CN107885854A - A kind of semi-supervised cross-media retrieval method of feature based selection and virtual data generation - Google Patents
A kind of semi-supervised cross-media retrieval method of feature based selection and virtual data generation Download PDFInfo
- Publication number
- CN107885854A CN107885854A CN201711124618.9A CN201711124618A CN107885854A CN 107885854 A CN107885854 A CN 107885854A CN 201711124618 A CN201711124618 A CN 201711124618A CN 107885854 A CN107885854 A CN 107885854A
- Authority
- CN
- China
- Prior art keywords
- msub
- mrow
- data
- msup
- feature
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
- G06F16/43—Querying
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The present invention proposes a kind of semi-supervised cross-media retrieval method of feature based selection and virtual data generation.This method proposes to generate some virtual data points according to the feature of training data to expand training data, while uses l during two pairs of projection matrixes are learnt2,1Norm carries out feature selecting.Specifically, its class center is asked to every a kind of image and text first, then generates new data point at random around it, form new training data；Then, two pairs of projection matrixes are learnt using new data, at the same time, using l2,1Norm constraint, carry out feature selecting；Finally retrieval result is assessed with mAP values.This method not only generates some random number strong points to improve the diversity of training data, while some can be selected more to have the feature of distinction and abundant information when two pairs of projection matrixes are learnt.Experimental result on three different pieces of information collection also indicates that the superiority of the method.
Description
Technical field
The present invention relates to cross-media retrieval method, is given birth to more specifically to a kind of selection of feature based and virtual data
Into semi-supervised cross-media retrieval method.
Background technology
With the development of multimedia technology, increasing data can be expressed as different mode, and different modalities
Data may have same semantic information.Therefore, how to explore these has identical semantic but shows as different modalities number
Relation between becomes particularly important.Wherein, it is nearest for many years in, cross-media retrieval technology increasingly causes researcher's
Concern.Cross-media retrieval just refers to retrieve other with identical semantic information by the use of a kind of data of mode as inquiry data
The data of mode.By taking the retrieval of picture and text as an example, picture can be used to go text of the retrieval with corresponding semantic information, letter
Claim：I2T；Or picture of the retrieval with corresponding semantic information is removed using text, referred to as：T2I.The present invention is between picture and text
Retrieval exemplified by analyzed and tested, but the retrieval that the method can be expanded between other different modalities.
In cross-media retrieval technology, most important problem is：The data of different modalities have different character representations, this
Different a bit to be characterized in different dimensional spaces, the similitude between such isomeric data cannot directly compare.Cause
This, the problem of cross-media retrieval field is primarily upon, is how to cross over this semantic gap.A kind of popular solution method is just
It is sub-space learning method.Sub-space learning method is intended to learn a potential semantic space, in this potential semantic space
The similitude of isomeric data can be directly measured.Traditional sub-space learning method is a pair of projection matrixes of study, passes through this
The data of different modalities are mapped in a potential semantic space by projection matrix can, such isomeric data it is similar
Property can is measured.A kind of popular method is：Canonical correlation analysis (Canonical Correlation
Analysis, CCA), CCA has learnt a pair of projection matrixes, is maximized when by the Feature Mapping of different modalities to semantic space
Correlation between isomeric data.Based on CCA, semantic relevant matches (Semantic Correlation Match, SCM) use
Logistic regression obtains semantic space.Another popular method is：Offset minimum binary (Partial Least Squares,
PLS), PLS is intended to learn two potential semantic spaces by maximizing the correlation between isomeric data.In addition,
Normalize multi-view analysis (generalized multi-view analysis, GMA) and GMLDA based on GMA and
GMMFA obtains the feature of multi-angle by using label information, and achieves more preferable effect.
However, there is directionality, i.e. image retrieval text (I2T) and text retrieval image in common cross-media retrieval task
(T2I), above method only learns a pair of projection matrixes, does not emphasize to inquire about the importance of data.Specifically, in I2T
In task, picture is more decisive for study projection matrix；Similarly, in T2I tasks, then the importance of text is more emphasized.
