the concept labelling of delay) are associated learns, to obtain certain from state to action
Mapping.The most all of training sample is all markd, but unlike supervised learning, labelling is to postpone, therefore
The ambiguousness of intensified learning is between supervised learning and unsupervised learning.
The nineties in 20th century middle and late stage, researchers are in the research to pharmaceutically active forecasting problem, it is proposed that many examples
The concept of study.In this type of learns, " bag " that training set is had concept labelling by several forms, and comprises some in each bag
There is no the example of concept labelling.If at least an example is under the jurisdiction of certain concept classification in a bag, then this is coated and is labeled as this
Classification；If a bag is not belonging to certain concept classification, then any one example in this bag is not the most subordinate to the category.By to training
The study of bag, it is desirable to the concept labelling of the bag outside training set is predicted by learning system as correctly as possible.
Compared with supervised learning, the training example in multi-instance learning does not has concept labelling, during this is with supervised learning
All training examples have concept labelling different；Compared with unsupervised learning, in multi-instance learning, training package is to have concept labelling
, these are the most different from not having any concept labelling in the training sample of unsupervised learning；And compared with intensified learning, many examples
Habit does not has again the concept that timeliness postpones.The more important thing is, in conventional various learning frameworks, a sample is exactly one and shows
Example, i.e. sample and example are one-to-one relationships；And in multi-instance learning, a sample (i.e. bag) contains multiple example,
I.e. sample and example is the corresponding relation of one-to-many.Therefore, the ambiguousness of training sample and supervised learning in multi-instance learning, non-
Supervised learning, intensified learning ambiguousness the most different, this allows for conventional learning method and is difficult to solve well problems.
Owing to multi-instance learning has the character of uniqueness and is widely applied prospect, belong to a blind area of conventional machine learning research,
Therefore it is considered as a kind of new learning framework.
The application scenarios of existing multi-instance learning algorithm is all classification kind and fixed number of static environment, rather than concept
The Open Dynamic environment that classification is variable.As, when building image classification system based on multi-instance learning, at the number of training stage
" elephant ", " fox " and " bird " three kinds of image concept classifications are only comprised according to concentrating.The system practical stage occurs newly the most most probably
Image category, such as Tiger.Now, existing multi-instance learning algorithm can only simply by belong to newly-increased classification sample (as
The picture of tiger) it is divided into a certain class (such as " fox ") of known class mistakenly, system thus can be made at Open Dynamic
Environment lost efficacy.Therefore, multi-instance learning needs a kind of Robust learning algorithm that can detect newly-increased classification.
Summary of the invention
Goal of the invention: current multi-instance learning algorithm all existing classification sample can only be carried out concept classification prediction/point
Class, under the scene having newly-increased classification to occur, simple for newly-increased classification sample and mistake can only be divided into known class by existing algorithm
One in not.For the problems referred to above, present invention firstly provides and solve the newly-increased classification detection under multi-instance learning scene
Task, turns to its form the framework of a metric learning, and proposes many examples new class detection learning algorithm of correspondence.Concrete next
Say, the crucial example extraction in many examples " being wrapped " first with crucial example detection algorithm more ripe in multi-instance learning
Out.These are so-called " crucial examples " refers in " bag " those examples that may decide that the corresponding concept labelling wrapped.Afterwards,
For each known class, the crucial example of its correspondence is combined into one " the super bag of classification ", is not identified as key simultaneously and shows
All examples of example form one " the super bag of unit ".Subsequently, wrapping the distance between super bag just can be by follow-up metric learning
Determine.Wherein, classification surpasses wraps the classification for concept classification, and the super Bao Ze of unit is for increasing the detection of classification newly.On training rank
Section, for the sample (bag) of known class, its some crucial example is wrapped the classification of its correspondence is super, additionally some
Example is in the super bag of unit.But in order to obtain more preferable discriminant classification ability, it is known that the bag of class surpasses the distance of bag with corresponding classification should
For minimum, the distance surpassing bag with unit is taken second place, and surpass the distance of bag with other classification should be bigger.In the system practical stage, for
Know the bag of classification, bag can be surpassed according to its closest classification and judge its concept classification；And for the bag of newly-increased classification, by
In the super bag of the classification that there is not its correspondence, should be the super bag of unit apart from its nearest super Bao Ze, so can come according to this situation
Determine that it is newly-increased classification.
Technical scheme: a kind of newly-increased category detection method based on multi-instance learning, including multi-instance learning disaggregated model
Training step and disaggregated model prediction steps；
Described many example classification model training step particularly as follows:
Step 1.1, on existing many sample datas, utilizes existing crucial example detection algorithm from each many examples bag
XiThe crucial example that middle extraction is corresponding；
Step 1.2, the crucial example that the bag being under the jurisdiction of identical concept classification is extracted composition " the super bag of classification " Sc(c
=1 ..., C).It addition, be not detected as all examples composition " the super bag of the unit " S of crucial example0；
Step 1.3, definition bag to the distance of super bag is: the weighted sum of each example in super bag to bag beeline, its
In, the beeline of example to bag is example to the minimum euclidean distance of example in bag, and in super bag, the weight of each example is by spending
Amount study determines according to training data training.Described super bag both can be the super bag of classification, it is also possible to be the super bag of unit.
Step 1.4, according to the distance of bag to super bag, utilization measure learns described weight.In order to ensure this system
There are enough discriminant classification abilities and new class power of test, in metric learning, need to meet following condition: (1) bag is to its corresponding class
The distance Chao do not wrapped is minimum；(2) bag is second little to the super distance wrapped of unit；(3) bag surpasses the distance of bag more than front to other classification
Both.On this basis, the training of model is carried out.
Many example classification model prediction step particularly as follows:
Step 2.1, the tolerance obtained in utilization measure learning process calculates the many examples bag in test set respectively to classification
Super bag and unit surpass the distance of bag；
Step 2.2, according to the concept classification of calculated range estimation test set many examples bag.
Described from each many examples bag XiThe crucial example that middle extraction is corresponding, particularly as follows: according to multi-instance learning it is assumed that
If bag XiIt is under the jurisdiction of in concept classification c, then XiIn at least an example be subordinate to c；If bag XiIt is not affiliated with in concept classification c, then
XiIn any one example be all not affiliated with c.After utilizing existing crucial example extraction algorithm, wrap XiIn example be i.e. divided into
Crucial example (indicates bag XiConcept classification ownership) and non-key example.
The described crucial example super bag of composition classification that the bag being under the jurisdiction of identical concept classification is extracted, particularly as follows:
After the bag belonging to concept classification c is extracted its crucial example, by these crucial example one set of composition, it is denoted asWhereinFor i-th key example, VcFor being under the jurisdiction of the number of the crucial example of classification c, ScIt is general
Read the super bag of classification corresponding to classification c.Similarly, the super package definition of all non-key examples composition is the super bag of unit, is denoted asWhereinFor the example of the super bag of composition unit, it is the non-key example of i-th in the super bag of unit.
Described bag is specially to the distance of super bag Wherein VkFor super bag SkIn example numbers, C represents the number of classification,ForAt bag XiIn arest neighbors,Table
Show super bag SkIn the v example, MkFor the parameter that obtained by metric learning (i.e. from XiTo SkTolerance).
Described metric learning learns weight, is by the object function of distance one belt restraining of writing of bag to super bag, logical
The solution crossing tradition alternative optimization solves and obtains desired parameters.
Described in model prediction step, the classifier calculated bag obtained during utilizing training, to the distance of super bag, has
Body is: tolerance M that will arrive at model training step learningkBring into I.e. can obtain the bag distance to each super bag.
The described concept classification according to calculated range estimation test set many examples bag, if particularly as follows: bag is to certain class
Chao not wrap ScClosest, then this bag is under the jurisdiction of concept classification c；If bag wraps S to unit is super0Closest, then this bag is under the jurisdiction of
Newly-increased concept classification.
Beneficial effect: compared with prior art, the super bag of utilization structure provided by the present invention and metric learning judge
Know that classification many examples bag detects newly-increased class method for distinguishing simultaneously, implementation process can make full use of flag data training many
Learn-by-example grader, it is adaptable to have the scene that newly-increased classification occurs, can classify by sample datas many to known class, the most permissible
The newly-increased classification of detection.It addition, under the scene occurred without newly-increased classification, the method for the invention still can obtain excellent performance.
Accompanying drawing explanation
Fig. 1 is many example classification model training workflow diagram of the embodiment of the present invention；
Fig. 2 is many example classification model prediction workflow diagram of the embodiment of the present invention.
Detailed description of the invention
Below in conjunction with specific embodiment, it is further elucidated with the present invention, it should be understood that these embodiments are merely to illustrate the present invention
Rather than restriction the scope of the present invention, after having read the present invention, the those skilled in the art's various equivalences to the present invention
The amendment of form all falls within the application claims limited range.
As shown in Figure 1-2, newly-increased category detection method based on multi-instance learning, instruct including multi-instance learning disaggregated model
Practice step and disaggregated model prediction steps；
As it is shown in figure 1, described many example classification model training step particularly as follows:
Step 1.1, on existing many sample datas, utilizes existing crucial example detection algorithm from each many examples bag
XiThe crucial example that middle extraction is corresponding；" crucial example " refers in " bag " that those may decide that showing of the corresponding concept labelling wrapped
Example.
From each many examples bag XiThe crucial example that middle extraction is corresponding, if particularly as follows: according to multi-instance learning it is assumed that bag
XiIt is under the jurisdiction of in concept classification c, then XiIn at least an example be subordinate to c；If bag XiIt is not affiliated with in concept classification c, then XiIn
Any one example is all not affiliated with c.After utilizing existing crucial example extraction algorithm, wrap XiIn example be i.e. divided into key
Example (indicates bag XiConcept classification ownership) and non-key example.
Step 1.2, the crucial example that the bag being under the jurisdiction of identical concept classification is extracted composition " the super bag of classification " Sc(c
=1 ..., C).It addition, be not detected as all examples composition " the super bag of the unit " S of crucial example0；
The crucial example super bag of composition classification that the bag being under the jurisdiction of identical concept classification is extracted, particularly as follows: to same
Belong to after the bag of concept classification c extracts its crucial example, by these crucial examples one set of composition, be denoted asWhereinFor i-th key example, VcFor being under the jurisdiction of the number of the crucial example of classification c, ScIt is general
Read the super bag of classification corresponding to classification c.Similarly, the super package definition of all non-key examples composition is the super bag of unit, is denoted as
Step 1.3, definition bag to the distance of super bag is: the weighted sum of each example in super bag to bag beeline, its
In, the beeline of example to bag is example to the minimum euclidean distance of example in bag, and in super bag, the weight of each example is by spending
Amount study determines according to training data training.
Described bag is specially to the distance of super bag Wherein VkFor super bag SkIn example numbers,ForAt bag XiIn arest neighbors, MkFor being obtained by metric learning
Parameter (i.e. from XiTo SkTolerance), C represents the number of classification,Represent super bag SkIn the v example.
Metric learning learns weight, is by the object function of distance one belt restraining of writing of bag to super bag, by passing
The solution of system alternative optimization solves and obtains desired parameters.
Step 1.4, according to the distance of bag to super bag, utilization measure learns described weight.In order to ensure this system
There are enough discriminant classification abilities and new class power of test, in metric learning, need to meet following condition: (1) bag is to its corresponding class
The distance Chao do not wrapped is minimum；(2) bag is second little to the super distance wrapped of unit；(3) bag surpasses the distance of bag more than front to other classification
Both.On this basis, the training of model is carried out.
As in figure 2 it is shown, many example classification model prediction step particularly as follows:
Step 2.1, the super bag of many examples bag that the tolerance obtained during utilizing training calculates in test set respectively to classification
With the distance that unit surpasses bag；Tolerance M that will arrive at model training step learningkBring into I.e. can obtain the bag distance to each super bag.
Step 2.2, according to the concept classification of calculated range estimation test set many examples bag: if bag surpasses to certain classification
Bag ScClosest, then this bag is under the jurisdiction of concept classification c；If bag wraps S to unit is super0Closest, then this bag is under the jurisdiction of newly-increased
Concept classification.
Claims (7)
1. a newly-increased category detection method based on multi-instance learning, it is characterised in that: include multi-instance learning disaggregated model
Training step and disaggregated model prediction steps；
Described many example classification model training step particularly as follows:
Step 1.1, on existing many sample datas, utilizes existing crucial example detection algorithm from each many examples bag XiIn
The crucial example that extraction is corresponding；
Step 1.2, the crucial example that the bag being under the jurisdiction of identical concept classification is extracted composition " the super bag of classification " Sc(c=
1,…,C)；It addition, be not detected as all examples composition " the super bag of the unit " S of crucial example0；
Step 1.3, definition bag to the distance of super bag is: each example in super bag, to the weighted sum of bag beeline, wherein, is shown
Example is example to the minimum euclidean distance of example in bag to the beeline of bag, and in super bag, the weight of each example is by metric learning
Determine according to training data training；
Step 1.4, according to the distance of bag to super bag, utilization measure learns described weight；In order to ensure that this system has foot
Enough discriminant classification abilities and new class power of test, need to meet following condition: (1) bag surpasses to its corresponding classification in metric learning
The distance of bag is minimum；(2) bag is second to the distance of the super bag of unit；(3) bag surpasses the distance of bag more than the above two to other classification；?
On the basis of this, carry out the training of model；
Many example classification model prediction step particularly as follows:
Step 2.1, the super bag of many examples bag that the tolerance obtained during utilizing training calculates in test set respectively to classification and unit
The distance of super bag；
Step 2.2, according to the concept classification of calculated range estimation test set many examples bag.
2. newly-increased category detection method based on multi-instance learning as claimed in claim 1, it is characterised in that: described from each
Many examples bag XiThe crucial example that middle extraction is corresponding, if particularly as follows: according to multi-instance learning it is assumed that bag XiIt is under the jurisdiction of concept class
In other c, then XiIn at least an example be subordinate to c；If bag XiIt is not affiliated with in concept classification c, then XiIn any one example all
It is not affiliated with c；After utilizing existing crucial example extraction algorithm, wrap XiIn example be i.e. divided into crucial example and non-key show
Example.
3. newly-increased category detection method based on multi-instance learning as claimed in claim 1, it is characterised in that: described will be subordinate to
In the crucial example super bag of composition classification that the bag of identical concept classification is extracted, particularly as follows: to belonging to concept classification c's
After its crucial example of bag extraction, by these crucial example one set of composition, it is denoted asWhereinFor i-th
Crucial example, VcFor being under the jurisdiction of the number of the crucial example of classification c, ScIt is the super bag of classification corresponding for concept classification c；Similarly,
The super package definition of all non-key examples composition is the super bag of unit, is denoted as
4. newly-increased category detection method based on multi-instance learning as claimed in claim 1, it is characterised in that: described bag is to super
The distance of bag is specially Wherein VkFor super bag SKIn
Example numbers,ForAt bag XiIn arest neighbors, MkFor the parameter obtained by metric learning.
5. newly-increased category detection method based on multi-instance learning as claimed in claim 1, it is characterised in that: described tolerance
Practise and learn weight, be by the object function of distance one belt restraining of writing of bag to super bag, by the solution of tradition alternative optimization
Method solves and obtains desired parameters.
6. newly-increased category detection method based on multi-instance learning as claimed in claim 1, it is characterised in that: described at model
In prediction steps, the classifier calculated bag obtained during utilizing training is to the distance of super bag, particularly as follows: will walk at model training
Tolerance M that rapid learning arriveskBring into The most available
Bag is to the distance of each super bag.
7. newly-increased category detection method based on multi-instance learning as claimed in claim 1, it is characterised in that: described according to meter
The concept classification of the range estimation test set many examples bag obtained, if particularly as follows: bag wraps S to certain classification is supercClosest,
Then this bag is under the jurisdiction of concept classification c；If bag wraps S to unit is super0Closest, then this bag is under the jurisdiction of newly-increased concept classification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610600041.3A CN106250924B (en) | 2016-07-27 | 2016-07-27 | A kind of newly-increased category detection method based on multi-instance learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610600041.3A CN106250924B (en) | 2016-07-27 | 2016-07-27 | A kind of newly-increased category detection method based on multi-instance learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106250924A true CN106250924A (en) | 2016-12-21 |
| CN106250924B CN106250924B (en) | 2019-07-16 |
Family
ID=57604236
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610600041.3A Active CN106250924B (en) | 2016-07-27 | 2016-07-27 | A kind of newly-increased category detection method based on multi-instance learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106250924B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109192314A (en) * | 2018-08-27 | 2019-01-11 | 山东大学 | A kind of cervical vertebra health score assigning device and its application based on multi-instance learning |
| CN110377636A (en) * | 2019-07-09 | 2019-10-25 | 南京工业大学 | A kind of Outliers Detection algorithm based on more example models |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1480158A1 (en) * | 2003-05-21 | 2004-11-24 | Microsoft Corporation | Systems and methods for adaptive handwriting recognition |
| CN102509084A (en) * | 2011-11-18 | 2012-06-20 | 中国科学院自动化研究所 | Multi-examples-learning-based method for identifying horror video scene |
| CN103325125A (en) * | 2013-07-03 | 2013-09-25 | 北京工业大学 | Moving target tracking method based on improved multi-example learning algorithm |
| CN104091038A (en) * | 2013-04-01 | 2014-10-08 | 太原理工大学 | Method for weighting multiple example studying features based on master space classifying criterion |
| CN104361059A (en) * | 2014-11-03 | 2015-02-18 | 中国科学院自动化研究所 | Harmful information identification and web page classification method based on multi-instance learning |
| CN104778457A (en) * | 2015-04-18 | 2015-07-15 | 吉林大学 | Video face identification algorithm on basis of multi-instance learning |
| CN105045907A (en) * | 2015-08-10 | 2015-11-11 | 北京工业大学 | Method for constructing visual attention-label-user interest tree for personalized social image recommendation |
| CN105117429A (en) * | 2015-08-05 | 2015-12-02 | 广东工业大学 | Scenario image annotation method based on active learning and multi-label multi-instance learning |
-
2016
- 2016-07-27 CN CN201610600041.3A patent/CN106250924B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1480158A1 (en) * | 2003-05-21 | 2004-11-24 | Microsoft Corporation | Systems and methods for adaptive handwriting recognition |
| CN102509084A (en) * | 2011-11-18 | 2012-06-20 | 中国科学院自动化研究所 | Multi-examples-learning-based method for identifying horror video scene |
| CN104091038A (en) * | 2013-04-01 | 2014-10-08 | 太原理工大学 | Method for weighting multiple example studying features based on master space classifying criterion |
| CN103325125A (en) * | 2013-07-03 | 2013-09-25 | 北京工业大学 | Moving target tracking method based on improved multi-example learning algorithm |
| CN104361059A (en) * | 2014-11-03 | 2015-02-18 | 中国科学院自动化研究所 | Harmful information identification and web page classification method based on multi-instance learning |
| CN104778457A (en) * | 2015-04-18 | 2015-07-15 | 吉林大学 | Video face identification algorithm on basis of multi-instance learning |
| CN105117429A (en) * | 2015-08-05 | 2015-12-02 | 广东工业大学 | Scenario image annotation method based on active learning and multi-label multi-instance learning |
| CN105045907A (en) * | 2015-08-10 | 2015-11-11 | 北京工业大学 | Method for constructing visual attention-label-user interest tree for personalized social image recommendation |
Non-Patent Citations (6)
| Title |
|---|
| GUOQING LIU: "Key Instance Detection in Multi-Instance Learning", 《JMLR: WORKSHOP AND CONFERENCE PROCEEDINGS》 * |
| XIU-SHEN WEI: "Scalable Multi-instance Learning", 《2014 IEEE INTERNATIONAL CONFERENCE ON DATA MINING》 * |
| YUE ZHU: "Learning with Augmented Multi-Instance View", 《JMLR: WORKSHOP AND CONFERENCE PROCEEDINGS》 * |
| 温超: "构建新包空间的多示例学习方法", 《西安交通大学学报》 * |
| 胡聪: "多示例学习在多ROI乳腺肿瘤分类中的应用", 《智能计算机与应用》 * |
| 邹晓川: "基于强模态辅助的弱模态图像概念分类及检测", 《南京大学学报》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109192314A (en) * | 2018-08-27 | 2019-01-11 | 山东大学 | A kind of cervical vertebra health score assigning device and its application based on multi-instance learning |
| CN109192314B (en) * | 2018-08-27 | 2021-08-31 | 山东大学 | Cervical vertebra health scoring device based on multi-example learning and application thereof |
| CN110377636A (en) * | 2019-07-09 | 2019-10-25 | 南京工业大学 | A kind of Outliers Detection algorithm based on more example models |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106250924B (en) | 2019-07-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Kong et al. | Deep sequential context networks for action prediction | |
| CN106022237B (en) | A kind of pedestrian detection method of convolutional neural networks end to end | |
| CN104200237B (en) | One kind being based on the High-Speed Automatic multi-object tracking method of coring correlation filtering | |
| CN104732208B (en) | Video human Activity recognition method based on sparse subspace clustering | |
| CN104330721B (en) | IC Hardware Trojan detecting method and system | |
| CN104992177A (en) | Internet porn image detection method based on deep convolution nerve network | |
| CN112580664A (en) | Small target detection method based on SSD (solid State disk) network | |
| CN107247956A (en) | A kind of fast target detection method judged based on grid | |
| CN107247938A (en) | A kind of method of high-resolution remote sensing image City Building function classification | |
| CN107316049A (en) | A kind of transfer learning sorting technique based on semi-supervised self-training | |
| CN109636846B (en) | Target positioning method based on cyclic attention convolution neural network | |
| CN106228109A (en) | A kind of action identification method based on skeleton motion track | |
| CN104537355A (en) | Remarkable object detecting method utilizing image boundary information and area connectivity | |
| CN113378656B (en) | Action recognition method and device based on self-adaptive graph convolution neural network | |
| CN109829428A (en) | Based on the video image pedestrian detection method and system for improving YOLOv2 | |
| CN107679362A (en) | The recognition methods of compound protein interaction affinity, system and device | |
| CN106250924A (en) | A kind of newly-increased category detection method based on multi-instance learning | |
| CN110008819A (en) | A kind of facial expression recognizing method based on figure convolutional neural networks | |
| CN112148875A (en) | Dispute detection method based on graph convolution neural network integration content and structure information | |
| Ma et al. | Location-aware box reasoning for anchor-based single-shot object detection | |
| CN105183659B (en) | Software systems abnormal behavior detection method based on multilevel mode prediction | |
| CN107491735B (en) | Tag and interaction relation joint learning method for human behavior recognition | |
| CN103258212A (en) | Semi-supervised integrated remote-sensing image classification method based on attractor propagation clustering | |
| CN105160285A (en) | Method and system for recognizing human body tumble automatically based on stereoscopic vision | |
| CN114743273A (en) | Human skeleton behavior identification method and system based on multi-scale residual error map convolutional network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB02 | Change of applicant information |
Address after: 210008 Hankou Road, Drum Tower District, Nanjing, Jiangsu Province, No. 22 Applicant after: Nanjing University Address before: 210046 Xianlin Avenue 163, Qixia District, Nanjing City, Jiangsu Province Applicant before: Nanjing University |
|
| CB02 | Change of applicant information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |