CN108038122A - A kind of method of trademark image retrieval - Google Patents
A kind of method of trademark image retrieval Download PDFInfo
- Publication number
- CN108038122A CN108038122A CN201711072145.2A CN201711072145A CN108038122A CN 108038122 A CN108038122 A CN 108038122A CN 201711072145 A CN201711072145 A CN 201711072145A CN 108038122 A CN108038122 A CN 108038122A
- Authority
- CN
- China
- Prior art keywords
- trademark
- trademark image
- image
- trade mark
- retrieved
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—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)
- Data Mining & Analysis (AREA)
- Library & Information Science (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of method of trademark image retrieval, and the trademark database of trademark image is included by generation;Trademark image in the trademark database is trained convolutional neural networks model, obtains stable convolutional neural networks model;Trademark image in the trademark database is inputted to the convolutional neural networks model of the stabilization, obtains trade mark feature database;Trademark image to be retrieved is inputted to the convolutional neural networks model of the stabilization, obtains trade mark feature to be retrieved;Candidate's trademark image is obtained in the trade mark feature database according to the trade mark feature to be retrieved;Obtain preceding L trademark images of candidate's trademark image, according to the trade mark feature of the preceding L trademark images, obtained and the matched trademark image of trade mark to be retrieved in the trade mark feature database, L is positive integer, reduced obtaining stable convolutional neural networks model on the basis of retrieval error, retrieved again, so as to improve the accuracy of retrieval.
Description
Technical field
The present invention relates to computer realm, more particularly to a kind of method of trademark image retrieval.
Background technology
In recent years, enterprise is continuously increased, and the number-of-registration of trade mark also doubles year by year.Gradual with trademark image increases
Add, trade-mark administration department even more increases the difficulty of trademark image management at double.Therefore, effectively differentiate that counterfeit trademarks seem outstanding
To be important.The traditional management business calibration method of trade-mark administration department is mainly the trademark image used so that " classification number " is keyword
Search modes, however the similarity between image or it is main manually identify, this mode efficiency is low and accuracy not
It is high.
After solving the problems, such as the model optimization of deep learning with Hinton even depth learning areas experts, deep learning skill
Art has obtained quick development, and new hope is also brought to field of image search.Deep learning can be by combining shallow-layer spy
The mode of sign produces more abstract profound image content features, utilizes the attribute of these profound character representation images
And content, therefore, depth learning technology had huge breakthrough in artificial intelligence field in recent years, and in machine vision and language
Sound identification etc. performance is outstanding.
Traditional content-based image retrieval is mainly retrieved using the visual signature of image shallow-layer.Such as:Shape
Feature, color characteristic, textural characteristics, HU moment characteristics etc..Although it is more to hold in the retrieval of the image of some simple trade marks
Easy finds similar image, but for some abstract, complicated, even more difficult trademark images, these tradition
Characteristics of image retrieval when be difficult to preferable retrieval effectiveness.Because the problem of these methods are maximum, which is still, to be overcome
The problem of " semantic gap " so that characteristics of image that computer obtains and people from the obtained characteristics of image of high-level semantics there is
Difference, so as to cause retrieval accuracy not high.
The content of the invention
The technical problems to be solved by the invention are:There is provided a kind of accuracy high trademark image retrieval method.
In order to solve the above-mentioned technical problem, a kind of technical solution for using of the present invention for:
A kind of method of trademark image retrieval, including step:
S1, generation include the trademark database of trademark image;
S2, the trademark image in the trademark database are trained convolutional neural networks model, are stablized
Convolutional neural networks model;
S3, the convolutional neural networks model by the trademark image input stabilization in the trademark database, obtain business
Mark feature database;
S4, the convolutional neural networks model by the trademark image input stabilization to be retrieved, it is special to obtain trade mark to be retrieved
Sign;
S5, according to the trade mark feature to be retrieved obtain candidate's trademark image in the trade mark feature database;
S6, preceding L trademark images for obtaining candidate's trademark image, it is special according to the trade mark of the preceding L trademark images
Sign, acquisition and the matched trademark image of trade mark to be retrieved in the trade mark feature database, L is positive integer.
The beneficial effects of the present invention are:The trademark database of trademark image is included by generation;According to the trade mark number
Convolutional neural networks model is trained according to the trademark image in storehouse, obtains stable convolutional neural networks model;By described in
Trademark image in trademark database inputs the convolutional neural networks model of the stabilization, obtains trade mark feature database;Will be to be retrieved
Trademark image input the convolutional neural networks model of the stabilization, obtain trade mark feature to be retrieved;According to the business to be retrieved
Mark feature obtains candidate's trademark image in the trade mark feature database;Preceding L trademark images of candidate's trademark image are obtained,
According to the trade mark feature of the preceding L trademark images, obtained in the trade mark feature database matched with the trade mark to be retrieved
Trademark image, L are positive integer, are reduced obtaining stable convolutional neural networks model on the basis of retrieval error, according to treating
L candidates trademark image reduces range of search before retrieval trademark image feature obtains, and is retrieved again, so as to improve the standard of retrieval
True property.
Brief description of the drawings
Fig. 1 is the method flow diagram of the trademark image retrieval of the embodiment of the present invention.
Embodiment
For the technology contents that the present invention will be described in detail, the objects and the effects, below in conjunction with embodiment and coordinate attached
Figure is explained.
The design of most critical of the present invention is:Using stable convolutional neural networks model extraction trademark image feature, drop
On the basis of low retrieval error, preceding L trademark images of candidate's trademark image are obtained, according to the business of the preceding L trademark images
Mark feature is retrieved again, and the trademark image with the trade mark matching degree higher to be retrieved is obtained in the trade mark feature database
Picture.
It refer to Fig. 1, a kind of method of trademark image retrieval, including step:
S1, generation include the trademark database of trademark image;
S2, the trademark image in the trademark database are trained convolutional neural networks model, are stablized
Convolutional neural networks model;
S3, the convolutional neural networks model by the trademark image input stabilization in the trademark database, obtain business
Mark feature database;
S4, the convolutional neural networks model by the trademark image input stabilization to be retrieved, it is special to obtain trade mark to be retrieved
Sign;
S5, according to the trade mark feature to be retrieved obtain candidate's trademark image in the trade mark feature database;
S6, preceding L trademark images for obtaining candidate's trademark image, it is special according to the trade mark of the preceding L trademark images
Sign, acquisition and the matched trademark image of trade mark to be retrieved in the trade mark feature database, L is positive integer.
As can be seen from the above description, the beneficial effects of the present invention are:The trademark database of trademark image is included by generation;
Trademark image in the trademark database is trained convolutional neural networks model, obtains stable convolutional Neural net
Network model;Trademark image in the trademark database is inputted to the convolutional neural networks model of the stabilization, obtains trade mark spy
Levy storehouse;Trademark image to be retrieved is inputted to the convolutional neural networks model of the stabilization, obtains trade mark feature to be retrieved;According to
The trade mark feature to be retrieved obtains candidate's trademark image in the trade mark feature database;Before obtaining candidate's trademark image
L trademark images, according to the trade mark feature of the preceding L trademark images, obtained in the trade mark feature database with it is described to be checked
The matched trademark image of rope trade mark, L are positive integer, and the basis of retrieval error is reduced obtaining stable convolutional neural networks model
On, L candidates trademark image reduces range of search before being obtained according to trademark image feature to be retrieved, is retrieved again, so that
Improve the accuracy of retrieval.
Further, step S1 is specifically included:
Trademark image is classified, generation includes the trademark database of different classes of trademark image;
Step S3 is specifically included:
S31, the convolutional Neural by the different classes of trademark image input stabilization included in the trademark database
Network model, obtains the feature vector of trademark image;
S32, the feature vector according to the trademark image, structure include the trade mark feature of different classes of trade mark feature
Storehouse.
As can be seen from the above description, trademark image is classified, generation includes the trade mark number of different classes of trademark image
According to storehouse, so that the trade mark feature database for including different classes of trade mark feature is built, in the trade mark corresponding to trademark image to be retrieved
Retrieved in feature database, reduce range of search, improve recall precision.
Further, step S2 is specifically included:
S21, by the trademark database trademark image input Alex convolutional neural networks models, obtain the Alex
The penalty values of convolutional neural networks model;
Whether S22, the penalty values according to the Alex convolutional neural networks model, judge the Alex convolutional networks model
Stablize, if unstable, return to step S21 is until the Alex convolutional networks model stability.
Further, in the step S22 according to the penalty values of the Alex convolutional neural networks model, described in judgement
Whether Alex convolutional networks model, which is stablized, specifically includes:
Judge whether the penalty values are less than the first preset value, if so, then judging the Alex convolutional neural networks model
Stablize, otherwise, judge that the Alex convolutional neural networks model is unstable.
