CN110489585A - Distributed image searching method based on supervised learning - Google Patents
Distributed image searching method based on supervised learning Download PDFInfo
- Publication number
- CN110489585A CN110489585A CN201910609588.3A CN201910609588A CN110489585A CN 110489585 A CN110489585 A CN 110489585A CN 201910609588 A CN201910609588 A CN 201910609588A CN 110489585 A CN110489585 A CN 110489585A
- Authority
- CN
- China
- Prior art keywords
- matrix
- node
- classification
- hash codes
- supervised learning
- 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
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2471—Distributed queries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/51—Indexing; Data structures therefor; Storage structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
Abstract
The invention discloses the distributed image searching methods based on supervised learning, classification marker is carried out to image, video, file first in the database of each node, initialize classification matrix, encoder matrix, Hash codes matrix and corresponding Lagrange multiplier, then it introduces and minimizes error in classification and reconstructed error building objective function, solve above-mentioned objective function, undated parameter matrix;Back end is communicated with central node, and judges whether the transition matrix of each node reaches unanimity, and updates Lagrange multiplier, finally carries out approximation search process;The problem of present invention solves large-scale data and is storing, and required scale is excessive when calculating, and centralization ground training algorithm model has been no longer appropriate for;And back end does not exchange raw information with central node communication, can effectively solve transmission and communicates excessive problem, while the data on node keep independence.
Description
Technical field
The present invention relates to a kind of image search method, specifically a kind of distributed image searching method belongs to machine
Learning areas.
Background technique
With the continuous development of social networks, e-commerce, mobile Internet etc., data need the scale for storing, handling
Increasing, one-of-a-kind system has been unable to satisfy growing demand.The Internet companies such as Google, Alibaba are successfully expedited the emergence of
Cloud computing and this two big hot topics field of big data, cloud computing and big data are all the applications constructed on distributed storage.
The core of cloud storage is the large-scale distributed storage system of rear end, and big data not only needs to store the data of magnanimity, also to be led to
It crosses suitable frame and tool to analyze these data, obtains wherein useful part, if without distributed storage
It is just far from being and big data is analyzed.Although many years have been carried out in the research of distributed system, until in recent years, mutually
The rise of networking big data applies distributed system on a large scale in engineering practice.Distributed system is to utilize more
Collaborative computer solves calculating, the storage problem that single computer cann't be solved, and distributed system is maximum with one-of-a-kind system
Difference is that the scale of problem.The system that it is made of multiple nodes, often will be on a server or server
One process is known as a node, these nodes are generally not isolated, but are communicated with each other by network, transmits information.
In addition, smart phone stores the letter such as a large amount of picture, text and video due to the fast development of the mobile terminals such as smart phone
Breath, smart phone also can be regarded as an independent node, divided between smart phone by base station or mutual pass through
Cloth cooperates to improve data-handling capacity.
Supervised learning (Supervised learning), is the algorithm in a kind of machine learning, can be by training data
A mode is acquired or establishes, and mode speculates new example according to this.Training data be by input object (usually vector) and
Anticipated output is formed.The output of function can be a continuous value (referred to as regression analysis), or one contingency table of prediction
Label.There are also a major class algorithm in machine learning, it is called unsupervised learning (Unsupervised learning), is directly to not having
Markd training data carries out modeling study, notices that data herein are that do not have markd data, most with supervised learning
Basic difference is that have one, label be not no label to data one of modeling.Compared to unsupervised learning, supervised learning
Advantage can exactly make full use of known mark information, merge more information into the model of building, effectively increase model
Reliability.
In addition, the development of the widely available and multimedia technology with internet, the data of all trades and professions sharply increase, it is existing
Have to handle huge database for information technology infrastructure.In fact, compared with carrying cost, in large scale database
Middle retrieval related content is the task of a more challenge, especially in searching multimedia data, such as audio, image and view
The retrieval of frequency content is the having more challenge of the task.Traditional nearest neighbor algorithm is in processing large-scale image search problem
When, up to thousands of dimensions, " dimension disaster " so will lead to memory space consumption greatly and examine the characteristic dimension of sample data
The slow-footed problem of rope.In recent years, hash algorithm can satisfy extensive inspection as a kind of representative nearest _neighbor retrieval technology
To the particular/special requirement of memory space and retrieval time in rope.The purpose of hash algorithm is expressed as image as one group of regular length
Binary-coding, i.e. Hash codes, usually used -1/1 or 0/1 indicates bit therein.Hash algorithm solves conventional retrieval
Problem for mass data storage space and retrieval time unwarranted demand so that it is for memory space and retrieval time
Demand be greatly lowered, while having and can obtain good retrieval effectiveness, therefore, it has become the sharp sword of processing big data problem,
Receive the extensive concern of computer vision field.However current most of hash algorithms be all it is centralized, there are single sections
The problems such as computationally intensive are put, it is an interesting problem that hash algorithm how is applied in distributed scene.
In conclusion not having still for how to search for problem using supervision hash algorithm realization distributed image in the prior art
There is disclosed disclosure.
Summary of the invention
The purpose of the present invention is to provide a kind of distributed image searching method based on supervised learning, is mainly used for solving
Image, video, text equal samples number is big, can not accurately find semantic neighbour, if concentrating in together training, transmission quantity and
The excessive problem of calculation amount, the main purpose of this method are to be obtained by distribution training with the training of lower computing cost complete
The encoder matrix of office's optimization, while the data independence of each node in distributed training is protected, and realize the neighbour of query sample
Search.
The present invention provides a kind of distributed image searching method based on supervised learning, comprising the following steps:
Step 1: classification marker being carried out to image, video, file in the database of each node;
Step 2: initialization classification matrix, encoder matrix, Hash codes matrix and corresponding Lagrange multiplier;
Step 3: introducing and minimize error in classification and reconstructed error building objective function;
Step 4: solving above-mentioned objective function, update classification matrix, encoder matrix, Hash codes matrix;
Step 5: back end is communicated with central node, judges whether the transition matrix of each node reaches unanimity, more
New Lagrange multiplier;
Step 6: carrying out approximation search process.
It is further limited as of the invention, in step 1, it is assumed that share N number of node, the corresponding database of each node
Xi, XiThe database for indicating i-th of node, the database in different nodes are independent from each other, and between different nodes not
Wish shared information, has c kind category label in each database, different labels is stamped to different samples.
It is further limited as of the invention, in step 2, on each node all to classification matrix, encoder matrix, Hash
Code matrix and corresponding Lagrange multiplier carry out Initialize installation, and setting initialization classification matrix is the unit square of d × r dimension
Battle array, corresponding initialization Lagrange multiplier are the full 0 matrix of d × r dimension, and initialization classification matrix is the unit square of r × c dimension
Battle array, corresponding initialization Lagrange multiplier are the full 0 matrix of r × c dimension, and initialization Hash codes matrix is each member of r × n dimension
The matrix that plain absolute value is 1, d indicate the dimension of sample original feature space, r presentation code digit, and c indicates class number, n table
Show number of samples.
It is further limited as of the invention, in step 3, is introduced in objective function and minimize error in classification and reconstruct mistake
Original feature space is mapped to Hash codes by encoder matrix, so that the classification accuracy based on Hash codes is as far as possible by difference
Height increases an orthogonality constraint to guarantee the validity of Hash codes, while in order to reduce the correlation between Hash codes, in order to
Lower quantization error increases a discrete constraint, i.e. pressure Hash codes are equal to 1 or -1.
The objective function of building is successively as follows:
X in above formulaiIndicate the sample of i-th of node, i.e. database Xi,Ci、BiRespectively indicate the coding square of i-th of node
Battle array and Hash codes matrix, ΠiIndicate dual variable, ρ is Lagrange multiplier, and Z is the global parameter that consistency introduces, in above formula
Constraint consists of two parts, and first part is the global coherency constraint of alternating direction multipliers method (ADMM), and second constraint is
The mutually independent constraint of Hash codes:
Y in above formulaiIndicate the sample labeling of i-th of node, Wi、BiRespectively indicate the classification matrix and Hash of i-th of node
Code matrix, λ are Lagrange multipliers, and U is the global parameter that alternating direction multipliers method (ADMM) consistency introduces, and constraint is global
Consistency constraint.
Y in above formulaiIndicate the sample labeling of i-th of node, Wi、Bi、CiIt respectively indicates the classification matrix of i-th of node, breathe out
Uncommon code matrix and encoder matrix, v are tradeoff parameters, and increased constraint is in order to which each median for guaranteeing Hash codes is discrete
's.
Further limited as of the invention, in step 4, when solving encoder matrix C, due to be related to solve one
The problem of trace of a matrix is minimized under conditions of orthogonality constraint needs to utilize Singular-value Decomposition Solution.
It is further limited as of the invention, in step 5, distributed optimization encoder matrix C and when classification matrix W, in addition to N
Except a back end, there are one central nodes, for updated to W the and C overall situation, and central node and back end
Between Transfer Parameters information, to guarantee the consistency of parameter.
It is further limited as of the invention, in step 6, it is broadcast to all sections by the query sample new for one
Point calculates the Hamming distance between new samples and node sample after encoder matrix maps, before taking wherein it is k the smallest away from
The result searched for from corresponding sample as approximation.
The invention adopts the above technical scheme compared with prior art, has following technical effect that
1, excessively unilateral when solving many traditional methods for neighbor search, it does not account for marking when finding neighbour
Information, pilot process do not carry out sliding-model control, so that these algorithms performance in the practical application of proximity search is bad
Problem;
2, it solving large-scale data storing, required scale is excessive when calculating, exceed single calculate node computing capability,
The problem of centralized ground training algorithm model has been no longer appropriate for;
3, the communication between node is carried out using parameter matrix, the communication between node does not exchange raw information, can effectively solve
Transmission communicates excessive problem, while realizing good performance.
Detailed description of the invention
Fig. 1 is the system framework figure of this method.
Fig. 2 is the distribution training flow chart of this method.
Fig. 3 is the neighbor search flow chart of this method.
Specific embodiment
Technical solution of the present invention is described in further detail with reference to the accompanying drawing:
The system framework figure of this method as shown in Figure 1, entire method process can be divided into distributed training process with it is approximate
Property search process, detailed process difference it is as shown in Figures 2 and 3, wherein the first step to the 5th step by the way of as shown in Figure 2,
6th step is by the way of as shown in Figure 3.
The first step carries out classification marker to image, video, file etc. in the database of each node.
Assuming that N number of node is shared, the corresponding database X of each nodei, XiThe database for indicating i-th of node, not
It is independent from each other with the database in node, and is not intended to shared information between different nodes, there is n in each database
Sample, while having c kind category label in each database, different labels is stamped to different samples.
Second step initializes classification matrix, encoder matrix and Lagrange multiplier matrix, initializes Hash codes matrix.
Classification matrix, encoder matrix and corresponding Lagrange multiplier are initialized in each node, initialize Hash codes
The initialization classification matrix C of i-th of node is arranged in matrixiIt is the unit matrix of d × r dimension, when i-th of node optimization C is corresponding
The full 0 matrix that Lagrange multiplier is d × r dimension is initialized, classification matrix W is initializediIt is the unit matrix of r × c dimension, optimizes W
Corresponding initialization Lagrange multiplier is the full 0 matrix of r × c dimension, and initialization Hash codes matrix B is each element of r × n dimension
The matrix that absolute value is 1.D indicates the dimension of sample original feature space, r presentation code digit, and c indicates that class number, n indicate
Number of samples.The transition matrix of all nodes and the initialization value of Lagrange multiplier are all.
Third step introduces and minimizes error in classification and reconstructed error and discretization and orthogonality constraint building objective function.
It should be noted that the utilization for focusing on obtaining encoder matrix and encoder matrix of the invention, so no longer
Emphasis lists original objective function, and to the specific optimization process of objective function, after only giving objective function optimization
As a result;The optimization C of i-th of node buildingiObjective function afterwards is as follows:
X in above formulaiIndicate the sample of i-th of node, Ci、BiRespectively indicate the encoder matrix and Hash codes square of i-th of node
Battle array, ρ is Lagrange multiplier, ΠiIndicate dual variable, Z is the global ginseng that alternating direction multipliers method (ADMM) consistency introduces
It counting, constrains and consist of two parts in above formula, first part is that the global coherency of alternating direction multipliers method (ADMM) constrains, second
A constraint is the mutually independent constraint of Hash codes.
The optimization W of i-th of node buildingiObjective function afterwards is as follows:
Y in above formulaiIndicate the sample labeling of i-th of node, Wi、BiRespectively indicate the classification matrix and Hash of i-th of node
Code matrix, λ are Lagrange multipliers, and U is the global parameter that alternating direction multipliers method (ADMM) consistency introduces, and constraint is global
Consistency constraint.
The optimization B of i-th of node buildingiObjective function afterwards is as follows:
Y in above formulaiIndicate the sample labeling of i-th of node, Wi、Bi、CiIt respectively indicates the classification matrix of i-th of node, breathe out
Uncommon code matrix and encoder matrix, v are tradeoff parameters, and increased constraint is in order to which each median for guaranteeing Hash codes is discrete
's.
4th step solves objective function, undated parameter matrix.
Three objective functions above are solved respectively, optimize Wi,CiIt is all made of alternating direction multipliers method (ADMM) progress
It solves, optimizes BiDirectly solved using the data information of each node, it is fully distributed when optimizing between node.
5th step, back end are communicated with central node, judge whether the transition matrix of each node reaches unanimity, more
New Lagrange multiplier.
Back end and global node, which carry out communicating the parameter matrix for referring to that each node calculates oneself, to be passed to
Central node carries out global optimization by central node, and then the parameter matrix of global optimization is traveled to each data section
Point can guarantee that training process meets the thought of consistency by the transmitting of this parameter to carry out next iteration update.
If the transition matrix of all nodes tends not to unanimously, iteration updates Lagrange multiplier, and repeats third step.
6th step, approximation search process.
Query sample x new for onec, it is input to all distributed nodes, it is assumed that enter i-th of section
Point, the encoder matrix C obtained using distributed training processi, to query sample xcIt is encoded with back end sample, then
Calculate xcFrom their Hamming distance (i.e. the corresponding position of Hash codes different quantity);Obtain query sample xcWith its in each node
After the distance of his sample, these distances are ranked up, k the smallest distances before obtaining, corresponding to preceding k the smallest distances
Sample is the result that our approximations are searched for.
The above, the only specific embodiment in the present invention, but scope of protection of the present invention is not limited thereto, appoints
What is familiar with the people of the technology within the technical scope disclosed by the invention, it will be appreciated that expects transforms or replaces, and should all cover
Within scope of the invention, therefore, the scope of protection of the invention shall be subject to the scope of protection specified in the patent claim.
Claims (7)
1. a kind of distributed image searching method based on supervised learning, which comprises the following steps:
Step 1: classification marker being carried out to image, video, file in the database of each node;
Step 2: initialization classification matrix, encoder matrix, Hash codes matrix and corresponding Lagrange multiplier;
Step 3: introducing and minimize error in classification and reconstructed error building objective function;
Step 4: solving above-mentioned objective function, update classification matrix, encoder matrix, Hash codes matrix;
Step 5: back end is communicated with central node, judges whether the transition matrix of each node reaches unanimity, and is updated and is drawn
Ge Lang multiplier;
Step 6: carrying out approximation search process.
2. the distributed image searching method according to claim 1 based on supervised learning, which is characterized in that in step 1,
Assuming that N number of node is shared, the corresponding database X of each nodei, XiThe database for indicating i-th of node, in different nodes
Database be independent from each other, and be not intended to shared information between different nodes, there is c kind classification mark in each database
Note, different labels is stamped to different samples.
3. the distributed image searching method according to claim 2 based on supervised learning, which is characterized in that in step 2,
All are carried out by initialization and is set for classification matrix, encoder matrix, Hash codes matrix and corresponding Lagrange multiplier on each node
It sets, setting initialization classification matrix is the unit matrix of d × r dimension, and corresponding initialization Lagrange multiplier is the full 0 of d × r dimension
Matrix, initialization classification matrix are the unit matrixs of r × c dimension, and corresponding initialization Lagrange multiplier is the full 0 square of r × c dimension
Battle array, initialization Hash codes matrix are the matrixes that each element absolute value of r × n dimension is 1, and d indicates sample original feature space
Dimension, r presentation code digit, c indicate that class number, n indicate number of samples.
4. the distributed image searching method according to claim 3 based on supervised learning, which is characterized in that in step 3,
It is introduced in objective function and minimizes error in classification and reconstructed error, original feature space is mapped to by Kazakhstan by encoder matrix
Uncommon code so that the classification accuracy based on Hash codes is as high as possible, to guarantee the validity of Hash codes, the objective function of building according to
It is secondary as follows:
X in above formulaiIndicate the sample of i-th of node, i.e., above-mentioned database Xi, Ci、BiRespectively indicate the volume of i-th of node
Code matrix and Hash codes matrix, ΠiIndicating dual variable, ρ is Lagrange multiplier, and Z is the global parameter that consistency introduces, on
Constraint consists of two parts in formula, and first part is the global coherency constraint of alternating direction multipliers method (ADMM), and second about
Beam is the mutually independent constraint of Hash codes:
Y in above formulaiIndicate the sample labeling of i-th of node, Wi、BiRespectively indicate the classification matrix and Hash codes square of i-th of node
Battle array, λ are Lagrange multipliers, and U is the global parameter that alternating direction multipliers method (ADMM) consistency introduces, and constraint is globally consistent
Property constraint.
Y in above formulaiIndicate the sample labeling of i-th of node, Wi、Bi、CiRespectively indicate the classification matrix, Hash codes of i-th of node
Matrix and encoder matrix, v are tradeoff parameters, and increased constraint is in order to which each median for guaranteeing Hash codes is discrete.
5. the distributed image searching method according to claim 4 based on supervised learning, which is characterized in that in step 4,
When solving encoder matrix C, the problem of minimum trace of a matrix, needed under conditions of orthogonality constraint due to being related to solving one
Utilize Singular-value Decomposition Solution.
6. the distributed image searching method according to claim 5 based on supervised learning, which is characterized in that in step 5,
When distributed optimization encoder matrix C and classification matrix W, other than N number of back end, there are one central node, be used into
Row updates W the and C overall situation, and Transfer Parameters information between central node and back end, to guarantee the consistency of parameter.
7. the distributed image searching method according to claim 6 based on supervised learning, which is characterized in that in step 6,
It is broadcast to all nodes by the query sample new for one, after encoder matrix maps, calculates new samples and node sample
Hamming distance between this, the result that sample corresponding to k the smallest distances is searched for as approximation before taking wherein.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910609588.3A CN110489585B (en) | 2019-07-08 | 2019-07-08 | Distributed image searching method based on supervised learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910609588.3A CN110489585B (en) | 2019-07-08 | 2019-07-08 | Distributed image searching method based on supervised learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110489585A true CN110489585A (en) | 2019-11-22 |
CN110489585B CN110489585B (en) | 2022-12-02 |
Family
ID=68546684
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910609588.3A Active CN110489585B (en) | 2019-07-08 | 2019-07-08 | Distributed image searching method based on supervised learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110489585B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111159443A (en) * | 2019-12-31 | 2020-05-15 | 深圳云天励飞技术有限公司 | Image characteristic value searching method and device and electronic equipment |
CN111553418A (en) * | 2020-04-28 | 2020-08-18 | 腾讯科技(深圳)有限公司 | Method and device for detecting neuron reconstruction errors and computer equipment |
CN111832637A (en) * | 2020-06-30 | 2020-10-27 | 南京邮电大学 | Distributed deep learning classification method based on alternative direction multiplier method ADMM |
CN111881928A (en) * | 2020-05-19 | 2020-11-03 | 杭州中奥科技有限公司 | Coding model training method and device, storage medium and electronic equipment |
CN112199520A (en) * | 2020-09-19 | 2021-01-08 | 复旦大学 | Cross-modal Hash retrieval algorithm based on fine-grained similarity matrix |
CN112965722A (en) * | 2021-03-03 | 2021-06-15 | 深圳华大九天科技有限公司 | Verilog-A model optimization method, electronic device and computer readable storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150234779A1 (en) * | 2014-02-20 | 2015-08-20 | Mitsubishi Electric Research Laboratories, Inc. | Method for Solving Quadratic Programs for Convex Sets with Linear Equalities by an Alternating Direction Method of Multipliers with Optimized Step Sizes |
CN107315765A (en) * | 2017-05-12 | 2017-11-03 | 南京邮电大学 | A kind of method of the concentrated-distributed proximity search of extensive picture |
-
2019
- 2019-07-08 CN CN201910609588.3A patent/CN110489585B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150234779A1 (en) * | 2014-02-20 | 2015-08-20 | Mitsubishi Electric Research Laboratories, Inc. | Method for Solving Quadratic Programs for Convex Sets with Linear Equalities by an Alternating Direction Method of Multipliers with Optimized Step Sizes |
CN107315765A (en) * | 2017-05-12 | 2017-11-03 | 南京邮电大学 | A kind of method of the concentrated-distributed proximity search of extensive picture |
Non-Patent Citations (2)
Title |
---|
YUE GAO 等: "Weakly Supervised Visual Dictionary Learning by Harnessing Image Attributes", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
谢辉: "搜索引擎中基于内容的图像重排序", 《计算机应用》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111159443A (en) * | 2019-12-31 | 2020-05-15 | 深圳云天励飞技术有限公司 | Image characteristic value searching method and device and electronic equipment |
CN111553418A (en) * | 2020-04-28 | 2020-08-18 | 腾讯科技(深圳)有限公司 | Method and device for detecting neuron reconstruction errors and computer equipment |
CN111881928A (en) * | 2020-05-19 | 2020-11-03 | 杭州中奥科技有限公司 | Coding model training method and device, storage medium and electronic equipment |
CN111881928B (en) * | 2020-05-19 | 2022-07-29 | 杭州中奥科技有限公司 | Coding model training method and device, storage medium and electronic equipment |
CN111832637A (en) * | 2020-06-30 | 2020-10-27 | 南京邮电大学 | Distributed deep learning classification method based on alternative direction multiplier method ADMM |
CN111832637B (en) * | 2020-06-30 | 2022-08-30 | 南京邮电大学 | Distributed deep learning classification method based on alternating direction multiplier method ADMM |
CN112199520A (en) * | 2020-09-19 | 2021-01-08 | 复旦大学 | Cross-modal Hash retrieval algorithm based on fine-grained similarity matrix |
CN112199520B (en) * | 2020-09-19 | 2022-07-22 | 复旦大学 | Cross-modal Hash retrieval algorithm based on fine-grained similarity matrix |
CN112965722A (en) * | 2021-03-03 | 2021-06-15 | 深圳华大九天科技有限公司 | Verilog-A model optimization method, electronic device and computer readable storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN110489585B (en) | 2022-12-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110489585A (en) | Distributed image searching method based on supervised learning | |
Wu et al. | Unsupervised Deep Hashing via Binary Latent Factor Models for Large-scale Cross-modal Retrieval. | |
CN103678431B (en) | A kind of recommendation method to be scored based on standard label and project | |
CN110677284B (en) | Heterogeneous network link prediction method based on meta path | |
CN111325326A (en) | Link prediction method based on heterogeneous network representation learning | |
CN104731962A (en) | Method and system for friend recommendation based on similar associations in social network | |
CN112199532B (en) | Zero sample image retrieval method and device based on Hash coding and graph attention machine mechanism | |
CN111310074B (en) | Method and device for optimizing labels of interest points, electronic equipment and computer readable medium | |
Cheng et al. | Bridging multimedia heterogeneity gap via graph representation learning for cross-modal retrieval | |
CN111259263A (en) | Article recommendation method and device, computer equipment and storage medium | |
CN111832637B (en) | Distributed deep learning classification method based on alternating direction multiplier method ADMM | |
CN112131261B (en) | Community query method and device based on community network and computer equipment | |
CN109919172A (en) | A kind of clustering method and device of multi-source heterogeneous data | |
CN112380299A (en) | Relational network construction method, device and storage medium | |
WO2022188646A1 (en) | Graph data processing method and apparatus, and device, storage medium and program product | |
CN108509651B (en) | The distributed approximation searching method with secret protection based on semantic consistency | |
CN112288154A (en) | Block chain service reliability prediction method based on improved neural collaborative filtering | |
CN115600017A (en) | Feature coding model training method and device and media object recommendation method and device | |
CN115601745A (en) | Multi-view three-dimensional object identification method facing application end | |
CN111459990B (en) | Object processing method, system, computer readable storage medium and computer device | |
Zhang et al. | End‐to‐end generation of structural topology for complex architectural layouts with graph neural networks | |
CN114925210A (en) | Knowledge graph construction method, device, medium and equipment | |
CN114398980A (en) | Cross-modal Hash model training method, encoding method, device and electronic equipment | |
CN115631008A (en) | Commodity recommendation method, commodity recommendation device, commodity recommendation equipment and commodity recommendation medium | |
CN112561599A (en) | Click rate prediction method based on attention network learning and fusing domain feature interaction |
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 |