CN107784122A - A kind of instance-level image search method represented based on multilayer feature - Google Patents
A kind of instance-level image search method represented based on multilayer feature Download PDFInfo
- Publication number
- CN107784122A CN107784122A CN201711176421.XA CN201711176421A CN107784122A CN 107784122 A CN107784122 A CN 107784122A CN 201711176421 A CN201711176421 A CN 201711176421A CN 107784122 A CN107784122 A CN 107784122A
- Authority
- CN
- China
- Prior art keywords
- picture
- feature
- value
- data
- dimension
- 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.)
- Withdrawn
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
- G06F16/50—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
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
Landscapes
- Engineering & Computer Science (AREA)
- Library & Information Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Processing Or Creating Images (AREA)
Abstract
The invention discloses a kind of instance-level image search method represented based on multilayer feature, comprise the following steps：Extracted region：One width picture is cut into multiple objects, picture in its entirety is described by multiple objects, in object aspect, the object that most pictures are included is not over double figures；Screen in region；Feature coding；Index construct, the present invention proposes the concept of the multilayer feature of depth structure, from the angle of biology, explain the mode that people observe object in natural world, well and this mode is extracted into well applied among the retrieval mode of picture, several different region representations are split into comprising multiple object pictures by one, from natural understanding of the people for image is semantically met naturally, more press close to user's custom of people again；On retrieval mode, the present invention proposes to be retrieved using the indexed mode based on KD trees, has analysed in depth KDTree in terms of index compared to the superiority and inferiority of linear directory.
Description
Technical field
The present invention relates to picture search technical field, specially a kind of instance-level picture search represented based on multilayer feature
Method.
Background technology
With developing rapidly for internet.Volatile growth is presented in the quantity of online picture, especially nearest mobile mutual
The fast development at networking end, picture are gradually hot as the carrier that its information is presented.With emerging in large numbers for a large amount of picture resources.Big
Image retrieval demand under the image scale of scale is prominent all the more.Requirement of the people for retrieval also more and more higher.Traditional is logical
Crossing word can not more meet people for Novel Internet to retrieve as keyword using the pattern of label mark picture
It is required that.More, picture as retrieving associated class by picture semanteme in itself is the developing direction of the picture retrieval in future.
How by extracting the semanteme of picture in itself.The well and information extracted by it is retrieved.
So how to design a kind of instance-level image search method represented based on multilayer feature, turn into us and currently
Solve the problems, such as.
The content of the invention
It is an object of the invention to provide a kind of instance-level image search method represented based on multilayer feature, with solution
State the problem of being proposed in background technology.
To achieve the above object, the present invention provides following technical scheme：A kind of instance-level figure represented based on multilayer feature
As searching method, comprise the following steps：
1) extracted region：One width picture is cut into multiple objects, picture in its entirety is described by multiple objects, in object layer
On face, not over double figures, this also to be adapted to some of adaptation on sift good the object that most pictures are included
Good operation method, therefore, we can take MAX function to extract provincial characteristics；
2) region is screened：In field of object detection, differentiate whether a feature represents an object using SVM, screen
During, the method with reference to used in SVM, greatly reduce in the quantity Jing Guo NMS processing, candidate frame, passing through this point
In analysis, rule of thumb we need to set a threshold value, if selecting the scoring of frame to be more than this threshold value can just be added into most
In candidate frame afterwards, final image result is finally obtained；
3) feature invariance is analyzed：Multi-stage characteristics are cut out in level, are cut out vertically, in terms of scaling, especially in terms of scaling
More there is good stability than single-stage feature, in scaling, the size of its Euclidean distance reaches when the size of size is 1.5
To extreme point；
4) feature coding：In order that quickly carry out image retrieval with the characteristic vector extracted, it would be desirable to obtaining
The characteristic vector that obtains carries out compression, coding work accordingly, can so reduce in retrieval when ask complexity, convolution
Neutral net is made up of convolutional layer, down-sampled layer, full articulamentum, and for picture when by convolutional layer, each characteristic pattern is by difference
Convolution kernel carry out convolution algorithm acquisition on picture, the different structures in the representative picture of each convolution kernel, Relu
It is an original place operation, activation value of the activation value for the node of negative value is directly disposed as into 0, it ensure that working as in hidden node
The nonnegativity of middle activation value, down-sampled, i.e. pond layer, by way of max-pooling, using getting covered region
Interior peak describes the region, and by way of this max-pooling, picture has for a certain degree of noise to be compared
Good resistance, is handled digital picture using PCA dimensionality reductions；
5) index construct：One Kd-Tree of structure is closed in a K dimension data collection to represent to the K dimension data set structures
Into one of K dimension spaces division, that is, each node in setting just corresponded to the hypermatrix region of a K dimension, and KD trees are one kind
Space partition tree, whole space is exactly divided into specific several parts, then carried out in the part of particular space related
Search operation.
According to above-mentioned technical proposal, the fundamental formular of MAX function is expressed as in the step 1)：
Wherein, what f () was represented is used phase knowledge and magnanimity metric function, bx iRepresent X i-th of candidate frame, by jRepresent
Scheme Y j-th candidates frame.
According to above-mentioned technical proposal, selected in the step 2) frame screen process among, except need filter out with height
Beyond the semantic part of level, we still need to consider 2 factors：
(1) picture includes advanced definition, but selects frame not cover, or SVM scorings are too low, it is believed that it can not
It is a subject area that can be regarded as；
(2) image in itself and does not include high-level semantics.
According to above-mentioned technical proposal, in the step 4) data carry out it is default after, by only selecting above some
Individual latitude reduces the purpose of data latitude to reach, and the use of the algorithm needs two processes, training and calculating：
(1) after taking sample set first, the covariance matrix C of the covariance matrix generation sample of calculating matrix；
(2) corresponding covariance matrix C characteristic vector { e is calculated1,e2,e3... } and characteristic value { λ1, λ2, λ3...；
(3) data for projection is projected in new feature space；
(4) k (k before choosing<N) individual characteristic vector represents the feature after dimensionality reduction.
According to above-mentioned technical proposal, the building course of step 5) the KD trees is as follows：
(1) split domains are determined, to all description subdatas (characteristic vector), count them on each dimension
Variance, exemplified by thinking SURF features, 64 variances can be calculated, pick out the maximum of wherein variance, the dimension where it
It is exactly the value in split domains, distribution of the maximum explanation data of data variance on this dimension is more dispersed, it may have relatively good
Identification, segmentation is carried out in this dimension can obtain relatively good data division；
(2) data field is determined, set of data points is ranked up according to the value in split domains, by being arranged to positioned at middle
Data point, this selection is to carry out binary chop in this dimension；
(3) by split domains value it is smaller than the value in selected data domain be divided into one group, equally, will be bigger than the value of data field
Be divided into one group, repeat step (1), a node port only included in sky is asked.
According to above-mentioned technical proposal, in the step 2) feature extraction contain two parts, a part is whole from one
Characteristic pattern is got among the picture opened, when calculating the feature of each candidate frame again after this, takes following strategy：
(1) it is { 480 by artwork progress scaling to minimum edge；576；688；864；1200 } each yardstick in, is calculated
Picture corresponds to character representation of the picture under conv5 layers；
(2) for each candidate frame, the element trees that it is included under each yardstick are calculated, and selected
The number of elements included under corresponding yardstick closest to 224 × 224 sizes yardstick as best scale.
Compared with prior art, the beneficial effects of the invention are as follows：The present invention proposes the general of the multilayer feature of depth structure
Read, from the angle of biology, explain the mode that people observe object in natural world, well and extract this mode
Out well has been applied among the retrieval mode of picture, and several different regions are split into comprising multiple object pictures by one
Represent, from natural understanding of the people for image is semantically met naturally, more press close to user's custom of people again；Retrieving
In mode, the present invention proposes to be retrieved using the indexed mode based on KD trees, has analysed in depth KDTree in terms of index
Compared to the superiority and inferiority of linear directory.
Brief description of the drawings
Fig. 1 is the image search method flow chart of the present invention；
Fig. 2 is the KD trees structure flow chart of the present invention；
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Site preparation describes, it is clear that described embodiment is only part of the embodiment of the present invention, rather than whole embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art are obtained every other under the premise of creative work is not made
Embodiment, belong to the scope of protection of the invention.
Fig. 1-2 is referred to, the present invention provides a kind of instance-level image search method represented based on multilayer feature, including such as
Lower step：
1) extracted region：One width picture is cut into multiple objects, picture in its entirety is described by multiple objects, in object layer
On face, not over double figures, this also to be adapted to some of adaptation on sift good the object that most pictures are included
Good operation method, therefore, we can take MAX function to extract provincial characteristics；
2) region is screened：In field of object detection, differentiate whether a feature represents an object using SVM, screen
During, the method with reference to used in SVM, greatly reduce in the quantity Jing Guo NMS processing, candidate frame, passing through this point
In analysis, rule of thumb we need to set a threshold value, if selecting the scoring of frame to be more than this threshold value can just be added into most
In candidate frame afterwards, final image result is finally obtained；
3) feature invariance is analyzed：Multi-stage characteristics are cut out in level, are cut out vertically, in terms of scaling, especially in terms of scaling
More there is good stability than single-stage feature, in scaling, the size of its Euclidean distance reaches when the size of size is 1.5
To extreme point；
4) feature coding：In order that quickly carry out image retrieval with the characteristic vector extracted, it would be desirable to obtaining
The characteristic vector that obtains carries out compression, coding work accordingly, can so reduce in retrieval when ask complexity, convolution
Neutral net is made up of convolutional layer, down-sampled layer, full articulamentum, and for picture when by convolutional layer, each characteristic pattern is by difference
Convolution kernel carry out convolution algorithm acquisition on picture, the different structures in the representative picture of each convolution kernel, Relu
It is an original place operation, activation value of the activation value for the node of negative value is directly disposed as into 0, it ensure that working as in hidden node
The nonnegativity of middle activation value, down-sampled, i.e. pond layer, by way of max-pooling, using getting covered region
Interior peak describes the region, and by way of this max-pooling, picture has for a certain degree of noise to be compared
Good resistance, is handled digital picture using PCA dimensionality reductions；
5) index construct：One Kd-Tree of structure is closed in a K dimension data collection to represent to the K dimension data set structures
Into one of K dimension spaces division, that is, each node in setting just corresponded to the hypermatrix region of a K dimension, and KD trees are one kind
Space partition tree, whole space is exactly divided into specific several parts, then carried out in the part of particular space related
Search operation.
According to above-mentioned technical proposal, the fundamental formular of MAX function is expressed as in step 1)：
Wherein, what f () was represented is used phase knowledge and magnanimity metric function, bx iRepresent X i-th of candidate frame, by jRepresent
Scheme Y j-th candidates frame.
According to above-mentioned technical proposal, selected in step 2) frame screen process among, except need filter out with advanced language
Beyond the part of justice, we still need to consider 2 factors：
(1) picture includes advanced definition, but selects frame not cover, or SVM scorings are too low, it is believed that it can not
It is a subject area that can be regarded as；
(2) image in itself and does not include high-level semantics.
According to above-mentioned technical proposal, in step 4) after data preset, by only selecting several latitudes above
Spend to reach the purpose of reduction data latitude, the use of the algorithm needs two processes, and training is with calculating：
(1) after taking sample set first, the covariance matrix C of the covariance matrix generation sample of calculating matrix；
(2) corresponding covariance matrix C characteristic vector { e is calculated1,e2,e3... } and characteristic value { λ1, λ2, λ3...；
(3) data for projection is projected in new feature space；
(4) k (k before choosing<N) individual characteristic vector represents the feature after dimensionality reduction.
According to above-mentioned technical proposal, the building course of step 5) KD trees is as follows：
(1) split domains are determined, to all description subdatas (characteristic vector), count them on each dimension
Variance, exemplified by thinking SURF features, 64 variances can be calculated, pick out the maximum of wherein variance, the dimension where it
It is exactly the value in split domains, distribution of the maximum explanation data of data variance on this dimension is more dispersed, it may have relatively good
Identification, segmentation is carried out in this dimension can obtain relatively good data division；
(2) data field is determined, set of data points is ranked up according to the value in split domains, by being arranged to positioned at middle
Data point, this selection is to carry out binary chop in this dimension；
(3) by split domains value it is smaller than the value in selected data domain be divided into one group, equally, will be bigger than the value of data field
Be divided into one group, repeat step (1), a node port only included in sky is asked.
According to above-mentioned technical proposal, in step 2) feature extraction contain two parts, a part is from one whole
Characteristic pattern is got among picture, when calculating the feature of each candidate frame again after this, takes following strategy：
(1) it is { 480 by artwork progress scaling to minimum edge；576；688；864；1200 } each yardstick in, is calculated
Picture corresponds to character representation of the picture under conv5 layers；
(2) for each candidate frame, the element trees that it is included under each yardstick are calculated, and selected
The number of elements included under corresponding yardstick closest to 224 × 224 sizes yardstick as best scale.
Based on above-mentioned, it is an advantage of the current invention that the present invention proposes the concept of the multilayer feature of depth structure, from biology
Angle is set out, and explains the mode that people observe object in natural world, well and this mode is extracted into well fortune
Use among the retrieval mode of picture, several different region representations are split into comprising multiple object pictures by one, from
Naturally semantically meet natural understanding of the people for image, more press close to user's custom of people again；On retrieval mode,
The present invention proposes to be retrieved using the indexed mode based on KD trees, has analysed in depth KDTree in terms of index compared to line
The superiority and inferiority that sex cords draws.
Although an embodiment of the present invention has been shown and described, for the ordinary skill in the art, can be with
A variety of changes, modification can be carried out to these embodiments, replace without departing from the principles and spirit of the present invention by understanding
And modification, the scope of the present invention is defined by the appended.
Claims (6)
- A kind of 1. instance-level image search method represented based on multilayer feature, it is characterised in that：Comprise the following steps：1) extracted region：One width picture is cut into multiple objects, picture in its entirety is described by multiple objects, in object aspect On, not over double figures, this also to be adapted to some of adaptation on sift good the object that most pictures are included Operation method, therefore, we can take MAX function to extract provincial characteristics；2) region is screened：In field of object detection, differentiate whether a feature represents an object, the mistake of screening using SVM Cheng Zhong, the method with reference to used in SVM, greatly reduces in the quantity Jing Guo NMS processing, candidate frame, in by this analysis, Rule of thumb we need to set a threshold value, if last time can be just added into by selecting the scoring of frame to be more than this threshold value Select in frame, finally obtain final image result；3) feature invariance is analyzed：Multi-stage characteristics are cut out in level, are cut out vertically, in terms of scaling, especially than single in terms of scaling Level feature more has good stability, and in scaling, when the size of size is 1.5, the size of its Euclidean distance reaches pole Value point；4) feature coding：In order that quickly carry out image retrieval with the characteristic vector extracted, it would be desirable to obtaining The characteristic vector that arrives carries out compression, coding work accordingly, can so reduce in retrieval when ask complexity, convolutional Neural Network is made up of convolutional layer, down-sampled layer, full articulamentum, and for picture when by convolutional layer, each characteristic pattern is by different volumes Product core carries out convolution algorithm acquisition on picture, and the different structures in picture representated by each convolution kernel, Relu is one Individual original place operation, activation value of the activation value for the node of negative value is directly disposed as into 0, it ensure that swashing among hidden node The nonnegativity of value living, down-sampled, i.e. pond layer, by way of max-pooling, using getting in covered region Peak describes the region, and by way of this max-pooling, picture has relatively good for a certain degree of noise Resistance, digital picture is handled using PCA dimensionality reductions；5) index construct：A K dimension data collection close structure one Kd-Tree represent to the K dimension datas set form K One division of dimension space, that is, each node in setting just have corresponded to the hypermatrix region of a K dimension, and KD trees are that a kind of space is drawn Divide tree, whole space is exactly divided into specific several parts, relevant search behaviour is then carried out in the part of particular space Make.
- A kind of 2. instance-level image search method represented based on multilayer feature according to claim 1, it is characterised in that：It is described The fundamental formular of MAX function is expressed as in step 1)：<mrow> <mi>S</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>)</mo> </mrow> <mo>=</mo> <mi>m</mi> <mi>a</mi> <mi>x</mi> <mrow> <mo>(</mo> <mi>f</mi> <mo>(</mo> <mrow> <msubsup> <mi>B</mi> <mi>x</mi> <mi>i</mi> </msubsup> <mo>,</mo> <msubsup> <mi>B</mi> <mi>y</mi> <mi>j</mi> </msubsup> </mrow> <mo>)</mo> <mo>)</mo> </mrow> </mrow>Wherein, what f () was represented is used phase knowledge and magnanimity metric function, bx iRepresent X i-th of candidate frame, by jRepresentative graph Y's J-th candidates frame.
- A kind of 3. instance-level image search method represented based on multilayer feature according to claim 1, it is characterised in that：It is described Among the process that frame screening is selected in step 2), in addition to needing to filter out with the part of high-level semantics, we still need Consider 2 factors：(1) picture includes advanced definition, but selects frame not cover, or SVM scorings are too low, it is believed that it can not can be regarded as It is a subject area；(2) image in itself and does not include high-level semantics.
- A kind of 4. instance-level image search method represented based on multilayer feature according to claim 1, it is characterised in that：It is described In step 4) after data preset, reach the mesh of reduction data latitude by only selecting several latitudes above , the use of the algorithm needs two processes, training and calculating：(1) after taking sample set first, the covariance matrix C of the covariance matrix generation sample of calculating matrix；(2) corresponding covariance matrix C characteristic vector { e is calculated1,e2,e3... } and characteristic value { λ1, λ2, λ3...；(3) data for projection is projected in new feature space；(4) k (k before choosing<N) individual characteristic vector represents the feature after dimensionality reduction.
- A kind of 5. instance-level image search method represented based on multilayer feature according to claim 1, it is characterised in that：It is described The building course of step 5) KD trees is as follows：(1) split domains are determined, to all description subdatas (characteristic vector), count their variances on each dimension, Exemplified by thinking SURF features, 64 variances can be calculated, pick out the maximum of wherein variance, the dimension where it is exactly The value in split domains, distribution of the maximum explanation data of data variance on this dimension are more dispersed, it may have relatively good distinguishes Knowledge and magnanimity, segmentation is carried out in this dimension can obtain relatively good data division；(2) data field is determined, set of data points is ranked up according to the value in split domains, will be arranged to data positioned at middle Point, this selection is to carry out binary chop in this dimension；(3) by split domains value it is smaller than the value in selected data domain be divided into one group, equally, by point bigger than the value of data field For one group, repeat step (1), a node port is only included in sky is asked.
- A kind of 6. instance-level image search method represented based on multilayer feature according to claim 1, it is characterised in that：It is described Feature extraction contains two parts in step 2), and a part is to get characteristic pattern among the picture of one whole, at this When calculating the feature of each candidate frame again afterwards, following strategy is taken：(1) it is { 480 by artwork progress scaling to minimum edge；576；688；864；1200 } each yardstick in, calculates picture Character representation of the corresponding picture under conv5 layers；(2) for each candidate frame, the element trees that it is included under each yardstick are calculated, and selected corresponding The number of elements included under yardstick closest to 224 × 224 sizes yardstick as best scale.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711176421.XA CN107784122A (en) | 2017-11-22 | 2017-11-22 | A kind of instance-level image search method represented based on multilayer feature |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711176421.XA CN107784122A (en) | 2017-11-22 | 2017-11-22 | A kind of instance-level image search method represented based on multilayer feature |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107784122A true CN107784122A (en) | 2018-03-09 |
Family
ID=61430592
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711176421.XA Withdrawn CN107784122A (en) | 2017-11-22 | 2017-11-22 | A kind of instance-level image search method represented based on multilayer feature |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107784122A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108537732A (en) * | 2018-04-10 | 2018-09-14 | 福州大学 | Fast image splicing method based on PCA-SIFT |
| CN108871760A (en) * | 2018-06-07 | 2018-11-23 | 广东石油化工学院 | A kind of high-efficient gear method of fault pattern recognition |
| CN109215034A (en) * | 2018-07-06 | 2019-01-15 | 成都图必优科技有限公司 | A kind of Weakly supervised image, semantic dividing method for covering pond based on spatial pyramid |
| CN111708920A (en) * | 2020-06-06 | 2020-09-25 | 谢国柱 | Internet big data processing method based on artificial intelligence and intelligent cloud service platform |
-
2017
- 2017-11-22 CN CN201711176421.XA patent/CN107784122A/en not_active Withdrawn
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108537732A (en) * | 2018-04-10 | 2018-09-14 | 福州大学 | Fast image splicing method based on PCA-SIFT |
| CN108537732B (en) * | 2018-04-10 | 2021-11-02 | 福州大学 | PCA-SIFT-based rapid image splicing method |
| CN108871760A (en) * | 2018-06-07 | 2018-11-23 | 广东石油化工学院 | A kind of high-efficient gear method of fault pattern recognition |
| CN109215034A (en) * | 2018-07-06 | 2019-01-15 | 成都图必优科技有限公司 | A kind of Weakly supervised image, semantic dividing method for covering pond based on spatial pyramid |
| CN111708920A (en) * | 2020-06-06 | 2020-09-25 | 谢国柱 | Internet big data processing method based on artificial intelligence and intelligent cloud service platform |
| CN111708920B (en) * | 2020-06-06 | 2021-01-08 | 广东和邦网络科技有限公司 | Internet big data processing method based on artificial intelligence and intelligent cloud service platform |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107784122A (en) | A kind of instance-level image search method represented based on multilayer feature | |
| CN108052593B (en) | Topic keyword extraction method based on topic word vector and network structure | |
| Wang et al. | Weakly supervised learning for attribute localization in outdoor scenes | |
| CN104834693B (en) | Visual pattern search method and system based on deep search | |
| Sanoja et al. | Block-o-matic: A web page segmentation framework | |
| Perret et al. | Evaluation of hierarchical watersheds | |
| Huang et al. | Unconstrained multimodal multi-label learning | |
| Nowak et al. | Overview of the CLEF 2009 large-scale visual concept detection and annotation task | |
| CN109344298A (en) | A kind of method and device converting unstructured data to structural data | |
| Mzoughi et al. | Advanced tree species identification using multiple leaf parts image queries | |
| CN109815357A (en) | A kind of remote sensing image retrieval method based on Nonlinear Dimension Reduction and rarefaction representation | |
| CN103440315B (en) | A kind of Web page cleaning method based on theme | |
| CN111177410B (en) | Knowledge graph storage and similarity retrieval method based on evolution R-tree | |
| Mamatha et al. | Content based image retrieval of satellite imageries using soft query based color composite techniques | |
| CN109145971A (en) | Based on the single sample learning method for improving matching network model | |
| Yan et al. | Change-based image cropping with exclusion and compositional features | |
| Kuric et al. | ANNOR: Efficient image annotation based on combining local and global features | |
| CN107145519A (en) | A kind of image retrieval and mask method based on hypergraph | |
| Koskela et al. | Comparison of techniques for content-based image retrieval | |
| Chen et al. | Bilinear CNN models for food recognition | |
| CN103049570B (en) | Based on the image/video search ordering method of relevant Preserving map and a sorter | |
| CN112214488A (en) | European style spatial data index tree and construction and retrieval method | |
| CN104156423B (en) | Multiple dimensioned video key frame extracting method based on integer programming | |
| CN109766468A (en) | A kind of implementation method and device retrieved in appearance patent image based on iamge description algorithm with management | |
| Pandey et al. | Clustering of hierarchical image database to reduce inter-and intra-semantic gaps in visual space for finding specific image semantics |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20180309 |