As can be seen from the above description, by the way that the trademark image in the trademark database is inputted Alex convolutional neural networks moulds
Type, obtains the penalty values of the Alex convolutional neural networks model, according to the penalty values of the Alex convolutional neural networks model,
Judge whether the Alex convolutional networks model is stablized, if unstable, continue to circulate, it is ensured that obtained Alex convolution
Network model is reliable and stable, ensure that error is retrieved during later retrieval trademark image to be fluctuated in less scope, accurately
Property higher.
Further, step S4 further includes the convolutional neural networks mould that trademark image to be retrieved is inputted to the stabilization
Type, according to the corresponding output image of the Alex convolutional neural networks model probability on different classes of trademark image point
Cloth, obtains the classification belonging to the trademark image to be retrieved;
The step S5 is specifically included:
Classification according to belonging to the trade mark feature to be retrieved and the trademark image to be retrieved is special in generic trade mark
Levy and candidate's trademark image is obtained in storehouse.
Further, candidate's trademark image is obtained in the step S5 in generic trade mark feature database to specifically include:
Calculate the feature vector of the trademark image to be retrieved and the feature of trademark image in generic trade mark feature database
The distance between vector, according to the distance, obtains the similarity of trademark image to be retrieved and the trademark image in trade mark feature;
Obtain similarity and be more than the trademark image of the second preset value as candidate's trademark image.
As can be seen from the above description, trademark image to be retrieved is first inputted to the convolutional neural networks model of the stabilization, is obtained
To the classification belonging to the trademark image to be retrieved, range of search is reduced, then by calculating the spy of the trademark image to be retrieved
The distance between feature vector of trademark image in the generic trade mark feature database of vector sum is levied, according to the distance, is treated
The similarity of trademark image and trademark image in trade mark feature database is retrieved, obtains the trademark image that similarity is more than the second preset value
As candidate's trademark image, range of search is further reduced by the second preset value, ensure that energy when follow-up progress is retrieved again
Obtain the trademark image of similarity higher.
Further, step S6 specifically includes step:
S61, be ranked up the L candidate's trademark images according to the similarity is descending, m candidates before obtaining
Trademark image and rear n candidate's trademark images, wherein, m+n≤L;
S62, the feature vector to rear n candidate's trademark images are summed and calculate average, are obtained vectorial y, are made
The feature vector of the trademark image to be retrieved is x;
S63, calculate the feature vector of each candidate's trademark image and vector in the preceding m candidate's trademark images respectively
The distance between x and vector y, are denoted as dx and dy, judge dx the and dy sizes, obtain positive sample trademark image respectively;
S63, the feature vector to the positive sample trademark image are summed and calculate average, obtain vectorial z;
S64, according to the vector z, obtained and the matched business of trademark image to be retrieved in the trade mark feature database
Logo image.
Further, judge dx the and dy sizes in the step S63, obtain positive sample trademark image and specifically include:
If dx is less than dy, candidate's trademark image is positive sample trademark image, and otherwise, candidate's trademark image is
Negative sample trademark image.
As can be seen from the above description, on the basis of reducing candidate image scope by the second preset value, L candidate's trade marks are obtained
Image, by being ranked up to the L candidate's trademark images according to the similarity is descending, m candidate quotients before obtaining
Logo image and rear n candidate's trademark images;Retrieved again, accuracy higher.
Embodiment one
A kind of method of trademark image retrieval, including step:
S1, generation include the trademark database of trademark image;
Step S1 is specifically included:
Trademark image is classified, generation includes the trademark database of different classes of trademark image;
Classify with specific reference to the shape of trademark image, profile and side number feature, can be by the trademark database point
It is respectively simple type, round, polygonal, character type, combined and complexity for six classes;
Wherein, the trademark image of the quantity on side or the quantity of angle point less than or equal to 4 is simple type;
Contour shape is that circular or ellipse trademark image is round;
The trademark image of the quantity on side or the quantity of angle point more than 4 is polygonal;
Trademark image is only word or letter composition, then is character type;
Trademark image is collectively constituted by word and pattern, and word and pattern do not merge, then to be combined;
Trademark image is that complex pattern or pattern and word mutually merge, then is complexity;
S2, the trademark image in the trademark database are trained convolutional neural networks model, are stablized
Convolutional neural networks model;
Step S2 is specifically included:
S21, by the trademark database trademark image input Alex convolutional neural networks models, obtain the Alex
The penalty values of convolutional neural networks model;
Whether S22, the penalty values according to the Alex convolutional neural networks model, judge the Alex convolutional networks model
Stablize, if unstable, return to step S21 is until the Alex convolutional networks model stability;
According to the penalty values of the Alex convolutional neural networks model in the step S22, the Alex convolution net is judged
Whether network model, which is stablized, specifically includes:
Judge whether the penalty values are less than the first preset value, first preset value is preferably 0.1, if so, then judging
The Alex convolutional neural networks model stability, otherwise, judges that the Alex convolutional neural networks model is unstable;
Alex convolutional neural networks models are chosen, and using Caffe deep learnings frame to Alex convolutional neural networks moulds
Type is trained, using six kinds of different classes of trademark images in the trademark database as Alex convolutional neural networks models
Input, obtains the penalty values (loss) of Alex convolutional neural networks models, and after continuous repetitive exercise, the penalty values will be by
Gradually reduce, when the penalty values are less than 0.1, then the Alex convolutional neural networks models judged are reliable and stable moulds
Type;
S3, the convolutional neural networks model by the trademark image input stabilization in the trademark database, obtain business
Mark feature database;
Step S3 is specifically included:
S31, the convolutional Neural by the different classes of trademark image input stabilization included in the trademark database
Network model, obtains the feature vector of trademark image;
S32, the feature vector according to the trademark image, structure include the trade mark feature of different classes of trade mark feature
Storehouse;
S4, the convolutional neural networks model by the trademark image input stabilization to be retrieved, it is special to obtain trade mark to be retrieved
Sign;
Step S4 further includes the convolutional neural networks model that trademark image to be retrieved is inputted to the stabilization, according to described
Probability distribution of the corresponding output image of Alex convolutional neural networks models on different classes of trademark image, obtains described treat
Retrieve the classification belonging to trademark image;
The feature extraction and the classification of trademark image to be retrieved that feature extraction to trademark image includes trademark database carry
Take two parts, in the characteristic extraction part of trademark database, by trademark image different classes of in the trademark database according to
Secondary input extracts network layer 7 output data as every trade mark into the Alex convolutional neural networks models of the stabilization
The feature vector of image, according to the feature vector of the trademark image, trade mark of the structure comprising different classes of trade mark feature is special
Levy storehouse;Part is extracted in the classification of trademark image to be retrieved, the trademark image to be retrieved is inputted to the Alex of the stabilization
In convolutional neural networks model, when input picture reaches Alex network output layers, the softmax graders in output layer obtain
Probability distribution of the trademark image to be retrieved on different classes of trademark image, and extract network layer 7 output data work
For the feature vector of the trademark image to be retrieved;
Wherein, the feature extraction in trademark image storehouse is carried out in offline part, i.e., just without again after need to only extracting once
Secondary extraction;And the classification extraction of trademark image to be retrieved is carried out in online part, when inputting trademark image to be retrieved every time
It is required for extraction once;By using the model of " offline+online ", by under the feature preservation of trademark image in trademark database
Come, will build trade mark feature database so as to avoid each retrieval trademark image, recall precision can be greatly promoted;
S5, according to the trade mark feature to be retrieved obtain candidate's trademark image in the trade mark feature database;
The step S5 is specifically included:
Classification according to belonging to the trade mark feature to be retrieved and the trademark image to be retrieved is special in generic trade mark
Levy and candidate's trademark image is obtained in storehouse;
Candidate's trademark image is obtained in the step S5 in generic trade mark feature database to specifically include:
Calculate the feature vector of the trademark image to be retrieved and the feature of trademark image in generic trade mark feature database
The distance between vector, according to the distance, obtains the similarity of trademark image to be retrieved and the trademark image in trade mark feature;
Obtain similarity and be more than the trademark image of the second preset value as candidate's trademark image;
The present invention calculates the distance between the feature vector of two trademark images size, wherein the feature of a trademark image
Vector is (x1,x2,...,xm), xi, i=1,2 ..., m, the feature vector of another trademark image is (y1,y2,...,ym),
yi, i=1,2 ..., m, using Euclidean distance formula, specific formula is as follows:
Then the similarity of two trade marks isD represents the distance between feature vector of two trademark images;
S6, preceding L trademark images for obtaining candidate's trademark image, it is special according to the trade mark of the preceding L trademark images
Sign, acquisition and the matched trademark image of trade mark to be retrieved in the trade mark feature database, L is positive integer;
Step S6 specifically includes step:
S61, be ranked up the L candidate's trademark images according to the similarity is descending, m candidates before obtaining
Trademark image and rear n candidate's trademark images, the value of L are more than or equal to m and add the sum of n, and the specific values of m are 10, n tools in the present embodiment
Body value is 50;
S62, the feature vector to rear 50 candidate's trademark images are summed and calculate average, are obtained vectorial y, are made
The feature vector of the trademark image to be retrieved is x;
S63, calculate respectively in preceding 10 candidate's trademark images the feature vector of each candidate's trademark image with to
The distance between x and vector y are measured, dx and dy is denoted as respectively, judges dx the and dy sizes, obtain positive sample trademark image;
S63, the feature vector to the positive sample trademark image are summed and calculate average, obtain vectorial z;
S64, according to the vector z, obtained and the matched business of trademark image to be retrieved in the trade mark feature database
Logo image;
Dx the and dy sizes are judged in the step S63, positive sample trademark image is obtained and specifically includes:
If dx is less than dy, candidate's trademark image is positive sample trademark image, and otherwise, candidate's trademark image is
Negative sample trademark image.
In conclusion a kind of method of trademark image retrieval provided by the invention, trademark image is classified, generation bag
Trademark database containing different classes of trademark image, so that the trade mark feature database for including different classes of trade mark feature is built,
Retrieved in the trade mark feature database corresponding to trademark image to be retrieved, reduce range of search, improve recall precision;It is logical
Cross and the trademark image in the trademark database is inputted into Alex convolutional neural networks models, obtain the Alex convolutional Neurals net
The penalty values of network model, according to the penalty values of the Alex convolutional neural networks model, judge the Alex convolutional networks model
Whether stablize, if unstable, continue to circulate, it is ensured that obtained Alex convolutional network models are reliable and stable, are protected
Retrieve error when having demonstrate,proved later retrieval trademark image to fluctuate in less scope, accuracy higher;First by trade mark to be retrieved
Image inputs the convolutional neural networks model of the stabilization, obtains the classification belonging to the trademark image to be retrieved, reduces retrieval
Scope, then by calculating the feature vector of the trademark image to be retrieved and the spy of trademark image in generic trade mark feature database
Levy the distance between vector, according to the distance, obtain trademark image to be retrieved in trade mark feature database trademark image it is similar
Degree, obtains trademark image of the similarity more than the second preset value as candidate's trademark image, is further contracted by the second preset value
Small range of search, ensure that the follow-up trademark image for carrying out that similarity higher can be obtained when retrieving again;The spy in trademark image storehouse
Sign extraction is carried out in offline part, i.e., just without extracting again after need to only extracting once;And point of trademark image to be retrieved
Class extraction is carried out in online part, and extraction is required for when inputting trademark image to be retrieved every time once;By using " offline
+ it is online " model, the feature of trademark image in trademark database is preserved, so as to avoid each retrieval trademark image
Trade mark feature database will be built, recall precision can be greatly promoted;The basis of candidate image scope is reduced by the second preset value
On, L candidate's trademark images are obtained, by being arranged according to the similarity is descending the L candidate's trademark images
Sequence, obtains preceding m candidate's trademark images and rear n candidate's trademark images;Retrieved again, accuracy higher.
The foregoing is merely the embodiment of the present invention, is not intended to limit the scope of the invention, every to utilize this hair
The equivalents that bright specification and accompanying drawing content are made, are directly or indirectly used in relevant technical field, similarly include
In the scope of patent protection of the present invention.
Claims (8)
- A kind of 1. method of trademark image retrieval, it is characterised in that including step:S1, generation include the trademark database of trademark image;S2, the trademark image in the trademark database are trained convolutional neural networks model, obtain the volume of stabilization Product neural network model;S3, the convolutional neural networks model by the trademark image input stabilization in the trademark database, obtain trade mark spy Levy storehouse;S4, the convolutional neural networks model by the trademark image input stabilization to be retrieved, obtain trade mark feature to be retrieved;S5, according to the trade mark feature to be retrieved obtain candidate's trademark image in the trade mark feature database;S6, preceding L trademark images for obtaining candidate's trademark image, the trade mark feature of trademark image is opened according to the preceding L, Acquisition and the matched trademark image of trade mark to be retrieved in the trade mark feature database, L is positive integer.
- 2. the method for trademark image retrieval according to claim 1, it is characterised in thatStep S1 is specifically included:Trademark image is classified, generation includes the trademark database of different classes of trademark image;Step S3 is specifically included:S31, the convolutional neural networks by the different classes of trademark image input stabilization included in the trademark database Model, obtains the feature vector of trademark image;S32, the feature vector according to the trademark image, structure include the trade mark feature database of different classes of trade mark feature.
- 3. the method for trademark image retrieval according to claim 1, it is characterised in thatStep S2 is specifically included:S21, by the trademark database trademark image input Alex convolutional neural networks models, obtain the Alex convolution The penalty values of neural network model;S22, the penalty values according to the Alex convolutional neural networks model, judge whether the Alex convolutional networks model is steady Fixed, if unstable, return to step S21 is until the Alex convolutional networks model stability.
- 4. the method for trademark image retrieval according to claim 3, it is characterised in thatAccording to the penalty values of the Alex convolutional neural networks model in the step S22, the Alex convolutional networks mould is judged Whether type, which is stablized, specifically includes:Judge whether the penalty values are less than the first preset value, if so, then judge the Alex convolutional neural networks model stability, Otherwise, judge that the Alex convolutional neural networks model is unstable.
- 5. the method for trademark image retrieval according to claim 2, it is characterised in thatStep S4 further includes the convolutional neural networks model that trademark image to be retrieved is inputted to the stabilization, according to the Alex Probability distribution of the corresponding output image of convolutional neural networks model on different classes of trademark image, obtains described to be retrieved Classification belonging to trademark image;The step S5 is specifically included:Classification according to belonging to the trade mark feature to be retrieved and the trademark image to be retrieved is in generic trade mark feature database Middle acquisition candidate's trademark image.
- 6. the method for trademark image retrieval according to claim 5, it is characterised in thatCandidate's trademark image is obtained in the step S5 in generic trade mark feature database to specifically include:Calculate the feature vector of the trademark image to be retrieved and the feature vector of trademark image in generic trade mark feature database The distance between, according to the distance, obtain the similarity of trademark image to be retrieved and the trademark image in trade mark feature;Obtain similarity and be more than the trademark image of the second preset value as candidate's trademark image.
- 7. the method for trademark image retrieval according to claim 6, it is characterised in thatStep S6 specifically includes step:S61, be ranked up the L candidate's trademark images according to the similarity is descending, m candidate's trade marks before obtaining Image and rear n candidate's trademark images, wherein, m+n≤L;S62, the feature vector to n candidate's trademark images after described are summed and calculate average, obtain vectorial y, described in order The feature vector of trademark image to be retrieved is x;S63, the feature vector for calculating in the preceding m candidate's trademark images each candidate's trademark image respectively and vector x and The distance between vectorial y, is denoted as dx and dy respectively, judges dx the and dy sizes, obtains positive sample trademark image;S63, the feature vector to the positive sample trademark image are summed and calculate average, obtain vectorial z;S64, according to the vector z, obtained and the matched trademark image of trademark image to be retrieved in the trade mark feature database Picture.
- 8. the method for trademark image retrieval according to claim 7, it is characterised in thatDx the and dy sizes are judged in the step S63, positive sample trademark image is obtained and specifically includes:If dx is less than dy, candidate's trademark image is positive sample trademark image, and otherwise, candidate's trademark image is negative sample This trademark image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711072145.2A CN108038122B (en) | 2017-11-03 | 2017-11-03 | Trademark image retrieval method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711072145.2A CN108038122B (en) | 2017-11-03 | 2017-11-03 | Trademark image retrieval method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108038122A true CN108038122A (en) | 2018-05-15 |
CN108038122B CN108038122B (en) | 2021-12-14 |
Family
ID=62093588
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711072145.2A Active CN108038122B (en) | 2017-11-03 | 2017-11-03 | Trademark image retrieval method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108038122B (en) |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109189888A (en) * | 2018-09-07 | 2019-01-11 | 平安科技(深圳)有限公司 | The method and storage medium that electronic device, infringement are analyzed |
CN109344313A (en) * | 2018-07-31 | 2019-02-15 | 中山大学 | A kind of Automatic identification method based on trademark image |
CN109376741A (en) * | 2018-09-10 | 2019-02-22 | 平安科技(深圳)有限公司 | Recognition methods, device, computer equipment and the storage medium of trademark infringement |
CN109784384A (en) * | 2018-12-28 | 2019-05-21 | 佛山科学技术学院 | A kind of method and device of the automatic discrimination trade mark true and false |
CN109815358A (en) * | 2018-12-26 | 2019-05-28 | 北京博鳌纵横网络科技有限公司 | A kind of pictorial trademark approximate query system and method |
CN109815394A (en) * | 2018-12-26 | 2019-05-28 | 北京博鳌纵横网络科技有限公司 | A kind of intellectual property mandatory system |
CN109934809A (en) * | 2019-03-08 | 2019-06-25 | 深慧视(深圳)科技有限公司 | A kind of paper labels character defect inspection method |
CN110033018A (en) * | 2019-03-06 | 2019-07-19 | 平安科技(深圳)有限公司 | Shape similarity judgment method, device and computer readable storage medium |
CN110110116A (en) * | 2019-04-02 | 2019-08-09 | 浙江工业大学 | A kind of trademark image retrieval method for integrating depth convolutional network and semantic analysis |
CN110674328A (en) * | 2019-09-27 | 2020-01-10 | 长城计算机软件与***有限公司 | Trademark image retrieval method, system, medium and equipment |
CN110851634A (en) * | 2019-11-18 | 2020-02-28 | 广东三维家信息科技有限公司 | Picture retrieval method and device and electronic equipment |
CN112036421A (en) * | 2019-05-16 | 2020-12-04 | 搜狗(杭州)智能科技有限公司 | Image processing method and device and electronic equipment |
CN112256910A (en) * | 2020-10-23 | 2021-01-22 | 蓝普金睛(北京)科技有限公司 | Content-based shop identification image retrieval method |
CN114428878A (en) * | 2022-04-06 | 2022-05-03 | 广东知得失网络科技有限公司 | Trademark image retrieval method and system |
CN114494730A (en) * | 2022-04-15 | 2022-05-13 | 深圳安牌信息技术有限公司 | Trademark automatic classification processing system based on image recognition |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2004064297A2 (en) * | 2003-01-09 | 2004-07-29 | University Of Rochester | Efficient room-temperature source of polarized single photons |
CN104156413A (en) * | 2014-07-30 | 2014-11-19 | 中国科学院自动化研究所 | Trademark density based personalized trademark matching recognition method |
CN105631037A (en) * | 2015-12-31 | 2016-06-01 | 北京恒冠网络数据处理有限公司 | Image retrieval method |
CN105989043A (en) * | 2015-02-04 | 2016-10-05 | 阿里巴巴集团控股有限公司 | Method and device for automatically acquiring trademark in commodity image and searching trademark |
US20170193097A1 (en) * | 2016-01-03 | 2017-07-06 | Gracenote, Inc. | Model-based media classification service using sensed media noise characteristics |
CN106951911A (en) * | 2017-02-13 | 2017-07-14 | 北京飞搜科技有限公司 | A kind of quick multi-tag picture retrieval system and implementation method |
CN107016677A (en) * | 2017-03-24 | 2017-08-04 | 北京工业大学 | A kind of cloud atlas dividing method based on FCN and CNN |
CN107256246A (en) * | 2017-06-06 | 2017-10-17 | 西安工程大学 | PRINTED FABRIC image search method based on convolutional neural networks |
-
2017
- 2017-11-03 CN CN201711072145.2A patent/CN108038122B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2004064297A2 (en) * | 2003-01-09 | 2004-07-29 | University Of Rochester | Efficient room-temperature source of polarized single photons |
CN104156413A (en) * | 2014-07-30 | 2014-11-19 | 中国科学院自动化研究所 | Trademark density based personalized trademark matching recognition method |
CN105989043A (en) * | 2015-02-04 | 2016-10-05 | 阿里巴巴集团控股有限公司 | Method and device for automatically acquiring trademark in commodity image and searching trademark |
CN105631037A (en) * | 2015-12-31 | 2016-06-01 | 北京恒冠网络数据处理有限公司 | Image retrieval method |
US20170193097A1 (en) * | 2016-01-03 | 2017-07-06 | Gracenote, Inc. | Model-based media classification service using sensed media noise characteristics |
CN106951911A (en) * | 2017-02-13 | 2017-07-14 | 北京飞搜科技有限公司 | A kind of quick multi-tag picture retrieval system and implementation method |
CN107016677A (en) * | 2017-03-24 | 2017-08-04 | 北京工业大学 | A kind of cloud atlas dividing method based on FCN and CNN |
CN107256246A (en) * | 2017-06-06 | 2017-10-17 | 西安工程大学 | PRINTED FABRIC image search method based on convolutional neural networks |
Non-Patent Citations (1)
Title |
---|
郑启财: "基于深度学习的图像检索技术的研究", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109344313A (en) * | 2018-07-31 | 2019-02-15 | 中山大学 | A kind of Automatic identification method based on trademark image |
CN109189888A (en) * | 2018-09-07 | 2019-01-11 | 平安科技(深圳)有限公司 | The method and storage medium that electronic device, infringement are analyzed |
CN109189888B (en) * | 2018-09-07 | 2023-03-21 | 平安科技(深圳)有限公司 | Electronic device, infringement analysis method, and storage medium |
WO2020052183A1 (en) * | 2018-09-10 | 2020-03-19 | 平安科技(深圳)有限公司 | Method and apparatus for identifying trademark infringement, and computer device and storage medium |
CN109376741A (en) * | 2018-09-10 | 2019-02-22 | 平安科技(深圳)有限公司 | Recognition methods, device, computer equipment and the storage medium of trademark infringement |
CN109815358A (en) * | 2018-12-26 | 2019-05-28 | 北京博鳌纵横网络科技有限公司 | A kind of pictorial trademark approximate query system and method |
CN109815394A (en) * | 2018-12-26 | 2019-05-28 | 北京博鳌纵横网络科技有限公司 | A kind of intellectual property mandatory system |
CN109784384A (en) * | 2018-12-28 | 2019-05-21 | 佛山科学技术学院 | A kind of method and device of the automatic discrimination trade mark true and false |
CN110033018A (en) * | 2019-03-06 | 2019-07-19 | 平安科技(深圳)有限公司 | Shape similarity judgment method, device and computer readable storage medium |
CN110033018B (en) * | 2019-03-06 | 2023-10-31 | 平安科技(深圳)有限公司 | Graph similarity judging method and device and computer readable storage medium |
CN109934809A (en) * | 2019-03-08 | 2019-06-25 | 深慧视(深圳)科技有限公司 | A kind of paper labels character defect inspection method |
CN110110116A (en) * | 2019-04-02 | 2019-08-09 | 浙江工业大学 | A kind of trademark image retrieval method for integrating depth convolutional network and semantic analysis |
CN110110116B (en) * | 2019-04-02 | 2021-04-06 | 浙江工业大学 | Trademark image retrieval method integrating deep convolutional network and semantic analysis |
CN112036421A (en) * | 2019-05-16 | 2020-12-04 | 搜狗(杭州)智能科技有限公司 | Image processing method and device and electronic equipment |
CN110674328A (en) * | 2019-09-27 | 2020-01-10 | 长城计算机软件与***有限公司 | Trademark image retrieval method, system, medium and equipment |
CN110851634A (en) * | 2019-11-18 | 2020-02-28 | 广东三维家信息科技有限公司 | Picture retrieval method and device and electronic equipment |
CN112256910A (en) * | 2020-10-23 | 2021-01-22 | 蓝普金睛(北京)科技有限公司 | Content-based shop identification image retrieval method |
CN114428878A (en) * | 2022-04-06 | 2022-05-03 | 广东知得失网络科技有限公司 | Trademark image retrieval method and system |
CN114494730A (en) * | 2022-04-15 | 2022-05-13 | 深圳安牌信息技术有限公司 | Trademark automatic classification processing system based on image recognition |
Also Published As
Publication number | Publication date |
---|---|
CN108038122B (en) | 2021-12-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108038122A (en) | A kind of method of trademark image retrieval | |
Klein et al. | Associating neural word embeddings with deep image representations using fisher vectors | |
CN104572940B (en) | A kind of image automatic annotation method based on deep learning and canonical correlation analysis | |
CN107944020A (en) | Facial image lookup method and device, computer installation and storage medium | |
CN106126581A (en) | Cartographical sketching image search method based on degree of depth study | |
CN106951911A (en) | A kind of quick multi-tag picture retrieval system and implementation method | |
CN106649275A (en) | Relation extraction method based on part-of-speech information and convolutional neural network | |
CN108228845B (en) | Mobile phone game classification method | |
CN103258037A (en) | Trademark identification searching method for multiple combined contents | |
Zhang et al. | Semantic classification of high-resolution remote-sensing images based on mid-level features | |
CN105243139A (en) | Deep learning based three-dimensional model retrieval method and retrieval device thereof | |
CN107515877A (en) | The generation method and device of sensitive theme word set | |
CN106055573B (en) | Shoe print image retrieval method and system under multi-instance learning framework | |
CN107463658A (en) | File classification method and device | |
CN107316042A (en) | A kind of pictorial image search method and device | |
CN106991386A (en) | A kind of gesture identification method based on depth residual error network | |
Xu et al. | Mental retrieval of remote sensing images via adversarial sketch-image feature learning | |
CN109582782A (en) | A kind of Text Clustering Method based on Weakly supervised deep learning | |
CN108009286A (en) | A kind of Sketch Searching method based on deep learning | |
CN104778283B (en) | A kind of user's occupational classification method and system based on microblogging | |
CN106445917B (en) | A kind of Chinese entity abstracting method of pattern-based bootstrapping | |
Zhu et al. | Visual relationship detection with object spatial distribution | |
CN104732247B (en) | A kind of human face characteristic positioning method | |
CN104820711A (en) | Video retrieval method for figure target in complex scene | |
CN113032613A (en) | Three-dimensional model retrieval method based on interactive attention convolution neural 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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |