CN106022389B - A kind of related feedback method actively selecting more example multiple labeling digital pictures - Google Patents
A kind of related feedback method actively selecting more example multiple labeling digital pictures Download PDFInfo
- Publication number
- CN106022389B CN106022389B CN201610369363.1A CN201610369363A CN106022389B CN 106022389 B CN106022389 B CN 106022389B CN 201610369363 A CN201610369363 A CN 201610369363A CN 106022389 B CN106022389 B CN 106022389B
- Authority
- CN
- China
- Prior art keywords
- image
- label
- user
- feedback
- correlation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- 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
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
- G06F18/2178—Validation; Performance evaluation; Active pattern learning techniques based on feedback of a supervisor
-
- 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
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
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 related feedback methods for actively selecting more example multiple labeling digital pictures.More example Multi-label learnings are the new machine learning frames of one kind proposed in recent years, and application of succeeding in many practical problems, realistic task can be advantageously applied to by inputting the automatic image annotation technology indicated based on more example multiple labelings.But with the enhancing of its ability to express, image automatic annotation method increased dramatically the demand of the training sample marked with becoming larger for representation space.The present invention is by combining more example Multi-label learnings and active learning techniques in machine learning, under the premise of not increasing user annotation cost, more fine mark information abundant is obtained during relevant feedback each time, to which a greater degree of lifting system marks precision, the participation cost of user is effectively reduced.
Description
Technical field
The invention belongs to digital picture automatic marking technical fields, and in particular to a kind of actively to select more example multiple labeling numbers
The related feedback method of word image.
Background technique
With the prevalence of the universal and all kinds of social network sites of digital product, digital picture becomes in more and more internets
The carrier of appearance.In order to efficiently utilize these digital image datas, it is to allow calculating mechanism that it is also most difficult task that one most crucial
The semanteme of image is solved, and automatic image annotation is then key technology therein.Current automatic image annotation technology often will figure
As being expressed as single example, and it is concerned only with the correlation of image with single semantic classes.But image often has complexity
Semanteme includes multiple object entities, the expression that example list single in this way marks will cause information loss, can not accurate description image
Semanteme, thus can not Accurate Prediction image tagged.More efficient method is based on more example multiple labeling machine learning (abbreviation MIML)
Frame, indicate piece image with the set that multiple examples of features form, and simultaneously forecast image and multiple semantic markers it
Between correlation.There are a small number of automatic image annotation technologies indicated based on MIML input at present, but they are to the instruction marked
The demand for practicing sample increased dramatically with becoming larger for representation space, and the marked data in realistic task often have very much
Limit, causes these technologies to be difficult to be widely used.One solution is to introduce Relevance Feedback, allows human user automatic
Some useful mark informations are provided in annotation process to improve system mark precision.Active Learning Method in machine learning
The sample that most worthy can be selected is marked to user query, so as to reduce the participation cost of user as far as possible.It is existing
Feedback Mechanism based on Active Learning selects the image of most worthy and requries the users the image and some semantic category
It is whether not related.The information that this feedback system obtains is very limited, the image that cannot be adapted under more example multiple labeling expressions
Automatic marking problem.
Summary of the invention
Goal of the invention: in order to overcome the deficiencies in the prior art, the present invention, which provides, a kind of actively selects more examples more
The related feedback method of reference numerals image,
Technical solution: to achieve the above object, the technical solution adopted by the present invention are as follows:
A kind of related feedback method actively selecting more example multiple labeling digital pictures, comprising the following steps:
(1) every image in database is calculated to score to the value of marking model, and select to score maximum image as
Candidate image X*；
(2) current marking model is calculated for the palm of all labels not fed back on image selected in step (1)
Degree is held, and selects the corresponding label of minimum value as candidates y*；
(3) by the candidate image X selected in step (1) (two)*With candidates y*It is supplied to user, allows user
Relevance feedback is provided；
(4) it is directed to the candidate image X*And candidates y*, according to including including whether related, most associated exemplary
User feedback result carries out gradient decline and updates marking model；
(5) return step (one) or terminate and export marking model.
Further, the step (1) image scores to the value of marking model method particularly includes:
Assuming that a total of K class label, uses Yi=[yi1..., yiK]TIndicate the i-th width image XiLabel vector, if this
Image is related to k-th of label, then yik=1, otherwise yik=-1；
Every image X in the image collection U without user feedback is calculated according to formulaiTo the value scoring C of marking model1
(Xi), the formula are as follows:
Wherein, p records image XiPredicted vectorIn 1 number,
I [a]=1 when a is true, otherwise I [a]=0,
Card () indicates the number of element in a set,
ξ ∈ (0,1).
Further, the step (2) selects candidates y*Method particularly includes:
Marking model is calculated to the tag set U (X without user feedback according to formula*) in label y Grasping level C2
(X*, y), the formula are as follows:
AC2(X*, y)=| fy(X*)-fyo(X*)|
Wherein, fy(X)=maxJ=1 ..., mf(xj) indicate image X and mark the degree of correlation of y,
y0For virtual tag, i.e., according to fy(X*) sequence of value from big to small is to U (X*) in tag sort, be located at image X*'s
Label between mark of correlation and uncorrelated label is y0；Grasping level C2(X*, y) and label corresponding to minimum value, use y*Table
Show.
Further, the step (4) user feedback result method particularly includes:
Firstly, user judges X*With y*It is whether related: if user judges X*With y*Correlation then needs to point out image X*In with label
y*Maximally related example, is denoted as x*, fed back for several times, image X*Mark of correlation set be denoted as Y+, uncorrelated tag set is denoted as
Y-, by y*It is added to Y+, y will be marked later*From X*Non- feedback flag set U (X*) in remove；If user judges X*With y*Not phase
It closes, then by y*It is added to Y-, y will be marked*From X*Non- feedback flag set U (X*) in remove；
Then, judge U (X*) in element number whether be zero: if it is, by X*Never tag image set U is moved into
Tag set L, is updated marking model；Otherwise directly it is updated marking model.
Further, marking model is updated in the step (4) method particularly includes:
Step 151: start program；
Step 152: initialization intermediate variable r=0,I=1, wherein r records image X*With label y*It is whether related,
Indicate related when r=1, s records image X*In with y*Maximally related example, i record are not put back to sampled images X*
Uncorrelated tag set Y-Number；
Step 153: the judging result of user is recorded, to r assignment；
Step 154: whether the value for judging r is 1, and 155 are entered step if r=1, otherwise enters step 159；
Step 155: by the most associated exemplary x of user feedback*It is assigned to s；
Step 156: assuming that image X*Share m example { x1..., xm, s corresponds to kth ∈ { 1 ..., m } a example, calculatesIts value sorts from large to small, wherein fyIt (x) is the degree of correlation of example x and label y, fy(x)=
Wy TVx, WyIndicate the y column of W；
Step 157: judging whether there is xjSo thatIt sets up, if it is, entering step 158, otherwise
Enter step 159；
Step 158: the y of model variable W is updated according to formula*Column and model variable V；The formula difference is as follows:
Wherein, P (X*, x*, y*) indicate to sort in step 156 to be located atCorresponding example number before；
Step 159: stochastical sampling X*A mark of correlation y；
Step 1510: judging whether the value of intermediate variable i is less than or equal to image X*Uncorrelated label number t=| Y-|,
If it is, entering step 1511, otherwise terminate；
Step 1511: one image X of stochastical sampling*Uncorrelated label
Step 1512: judgementIt is whether true, if set up, 1513 are entered step, otherwise directly
Enter step 1514；
Step 1513: model variable is updated according to formula, the formula difference is as follows:
Wherein,It indicates and image X*Degree of correlation greater than flag y label number, x andRespectively
Indicate image X*Label y andOn representative example；
Step 1514: i will be assigned to again after i plus 1, later return step 1510.
The utility model has the advantages that the related feedback method provided by the invention for actively selecting more example multiple labeling digital pictures, in conjunction with
More example Multi-label learnings and active learning techniques in machine learning, it is each under the premise of not increasing user annotation cost
More fine mark information abundant is obtained during secondary relevant feedback, so that a greater degree of lifting system marks precision.
