CN110135440A - A kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval - Google Patents
A kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval Download PDFInfo
- Publication number
- CN110135440A CN110135440A CN201910401697.6A CN201910401697A CN110135440A CN 110135440 A CN110135440 A CN 110135440A CN 201910401697 A CN201910401697 A CN 201910401697A CN 110135440 A CN110135440 A CN 110135440A
- Authority
- CN
- China
- Prior art keywords
- characteristic pattern
- image
- size
- characteristic
- grid
- 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
- G06F16/50—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
- G06F16/55—Clustering; Classification
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Library & Information Science (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image characteristic extracting methods suitable for magnanimity Cultural Relics Image Retrieval, comprising: step 1. carries out propagated forward processing to input picture I with the depth convolutional neural networks for removing full articulamentum;Step 2. extracts characteristic pattern F from the output end of used depth convolutional neural networks respectively3、F4And F5;Step 3. is to characteristic pattern F3、F4Pond processing is carried out, characteristic pattern F ' is respectively obtained3With characteristic pattern F '4, concatenate characteristic pattern F '3, F '4, the result after concatenation is handled by convolution algorithm to obtain output characteristic pattern F 'o;Step 4. is by F 'oIt is processed into single pass characteristic pattern FoObtain the adaptive aggregate weight distribution map of local feature;Step 5. is by F5In each grid in element according to the adaptive aggregate weight distribution map F of local featureoCarry out aminated polyepichlorohydrin;Step 6. carries out model training to step 1-5 by loss function.
Description
Technical field
The present invention relates to field of image processings more particularly to a kind of characteristics of image suitable for magnanimity Cultural Relics Image Retrieval to mention
Take method.
Background technique
China has a long history culture, and the history of 5,000-year and down leaves the abundant and valuable legacy of history.?
China has museums at different levels up to more than 5100 families, and all kinds of historical relic quantity of collection have million.At present only about the introduction of historical relic
It is limited to the historical relic brief introduction of museum's offer, information contained amount is extremely limited.If can be by Cultural Relics Image Retrieval method, use
The cultural relic images of family shooting are associated with corresponding target historical relic, and spectators are understood with the story of historical relic behind, carries forward national tradition
Cultural significance is great.
However, the collection work of cultural relic images has due to each museum's cultural relics in the collection of cultural institution enormous amount and widely distributed
Great challenge.Many historical relics can only collect the image under very harsh qualifications, therefore generate separating capacity
Strong characteristics of image is the important prerequisite for realizing effective image retrieval.Current image characteristic extracting method can be divided mainly into two
Class.One kind is global characteristics extracting method, and entire image is expressed as a vector by this method, and the image that this method obtains is special
It is lower to levy discrimination, it is difficult to obtain good retrieval rate.Another method uses local feature, and this method is from image
Expression of the feature as image is extracted in regional area.The problem of such method, is, if the quantity of local feature is too small
It is not sufficient enough to indicate the content of image, and excessive bear can be brought to storage and calculating if the quantity of local feature is excessive
Load.Meanwhile the local feature extracted from image cannot be guaranteed there is sufficient information representation ability, this meeting greatly influence diagram
As the accuracy of retrieval.
The characteristics of in view of cultural relic images, and be currently used in the feature extracting method of image retrieval there are the problem of, this hair
It is bright to propose a kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval.
Summary of the invention
Purpose to realize the present invention, is achieved using following technical scheme:
A kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval, this method include following in the training stage
Step:
Step 1. is W to size with the ResNet-50 depth convolutional neural networks for removing full articulamentumI×HI× 3 it is defeated
Enter image I and carries out propagated forward processing, WIFor picture traverse, HI3 channels are indicated for picture altitude, 3;
Third convolution module, four convolution module and fiveth of the step 2. from used depth convolutional neural networks
It is W that the output end of a convolution module extracts size respectively3×H3×C3、W4×H4×C4And W5×H5×C5Characteristic pattern F3、F4With
F5, wherein W, H and C are respectively the width, height and port number of characteristic pattern;
Step 3. is to characteristic pattern F3、F4Pond processing is carried out, respectively obtaining size is W '3×H’3×C’3Characteristic pattern F '3
It is W ' with size4×H’4×C’4Characteristic pattern F '4, concatenate characteristic pattern F '3, F '4, by convolution algorithm to the result after concatenation into
It is W ' that row processing, which obtains size,o×H’o×C’oOutput characteristic pattern F 'o;
Step 4. is by F 'oIt is processed into single pass characteristic pattern Fo, by FoIn numerical value be transformed between 0-1, obtain local spy
Levy adaptive aggregate weight distribution map;
Step 5. is by characteristic pattern F5It is cut into the grid of r × r of rule, by F5In each grid in element, according to office
The adaptive aggregate weight distribution map F of portion's featureoCarry out aminated polyepichlorohydrin;
Step 6. carries out model training to step 1-5 by loss function.
The image characteristic extracting method, wherein in step 2: W5=WI/ 32, H5=HI/ 32, W3=2W4=4W5, H3=
2H4=4H5。
The image characteristic extracting method, wherein step 3 specifically: use the core size for 4 × 4, step-length is 4 to be averaged
Pondization is operated to F3Pond is carried out, the convolution algorithm that core size is 1 × 1 is then reused and the result of Chi Huahou is handled;
Simultaneously use size be 2 × 2, step-length for 2 average pondization operation to F4Pond is carried out, then reusing core size is 1 × 1
Convolution algorithm handles the result of Chi Huahou, and respectively obtaining size is W '3×H’3×C’3Characteristic pattern F '3It is with size
W’4×H’4×C’4Characteristic pattern F '4, characteristic pattern F '3And F '4Meet W '3=W '4=W5, H '3=H '4=H5, on channel dimension
Concatenate characteristic pattern F '3, F '4, and using core size be 1 × 1 convolution algorithm the result after concatenation is handled to obtain size be
W’o×H’o×C’oOutput characteristic pattern F 'o。
The image characteristic extracting method, wherein step 4 specifically: the convolution algorithm for being 1 × 1 with core size is by F 'o
It is processed into single pass characteristic pattern Fo, sigmoid function is applied to characteristic pattern Fo, by FoIn numerical value be transformed between 0-1,
Obtain the adaptive aggregate weight distribution map of local feature.
The image characteristic extracting method wherein carries out aminated polyepichlorohydrin as the following formula in step 5:
Wherein, r is natural number;Φi,jFor by characteristic pattern F5It is i row, j column corresponding to coordinate after being cut into regular grids
The character representation of grid, f5 m,nThen correspond to characteristic pattern F5In coordinated indexing be m row, n column element vector, Fo(m, n) is office
The adaptive aggregate weight distribution map F of portion's featureoMiddle index is m row, the value of the element of n column, by F5In characteristic pattern, in identical grid
Element polymerize by above formula, using the feature vector after polymerization as the character representation of corresponding grid.
The image characteristic extracting method wherein defines the Ranking Triplet based on local feature in step 6
Loss loss function carries out model training, keeps the Euclidean distance minimum value of the local feature of the image from same historical relic sample small
In the Euclidean distance minimum value of the local feature of the image from different samples,
Loss function is defined as follows:
Wherein, IqFor thumbnail, I+For the image from historical relic sample identical as thumbnail, I-For from index
The image of image difference historical relic sample, Φi,j(I) coordinated indexing of the image I to be cut into regular grids is i row, j column
The feature of grid, Φu,v(I+) and Φu,v(I-) respectively indicate image I+And I-Coordinated indexing be u row, v column grid spy
Sign, g is a scalar parameter, for controlling IqLocal feature and I+And I-Local feature between distance surplus.
A kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval, this method include in feature extraction phases
Following steps:
Step 1. is W to size with the ResNet-50 depth convolutional neural networks for removing full articulamentumI×HI× 3 it is defeated
Enter image I and carries out propagated forward processing, WIFor picture traverse, HI3 channels are indicated for picture altitude, 3;
Third convolution module, four convolution module and fiveth of the step 2. from used depth convolutional neural networks
It is W that the output end of a convolution module extracts size respectively3×H3×C3、W4×H4×C4And W5×H5×C5Characteristic pattern F3、F4With
F5, wherein W, H and C are respectively the width, height and port number of characteristic pattern;
Step 3. is to characteristic pattern F3、F4Pond processing is carried out, respectively obtaining size is W '3×H’3×C’3Characteristic pattern F '3
It is W ' with size4×H’4×C’4Characteristic pattern F '4, concatenate characteristic pattern F '3, F '4, by convolution algorithm to the result after concatenation into
It is W ' that row processing, which obtains size,o×H’o×C’oOutput characteristic pattern F 'o;
Step 4. is by F 'oIt is processed into single pass characteristic pattern Fo, by FoIn numerical value be transformed between 0-1, obtain local spy
Levy adaptive aggregate weight distribution map;
Step 5. is by characteristic pattern F5It is cut into the grid of r × r of rule, by F5In each grid in element, according to office
The adaptive aggregate weight distribution map F of portion's featureoCarry out aminated polyepichlorohydrin.
The image characteristic extracting method, wherein in step 2: W5=WI/ 32, H5=HI/ 32, W3=2W4=4W5, H3=
2H4=4H5。
The image characteristic extracting method, wherein step 3 specifically: use the core size for 4 × 4, step-length is 4 to be averaged
Pondization is operated to F3Pond is carried out, the convolution algorithm that core size is 1 × 1 is then reused and the result of Chi Huahou is handled;
Simultaneously use size be 2 × 2, step-length for 2 average pondization operation to F4Pond is carried out, then reusing core size is 1 × 1
Convolution algorithm handles the result of Chi Huahou, and respectively obtaining size is W '3×H’3×C’3Characteristic pattern F '3It is with size
W’4×H’4×C’4Characteristic pattern F '4, characteristic pattern F '3And F '4Meet W '3=W '4=W5, H '3=H '4=H5, on channel dimension
Concatenate characteristic pattern F '3, F '4, and using core size be 1 × 1 convolution algorithm the result after concatenation is handled to obtain size be
W’o×H’o×C’oOutput characteristic pattern F 'o。
The image characteristic extracting method, wherein step 4 specifically: the convolution algorithm for being 1 × 1 with core size is by F 'o
It is processed into single pass characteristic pattern Fo, sigmoid function is applied to characteristic pattern Fo, by FoIn numerical value be transformed between 0-1,
Obtain the adaptive aggregate weight distribution map of local feature.
The image characteristic extracting method wherein carries out aminated polyepichlorohydrin as the following formula in step 5:
Wherein, r is natural number;Φi,jFor by characteristic pattern F5It is i row, j column corresponding to coordinate after being cut into regular grids
The character representation of grid, f5 m,nThen correspond to characteristic pattern F5In coordinated indexing be m row, n column element vector, Fo(m, n) is office
The adaptive aggregate weight distribution map F of portion's featureoMiddle index is m row, the value of the element of n column, by F5In characteristic pattern, in identical grid
Element polymerize by above formula, using the feature vector after polymerization as the character representation of corresponding grid.
Detailed description of the invention
Fig. 1 is image characteristic extracting method training stage flow chart of the invention;
Fig. 2 is that image characteristic extracting method of the invention actually uses phase flow figure.
Specific embodiment
Invention is further explained with reference to the accompanying drawings and examples.
As shown in Figure 1, the method for the training stage of image characteristic extracting method of the invention includes:
The first step is W to size with the ResNet-50 depth convolutional neural networks for removing full articulamentumI×HI× 3
Input picture I carries out propagated forward processing, WIFor picture traverse, HI3 channels are indicated for picture altitude, 3;Image I is database
In the user of the image prestored or input shooting image.
Second step, from the third convolution module of used depth convolutional neural networks, the 4th convolution module and
It is W that the output end of five convolution modules extracts size respectively3×H3×C3、W4×H4×C4And W5×H5×C5Characteristic pattern F3、F4
And F5.Wherein, W, H and C are respectively the width, height and port number of characteristic pattern, wherein preferred C3=512, C4=1024, C5=
2048, and meet W5=WI/ 32, H5=HI/ 32, W3=2W4=4W5, H3=2H4=4H5。
Third step uses core size for 4 × 4, and the average pondization operation that step-length is 4 is to F3Pond is carried out, is then reused
The convolution algorithm that core size is 1 × 1 handles the result of Chi Huahou;Use the size for 2 × 2 simultaneously, step-length is averaged for 2
Pondization is operated to F4Pond is carried out, the convolution algorithm that core size is 1 × 1 is then reused and the result of Chi Huahou is handled.
Respectively obtaining size is W '3×H’3×C’3Characteristic pattern F '3It is W ' with size4×H’4×C’4Characteristic pattern F '4.Characteristic pattern F '3
And F '4Meet W '3=W '4=W5, H '3=H '4=H5.Characteristic pattern F ' is concatenated on channel dimension3, F '4, and the use of core size is 1
× 1 convolution algorithm is handled to obtain size to be W ' to the result after concatenationo×H’o×C’oOutput characteristic pattern F 'o。
4th step, the convolution algorithm for being 1 × 1 with core size is by F 'oIt is processed into single pass characteristic pattern Fo, by sigmoid letter
Number is applied to characteristic pattern Fo, by FoIn numerical value be transformed between 0-1, obtain the adaptive aggregate weight distribution map of local feature.
5th step, by characteristic pattern F5It is cut into the grid of r × r of rule, then, by F5In each grid in element,
According to the adaptive aggregate weight distribution map F of local featureo, it is polymerize as the following formula:
Wherein, r is natural number, such as can be 4;Φi,jFor by characteristic pattern F5After being cut into regular grids, corresponds to and sit
It is designated as the character representation of i row, the grid that j is arranged.f5 m,nCorresponding to characteristic pattern F5In coordinated indexing be m row, n column element vector.
Fo(m, n) is the adaptive aggregate weight distribution map F of local featureoMiddle index is m row, the value of the element of n column.By F5In characteristic pattern,
Element in identical grid is polymerize by above formula, using the feature vector after polymerization as the character representation of corresponding grid.
6th step defines the Ranking Triplet Loss loss function based on local feature and carries out model training.Make
The Euclidean distance minimum value of the local feature of image from same historical relic sample is less than the part of the image from different samples
The Euclidean distance minimum value of feature.
Loss function is defined as follows:
Wherein, IqFor thumbnail (i.e. trained target image), I+For the figure from historical relic sample identical as thumbnail
Picture, I-For from the image with thumbnail difference historical relic sample.Φi,jIt (I) is the seat for being cut into the image I of regular grids
Mark index is i row, the feature of the grid of j column.Φu,v(I+) and Φu,v(I-) respectively indicate image I+And I-Coordinated indexing be u
Row, the feature of the grid of v column.G is a scalar parameter, for controlling IqLocal feature and I+And I-Local feature spacing
From surplus.
After completing model training, it can extract local feature from image with model in the actual use stage.In reality
The local shape factor step of service stage is identical as the treatment process of above-mentioned 1 to 5th step, as shown in Fig. 2, image of the invention
Feature extracting method actual use the stage feature extracting method include:
The first step is W to size with the ResNet-50 depth convolutional neural networks for removing full articulamentumI×HI× 3
Input picture I carries out propagated forward processing, WIFor picture traverse, HI3 channels are indicated for picture altitude, 3;Image I is database
In the user of the image prestored or input shooting image.
Second step, from the third convolution module of used depth convolutional neural networks, the 4th convolution module and
It is W that the output end of five convolution modules extracts size respectively3×H3×C3、W4×H4×C4And W5×H5×C5Characteristic pattern F3、F4
And F5.Wherein, W, H and C are respectively the width, height and port number of characteristic pattern, and meet W5=WI/ 32, H5=HI/ 32, W3=2W4=
4W5, H3=2H4=4H5。
Third step uses core size for 4 × 4, and the average pondization operation that step-length is 4 is to F3Pond is carried out, is then reused
The convolution algorithm that core size is 1 × 1 handles the result of Chi Huahou;Use the size for 2 × 2 simultaneously, step-length is averaged for 2
Pondization is operated to F4Pond is carried out, the convolution algorithm that core size is 1 × 1 is then reused and the result of Chi Huahou is handled.
Respectively obtaining size is W '3×H’3×C’3Characteristic pattern F '3It is W ' with size4×H’4×C’4Characteristic pattern F '4.Characteristic pattern F '3
And F '4Meet W '3=W '4=W5, H '3=H '4=H5.Characteristic pattern F ' is concatenated on channel dimension3, F '4, and the use of core size is 1
× 1 convolution algorithm is handled to obtain size to be W ' to the result after concatenationo×H’o×C’oOutput characteristic pattern F 'o。
4th step, the convolution algorithm for being 1 × 1 with core size is by F 'oIt is processed into single pass characteristic pattern Fo, by sigmoid letter
Number is applied to characteristic pattern Fo, by FoIn numerical value be transformed between 0-1, obtain the adaptive aggregate weight distribution map of local feature.
5th step, by characteristic pattern F5It is cut into the grid of r × r of rule, then, by F5In each grid in element,
According to the adaptive aggregate weight distribution map F of local featureo, it is polymerize as the following formula:
Wherein, r is natural number, such as can be 4;Φi,jFor by characteristic pattern F5After being cut into regular grids, corresponds to and sit
It is designated as the character representation of i row, the grid that j is arranged.f5 m,nThen correspond to characteristic pattern F5In coordinated indexing be m row, n column element to
Amount.Fo(m, n) is the adaptive aggregate weight distribution map F of local featureoMiddle index is m row, the value of the element of n column.By F5Characteristic pattern
In, the element in identical grid is polymerize by above formula, using the feature vector after polymerization as the character representation of corresponding grid.
Using cultural relic images feature extracting method proposed by the present invention, text can be effectively indicated with less characteristics of image
The content of object image.Cultural Relics Image Retrieval is carried out using the feature extracted based on the mentioned method of the present invention, historical relic can improved
While image retrieval accuracy, the requirement to memory space and computing resource is reduced, to realize high speed, high-precision text
Object image retrieval.
Claims (4)
1. a kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval, which is characterized in that this method is in training rank
Section the following steps are included:
Step 1. is W to size with the ResNet-50 depth convolutional neural networks for removing full articulamentumI×HI× 3 input figure
As I carries out propagated forward processing, WIFor picture traverse, HI3 channels are indicated for picture altitude, 3;
Step 2. is rolled up from the third convolution module of used depth convolutional neural networks, the 4th convolution module and the 5th
It is W that the output end of volume module extracts size respectively3×H3×C3、W4×H4×C4And W5×H5×C5Characteristic pattern F3、F4And F5,
In, W, H and C are respectively the width, height and port number of characteristic pattern;
Step 3. is to characteristic pattern F3、F4Pond processing is carried out, respectively obtaining size is W '3×H’3×C’3Characteristic pattern F '3With it is big
Small is W '4×H’4×C’4Characteristic pattern F '4, concatenate characteristic pattern F '3, F '4, by convolution algorithm to the result after concatenation at
It is W ' that reason, which obtains size,o×H’o×C’oOutput characteristic pattern F 'o;
Step 4. is by F 'oIt is processed into single pass characteristic pattern Fo, by FoIn numerical value be transformed between 0-1, obtain local feature from
Adapt to aggregate weight distribution map;
Step 5. is by characteristic pattern F5It is cut into the grid of r × r of rule, by F5In each grid in element, according to local spy
Levy adaptive aggregate weight distribution map FoCarry out aminated polyepichlorohydrin;
Step 6. carries out model training to step 1-5 by loss function.
2. image characteristic extracting method according to claim 1, it is characterised in that in step 2: W5=WI/ 32, H5=HI/
32, W3=2W4=4W5, H3=2H4=4H5。
3. a kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval, which is characterized in that this method is mentioned in feature
Take the stage the following steps are included:
Step 1. is W to size with the ResNet-50 depth convolutional neural networks for removing full articulamentumI×HI× 3 input figure
As I carries out propagated forward processing, WIFor picture traverse, HI3 channels are indicated for picture altitude, 3;
Step 2. is rolled up from the third convolution module of used depth convolutional neural networks, the 4th convolution module and the 5th
It is W that the output end of volume module extracts size respectively3×H3×C3、W4×H4×C4And W5×H5×C5Characteristic pattern F3、F4And F5,
In, W, H and C are respectively the width, height and port number of characteristic pattern;
Step 3. is to characteristic pattern F3、F4Pond processing is carried out, respectively obtaining size is W '3×H’3×C’3Characteristic pattern F '3With it is big
Small is W '4×H’4×C’4Characteristic pattern F '4, concatenate characteristic pattern F '3, F '4, by convolution algorithm to the result after concatenation at
It is W ' that reason, which obtains size,o×H’o×C’oOutput characteristic pattern F 'o;
Step 4. is by F 'oIt is processed into single pass characteristic pattern Fo, by FoIn numerical value be transformed between 0-1, obtain local feature from
Adapt to aggregate weight distribution map;
Step 5. is by characteristic pattern F5It is cut into the grid of r × r of rule, by F5In each grid in element, according to local spy
Levy adaptive aggregate weight distribution map FoCarry out aminated polyepichlorohydrin.
4. image characteristic extracting method according to claim 3, it is characterised in that in step 2: W5=WI/ 32, H5=HI/ 32,
W3=2W4=4W5, H3=2H4=4H5。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910401697.6A CN110135440A (en) | 2019-05-15 | 2019-05-15 | A kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910401697.6A CN110135440A (en) | 2019-05-15 | 2019-05-15 | A kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110135440A true CN110135440A (en) | 2019-08-16 |
Family
ID=67573971
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910401697.6A Pending CN110135440A (en) | 2019-05-15 | 2019-05-15 | A kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110135440A (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107577758A (en) * | 2017-08-31 | 2018-01-12 | 桂林电子科技大学 | A kind of generation method for the image convolution feature for intersecting weights based on multizone |
US10102444B2 (en) * | 2016-11-22 | 2018-10-16 | Lunit Inc. | Object recognition method and apparatus based on weakly supervised learning |
CN108829711A (en) * | 2018-05-04 | 2018-11-16 | 上海得见计算机科技有限公司 | A kind of image search method based on multi-feature fusion |
CN109344840A (en) * | 2018-08-07 | 2019-02-15 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment, storage medium, program product |
-
2019
- 2019-05-15 CN CN201910401697.6A patent/CN110135440A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10102444B2 (en) * | 2016-11-22 | 2018-10-16 | Lunit Inc. | Object recognition method and apparatus based on weakly supervised learning |
CN107577758A (en) * | 2017-08-31 | 2018-01-12 | 桂林电子科技大学 | A kind of generation method for the image convolution feature for intersecting weights based on multizone |
CN108829711A (en) * | 2018-05-04 | 2018-11-16 | 上海得见计算机科技有限公司 | A kind of image search method based on multi-feature fusion |
CN109344840A (en) * | 2018-08-07 | 2019-02-15 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment, storage medium, program product |
Non-Patent Citations (5)
Title |
---|
OMAR SEDDATI ET AL: "Towards Good Practices for Image Retrieval Based on CNN Features", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION WORKSHOPS (ICCVW)》 * |
YUNCHAO GONG ET AL: "Multi-Scale Orderless Pooling of Deep Convolutional Activation Features", 《ARXIV:1403.1840V3》 * |
张京: "智能监控中行人序列检索关键技术研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
时璇等: "联合加权聚合深度卷积特征的图像检索方法", 《西安交通大学学报》 * |
董荣胜: "用于图像检索的多区域交叉加权聚合深度卷积特征", 《计算机辅助设计与图形学学报》 * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103646247B (en) | A kind of music score recognition method | |
CN107464210A (en) | A kind of image Style Transfer method based on production confrontation network | |
CN108197606A (en) | The recognition methods of abnormal cell in a kind of pathological section based on multiple dimensioned expansion convolution | |
CN110929602A (en) | Foundation cloud picture cloud shape identification method based on convolutional neural network | |
CN108920720A (en) | The large-scale image search method accelerated based on depth Hash and GPU | |
CN105678293A (en) | Complex image and text sequence identification method based on CNN-RNN | |
CN105243154B (en) | Remote sensing image retrieval method based on notable point feature and sparse own coding and system | |
CN104137119B (en) | Image processing apparatus and image processing method | |
Termritthikun et al. | NU-InNet: Thai food image recognition using convolutional neural networks on smartphone | |
CN111210432B (en) | Image semantic segmentation method based on multi-scale multi-level attention mechanism | |
CN103377237B (en) | The neighbor search method of high dimensional data and fast approximate image searching method | |
CN109871461A (en) | The large-scale image sub-block search method to be reordered based on depth Hash network and sub-block | |
CN106649782A (en) | Picture retrieval method and system | |
CN110909874A (en) | Convolution operation optimization method and device of neural network model | |
CN109636764A (en) | A kind of image style transfer method based on deep learning and conspicuousness detection | |
CN110674326A (en) | Neural network structure retrieval method based on polynomial distribution learning | |
CN110377659A (en) | A kind of intelligence chart recommender system and method | |
CN114330516A (en) | Small sample logo image classification based on multi-graph guided neural network model | |
CN107180079A (en) | The image search method of index is combined with Hash based on convolutional neural networks and tree | |
CN109784360A (en) | A kind of image clustering method based on depth multi-angle of view subspace integrated study | |
CN104881668B (en) | A kind of image fingerprint extracting method and system based on representative local mode | |
CN107506362A (en) | Image classification based on customer group optimization imitates brain storage method | |
CN111125396A (en) | Image retrieval method of single-model multi-branch structure | |
CN110135440A (en) | A kind of image characteristic extracting method suitable for magnanimity Cultural Relics Image Retrieval | |
CN103793714B (en) | Many Classification and Identification device generating means and its method, data identification means and its method |
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 |
Application publication date: 20190816 |
|
WD01 | Invention patent application deemed withdrawn after publication |