Therefore, the method for learning a pair of projection matrixes is extremely difficult to optimal effect.In order to emphasize that the weight of data is inquired about in different task
The property wanted, cross-media retrieval (Modality-dependent Cross-media Retrieval, MDCR) method based on mode
Two pairs of projection matrixes of study are proposed, i.e., a pair of projection matrixes is learnt respectively to I2T and T2I tasks, can thus taken into full account
The importance of data is inquired about, therefore the precision retrieved also is greatly improved.
But above method all simply has the method for supervision, is only trained and have ignored using only markd data
Unlabelled data, while cannot more expand intrinsic data set.Secondly, current method only from how to measure isomeric data it
Between the angle of similitude set out, it is intended to learn more effective projection matrix, it is more accurate always to be obtained in semantic space
Comparison, still, they all have ignored when learning projection matrix it is more rich to information content, more distinction feature choosing
Select.Therefore, we have invented a kind of semi-supervised method that can generate virtual data point at random based on MDCR, while use l2,1
Norm carries out feature selecting.
The content of the invention
The invention provides the semi-supervised cross-media retrieval technology of a kind of selection of feature based and pseudo-random data generation.Pass
The cross-media retrieval method of system, otherwise it is only to be used only to have what flag data was trained to have measure of supervision, otherwise it is to select one
Part Unlabeled data adds the semi-supervised method of training.The present invention propose, generated in markd data basis some with
Correlation pseudorandom virtual data point, so not only it is contemplated that unlabelled data, it is related that some can also be increased
Virtual data point is to improve the precision of training.Meanwhile unlike traditional search method：Our method is for different
The different projection matrix of tasking learning, l is used when learning projection matrix2,1Norm carries out feature selecting.On the whole, we
Method simultaneously consider the diversity of training data and the selection of validity feature.
The concrete technical scheme of the present invention is as follows：
A kind of semi-supervised cross-media retrieval technology of feature based selection and virtual data generation, comprises the following steps：
Step 1：Data-oriented collectionN represents the sum of data pair, xiRepresentative picture feature, tiRepresent text
Eigen, then, picture and text feature matrix can be expressed as：XG=[x1,x2,...,xn-1,xn] and TG=[t1,
t2,t3,...,tn-1,tn]；
Step 2：Pseudorandom virtual data point is generated, legacy data collection is expanded, specific method is：Calculate XGAnd TG
Per a kind of class center, i.e., for every a kind of data, calculate average of such data per dimension, obtained each dimension average structure
Into it is new vector as such class center；Then, using the average of every dimension as center, at it, generation n' is individual at random up and down
Numerical value, the sets of random values in all dimensions is combined n' pseudorandom virtual data of generation, by the puppet
Random virtual data point adds the data set G after legacy data collection is expandedall={ G, G'}, picture and text after expansion
Eigenmatrix is expressed as：X=[x1,...xn,x1',x'2...x'n] and T=[t1,...,tn,t1',t'2,...,t'n]；
Step 3：Build object function：
Objective function：
Wherein, U, V represent this method a pair of projection matrixes to be learnt, and C (U, V) is correlation analysis item so that multimode
The data of state can be kept into potential semantic space to neighbor relationships；L (U, V) is from image or text modality feature
Space to semantic space linear regression item, for keep with identical semantic different modalities data neighbor relationships；N(U,
V it is) regular terms, the selection for feature；
According to formula (1), image inspection text I2T and the object function of text inspection image T2I retrieval tasks are respectively obtained, such as
Under：
(1) I2T object function is：
Wherein, U1,V1It is the projection matrix that learn to obtain in I2T tasks, the U corresponded respectively in formula (1), V, β are
Coefficient of balance and 0≤β≤1, Y are semantic matrixes；
(2) T2I object function is：
Wherein, U2,V2It is the projection matrix that learn to obtain in T2I tasks, the U, V corresponded respectively in formula (1)；
Step 4：By iterative method, optimal projection matrix is obtained：
Because formula (2) and (3) are non-convex, therefore use the method for control variable to solve, i.e., local derviation asked to U and V respectively,
And make it be equal to zero, projection matrix U and V value can be obtained；Then pass through continuous iteration, until convergence, try to achieve projection matrix U and
V optimal value.
Especially, in step 3, N (U, V)=λ1||U||2,1+λ2||V||2,1, wherein λ1, λ2For balancing two canonicals
, and all be positive number, this bound term is used to select when projection matrix learn more with distinction and abundant information
Feature.
Brief description of the drawings
Fig. 1 is the inventive method flow chart.
Embodiment
1. data set is handled：
Wikipedia, altogether comprising 10 classes, 2866 picture-texts pair.We select 2173 picture-texts pair
As initial training data, remainder is test data.Its picture feature be 4096 dimension CNN features, text feature 100
Tie up LDA features.
Pascal Sentence, 20 classes altogether, per 50 picture-texts pair of class.We select 30 figures in every class
Picture-text is to as initial training data, remaining is test data.Its picture feature be 4096 dimension CNN features, text feature
For 100 dimension LDA features.
INRIA-Websearch, altogether 353 class, 71478 image-texts pair.We randomly choose therein 70% and made
For initial training data, remaining is test data.Its picture feature is the CNN features of 4096 dimensions, and text feature is 1000 dimension LDA
Feature.
2. specific implementation step of the present invention：
Step 1：Data-oriented collectionN represents the sum of data pair, xiRepresentative picture feature, tiRepresent text
Eigen, then, picture and text feature matrix can be expressed as：XG=[x1,x2,...,xn-1,xn] and TG=[t1,
t2,t3,...,tn-1,tn]。
Step 2：Pseudorandom virtual data point is generated, legacy data collection is expanded, specific method is：Calculate XGAnd TG
Per a kind of class center, i.e., for every a kind of data, calculate average of such data per dimension, obtained each dimension average structure
Into it is new vector as such class center；Then, using the average of every dimension as center, at it, generation n' is individual at random up and down
Numerical value, the sets of random values in all dimensions is combined n' pseudorandom virtual data of generationBy described in
Pseudorandom virtual data point adds the data set G after legacy data collection is expandedall={ G, G'}, picture and text after expansion
Eigen matrix is expressed as：X=[x1,...xn,x1',x'2...x'n] and T=[t1,...,tn,t′1,t′2,...,t
′n]。
Step 3：Build object function：
Objective function：
Wherein, U, V represent this method a pair of projection matrixes to be learnt, and C (U, V) is correlation analysis item so that multimode
The data of state can be kept into potential semantic space to neighbor relationships；L (U, V) is from image or text modality feature
Space to semantic space linear regression item, for keep with identical semantic different modalities data neighbor relationships；N(U,
V it is) regular terms, the selection for feature；
According to formula (1), image inspection text I2T and the object function of text inspection image T2I retrieval tasks are respectively obtained, such as
Under：
(1) I2T object function is：
Wherein, U1,V1It is the projection matrix that learn to obtain in I2T tasks, β is coefficient of balance and 0≤β≤1, Y are semantemes
Matrix, and：N(U1,V1)=λ1||U1||2,1+λ2||V1||2,1, wherein λ1, λ2For balancing two regular terms, and all it is positive number；
(2) T2I object function is：
Wherein, U2,V2It is the projection matrix N (U that learn to obtain in T2I tasks2,V2)=λ1||U2||2,1+λ2||V2|
|2,1；
Step 4：By iterative method, optimal projection matrix is obtained：
Because formula (2) and (3) are non-convex, therefore use the method for control variable to solve, i.e., local derviation asked to U and V respectively,
And make it be equal to zero, projection matrix U and V value can be obtained；Then pass through continuous iteration, until convergence, try to achieve projection matrix U and
V optimal value.
Especially, for l2,1Norm can using mark come derivation, such as：Matrix U is defined, then：||U||2,1=Tr (UTRU),
R is a diagonal matrix,uiU every a line is represented, ε is a minimum real number.
3. evaluation criteria (mAP)
We use mean accuracy average (mean average precision, mAP) evaluation criteria, last to evaluate
Retrieval effectiveness.We define the average precision inquired about each time first：
Wherein, N represents the sum of sample in test data, has and corresponding category when in the sort result of ith retrieval
Rel (i)=1 when signing identical, otherwise rel (i)=0.P (i) represents the precision of ith retrieval ordering result.So, all inquiries
AP values be averagely exactly mAP.
4. algorithm is realized
(1)I2T:
Input：Picture feature matrix XGWith text feature matrix TG, sample labeling matrix Y, parameter lambda1, λ2, β
Generate virtual data：For every a kind of data, the average per dimension is calculated first, as in this kind of class
The heart,
Then, using the average of every dimension as center, in its n' numerical value of generation at random up and down, in all dimensions with
Machine value, which is combined, can be formed by the virtual data of n', finally, the virtual data of generation add input picture and
Text feature matrix, obtain new training picture feature matrix X and text feature matrix T.
Initialization：Initial projection matrix U1, V1For unit matrix.
Solve optimal solution：According to the U tried to achieve1=(XXT+λ1R11)-1[βXY+(1-β)XTTV1] and
V1=[(1- β) TTT+λ2R12]-1(1-β)TXTU1, by constantly iteration, until result restrain to obtain it is optimal
U1, V1。
This process false code is as follows：
(2)T2I：
It is similar with I2T tasks, finally obtain optimal projection matrix U2, V2
5. results contrast
We are tested on three data sets respectively, and compared for other current popular 7 kinds of methods (PLS,
CCA, SM, SCM, GMMFA, GMLDA, MDCR), following table shows that the inventive method all shows preferably to examine on different pieces of information collection
Suo Xiaoguo.
Claims (2)
1. a kind of feature based selection and the semi-supervised cross-media retrieval technology of virtual data generation, comprise the following steps：
Step 1：Data-oriented collectionN represents the sum of data pair, xiRepresentative picture feature, tiRepresent text spy
Sign, then, picture and text feature matrix can be expressed as：XG=[x1,x2,...,xn-1,xn] and TG=[t1,t2,
t3,...,tn-1,tn]；
Step 2：Pseudorandom virtual data point is generated, legacy data collection is expanded, specific method is：Calculate XGAnd TGIt is each
The class center of class, i.e., for every a kind of data, average of such data per dimension is calculated, what obtained each dimension average was formed
New vector is as such class center；Then, using the average of every dimension as center, n' number is generated at random up and down at it
Value, the sets of random values in all dimensions is combined n' pseudorandom virtual data of generation By the puppet
Random virtual data point adds the data set G after legacy data collection is expandedall={ G, G'}, picture and text after expansion
Eigenmatrix is expressed as：X=[x1,...xn,x1',x'2...x'n] and T=[t1,...,tn,t1',t'2,...,t'n]；
Step 3：Build object function：
Objective function：
Wherein, U, V represent this method a pair of projection matrixes to be learnt, and C (U, V) is correlation analysis item so that multi-modal
Data can be kept into potential semantic space to neighbor relationships；L (U, V) is from image or text modality feature space
To the linear regression item of semantic space, for keeping the neighbor relationships with identical semantic different modalities data；N (U, V) is
Regular terms, the selection for feature；
According to formula (1), image inspection text I2T and the object function of text inspection image T2I retrieval tasks are respectively obtained, it is as follows：
(1) I2T object function is：
<mrow>
<munder>
<mrow>
<mi>m</mi>
<mi>i</mi>
<mi>n</mi>
</mrow>
<mrow>
<msub>
<mi>U</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>V</mi>
<mn>1</mn>
</msub>
</mrow>
</munder>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>U</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>V</mi>
<mn>1</mn>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mi>&beta;</mi>
<mo>|</mo>
<mo>|</mo>
<msup>
<mi>X</mi>
<mi>T</mi>
</msup>
<msub>
<mi>U</mi>
<mn>1</mn>
</msub>
<mo>-</mo>
<mi>Y</mi>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
<mo>+</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<mi>&beta;</mi>
<mo>)</mo>
</mrow>
<mo>|</mo>
<mo>|</mo>
<msup>
<mi>X</mi>
<mi>T</mi>
</msup>
<msub>
<mi>U</mi>
<mn>1</mn>
</msub>
<mo>-</mo>
<msup>
<mi>T</mi>
<mi>T</mi>
</msup>
<msub>
<mi>V</mi>
<mn>1</mn>
</msub>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
<mo>+</mo>
<mi>N</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>U</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>V</mi>
<mn>1</mn>
</msub>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, U1,V1It is the projection matrix that learn to obtain in I2T tasks, the U corresponded respectively in formula (1), V, β are balances
Coefficient and 0≤β≤1, Y are semantic matrixes；
(2) T2I object function is：
<mrow>
<munder>
<mrow>
<mi>m</mi>
<mi>i</mi>
<mi>n</mi>
</mrow>
<mrow>
<msub>
<mi>U</mi>
<mn>2</mn>
</msub>
<mo>,</mo>
<msub>
<mi>W</mi>
<mn>2</mn>
</msub>
</mrow>
</munder>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>U</mi>
<mn>2</mn>
</msub>
<mo>,</mo>
<msub>
<mi>V</mi>
<mn>2</mn>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mi>&beta;</mi>
<mo>|</mo>
<mo>|</mo>
<msup>
<mi>T</mi>
<mi>T</mi>
</msup>
<msub>
<mi>V</mi>
<mn>2</mn>
</msub>
<mo>-</mo>
<mi>Y</mi>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
<mo>+</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<mi>&beta;</mi>
<mo>)</mo>
</mrow>
<mo>|</mo>
<mo>|</mo>
<msup>
<mi>X</mi>
<mi>T</mi>
</msup>
<msub>
<mi>U</mi>
<mn>2</mn>
</msub>
<mo>-</mo>
<msup>
<mi>T</mi>
<mi>T</mi>
</msup>
<msub>
<mi>V</mi>
<mn>2</mn>
</msub>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>F</mi>
<mn>2</mn>
</msubsup>
<mo>+</mo>
<mi>N</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>U</mi>
<mn>2</mn>
</msub>
<mo>,</mo>
<msub>
<mi>V</mi>
<mn>2</mn>
</msub>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>3</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, U2,V2It is the projection matrix that learn to obtain in T2I tasks, the U, V corresponded respectively in formula (1)；
Step 4：By iterative method, optimal projection matrix is obtained：
Because formula (2) and (3) are non-convex, therefore use the method for control variable to solve, i.e., local derviation is asked to U and V respectively, and make
It is equal to zero, can obtain projection matrix U and V value；Then continuous iteration is passed through, until convergence, tries to achieve projection matrix U and V
Optimal value.
2. feature based selection as claimed in claim 1 and the semi-supervised cross-media retrieval technology of virtual data generation, it is special
Sign is：In step 3, N (U, V)=λ1||U||2,1+λ2||V||2,1, wherein λ1, λ2For balancing two regular terms, and all it is
Positive number, this bound term are used for the feature that selection when projection matrix is learnt more has distinction and abundant information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711124618.9A CN107885854B (en) | 2017-11-14 | 2017-11-14 | semi-supervised cross-media retrieval method based on feature selection and virtual data generation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711124618.9A CN107885854B (en) | 2017-11-14 | 2017-11-14 | semi-supervised cross-media retrieval method based on feature selection and virtual data generation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107885854A true CN107885854A (en) | 2018-04-06 |
| CN107885854B CN107885854B (en) | 2020-01-31 |
Family
ID=61776703
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711124618.9A Active CN107885854B (en) | 2017-11-14 | 2017-11-14 | semi-supervised cross-media retrieval method based on feature selection and virtual data generation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107885854B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109784405A (en) * | 2019-01-16 | 2019-05-21 | 山东建筑大学 | Cross-module state search method and system based on pseudo label study and semantic consistency |
| CN109857892A (en) * | 2018-12-29 | 2019-06-07 | 西安电子科技大学 | Semi-supervised cross-module state Hash search method based on category transmitting |
| CN112419324A (en) * | 2020-11-24 | 2021-02-26 | 山西三友和智慧信息技术股份有限公司 | Medical image data expansion method based on semi-supervised task driving |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105069173A (en) * | 2015-09-10 | 2015-11-18 | 天津中科智能识别产业技术研究院有限公司 | Rapid image retrieval method based on supervised topology keeping hash |
| US9332137B2 (en) * | 2012-09-28 | 2016-05-03 | Interactive Memories Inc. | Method for form filling an address on a mobile computing device based on zip code lookup |
| CN106462642A (en) * | 2014-06-24 | 2017-02-22 | 谷歌公司 | Methods, Systems And Media For Performing Personalized Actions On Mobile Devices Associated With A Media Presentation Device |
-
2017
- 2017-11-14 CN CN201711124618.9A patent/CN107885854B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9332137B2 (en) * | 2012-09-28 | 2016-05-03 | Interactive Memories Inc. | Method for form filling an address on a mobile computing device based on zip code lookup |
