CN108446307A - A kind of the binary set generation method and image, semantic similarity search method of multi-tag image - Google Patents
A kind of the binary set generation method and image, semantic similarity search method of multi-tag image Download PDFInfo
- Publication number
- CN108446307A CN108446307A CN201810111604.1A CN201810111604A CN108446307A CN 108446307 A CN108446307 A CN 108446307A CN 201810111604 A CN201810111604 A CN 201810111604A CN 108446307 A CN108446307 A CN 108446307A
- Authority
- CN
- China
- Prior art keywords
- picture
- image
- neural networks
- convolutional neural
- networks model
- 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.)
- Pending
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
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/951—Indexing; Web crawling techniques
-
- 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
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Library & Information Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of binary set generation method of multi-tag image and image, semantic similarity search methods.The present invention is trained using training data set pair convolutional neural networks model, until the penalty values of the convolutional neural networks model tend towards stability；Then the picture in image data base is calculated using the convolutional neural networks model after training, obtain the binary set of every pictures and is stored；The binary set of picture to be checked is calculated using the convolutional neural networks model after training, and the binary set of itself and storage is subjected to similarity calculation, is returned and the most like several pictures of the picture to be checked according to similarity calculation result.The present invention substantially increases the storage efficiency and search efficiency of image；And Target Photo can be ranked up according to the multistage semantic similarity with inquiry picture.
Description
Technical field
Present invention is mainly applied to field of image search, are related to a kind of the binary set generation method and needle of multi-tag image
To the similarity search method of multi-tag image, semantic feature.
Background technology
In recent years, with the fast development of network technology, thousands of picture is uploaded in internet daily, such as
What rapidly and accurately retrieved according to different user demands from the picture of magnanimity picture concerned have become research hot spot and
Difficult point.For example, content-based image retrieval needs retrieve and inquire the similar Target Photo of image content, " phase here
Seemingly " finger vision or semantic similar.Image higher-dimension primitive character is mapped to low-dimensional by the hash algorithm towards image, semantic feature
Image, semantic information is remained while binary features, therefore is received significant attention.
The hash algorithm towards image, semantic feature of mainstream being capable of Simultaneous Extracting Image using depth learning technology at present
Semantic feature simultaneously learns hash function, but still has following limitation：(1) most of hash algorithm can only simple zones split-phase Sihe
Dissimilar picture, it is bad for the multi-tag image effect containing multistage semantic similarity.(2) Kazakhstan towards multi-tag image
Uncommon algorithm can not the effective district multi-tag picture that divides similarity different.
Invention content
For the technical problems in the prior art, the purpose of the present invention is to provide a kind of two-values of multi-tag image
Vector generation method and extensive multi-tag image, semantic similarity search method.The present invention is based on convolutional neural networks model,
By well-designed loss function learning model parameter, realizes Simultaneous Extracting Image primitive character and learn hash function, most
The image binary coding exported eventually has the following properties that：
● it is formed by 1, -1, while improving storage and search efficiency；
● Target Photo can be ranked up according to the multistage semantic similarity with inquiry picture；
● in the case where characteristics of image coding is shorter, remain to according to semantic similarity effective district partial objectives for picture, especially
It is to return the result former accuracys rate compared with mainstream algorithm higher.
The technical scheme is that：
A kind of binary set generation method of multi-tag image, step include：
1) it is trained using training data set pair convolutional neural networks model, until the damage of the convolutional neural networks model
Mistake value tends towards stability；Wherein, every time training when training dataset include N to picture, for i-th pair picture II, 1、II, 2If figure
Piece II, 1Number of labels be nI, 1, picture I in i-th pair imageI, 1With picture II, 2Common associated number of labels is nI, 2, calculate
The loss function η of the penalty values is
Wherein, nI, 1=nI, 2When, yi=0, otherwise yi=1；Ones representative elements are all 1 vector, | | | |1Indicate to
First norm of amount,Indicate the Euclidean distance between vector, | | it indicates to carry out absolute value operation, α per bit element to vector
It is the parameter for controlling quantization loss size, w is the weight vectors of the Hash layer of convolutional neural networks model, f (I；W) it is volume
The positions the k binary set of the picture I of product neural network model output, m refers to Hamming distance threshold parameter；
2) binary set of picture is calculated using the convolutional neural networks model after training.
A kind of multi-tag image, semantic similarity search method, step include：
1) it is trained using training data set pair convolutional neural networks model, until the damage of the convolutional neural networks model
Mistake value tends towards stability；Wherein, every time training when training dataset include N to picture, for i-th pair picture II, 1、II, 2If figure
Piece II, 1Number of labels be nI, 1, picture I in i-th pair imageI, 1With picture II, 2Common associated number of labels is nI, 2, calculate
The loss function η of the penalty values is
Wherein, nI, 1=nI, 2When, yi=0, otherwise yi=1；Ones representative elements are all 1 k bit vectors, | | | |1It indicates
First norm of vector,Indicate the Euclidean distance between vector, | | it indicates to carry out absolute value operation per bit element to vector,
α is the parameter for controlling quantization loss size, and w is the weight vectors of the Hash layer of convolutional neural networks model, f (I；W) it is
The positions the k binary set of the picture I of convolutional neural networks model output, m refers to Hamming distance threshold parameter；
2) picture in image data base is calculated using the convolutional neural networks model after training, obtains every figure
The binary set of piece is simultaneously stored；
3) binary set of picture to be checked is calculated using the convolutional neural networks model after training, and by itself and step 2)
Obtained binary set carries out similarity calculation, is returned and most like several of the picture to be checked according to similarity calculation result
Picture.
Further, the convolutional neural networks model is trained using minimum batch gradient descent method, can minimized
The value of the loss function.
Further, which includes sequentially connected first convolutional layer, maximum pond layer, volume Two
Lamination, maximum pond layer, third convolutional layer, Volume Four lamination, the 5th convolutional layer, maximum pond layer, the first full articulamentum, second
Full articulamentum and Hash layer.
Further, the hash function of the Hash layer is h (x；W)=sign (f (x；w))；Wherein, f (x；W)=wTf′
(x), f ' (x) is the output vector of the second full articulamentum；The Hash layer is equipped with k node, and a weight is arranged in each node, this
A little weights constitute weight vectors w.
Further, α=0.01.
Further, the Hamming distance of the binary set obtained according to the binary set of picture to be checked and step 2), really
The fixed similarity.
The invention mainly comprises the following contents：
1) image, semantic feature extraction and Hash letter can be carried out at the same time by being based on convolutional neural networks modelling one kind
The frame that mathematics is practised.Using the frame, multi-tag image can be mapped to binary set and retained more between image by the present invention
Grade semantic similarity.
2) a kind of well-designed loss function based on image tag pair.It can be to each layer of model based on the loss function
Parameter is learnt.
Compared with prior art, the positive effect of the present invention is：
The present invention substantially increases the storage efficiency and search efficiency of image；And it can be according to the multistage with inquiry picture
Semantic similarity is ranked up Target Photo.The present invention is remained in the case where characteristics of image coding is shorter according to semantic phase
Like degree effective district partial objectives for picture, former accuracys rate are especially returned the result compared with mainstream algorithm higher.
Description of the drawings
Fig. 1 is the frame construction drawing of the present invention.
Specific implementation mode
Present invention will be further explained below with reference to the attached drawings and examples.
One：Hash function
Define hash function h (x；w)：
h(x；W)=sign (f (x；w)) (1)
f(x；W)=wTF ' (x), w are the weight vectors of Hash layer, and f ' (x) is the output vector of full articulamentum seven.Hash
Layer is made of a full articulamentum, and Hash node layer number is equal with the binary set digit k generated is finally needed, and k values are advance
It sets.
Two：Loss function
By way of optimizing loss function, the study of each layer parameter of convolutional neural networks model is realized.For training mould
The training data of type is concentrated, and every image corresponds to different labels, and label is that artificial mark obtains, can be according to image when training
Label information judges the similarity degree of image.If i-th pair picture Ii1,Ii2Respectively with p1,p2A label association, enables Ii1's
Number of labels is ni1, then ni1=| p1|, enable Ii1,Ii2Common associated number of labels is ni2, then ni2=| p1∩p2|, work as ni1=
ni2When, enable variable yi=0, otherwise yi=1, for Ii1,Ii2Loss function be defined as follows：
DH() is the Hamming distance of two binary sets, and m is threshold parameter (m>0, taking for m can be discussed in detail hereinafter
Value).
Loss function consists of two parts, and is to separate with plus sige.n1=n2When, y=0, at this time it is considered herein that two pictures
Closely similar, as long as two image binary features have differences being presented as in loss function, the present invention implements to punish；n1
≠n2When, y=1 is presented as two at this time it is considered herein that two pictures are general similar or dissimilar in loss function
The similarity of image is different, and the Hamming distance between binary feature vector should change therewith.When N to picture as training set
When, the present invention finally needs the loss function minimized to be：
Three：The conversion of loss function
Hamming distance in equation (2) is presented in a discrete fashion, it is difficult to solution is directly optimized, it is of the invention right thus
Equation (2) is converted.Specifically, the present invention converts the Hamming distance in equation (2) to Euclidean distance, meanwhile, in order to
So that the output f (I of convolutional neural networks model；W) 1 or -1 is approached, invention introduces quantizations to lose.At this point, peer-to-peer
(2) optimization can be approximated to be the optimization to following formula：
Loss function is made of three parts, is to separate with plus sige, preceding two parts act on, Part III identical as equation (2)
Effect be make picture feature vector each approach 1 or -1, wherein Ones representative elements are all 1 vector, vectorial length
Degree is k (identical with the binary set length of output),Indicate the Euclidean distance between vector, | | | |1Indicate the first of vector
Norm, | | it indicates to carry out absolute value operation per bit element to vector, α (0 α≤1 ＜) is for controlling quantization loss size
Parameter.Bringing equation (4) into equation (3) can obtain：
Wherein, II, 1、II, 2It is two images in i-th pair image, nI, 1It is image I in i-th pair imageI, 1Number of labels,
nI, 2It is image I in i-th pair imageI, 1、II, 2Common associated number of labels, nI, 1=nI, 2When, yi=0, otherwise yi=1；M is threshold
Value parameter.
Four：The study of model parameter
The parameter of convolutional neural networks model includes the connection weight of convolution nuclear parameter and full articulamentum, it is determined that parameter
It could finally determine model, the two-value semantic feature of image can be extracted using model, the final quick phase for realizing large nuber of images
Like degree match query.The learning algorithm of model parameter utilizes the thought of backpropagation, specific to utilize minimum batch gradient descent method
Training neural network, can minimize the value of loss function i.e. formula (5), and the extraction of subsequent pictures feature will all use training
The modular form (5) obtained afterwards can be separated into three (being respectively Term1, Term2, Regularizer) according to plus sige, each is right
In fi,jDerivative it is as follows：
When -1≤x≤0 or x >=1, δ (x)=1, otherwise, δ (x)=0.
Wherein, II, jIt is an image in i-th pair image, the value of j is 1 or 2, nI, 1It is an image in i-th pair image
Number of labels, nI, 2It is I in i-th pair imageI, 1、II, 2Common associated number of labels, nI, 1=nI, 2When, yi=0, otherwise yi=
1, i value range is 1~N.
Five, details is realized
Algorithm is realized based on Caffe deep learning frames, as shown in Figure 1, after convolutional layer one, convolutional layer two and convolutional layer five
There is maximum pond layer (ReLU layers) in face, and in model training, minimum batch gradient descent algorithm parameter setting is as follows：
Batch size=32, momentum=0.9, weight decay=0.004.
The present invention compared the experiment effect of α={ 0.1,0.01,0.001 } respectively, the results showed that, as α=0.01, inspection
Rope best results.
The final output of algorithm, when the label of two pictures is entirely different, is enabled by the positions the k binary set of 1, -1 composition
Hamming distance between binary set is at leastWhen label overlaps, the Hamming distance of binary set is enabled to be at least
The specific calculations of threshold parameter m are following (m* initial values are 2k)：
Function：Calculate threshold parameter m
Input:m*,n1,k.
Output:m.
1:M=0；
2:while m<m*do
3:M+=4n1；
4:end while
5:if m>4k then
6:M=4k；
7:end if
8:return m。
To make full use of computing resource and memory space, the present invention to generate image tag online from each minimum batch
It is right.For the image pair for obtaining between different batches, the training of a full dose data set is often completed, the present invention can upset training set
Picture sequence, the following (Labels (I of specific implementation1) it is picture I1Associated number of labels)：
Function：It is online to generate image pair
Input:a batch S of 32training images
Output:a set T of 2-tuples
1：
2：For every 2-tuple t=(I1；I2)that I1∈S；I2∈S do
3：if Labels(I1)>0then
4：T←T∪t
5：end if
6：end for
7：return T.
Embodiment 1
Assuming that possessing the artificial V pictures marked now, there is at least one label to be used for characterizing the figure per pictures
The semantic information of piece, these pictures will train neural network model as training set.
It is convolutional layer one, maximum pond layer one, volume successively from front to back 1. building network model according to the structure in Fig. 1
Lamination two, maximum pond layer two, convolutional layer three, convolutional layer four, convolutional layer five, maximum pond layer, full articulamentum six, full articulamentum
Seven and Hash layer, wherein each channel of convolutional layer a pair of image carries out convolution operation, behind each convolutional layer to last layer
Output carry out convolution operation, maximum pond layer to the output of last layer carries out that maximum regional value is taken to operate, and full articulamentum is to upper
One layer of output carries out full attended operation, and the input of loss function is the output of the last one Hash layer.
2. the v inputs as network model are chosen in order from V pictures, according to the loss of formula (5) computation model
Value, wherein.By minimum batch gradient descent method, model is trained, obtains the value of each parameter in model,
Model parameter includes convolution nuclear parameter, each Node connectedness weight of full articulamentum.
3. random upset picture sequence, return to step 2 is until the penalty values being calculated according to formula (5) tend towards stability.Really
The value of all parameters in cover half type.
4. all pictures in image data base are sent into model, obtain the binary set of every pictures and stored.
Embodiment 2
Assuming that there is picture to be checked, which need not possess label information, and the present invention wishes the picture number in magnanimity
According to finding what semantic most like pictures in library：
1. as input, it is input in the model that training obtains, the value of its binary set is calculated.
2. then calculate Hamming distance with the binary sets of all pictures in database, the smaller picture of Hamming distance with
It is more similar to inquire picture.
3. sorting according to Hamming distance, image results collection is returned by sequence from small to large.
It is above to implement to be merely illustrative of the technical solution of the present invention rather than be limited, the ordinary skill people of this field
Member can be modified or replaced equivalently technical scheme of the present invention, without departing from the spirit and scope of the present invention, this hair
Bright protection domain should be subject to described in claims.
Claims (7)
1. a kind of binary set generation method of multi-tag image, step include：
1) it is trained using training data set pair convolutional neural networks model, until the penalty values of the convolutional neural networks model
It tends towards stability；Wherein, every time training when training dataset include N to picture, for i-th pair picture II, 1、II, 2If picture
II, 1Number of labels be nI, 1, picture I in i-th pair imageI, 1With picture II, 2Common associated number of labels is nI, 2, calculate institute
The loss function η for stating penalty values is
Wherein, nI, 1=nI, 2When, yi=0, otherwise yi=1；Ones representative elements are all 1 vector, | | | |1Indicate vector
First norm,Indicate the Euclidean distance between vector, | | it indicates to carry out absolute value operation per bit element to vector, α is to use
Come control quantization loss size parameter, w be convolutional neural networks model Hash layer weight vectors, f (I；W) it is convolution god
The positions the k binary set of picture I through network model output, m refers to Hamming distance threshold parameter；
2) binary set of picture is calculated using the convolutional neural networks model after training.
2. a kind of multi-tag image, semantic similarity search method, step include：
1) it is trained using training data set pair convolutional neural networks model, until the penalty values of the convolutional neural networks model
It tends towards stability；Wherein, every time training when training dataset include N to picture, for i-th pair picture II, 1、II, 2If picture
II, 1Number of labels be nI, 1, picture I in i-th pair imageI, 1With picture II, 2Common associated number of labels is nI, 2, calculate institute
The loss function η for stating penalty values is
Wherein, nI, 1=nI, 2When, yi=0, otherwise yi=1；Ones representative elements are all 1 k bit vectors, | | | |1Indicate vector
The first norm,Indicate the Euclidean distance between vector, | | it indicates to carry out absolute value operation per bit element to vector, α is
For controlling the parameter of quantization loss size, w is the weight vectors of the Hash layer of convolutional neural networks model, f (I；W) it is convolution
The positions the k binary set of the picture I of neural network model output, m refers to Hamming distance threshold parameter；
2) picture in image data base is calculated using the convolutional neural networks model after training, obtains every pictures
Binary set is simultaneously stored；
3) binary set of picture to be checked is calculated using the convolutional neural networks model after training, and itself and step 2) are obtained
Binary set carry out similarity calculation, according to similarity calculation result return with the most like several figures of the picture to be checked
Piece.
3. method as claimed in claim 1 or 2, which is characterized in that train the convolution refreshing using minimum batch gradient descent method
Through network model, the value of the loss function can be minimized.
4. method as claimed in claim 1 or 2, which is characterized in that the convolutional neural networks model includes sequentially connected
One convolutional layer, maximum pond layer, the second convolutional layer, maximum pond layer, third convolutional layer, Volume Four lamination, the 5th convolutional layer, most
Great Chiization layer, the first full articulamentum, the second full articulamentum and Hash layer.
5. method as claimed in claim 4, which is characterized in that the hash function of the Hash layer is h (x；W)=sign (f
(x；w))；Wherein, f (x；W)=wTF ' (x), f ' (x) are the output vector of the second full articulamentum；The Hash layer is equipped with k section
A weight is arranged in point, each node, these weights constitute weight vectors w.
6. method as claimed in claim 1 or 2, which is characterized in that α=0.01.
7. method as claimed in claim 2, which is characterized in that obtained according to the binary set of picture to be checked and step 2)
The Hamming distance of binary set determines the similarity.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810111604.1A CN108446307A (en) | 2018-02-05 | 2018-02-05 | A kind of the binary set generation method and image, semantic similarity search method of multi-tag image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810111604.1A CN108446307A (en) | 2018-02-05 | 2018-02-05 | A kind of the binary set generation method and image, semantic similarity search method of multi-tag image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108446307A true CN108446307A (en) | 2018-08-24 |
Family
ID=63191589
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810111604.1A Pending CN108446307A (en) | 2018-02-05 | 2018-02-05 | A kind of the binary set generation method and image, semantic similarity search method of multi-tag image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108446307A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197721A (en) * | 2019-05-06 | 2019-09-03 | 平安科技（深圳）有限公司 | Tendon condition evaluation method, apparatus and storage medium based on deep learning |
| CN110796182A (en) * | 2019-10-15 | 2020-02-14 | 西安网算数据科技有限公司 | Bill classification method and system for small amount of samples |
| CN112685590A (en) * | 2020-12-29 | 2021-04-20 | 电子科技大学 | Image retrieval method based on convolutional neural network regularization processing |
| CN113254695A (en) * | 2021-05-27 | 2021-08-13 | 支付宝(杭州)信息技术有限公司 | Image retrieval method and device, and training method and device of image characterization network |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106897366A (en) * | 2017-01-12 | 2017-06-27 | 华南理工大学 | Image search method based on face convolutional neural networks and random kd trees forest |
| CN107092661A (en) * | 2017-03-28 | 2017-08-25 | 桂林明辉信息科技有限公司 | A kind of image search method based on depth convolutional neural networks |
| CN108399185A (en) * | 2018-01-10 | 2018-08-14 | 中国科学院信息工程研究所 | A kind of the binary set generation method and image, semantic similarity search method of multi-tag image |
-
2018
- 2018-02-05 CN CN201810111604.1A patent/CN108446307A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106897366A (en) * | 2017-01-12 | 2017-06-27 | 华南理工大学 | Image search method based on face convolutional neural networks and random kd trees forest |
| CN107092661A (en) * | 2017-03-28 | 2017-08-25 | 桂林明辉信息科技有限公司 | A kind of image search method based on depth convolutional neural networks |
| CN108399185A (en) * | 2018-01-10 | 2018-08-14 | 中国科学院信息工程研究所 | A kind of the binary set generation method and image, semantic similarity search method of multi-tag image |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197721A (en) * | 2019-05-06 | 2019-09-03 | 平安科技（深圳）有限公司 | Tendon condition evaluation method, apparatus and storage medium based on deep learning |
| CN110197721B (en) * | 2019-05-06 | 2023-06-06 | 平安科技（深圳）有限公司 | Tendon condition assessment method, device and storage medium based on deep learning |
| CN110796182A (en) * | 2019-10-15 | 2020-02-14 | 西安网算数据科技有限公司 | Bill classification method and system for small amount of samples |
| CN112685590A (en) * | 2020-12-29 | 2021-04-20 | 电子科技大学 | Image retrieval method based on convolutional neural network regularization processing |
| CN112685590B (en) * | 2020-12-29 | 2022-10-14 | 电子科技大学 | Image retrieval method based on convolutional neural network regularization processing |
| CN113254695A (en) * | 2021-05-27 | 2021-08-13 | 支付宝(杭州)信息技术有限公司 | Image retrieval method and device, and training method and device of image characterization network |
| CN113254695B (en) * | 2021-05-27 | 2022-06-07 | 支付宝(杭州)信息技术有限公司 | Image retrieval method and device, and training method and device of image characterization network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108399185A (en) | A kind of the binary set generation method and image, semantic similarity search method of multi-tag image | |
| Zhang et al. | Improved deep hashing with soft pairwise similarity for multi-label image retrieval | |
| Lai et al. | Instance-aware hashing for multi-label image retrieval | |
| CN109299342B (en) | Cross-modal retrieval method based on cycle generation type countermeasure network | |
| CN110188227B (en) | Hash image retrieval method based on deep learning and low-rank matrix optimization | |
| CN111291836B (en) | Method for generating student network model | |
| Wang et al. | Robust and flexible discrete hashing for cross-modal similarity search | |
| Wang et al. | Effective multi-query expansions: Collaborative deep networks for robust landmark retrieval | |
| CN107480261B (en) | Fine-grained face image fast retrieval method based on deep learning | |
| CN109241317B (en) | Pedestrian Hash retrieval method based on measurement loss in deep learning network | |
| CN108038122B (en) | Trademark image retrieval method | |
| CN110046656B (en) | Multi-mode scene recognition method based on deep learning | |
| CN110110122A (en) | Image based on multilayer semanteme depth hash algorithm-text cross-module state retrieval | |
| CN108446307A (en) | A kind of the binary set generation method and image, semantic similarity search method of multi-tag image | |
| CN109783666B (en) | Image scene graph generation method based on iterative refinement | |
| CN110826338B (en) | Fine-grained semantic similarity recognition method for single-selection gate and inter-class measurement | |
| CN107330355B (en) | Deep pedestrian re-identification method based on positive sample balance constraint | |
| CN105718960A (en) | Image ordering model based on convolutional neural network and spatial pyramid matching | |
| CN104156464B (en) | Micro- video retrieval method and device based on micro- video frequency feature data storehouse | |
| US11860932B2 (en) | Scene graph embeddings using relative similarity supervision | |
| Lee et al. | Photo aesthetics analysis via DCNN feature encoding | |
| CN110457514A (en) | A kind of multi-tag image search method based on depth Hash | |
| Seng et al. | Big feature data analytics: Split and combine linear discriminant analysis (SC-LDA) for integration towards decision making analytics | |
| CN108595546B (en) | Semi-supervision-based cross-media feature learning retrieval method | |
| CN110598022B (en) | Image retrieval system and method based on robust deep hash network |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180824 |