It is looked into specifically, the present invention picks out every time for promoting the most helpful sub-picture of annotation equipment precision and a label
It askes, and user is required to make distinctive feedback.When selecting image and label, the present invention has combined annotation equipment currently
The difference degree of other images in information and candidate image and whole image database through grasping, picks out most worthy
One width candidate image and a label, allow user discriminatively to feed back.When selected candidate image is related to label, use
Family be also pointed out that in image with the maximally related example of the label, believe to provide more fine image to annotation equipment
Breath；Otherwise user need to only tell that image selected by annotation equipment is uncorrelated to label.Due to determining that image is relevant to marking whether
The process of process actually implicit user determination most associated exemplary, therefore such feedback mechanism can't be brought to user
Additional cost.To not only ensure that the lower participation of user, but also it can greatly improve mark precision.And in order to abundant
Ground utilizes this fine feedback information, more example Multi-label learning frames in present invention combination machine learning, so that for
Any piece image in training data set, before relevant label should be come uncorrelated label by marking model, simultaneously
For any one label of image, marking model should be come maximally related example before other examples.If current mould
Before mark of correlation is come uncorrelated label by type, or before most associated exemplary comes most, then not more new model, otherwise to mould
Shape parameter carries out gradient decline, so that the sequence of mark of correlation and most associated exemplary shifts to an earlier date, without mark of correlation and other shows
Example sequence can then step back.
Detailed description of the invention
Fig. 1 is digital picture automatic marking device work flow diagram；
Fig. 2 is the flow chart of trained initial marking model；
Fig. 3 is the flow chart of mechanism of the present invention；
Fig. 4 is the flow chart for updating marking model.
Specific embodiment
The present invention will be further explained with reference to the accompanying drawing.
Embodiment
It is as shown in Figure 1 digital picture automatic marking device work flow diagram.Assuming that training image data acquisition system is by two
It is grouped as, a part is to have marked, it is assumed that shared N1Width image, is indicated with L；Another part is anti-without user annotation
Feedback, it is assumed that shared N2Width image, is indicated with U.Device meets more example Multi-label learnings to the image zooming-out in data acquisition system
The feature of input, every piece image are indicated that each feature vector is known as an example by one group of feature vector.Feature extraction can be with
Applicable characteristics of image is generated using the classical way in machine learning textbook, such as first carries out image segmentation, then to each
Image block extracts the features such as color, texture, shape.First according to the feature vector of image in L (assuming that using XLIndicate) and figure
Picture is with the relevant information between label (assuming that using YL, being worth when being 1 indicates related, being worth indicates uncorrelated when being -1) and training is initial
Marking model, detailed process are shown in Fig. 2.Later, device carries out same feature extraction to the unmarked image of input, and with marking
Model predicted, output token result.If user is satisfied to annotation results, annotation process terminates；Otherwise transposition will
It selects several images and label to be supplied to user, allows user to judge correlation and provide distinguishing feedback.Mark dress
The information update model using user feedback is set, is recycled hence into a new round.
Fig. 2 show the flow chart of the initial marking model of training.By step 01, step 02 carries out just marking model
Beginningization mainly includes assigning initial value to two matrix Ws and V.Assuming that the characteristics of image dimension extracted is d, a total of L possible marks
Note, then W is the matrix of 100 × L size, and V is the matrix of 100 × d size.Here numerical value 100 can basis
User needs to change into other values, and mark accuracy can be improved by generally selecting biggish value, and lesser value can then accelerate speed.W
Random value is endowed with V and guarantees that the mean value of their every a line is 0, and standard deviation isStep 03 is from training data set
Middle stochastical sampling piece image X, it is assumed that its mark of correlation collection is combined into Y, and uncorrelated tag set isDevice is randomly choosed from Y
One label, it is assumed that be y, and be according to the following formula y from set X={ x1..., xmIn select representative example x:
X=argmax fy(xi), i=1...m (1)
Wherein, fy(x)=Wy TVx, WyIndicate the y column of W, fy(x) it can be understood as example x and mark the degree of correlation of y.
Uncorrelated tag set of the step 04 from this imageIn one by one stochastical sampling label, encountered until in the v times sampling
One is come uncorrelated label before yRepresentative example is determined also according to formula (1) for itThen to be calculated with following formula
Value come estimate ranking be located at label y before uncorrelated label total number:
WhereinForIn label number,Symbol indicates to be rounded downwards.Step 05 carry out gradient decline respectively according to
Formula (3) updates the y column of W, by the of formula (4) update WColumn, and V is updated by formula (5).
Whether step 06 judgment models reach requirement, are, terminate training process, otherwise return to step 03.Here judge mould
Whether type, which reaches standard, can reach use using common method in machine learning or pattern-recognition textbook, such as iteration wheel number
The specified number in family.
More example Multi-label learnings are the new machine learning frames of one kind proposed in recent years, in many practical problems
It succeeds application, the automatic image annotation technology indicated based on the input of more example multiple labelings can be advantageously applied to reality and appoint
Business.But with the enhancing of its ability to express, image automatic annotation method is to the demand of the training sample marked with table
Show becoming larger for space and increased dramatically.The present invention is by combining more example Multi-label learnings and Active Learning skill in machine learning
Art obtains more fine label abundant under the premise of not increasing user annotation cost each time during relevant feedback
Information effectively reduces the participation cost of user so that a greater degree of lifting system marks precision.
Fig. 3 show image labeling Feedback Mechanism of the invention.Step 1 is origination action.In step 2, it is assumed that total
Shared K class label, then can be used Yi=[yi1..., yiK]TIndicate the i-th width image XiLabel vector, if this image with k-th
Label is related, then yik=1, otherwise yik=-1.Every image X in the image collection U without user feedback is calculated according to formula (6)i
To the value scoring C of marking model f1(Xi):
Wherein, p records image XiPredicted vectorIn 1 number, that is, current marking model predict XiRelated mark
Remember number.I [a]=1 when a is true, otherwise I [a]=0, that is to say, that q is indicated
The average mark of correlation number of each image in labeled good image collection L.Card () indicates element in a set
Number, so K-card (U (Xi)) it is image XiBy the label number of user feedback, and ξ ∈ (0,1) be in order to avoid
The constant that denominator is zero in formula (6).It is worth it should be noted that C1(Xi) value it is bigger, then molecule is bigger, and denominator is smaller.
Molecule is bigger, i.e. image XiPrediction mark of correlation number and marked image average mark of correlation difference it is bigger, illustrate to mark
Injection molding type is higher to the uncertainty degree of the image.Denominator is bigger, i.e. image XiThe number fed back is fewer, illustrates that it may be wrapped
Containing more unknown messages.In brief, with the image X of maximum value*, maximum otherness and uncertainty are imply, is answered
This is selected as candidate image.The value marking and queuing that step 3 calculates step 2 sorts according to sequence from big to small.
Step 4 selects image corresponding to maximum scores, uses X*It indicates.Step 5 is according to formula (7) computation model to without user feedback
Tag set U (X*) in label y Grasping level C2(X*, y):
C2(X*, y)=| fy(X*)-fy0(X*)| (7)
Wherein, fy(X)=maxJ=1 ..., mf(xj) indicate image X and mark the degree of correlation of y.y0Referred to as virtual tag,
I.e. according to fy(X*) sequence of value from big to small is to U (X*) in tag sort, be located at image X*Mark of correlation and uncorrelated label
Between label be y0.Therefore, closer apart from virtual tag, i.e. fy(X*) and fy0(X*) difference it is smaller, illustrate model to label
Y is more uncertain, it should be selected as candidates.The Grasping level that step 6 calculates step 5 sorts, according to from small
It sorts to big sequence.Step 7 selects label corresponding to minimum value, uses y*It indicates.Step 8 selects step 4 and step 7
X*And y*It is presented to the user, user is allowed to judge image X*Whether with label y*It is related.If user judges X*With y*Correlation then enters
Step 9, user need to point out image X*In with label y*Maximally related example, is denoted as x*.Assuming that by repeatedly feedback, image X*'s
Mark of correlation set is denoted as Y+, uncorrelated tag set is denoted as Y-.Step 10 is by y*It is added to Y+, 12 are entered step later.If
User judges X*With y*It is uncorrelated, then 11 are entered step by y*It is added to Y-.Step 12 will mark y*From X*Non- feedback flag collection
Close U (X*) in remove.Step 13 judges U (X*) in element number whether be zero, i.e. image X*All labels fed back
It crosses.If it is, 14 are entered step, by X*Never tag image set U moves into marked set L, enters step 15 later；It is no
It is then directly entered step 15, updates marking model, detailed process is shown in Fig. 4.Step 16 terminates Feedback Mechanism.
It is the detailed process of step 15 in Fig. 3 shown in Fig. 4.By step 151, step 152 initializes intermediate variable r=
0,I=1, wherein r records selected image X in Fig. 3*Whether with label y*Correlation indicates related when r=1.s
Record image X*In with y*Maximally related example.I record is not put back to sampled images X*Uncorrelated tag set Y-Time
Number.Step 153 records the judging result of user, to r assignment.Step 154 judges whether the value of r is 1.If r=1 enters
Step 155 is by the most associated exemplary x of user feedback*It is assigned to s；Otherwise 159 are entered step.In step 156, it is assumed that image X*
Share m example { x1..., xm}.It calculates first(assuming that s corresponds to kth ∈ { 1 ..., m } a example),
Then it sorts according to the sequence of its value from big to small.Step 157 judges whether there is xjSo thatIt sets up, i.e., extremely
There is an example x lessj(when there are the example of multiple violation sequences xjIndicate it is therein any one), it and label y*'s
Degree of correlation ratio X*Most associated exemplary s and label y*Degree of correlation it is also big.This obviously violates the target in the present invention, will be every
The most associated exemplary of width image comes foremost.If it is present entering step 158 y for updating model variable W according to formula (8)*
Column update model variable V by formula (9)；Otherwise 159 are entered step.
Wherein, P (X*, x*, y*) indicate to sort in step 156 to be located atCorresponding example number before is disobeyed
The example total number of reverse-order.Step 159 stochastical sampling X*A mark of correlation y.Step 1510 judges the value of intermediate variable i
Whether image X is less than or equal to*Uncorrelated label number t=| Y-|.If set up, 1511 are entered step, stochastical sampling one
A image X*Uncorrelated labelOtherwise terminate.Step 1512 judgementIt is whether true.If set up,
1513 are entered step, updates model variable according to formula (10)-(12)；Otherwise it is directly entered step 1514.
Wherein,It indicates and image X*Degree of correlation greater than flag y label number, x andRespectively
Indicate image X*Label y andOn representative example.Step 1514 is assigned to i again after i is added 1, later return step
1510。
The above is only a preferred embodiment of the present invention, it should be pointed out that: for the ordinary skill people of the art
For member, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also answered
It is considered as protection scope of the present invention.
Claims (4)
1. a kind of related feedback method for actively selecting more example multiple labeling digital pictures, it is characterised in that: the following steps are included:
(1) it calculates every image in database to score to the value of marking model, and the maximum image that selects to score is as candidate
Image X*；
(2) current marking model is calculated for the grasp journey of all labels not fed back on image selected in step (1)
Degree, and select the corresponding label of minimum value as candidates y*；
(3) by the candidate image X selected in step (1) (two)*With candidates y*It is supplied to user, user is allowed to provide
Relevance feedback；
(4) it is directed to the candidate image X*And candidates y*, according to including whether user related, including most associated exemplary
Feedback result carries out gradient decline and updates marking model；Wherein update marking model method particularly includes:
Step 151: start program；
Step 152: initialization intermediate variable r=0,I=1, wherein r records image X*With label y*It is whether related, when and
Only indicate related as r=1, s records image X*In with y*Maximally related example, i record are not put back to sampled images X*No
Mark of correlation set Y-Number；
Step 153: the judging result of user is recorded, to r assignment；
Step 154: whether the value for judging r is 1, and 155 are entered step if r=1, otherwise enters step 159；
Step 155: by the most associated exemplary x of user feedback*It is assigned to s；
Step 156: assuming that image X*Share m example { x1..., xm, s corresponds to kth ∈ { 1 ..., m } a example, calculatesIts value sorts from large to small, wherein fyIt (x) is the degree of correlation of example x and label y, fy(x)=
Wy TVx, WyIndicate the y column of W；
Step 157: judging whether there is xjSo thatIt sets up, if it is, entering step 158, otherwise enters
Step 159；
Step 158: the y of model variable W is updated according to formula*Column and model variable V；The formula difference is as follows:
Wherein, P (X*, x*, y*) indicate to sort in step 156 to be located atCorresponding example number before；
Step 159: stochastical sampling X*A mark of correlation y；
Step 1510: judging whether the value of intermediate variable i is less than or equal to image X*Uncorrelated label number t=| Y- |, if
It is then to enter step 1511, otherwise terminates；
Step 1511: one image X of stochastical sampling*Uncorrelated label
Step 1512: judgementIt is whether true, if set up, 1513 are entered step, step is otherwise directly entered
Rapid 1514；
Step 1513: model variable is updated according to formula, the formula difference is as follows:
Wherein,It indicates and image X*Degree of correlation greater than flag y label number, x andIt respectively indicates
Image X*Label y andOn representative example；
Step 1514: i will be assigned to again after i plus 1, later return step 1510；
(5) return step (one) or terminate and export marking model.
2. the related feedback method according to claim 1 for actively selecting more example multiple labeling digital pictures, feature exist
In: the step (1) image scores to the value of marking model method particularly includes:
Assuming that a total of K class label, uses Yi=[yi1..., yiK]TIndicate the i-th width image XiLabel vector, if this image with
K-th of label is related, then yik=1, otherwise yik=-1；
Every image X in the image collection U without user feedback is calculated according to formulaiTo the value scoring C of marking model1(Xi),
The formula are as follows:
Wherein, p records image XiPredicted vectorIn 1 number,
I [a]=1 when a is true, otherwise I [a]=0,
Card () indicates the number of element in a set,
ξ ∈ (0,1).
3. the related feedback method according to claim 1 for actively selecting more example multiple labeling digital pictures, feature exist
In: the step (2) selects candidates y*Method particularly includes:
Marking model is calculated to the tag set U (X without user feedback according to formula*) in label y Grasping level C2(X*,
Y), the formula are as follows:
C2(X*, y)=| fy(X*)-fy0(X*)|
Wherein, fy(X)=maxJ=1..., mf (xj) indicate image X and mark the degree of correlation of y,
y0For virtual tag, i.e., according to fy(X*) sequence of value from big to small is to U (X*) in tag sort, be located at image X*Correlation
Label between label and uncorrelated label is y0；Grasping level C2(X*, y) and label corresponding to minimum value, use y*It indicates.
4. the related feedback method according to claim 1 for actively selecting more example multiple labeling digital pictures, feature exist
In: the step (4) user feedback result method particularly includes:
Firstly, user judges X*With y*It is whether related: if user judges X*With y*Correlation then needs to point out image X*In with label y*Most
Relevant example, is denoted as x*, fed back for several times, image X*Mark of correlation set be denoted as Y+, uncorrelated tag set is denoted as Y-,
By y*It is added to Y+, y will be marked later*From X*Non- feedback flag set U (X*) in remove；If user judges X*With y*Not phase
It closes, then by y*It is added to Y-, y will be marked*From X*Non- feedback flag set U (X*) in remove；
Then, judge U (X*) in element number whether be zero: if it is, by X*Never tag image set U moves into marked
Set L, is updated marking model；Otherwise directly it is updated marking model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610369363.1A CN106022389B (en) | 2016-05-30 | 2016-05-30 | A kind of related feedback method actively selecting more example multiple labeling digital pictures |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610369363.1A CN106022389B (en) | 2016-05-30 | 2016-05-30 | A kind of related feedback method actively selecting more example multiple labeling digital pictures |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106022389A CN106022389A (en) | 2016-10-12 |
| CN106022389B true CN106022389B (en) | 2019-05-14 |
Family
ID=57091789
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610369363.1A Active CN106022389B (en) | 2016-05-30 | 2016-05-30 | A kind of related feedback method actively selecting more example multiple labeling digital pictures |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106022389B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108009589A (en) * | 2017-12-12 | 2018-05-08 | 腾讯科技（深圳）有限公司 | Sample data processing method, device and computer-readable recording medium |
| CN113139568B (en) * | 2021-02-22 | 2022-05-10 | 杭州深睿博联科技有限公司 | Class prediction model modeling method and device based on active learning |
| CN113313178B (en) * | 2021-06-03 | 2024-05-03 | 南京航空航天大学 | Cross-domain image example level active labeling method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101226521A (en) * | 2008-02-18 | 2008-07-23 | 南京大学 | Machine learning method for ambiguity data object estimation modeling |
| CN103116893A (en) * | 2013-03-15 | 2013-05-22 | 南京大学 | Digital image labeling method based on multi-exampling multi-marking learning |
| CN103198333A (en) * | 2013-04-15 | 2013-07-10 | 中国科学院电子学研究所 | Automatic semantic labeling method of high resolution remote sensing image |
| CN105069481A (en) * | 2015-08-19 | 2015-11-18 | 西安电子科技大学 | Multi-label natural scene classification method based on spatial pyramid and sparse coding |
-
2016
- 2016-05-30 CN CN201610369363.1A patent/CN106022389B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101226521A (en) * | 2008-02-18 | 2008-07-23 | 南京大学 | Machine learning method for ambiguity data object estimation modeling |
| CN103116893A (en) * | 2013-03-15 | 2013-05-22 | 南京大学 | Digital image labeling method based on multi-exampling multi-marking learning |
| CN103198333A (en) * | 2013-04-15 | 2013-07-10 | 中国科学院电子学研究所 | Automatic semantic labeling method of high resolution remote sensing image |
| CN105069481A (en) * | 2015-08-19 | 2015-11-18 | 西安电子科技大学 | Multi-label natural scene classification method based on spatial pyramid and sparse coding |
Non-Patent Citations (1)
| Title |
|---|
| Active Query Driven by Uncertainty and Diversity for Incremental Multi-Label Learning;Sheng-Jun Huang 等;《ICDM13》;20140203;第1-6页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106022389A (en) | 2016-10-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110598203B (en) | Method and device for extracting entity information of military design document combined with dictionary | |
| CN107247947A (en) | Face character recognition methods and device | |
| CN109165385A (en) | Multi-triple extraction method based on entity relationship joint extraction model | |
| CN106776538A (en) | The information extracting method of enterprise's noncanonical format document | |
| CN105069129B (en) | Adaptive multi-tag Forecasting Methodology | |
| CN110188197B (en) | Active learning method and device for labeling platform | |
| CN106355446B (en) | A kind of advertisement recommender system of network and mobile phone games | |
| CN106815194A (en) | Model training method and device and keyword recognition method and device | |
| CN109710741A (en) | A kind of mask method the problem of study based on deeply towards online answer platform | |
| CN103116893B (en) | Digital image labeling method based on multi-exampling multi-marking learning | |
| CN106022389B (en) | A kind of related feedback method actively selecting more example multiple labeling digital pictures | |
| CN110097094A (en) | It is a kind of towards personage interaction multiple semantic fusion lack sample classification method | |
| CN108804577B (en) | Method for estimating interest degree of information tag | |
| CN107679110A (en) | The method and device of knowledge mapping is improved with reference to text classification and picture attribute extraction | |
| CN107346327A (en) | The zero sample Hash picture retrieval method based on supervision transfer | |
| CN110287952A (en) | A kind of recognition methods and system for tieing up sonagram piece character | |
| CN111539452A (en) | Image recognition method and device for multitask attributes, electronic equipment and storage medium | |
| CN106228139A (en) | A kind of apparent age prediction algorithm based on convolutional network and system thereof | |
| CN107239564A (en) | A kind of text label based on supervision topic model recommends method | |
| CN106777336A (en) | A kind of exabyte composition extraction system and method based on deep learning | |
| CN106951565A (en) | File classification method and the text classifier of acquisition | |
| CN110019703A (en) | Data markers method and device, intelligent answer method and system | |
| CN103020286B (en) | Based on the internet ranking list grasping system of rank website | |
| CN113590810A (en) | Abstract generation model training method, abstract generation device and electronic equipment | |
| CN106874397A (en) | A kind of automatic semanteme marking method of internet of things oriented equipment |
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 | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20161012 Assignee: Jiangsu Fenghuang Intelligent Education Research Institute Co.,Ltd. Assignor: Nanjing University of Aeronautics and Astronautics Contract record no.: X2020980003896 Denomination of invention: Related feedback method for actively selecting multi-instance multi-mark digital image Granted publication date: 20190514 License type: Common License Record date: 20200708 |
|
| EE01 | Entry into force of recordation of patent licensing contract |