CN109871454A - A kind of discrete across media Hash search methods of supervision of robust - Google Patents
A kind of discrete across media Hash search methods of supervision of robust Download PDFInfo
- Publication number
- CN109871454A CN109871454A CN201910096204.2A CN201910096204A CN109871454A CN 109871454 A CN109871454 A CN 109871454A CN 201910096204 A CN201910096204 A CN 201910096204A CN 109871454 A CN109871454 A CN 109871454A
- Authority
- CN
- China
- Prior art keywords
- sample
- image
- similarity matrix
- indicate
- text
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention discloses a kind of discrete across the media Hash search methods of supervision of robust, similarity matrix excavates the semantic association between isomery sample between sample two-by-two by learning a robust, the cross-media retrieval based on content can be realized using this method, method includes the following steps: establishing image and text data set, and vision and text feature are extracted respectively to the image and samples of text of data concentration;Similarity matrix between sample two-by-two is constructed respectively using class label, image and the text feature of sample, and learns using the sparse characteristic of the low-rank of similarity matrix and sample noise between sample two-by-two similarity matrix between the sample two-by-two of a robust;And then learn the better Hash codes of distinction using similarity matrix between the robust two-by-two sample;Hash function is appliedNorm canonical item constraint, to learn the hash function of more robust;It proposes a kind of discrete iteration optimization algorithm, directly obtains the discrete solution of Hash codes;Similarity matrix can effectively resist noise that may be present in sample to one robust of study of the method for the present invention between sample two-by-two, to greatly improve the performance of multimedia retrieval.
Description
Technical field:
The present invention relates to a kind of discrete supervision cross-module state Hash search methods of robust, belong to multimedia retrieval and machine learning
Field.
Background technique:
In recent years, a large amount of data can be all generated daily on internet, this brings huge to multimedia retrieval task
Challenge, how efficiently and effectively to search approximate sample becomes urgent need.Hash method is by learning one group of hash function for sample
This is mapped to Hamming space from original feature space, since its calculating speed in large-scale application is empty with saving storage fastly
Between, cause the great concern of researcher.Hash codes are more much lower than the carrying cost of primitive character, while passing through Hamming space
The middle similarity that can be rapidly calculated using XOR operation between sample.Extensive research has been obtained in hash method, but big
Majority research is concerned only with a kind of mode, however the sample of identical semanteme can be typically expressed as multiple mode on the internet, this leads
Cause the isomery semantic gap between different modalities.For example, image can be by vision and corresponding Text Representation.In addition, working as
When user submits query sample to search engine, user prefers to search for the similar sample that engine returns to multiple modalities.Therefore, across
Media retrieval causes more and more concerns.Target across media hash method is that isomery sample is mapped to a shared Chinese
Prescribed space, and the similar structure of sample is kept in this space.Specifically, for similar isomery sample, in shared Hamming space
Middle Hamming distance wants small, and vice versa.According to class label whether is used in the training process, across media hash methods usually can be with
It is divided into two classes: unsupervised and measure of supervision.Similitude learns to breathe out between mode in the former mode usually by keeping sample
Uncommon code, and the latter can be further combined with the better Hash codes of dividing property of class label learning region.Nearest work shows in conjunction with sample
Class label retrieval performance can be improved.
Although many supervision cross-module state hash methods are it has been proposed that and achieve satisfactory as a result, however still having one
A little problems need further to solve.Firstly, sample may contain noise in real world.But most of supervision cross-module states
The class label configurations of training data similarity matrix between sample two-by-two is used only in hash method, without considering making an uproar in sample
Sound, such as: outlier.Obviously, these noise samples can seriously damage the structure of similarity matrix between sample two-by-two, to mislead
The study of Hash codes, causes retrieval performance to reduce.Secondly, the discrete constraint of Hash codes leads to mixed integer optimization problem usually very
It is difficult to resolve certainly, most methods loosen the discrete constraint of Hash codes first, obtain continuous solution, and then quantization generates Hash codes.However,
Quantization will lead to information loss, so that the differentiation reduced performance of Hash codes.
Summary of the invention:
One kind is provided it is an object of the invention to overcome the shortcomings of above-mentioned prior art with the better Hash of learning performance
Code, the performance of boosting algorithm improve the separating capacity of Hash codes preferably to resist noise, are suitable for real network data
Cross-media retrieval the discrete supervision cross-module state Hash search method of robust.
The purpose of the present invention can provide the measure that reach: a kind of discrete supervision cross-module state Hash retrieval side of robust
Method, which is characterized in that method includes the following steps:
Step 1: collecting image and samples of text pair containing class label, image, the one-to-one cross-module state of text are constituted
The image, text and data collection of retrieval;
Step 2: respectively to image and text modality sample extraction feature, and respectively to image and text modality sample
Feature goes mean value, makes the characteristic mean value 0 of two mode samples;
Step 3: being training set and test set to random division by all samples in data set;
Step 4: constructing two respectively using the sample characteristics of the class label of sample pair, image and text modality in training set
Similarity matrix between two samples, and the low-rank characteristic of similarity matrix and the sparse characteristic of noise sample between sample two-by-two are utilized,
Learn similarity matrix between the sample two-by-two an of robust;The feature of training sample pair is set as X, X={ X(1), X(2), wherein X(1)
Indicate the sample characteristics of image modalities in training set, X(2)Indicate the sample characteristics of text modality in training set,Wherein d1And d2Respectively indicate image and text
The dimension of mode sample characteristics, N indicate that image or text modality sample size in training set, the class label of sample pair are indicated with L,C indicates the quantity of sample class, li∈ { 0,1 }cIf lij=1, indicate that i-th of sample belongs to
Jth class;, whereas if lij=0, indicate that i-th of sample is not belonging to jth class;Learn robust similarity matrix between sample two-by-two
Objective function the following steps are included:
(1) similarity moment between the sample two-by-two based on image modalities feature is calculated using the sample characteristics of image modalities
Battle array, is defined as follows:
Wherein | | | |FIndicate Frobenius norm, S(1)Similarity matrix between the sample two-by-two of expression image modalities,Indicate the similarity of i-th of image pattern and j-th of image pattern, σ1For scale parameter;
(2) sample characteristics of text modality is utilized to calculate similarity matrix between the sample two-by-two based on text modality feature,
It is defined as follows:
Wherein S(2)Similarity matrix between the sample two-by-two of expression text modality,Indicate i-th of samples of text and j-th
The similarity of samples of text, σ2For scale parameter;
(3) similarity matrix calculating the sample two-by-two based on class label using the class label of sample between, is defined as follows:
Wherein S(3)Indicate sample to the similarity matrix two-by-two of label,Indicate i-th of sample to label and j-th of sample
This similarity to label;
(4) objective function of similarity matrix is defined as follows study robust between sample two-by-two:
s.t.S(i)=S+ | | E(i)||0
Wherein S indicates similarity matrix two-by-two between the robust sample of study, E(i)It indicates i-th two-by-two in similarity matrix
Noise, the order of rank () representing matrix, | | | |0Indicate l0Norm;
(5) since there are discrete low-rank and l for the objective function in above-mentioned (4)0The constraint of norm, so problem is difficult directly
It solves, the two constraint conditions can be loosened, obtain the approximate solution of problem, institute's above formula can be rewritten as
s.t.S(i)=S+ | | E(i)||1
Wherein | | | |*Indicate nuclear norm, | | | |1Indicate l1Norm,
(6) this problem is solved using augmented vector approach, obtain robust similarity matrix between sample two-by-two;
Step 5: construction objective function, specifically includes the following steps:
(1) similitude based on similarity matrix between robust two-by-two sample is kept in Hamming space, and due to image text
This sample is identical to class label, therefore their distance should be as small as possible, so the objective function of Hash codes study is defined as follows:
Wherein k indicates the length of Hash codes, B1For the Hash codes of image modalities sample, B2For the Hash of text modality sample
Code, λ is weight parameter;
(2) using Linear Mapping as hash function, and l is utilized2,1Norm constrains image and text modality as regular terms
The study of hash function, to enhance it to antimierophonic ability, therefore the objective function definition of each mode hash function study is such as
Under:
Wherein W1, W2The hash function of image modalities and text modality is respectively indicated, Reg () indicates that regular terms prevented
Fitting, hereinβiIt is weight parameter with μ;
(3) objective function that Hash codes and hash function learn is added as to the objective function of this method, is defined as follows:
Wherein βiFor weight parameter;
Step 6: objective function is very since objective function includes the discrete constraint of multiple known variables and Hash codes
Solution hard to find, but can be found through observation, it is convex optimization problem when fixing its dependent variable and solving wherein some variable, therefore
Can use iteration optimization algorithms solution, solution procedure the following steps are included:
(1) fixed W1, W2And B2, solve B1:
Constant term is removed, objective function is writeable are as follows:
Due to B1Be it is discrete, problem is difficult direct solution, can be solved with sample-by-sample herein, enable b1iIndicate B1I-th column,
b2jIndicate B2Jth column, removal constant term objective function it is writeable are as follows:
This problem is still difficult direct solution, is solved herein using cyclic coordinate gradient descent method by bit, if b1imTable
Show b1iM bit,Indicate b1iThe vector that other bits other than m bit are constituted, then b1imIt can be obtained by following formula:
Above formula is repeated until having solved the Hash codes of all image modalities samples;
(2) fixed W1, W2And B1, solve B2:
With solution B1It is similar, it can obtain
Above formula is repeated until having solved the Hash codes of all text modality samples;
(3) fixed W2, B1And B2, solve W1:
Constant term is removed, objective function is writeable are as follows:
There are closed solutions for this problem
Wherein D1For diagonal matrix,
(4) fixed W1, B1And B2, solve W2:
With solution W1It is similar, W2There are closed solutions
Wherein D2For diagonal matrix,
(5) (1)-(4) are repeated to algorithmic statement or reach maximum number of iterations;
Step 7: user input query sample, extracts its feature, and go mean value to the feature of extraction;
Step 8: generating the Hash codes of query sample using the hash function learnt:
Step 9: calculating the Hamming distance of query sample and target (training) concentration isomery sample, and Hamming distance is pressed
Ascending order arrangement, the corresponding sample of preceding r Hamming distance is search result.
The present invention can produce following good effect compared with the prior art: the method for the present invention by by class label, image and
Similarity matrix between the sample two-by-two of one frame one robust of study of feature involvement of text modality, it is better with learning performance
Hash codes, the performance of boosting algorithm;It proposes to apply l2,1Study of the norm as canonical item constraint hash function, preferably to support
Antinoise;A kind of discrete optimization algorithm is proposed, discrete Hash codes can be directly obtained, improves the separating capacity of Hash codes,
The invention is suitable for the cross-media retrieval of real network data.
Detailed description of the invention:
Fig. 1 is the flow chart of the discrete supervision cross-module state Hash search method of robust of the present invention.
Specific embodiment:
To be more clearly understood that technical solution of the present invention, the present invention is further retouched in detail below in conjunction with specific embodiment
It states, and is not the limitation to its protection scope.
Embodiment: a kind of discrete supervision cross-module state Hash search method of robust comprising following steps:
Step 1: collecting image and samples of text pair containing class label, image, the one-to-one cross-module state of text are constituted
The image, text and data collection of retrieval;
Step 2: extracting the feature of image and text, wherein image modalities sample is indicated with the textural characteristics of 150 dimensions, text
BOW (Bag Of Words) character representation of 500 dimensions of this mode sample, and mean value is gone to feature, make two mode samples
Characteristic mean value is 0;
Step 3: being training set and test set to random division by all samples in data set;
Step 4: constructing two respectively using the sample characteristics of the class label of sample pair, image and text modality in training set
Similarity matrix between two samples, and the low-rank characteristic of similarity matrix and the sparse characteristic of noise sample between sample two-by-two are utilized,
Learn similarity matrix between the sample two-by-two an of robust;The feature of training sample pair is set as X, X={ X(1), X(2), wherein X(1)
Indicate the sample characteristics of image modalities in training set, X(2)Indicate the sample characteristics of text modality in training set,Wherein d1And d2Respectively indicate image and text
The dimension of mode sample characteristics, N indicate that image or text modality sample size in training set, the class label of sample pair are indicated with L,C indicates the quantity of sample class, li∈ { 0,1 }cIf lij=1, indicate that i-th of sample belongs to
Jth class;, whereas if lij=0, indicate that i-th of sample is not belonging to jth class;D herein1=150, d2=500;
Learn robust two-by-two between sample similarity matrix objective function the following steps are included:
(1) similarity moment between the sample two-by-two based on image modalities feature is calculated using the sample characteristics of image modalities
Battle array, is defined as follows:
Wherein | | | |FIndicate Frobenius norm, S(1)Similarity matrix between the sample two-by-two of expression image modalities,
Indicate the similarity of i-th of image pattern and j-th of image pattern, σ1For scale parameter, σ herein1=0.8;
(2) sample characteristics of text modality is utilized to calculate similarity matrix between the sample two-by-two based on text modality feature,
It is defined as follows:
Wherein S(2)Similarity matrix between the sample two-by-two of expression text modality,Indicate i-th of samples of text and j-th
The similarity of samples of text, σ2For scale parameter, σ herein2=0.3;
(3) similarity matrix calculating the sample two-by-two based on class label using the class label of sample between, is defined as follows:
Wherein S(3)Indicate sample to the similarity matrix two-by-two of label,Indicate i-th of sample to label and j-th of sample
This similarity to label;
(4) objective function of similarity matrix is defined as follows study robust between sample two-by-two:
s.t.S(i)=S+ | | E(i)||0
Wherein S indicates similarity matrix two-by-two between the robust sample of study, E(i)It indicates i-th two-by-two in similarity matrix
Noise, the order of rank () representing matrix, | | | |0Indicate l0Norm;
(5) since there are discrete low-rank and l for the objective function in above-mentioned (4)0The constraint of norm, so problem is difficult directly
It solves, the two constraint conditions can be loosened, obtain the approximate solution of problem, institute's above formula can be rewritten as
s.t.S(i)=S+ | | E(i)||1
Wherein | | | |*Indicate nuclear norm, | | | |1Indicate l1Norm,
(6) this problem is solved using augmented vector approach, obtain robust similarity matrix between sample two-by-two;
Step 5: construction objective function, specifically includes the following steps:
(1) similitude based on similarity matrix between robust two-by-two sample is kept in Hamming space, and due to image text
This sample is identical to class label, therefore their distance should be as small as possible, so the objective function of Hash codes study is defined as follows:
Wherein k indicates the length of Hash codes, B1For the Hash codes of image modalities sample, B2For the Hash of text modality sample
Code, λ is weight parameter, herein λ=1;
(2) using Linear Mapping as hash function, and l is utilized2,1Norm constrains image and text modality as regular terms
The study of hash function, to enhance it to antimierophonic ability, therefore the objective function definition of each mode hash function study is such as
Under:
Wherein W1, W2The hash function of image modalities and text modality is respectively indicated, Reg () indicates that regular terms prevented
Fitting, hereinβiIt is weight parameter, herein β with μ1=10, β2=10, μ=
0.1:
(3) objective function that Hash codes and hash function learn is added as to the objective function of this method, is defined as follows:
Step 6: objective function is very since objective function includes the discrete constraint of multiple known variables and Hash codes
Solution hard to find, but can be found through observation, it is convex optimization problem when fixing its dependent variable and solving wherein some variable, therefore
Can use iteration optimization algorithms solution, solution procedure the following steps are included:
(1) fixed W1, W2And B2, solve B1:
Constant term is removed, objective function is writeable are as follows:
Due to B1Be it is discrete, problem is difficult direct solution, can be solved with sample-by-sample herein, enable b1iIndicate B1I-th column,
b2jIndicate B2Jth column, removal constant term objective function it is writeable are as follows:
This problem is still difficult direct solution, is solved herein using cyclic coordinate gradient descent method by bit, if b1imTable
Show b1iM bit,Indicate b1iThe vector that other bits other than m bit are constituted, then b1imIt can be obtained by following formula:
Above formula is repeated until having solved the Hash codes of all image modalities samples;
(2) fixed W1, W2And B1, solve B2:
With solution B1It is similar, it can obtain
Above formula is repeated until having solved the Hash codes of all text modality samples;
(3) fixed W2, B1And B2, solve W1:
Constant term is removed, objective function is writeable are as follows:
There are closed solutions for this problem
Wherein D1For diagonal matrix,
(4) fixed W1, B1And B2, solve W2:
With solution W1It is similar, W2There are closed solutions
Wherein D2For diagonal matrix,
(5) (1)-(4) are repeated, if the mistake absolute value of the difference of iteration is less than 0.01 twice recently or the number of iterations is big
Terminate in 20;
Step 7: user input query sample, extracts its feature, and go mean value to the feature of extraction;
Step 8: generating the Hash codes of query sample using the hash function learnt:
Step 9: calculating the Hamming distance of query sample and target (training) concentration isomery sample, and Hamming distance is pressed
Ascending order arrangement, the corresponding sample of preceding r Hamming distance is search result, herein r=100.
In order to verify effectiveness of the invention, the present embodiment is by taking public data collection Mirflickr25K as an example, notebook data collection
Comprising 20015 image texts pair, all samples are to can be divided into 24 classifications;Randomly select a sample pair of 15011 (75%)
Composing training collection, and remaining a sample of 5004 (25%) is to composition test set;The textural characteristics of 150 dimensions of image modalities sample
It indicates, BOW (Bag Of Words) character representation of 500 dimensions of text modality sample, and mean value, two mode samples is gone to feature
This characteristic mean value is 0;In order to objectively evaluate the retrieval performance of the method for the present invention, with Average Accuracy (Mean
Average Precision, MAP) it is used as evaluation criterion, on Mirflickr25K data set, the MAP of different Hash code length p
The results are shown in Table 1.
MAP result of the table 1 on Mirflickr25K data set
P=16 | P=32 | P=64 | P=96 | |
Image retrieval text | 0.6718 | 0.6785 | 0.6843 | 0.6918 |
Text retrieval image | 0.6813 | 0.6953 | 0.6977 | 0.7045 |
It should be understood that the part that this specification does not elaborate belongs to the prior art.It is above-mentioned to implement for preferable
The description of example is more careful, but cannot be therefore, it is considered that being the limitation to the invention patent protection scope.
Claims (1)
1. a kind of discrete across media Hash search methods of supervision of robust, which is characterized in that this method comprises the following steps:
Step 1: collecting image and samples of text pair containing class label, the one-to-one cross-module state retrieval of image, text is constituted
Image, text and data collection;
Step 2: respectively to image and text modality sample extraction feature, and respectively to the feature of image and text modality sample
Mean value is gone, the characteristic mean value 0 of two mode samples is made;
Step 3: being training set and test set to random division by all samples in data set;
Step 4: constructing two differences respectively using the sample characteristics of the class label of sample pair, image and text modality in training set
This similarity matrix, and utilize the low-rank characteristic of similarity matrix and the sparse characteristic of noise sample between sample two-by-two, study
Similarity matrix between the sample two-by-two of one robust;The feature of training sample pair is set as X, X={ X(1), X(2), wherein X(1)It indicates
The sample characteristics of image modalities, X in training set(2)Indicate the sample characteristics of text modality in training set,Wherein d1And d2Respectively indicate image and text
The dimension of mode sample characteristics, N indicate that image or text modality sample size in training set, the class label of sample pair are indicated with L,C indicates the quantity of sample class, li∈ { 0,1 }cIf lij=1, indicate i-th of sample category
In jth class;, whereas if lij=0, indicate that i-th of sample is not belonging to jth class;Learn robust similarity matrix between sample two-by-two
Objective function the following steps are included:
(1) similarity matrix between the sample two-by-two based on image modalities feature is calculated using the sample characteristics of image modalities, it is fixed
Justice is as follows:
Wherein | | | |FIndicate Frobenius norm, S(1)Similarity matrix between the sample two-by-two of expression image modalities,It indicates
The similarity of i-th of image pattern and j-th of image pattern, σ1For scale parameter;
(2) similarity matrix, definition between sample two-by-two of the sample characteristics calculating based on text modality feature of text modality are utilized
It is as follows:
Wherein S(2)Similarity matrix between the sample two-by-two of expression text modality,Indicate i-th of samples of text and j-th of text
The similarity of sample, σ2For scale parameter;
(3) similarity matrix calculating the sample two-by-two based on class label using the class label of sample between, is defined as follows:
Wherein S(3)Indicate sample to the similarity matrix two-by-two of label,Indicate i-th of sample to label and j-th of sample pair
The similarity of label;
(4) objective function of similarity matrix is defined as follows study robust between sample two-by-two:
s.t.S(i)=S+ | | E(i)||0
Wherein S indicates similarity matrix two-by-two between the robust sample of study, E(i)Indicate i-th of making an uproar in similarity matrix two-by-two
Sound, the order of rank () representing matrix, | | | |0Indicate l0Norm;
(5) there are discrete low-rank and l for the objective function in above-mentioned (4)0The constraint of norm, above formula can be rewritten as
s.t.S(i)=S+ | | E(i)||1
Wherein | | | |*Indicate nuclear norm, | | | |1Indicate l1Norm,
(6) this problem is solved using augmented vector approach, obtain robust similarity matrix between sample two-by-two;
Step 5: construction objective function, specifically includes the following steps:
(1) similitude based on similarity matrix between robust two-by-two sample, the target letter of Hash codes study are kept in Hamming space
Number is defined as follows:
Wherein k indicates the length of Hash codes, B1For the Hash codes of image modalities sample, B2For the Hash codes of text modality sample, λ
For weight parameter;
(2) using Linear Mapping as hash function, and l is utilized2,1Norm constrains image and text modality Hash as regular terms
The objective function of the study of function, each mode hash function study is defined as follows:
Wherein W1, W2The hash function of image modalities and text modality is respectively indicated, Reg () indicates that regular terms prevents over-fitting,
HereinβiIt is weight parameter with μ;
(3) objective function that Hash codes and hash function learn is added as to the objective function of this method, is defined as follows:
Wherein βiFor weight parameter;
Step 6: objective function using iteration optimization algorithms solve, solution procedure the following steps are included:
(1) fixed W1, W2And B2, solve B1:
Constant term is removed, objective function is writeable are as follows:
It can be solved herein with sample-by-sample, enable b1iIndicate B1I-th column, b2jIndicate B2Jth column, removal constant term objective function can
It is written as:
It is solved herein using cyclic coordinate gradient descent method by bit, if b1imIndicate b1iM bit,Indicate b1iIn addition to
The vector that other bits outside m bit are constituted, then b1imIt can be obtained by following formula:
Above formula is repeated until having solved the Hash codes of all image modalities samples;
(2) fixed W1, W2And B1, solve B2:
With solution B1It is similar, it can obtain
Above formula is repeated until having solved the Hash codes of all text modality samples;
(3) fixed W2, B1And B2, solve W1:
Constant term is removed, objective function is writeable are as follows:
There are closed solutions for this problem
Wherein D1For diagonal matrix,
(4) fixed W1, B1And B2, solve W2:
With solution W1It is similar, W2There are closed solutions
Wherein D2For diagonal matrix,
(5) (1)-(4) are repeated to algorithmic statement or reach maximum number of iterations;
Step 7: user input query sample, extracts its feature, and go mean value to the feature of extraction;
Step 8: generating the Hash codes of query sample using the hash function learnt:
Step 9: calculating the Hamming distance of query sample and target (training) concentration isomery sample, and ascending order is pressed to Hamming distance
Arrangement, the corresponding sample of preceding r Hamming distance is search result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910096204.2A CN109871454B (en) | 2019-01-31 | 2019-01-31 | Robust discrete supervision cross-media hash retrieval method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910096204.2A CN109871454B (en) | 2019-01-31 | 2019-01-31 | Robust discrete supervision cross-media hash retrieval method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109871454A true CN109871454A (en) | 2019-06-11 |
CN109871454B CN109871454B (en) | 2023-08-29 |
Family
ID=66918414
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910096204.2A Active CN109871454B (en) | 2019-01-31 | 2019-01-31 | Robust discrete supervision cross-media hash retrieval method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109871454B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111078952A (en) * | 2019-11-20 | 2020-04-28 | 重庆邮电大学 | Cross-modal variable-length Hash retrieval method based on hierarchical structure |
CN111368176A (en) * | 2020-03-02 | 2020-07-03 | 南京财经大学 | Cross-modal Hash retrieval method and system based on supervision semantic coupling consistency |
CN111680173A (en) * | 2020-05-31 | 2020-09-18 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | CMR model for uniformly retrieving cross-media information |
CN112214623A (en) * | 2020-09-09 | 2021-01-12 | 鲁东大学 | Image-text sample-oriented efficient supervised image embedding cross-media Hash retrieval method |
CN112836068A (en) * | 2021-03-24 | 2021-05-25 | 南京大学 | Unsupervised cross-modal Hash retrieval method based on noisy label learning |
CN113961727A (en) * | 2021-09-13 | 2022-01-21 | 哈尔滨工业大学(深圳) | Cross-media Hash retrieval method, device, terminal and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107256271A (en) * | 2017-06-27 | 2017-10-17 | 鲁东大学 | Cross-module state Hash search method based on mapping dictionary learning |
WO2017210949A1 (en) * | 2016-06-06 | 2017-12-14 | 北京大学深圳研究生院 | Cross-media retrieval method |
-
2019
- 2019-01-31 CN CN201910096204.2A patent/CN109871454B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2017210949A1 (en) * | 2016-06-06 | 2017-12-14 | 北京大学深圳研究生院 | Cross-media retrieval method |
CN107256271A (en) * | 2017-06-27 | 2017-10-17 | 鲁东大学 | Cross-module state Hash search method based on mapping dictionary learning |
Non-Patent Citations (1)
Title |
---|
姚涛: "基于哈希方法的跨媒体检索研究", 万方中国学位论文 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111078952A (en) * | 2019-11-20 | 2020-04-28 | 重庆邮电大学 | Cross-modal variable-length Hash retrieval method based on hierarchical structure |
CN111078952B (en) * | 2019-11-20 | 2023-07-21 | 重庆邮电大学 | Cross-modal variable-length hash retrieval method based on hierarchical structure |
CN111368176A (en) * | 2020-03-02 | 2020-07-03 | 南京财经大学 | Cross-modal Hash retrieval method and system based on supervision semantic coupling consistency |
CN111368176B (en) * | 2020-03-02 | 2023-08-18 | 南京财经大学 | Cross-modal hash retrieval method and system based on supervision semantic coupling consistency |
CN111680173A (en) * | 2020-05-31 | 2020-09-18 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | CMR model for uniformly retrieving cross-media information |
CN111680173B (en) * | 2020-05-31 | 2024-02-23 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | CMR model for unified searching cross-media information |
CN112214623A (en) * | 2020-09-09 | 2021-01-12 | 鲁东大学 | Image-text sample-oriented efficient supervised image embedding cross-media Hash retrieval method |
CN112836068A (en) * | 2021-03-24 | 2021-05-25 | 南京大学 | Unsupervised cross-modal Hash retrieval method based on noisy label learning |
CN112836068B (en) * | 2021-03-24 | 2023-09-26 | 南京大学 | Unsupervised cross-modal hash retrieval method based on noisy tag learning |
CN113961727A (en) * | 2021-09-13 | 2022-01-21 | 哈尔滨工业大学(深圳) | Cross-media Hash retrieval method, device, terminal and storage medium |
CN113961727B (en) * | 2021-09-13 | 2022-10-21 | 哈尔滨工业大学(深圳) | Cross-media Hash retrieval method, device, terminal and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN109871454B (en) | 2023-08-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109871454A (en) | A kind of discrete across media Hash search methods of supervision of robust | |
CN108334574B (en) | Cross-modal retrieval method based on collaborative matrix decomposition | |
Shao et al. | Online multi-view clustering with incomplete views | |
CN105469096B (en) | A kind of characteristic bag image search method based on Hash binary-coding | |
CN110321925B (en) | Text multi-granularity similarity comparison method based on semantic aggregated fingerprints | |
CN110765260A (en) | Information recommendation method based on convolutional neural network and joint attention mechanism | |
US10867169B2 (en) | Character recognition using hierarchical classification | |
CN110222218B (en) | Image retrieval method based on multi-scale NetVLAD and depth hash | |
CN108595688A (en) | Across the media Hash search methods of potential applications based on on-line study | |
CN108491430B (en) | Unsupervised Hash retrieval method based on clustering characteristic directions | |
CN105718532A (en) | Cross-media sequencing method based on multi-depth network structure | |
CN115098620B (en) | Cross-modal hash retrieval method for attention similarity migration | |
CN112633431B (en) | Tibetan-Chinese bilingual scene character recognition method based on CRNN and CTC | |
Huang et al. | Large-scale heterogeneous feature embedding | |
CN108595546B (en) | Semi-supervision-based cross-media feature learning retrieval method | |
CN112214623A (en) | Image-text sample-oriented efficient supervised image embedding cross-media Hash retrieval method | |
Wei et al. | Representing word image using visual word embeddings and RNN for keyword spotting on historical document images | |
CN109857892B (en) | Semi-supervised cross-modal Hash retrieval method based on class label transfer | |
CN114782965A (en) | Visual rich document information extraction method, system and medium based on layout relevance | |
CN108182443B (en) | Automatic image labeling method and device based on decision tree | |
Guo et al. | Saliency detection on sampled images for tag ranking | |
Wei et al. | Word image representation based on visual embeddings and spatial constraints for keyword spotting on historical documents | |
CN116861022A (en) | Image retrieval method based on combination of deep convolutional neural network and local sensitive hash algorithm | |
Dhoot et al. | Efficient Dimensionality Reduction for Big Data Using Clustering Technique | |
CN106529601A (en) | Image classification prediction method based on multi-task learning in sparse subspace |
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 |