| CN106462642A (en) * | 2014-06-24 | 2017-02-22 | 谷歌公司 | Methods, Systems And Media For Performing Personalized Actions On Mobile Devices Associated With A Media Presentation Device |
| CN105069173A (en) * | 2015-09-10 | 2015-11-18 | 天津中科智能识别产业技术研究院有限公司 | Rapid image retrieval method based on supervised topology keeping hash |
Non-Patent Citations (1)
| Title |
|---|
| YUNCHAO WEI等: "Modality-Dependent Cross-Media Retrieval", 《ACM TRANSACTIONS ON INTELLIGENT SYSTEMS AND TECHNOLOGY》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109857892A (en) * | 2018-12-29 | 2019-06-07 | 西安电子科技大学 | Semi-supervised cross-module state Hash search method based on category transmitting |
| CN109857892B (en) * | 2018-12-29 | 2022-12-02 | 西安电子科技大学 | Semi-supervised cross-modal Hash retrieval method based on class label transfer |
| CN109784405A (en) * | 2019-01-16 | 2019-05-21 | 山东建筑大学 | Cross-module state search method and system based on pseudo label study and semantic consistency |
| CN109784405B (en) * | 2019-01-16 | 2020-09-08 | 山东建筑大学 | Cross-modal retrieval method and system based on pseudo-tag learning and semantic consistency |
| CN112419324A (en) * | 2020-11-24 | 2021-02-26 | 山西三友和智慧信息技术股份有限公司 | Medical image data expansion method based on semi-supervised task driving |
| CN112419324B (en) * | 2020-11-24 | 2022-04-19 | 山西三友和智慧信息技术股份有限公司 | Medical image data expansion method based on semi-supervised task driving |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107885854B (en) | 2020-01-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chen et al. | Msdn: Mutually semantic distillation network for zero-shot learning | |
| CN107402993B (en) | The cross-module state search method for maximizing Hash is associated with based on identification | |
| CN111639679B (en) | Small sample learning method based on multi-scale metric learning | |
| CN106649715B (en) | A kind of cross-media retrieval method based on local sensitivity hash algorithm and neural network | |
| CN106339756A (en) | Training data generation method and device and searching method and device | |
| CN104573669A (en) | Image object detection method | |
| CN110097095B (en) | Zero sample classification method based on multi-view generation countermeasure network | |
| CN109871454B (en) | Robust discrete supervision cross-media hash retrieval method | |
| CN111666427A (en) | Entity relationship joint extraction method, device, equipment and medium | |
| CN112418351B (en) | Zero sample learning image classification method based on global and local context sensing | |
| CN107885854A (en) | A kind of semi-supervised cross-media retrieval method of feature based selection and virtual data generation | |
| CN114186084B (en) | Online multi-mode Hash retrieval method, system, storage medium and equipment | |
| CN104424254A (en) | Method and device for obtaining similar object set and providing similar object set | |
| Nascimento et al. | A robust indoor scene recognition method based on sparse representation | |
| CN114117153A (en) | Online cross-modal retrieval method and system based on similarity relearning | |
| Chen et al. | An Improved Deep Fusion CNN for Image Recognition. | |
| Liu et al. | Incorporating domain and sentiment supervision in representation learning for domain adaptation | |
| CN105740917B (en) | The semi-supervised multiple view feature selection approach of remote sensing images with label study | |
| Ren et al. | Improving cross-domain recommendation through probabilistic cluster-level latent factor model | |
| CN105701225A (en) | Cross-media search method based on unification association supergraph protocol | |
| CN114780723B (en) | Portrayal generation method, system and medium based on guide network text classification | |
| Xu et al. | DeMT: Deformable mixer transformer for multi-task learning of dense prediction | |
| Dou et al. | Function-words enhanced attention networks for few-shot inverse relation classification | |
| CN104036021A (en) | Method for semantically annotating images on basis of hybrid generative and discriminative learning models | |
| CN106250918A (en) | A kind of mixed Gauss model matching process based on the soil-shifting distance improved |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |