CN105701501B - A kind of trademark image recognition methods - Google Patents
A kind of trademark image recognition methods Download PDFInfo
- Publication number
- CN105701501B CN105701501B CN201610004214.5A CN201610004214A CN105701501B CN 105701501 B CN105701501 B CN 105701501B CN 201610004214 A CN201610004214 A CN 201610004214A CN 105701501 B CN105701501 B CN 105701501B
- Authority
- CN
- China
- Prior art keywords
- point
- trade mark
- key point
- feature
- matching
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of trademark image recognition methods, comprising the following steps: the training sample for preparing trade mark to be identified extracts its local feature and clustered and quantified;Feature selecting is carried out using mutual information;Local feature identical with training sample is extracted to test image, and the cluster centre obtained before quantifies local feature;The feature remained using feature selecting is filtered the key point of test picture, then carries out key point matching to test picture and positive sample;Matching is constrained using spatial topotaxy, removes the key point pair of erroneous matching, similarity of number of the statistics with point centering difference visual word as test image and positive sample;Determine whether test image includes the trade mark according to similarity size.The present invention is accounted for from filtering uncorrelated features point and in terms of eliminating erroneous matching two, is a kind of relationship of complementation between them, can be mutually promoted, to preferably improve brand recognition accuracy rate.
Description
Technical field
The invention belongs to sensation target detections and identification technology field, and in particular to a kind of trademark image recognition methods.
Background technique
In recent years, with the rapid development of Internet technology and universal, the especially continuous promotion and application of social network sites,
Image on network presents a rapidly rising trend, this makes the network media become most potential advertisement and commercial podium.Brand with
Track is a kind of service occurred in recent years, is commented by analyzing the evaluation of frequent degree and user that brand exposes in the media
Estimate the growth of brand.There are two types of traditional analysis methods, one is relying on manually to be analyzed and counted, due to network media number
Measure huge, it will take a lot of manpower and time for this mode;The second way is the keyword search function by website
Relevant image is retrieved, however keyword might not be closely related with content: containing much noise in one side search result,
Although, often to brand and uncorrelated, on the other hand can not largely be retrieved containing brand-name image containing keyword.Therefore
It designs a kind of pair of picture material to analyze, automatic identification quotient's calibration method contained therein, there is important commercial application value.
In recent years, researcher proposes some brand recognition methods based on " bag of words ": mentioning from image first
SIFT or SURF feature is taken, then these features are clustered, visual word is formed, then is quantified to obtain by primitive character
These visual words, so every image can be expressed as the set of visual word.With directly being compared using original feature, quantify
Characteristic dimension afterwards substantially reduces, and can be used for large-scale brand recognition task.But the method based on " bag of words " is deposited
In two o'clock deficiency: first is that quantizing process also reduces primitive character while reducing primitive character dimension to a certain extent
Expression ability, the feature for causing two scripts different may be considered as after quantization it is identical, referred to as " error hiding ".For
These error hidings are reduced as far as, researchers propose serial of methods.Masterpiece includes S.Romberg et al. 2013
Year considers adjacent multiple passes while proposition in document " Bundle min-hashing for logo recognition "
The method of key point and C.Wan et al. 2013 in document " Tree-based shape descriptor for scalable
Meet certain condition spatial position four key point proposed in logo detection " is combined into a kind of tree structure
Method.These methods can alleviate the problem of error hiding to a certain extent, but their affine transformations and key to image
There are also insufficient for the robustness of point missing.
Second of method based on " bag of words " is disadvantageous in that in image in the presence of largely incoherent with trade mark
Key point, these key points can not only interfere the correct identification of trade mark, while can also seriously affect recognition speed.In order to remove this
A little incoherent key points are those of only to consider to appear in trade mark region key point than relatively straightforward way, still, for
Different types of trade mark, the keypoint quantity that detected is widely different, if will affect recognition effect very little, if too many meeting
Reduce recognition speed;In addition, those key points for appearing in trade mark region there are many, they and trade mark relevant to background
There is no what correlations for itself.So only considering that the method for appearing in the key point in trade mark region can not be gone well
Except the unrelated key point of those trade marks.
In conclusion there are two shortcomings for existing brand recognition technology: first is that without a kind of pair of affine change of image
Change the error hiding removing method that there is preferable robustness with key point missing;Second is that can effectively filter that not in identification process
The method of key point uncorrelated to trade mark a bit.
Summary of the invention
In view of the deficiencies of the prior art, the invention proposes a kind of new trademark image recognition methods, the technology is using first
It is first filtered out using feature selection approach based on mutual information largely with the incoherent characteristic point of trade mark, it is then new using one kind
Topological constraints method eliminate error hiding, by the combination of the two, can achieve the effect that recognition speed is fast, discrimination is high.
A kind of trademark image recognition methods of the invention, includes the following steps:
It is filtered out first using feature selection approach based on mutual information largely with the incoherent characteristic point of trade mark, includes
Following steps:
(1) for every kind of trade mark to be identified, prepare the training sample of certain amount (no less than 5), Certification Mark exists
At least occur in every sample primary;
(2) local feature is extracted to every training sample, such as SIFT (Scale-Invariant Feature
Transform), SURF (Speeded Up Robust Features) feature, and feature is clustered and quantified, by every
Training sample is expressed as vision set of words;
(3) for every kind of trade mark, using the sample comprising the trade mark as positive sample, the sample not comprising the trade mark is as negative
Sample is calculated the mutual information for all visual words that positive sample is included, is ranked up based on mutual information to these visual words, is chosen
The maximum preceding n (default 100) of association relationship is a as the feature for identifying this trade mark.
Further, based on select come feature to test picture carry out brand recognition comprising the steps of:
(4) identical local feature is extracted to every test image, the cluster centre then obtained using step (2) is to this
A little features are quantified, and test image is expressed as to the set of visual word;
(5) key point matches: first with the preceding n visual word feature selected in step (3) to test image
Key point is filtered, and only retaining those includes the key point in this n vision set of words.To test image and quotient to be identified
Every positive sample of target is matched, and initial matching point pair, and the location information of record matching key point pair are obtained.
Further, the method for error hiding is eliminated using the topological relation of characteristic point, the specific steps are as follows:
(6) for each pair of match point obtained in step (5), their the nearest k of distance (defaults in respective image are found
10) a key point respectively adds the symmetric points about central point (match point) to this k key point, then closes against 2k
Key point is ranked up in the direction of the clock, respectively obtains two 2k length, and end to end sequence;
(7) longest common subsequence (LCS) is asked to two sequences obtained in step (6), then finds out the public sub- sequence of longest
Column account for the ratio of all 2k key points, as the matching degree between matching double points, if matching degree is less than threshold alpha (default
0.6), then it is assumed that be error hiding, they are rejected from matching double points;The method of elimination error hiding in the present invention can also be single
It solely uses, for removing the key point pair of erroneous matching;
(8) number of different visual words in the matching double points remained is counted as the phase between test image and sample
Like degree;
(9) maximum similarity between test image and all positive samples of trade mark to be identified is calculated, is tested as judgement
Image whether include the trade mark confidence level, if confidence level be greater than threshold value beta, then it is assumed that test image include this trade mark.
The beneficial effects of the present invention are: it whether can rapidly and accurately be automatically identified from image comprising certain quotient
Mark identifies that the speed of a kind of trade mark is about 20ms/ in ordinary PC, and accuracy of identification can achieve 90% or more, reach
Actual application level.Why the present invention has said effect, and reason is: the present invention is filtered out greatly using feature selecting
Amount with the incoherent characteristic point of target trade mark, reduce their interference to recognition result, at the same reduce subsequent step when
Between complexity;On this basis, using a kind of pair of image affine transformation and key point missing all with the topology of preferable robustness
Constraint can further increase accuracy of identification to eliminate erroneous matching.
Detailed description of the invention
Fig. 1 is techniqueflow chart of the invention.
Fig. 2 is the techniqueflow chart using specific image.
Fig. 3 is feature selecting effect picture.
Fig. 4 is addition symmetric points schematic diagram.
Fig. 5 is topological constraints schematic diagram.
Fig. 6 is to eliminate error hiding effect picture.
Specific embodiment
The present invention is described in further detail with specific example with reference to the accompanying drawing.
The present invention is a kind of trademark image recognition methods, and techniqueflow is as depicted in figs. 1 and 2, specifically includes following step
It is rapid:
(1) prepare positive sample, extract visual signature
A certain number of training samples are prepared to every kind of trade mark to be identified, and a kind of part is extracted to these training samples
Then feature is clustered and is quantified to the feature extracted such as SIFT feature or SURF feature, the view of key point is obtained
Feel that word indicates, as shown in formula one:
Formula one: k={ P (k), S (k), I (k) }
Wherein, k indicates k-th of key point, and P (k) indicates position, and S (k) indicates that scale, I (k) indicate in nearest cluster
The heart.
(2) feature selecting based on mutual information
The trade mark is not included using the sample in all training samples comprising the trade mark as positive sample for every kind of trade mark
Sample as negative sample, calculate the mutual information of all visual words in positive sample, calculation method is as shown in formula two:
Formula two:
Wherein, c indicates that positive sample generic, t indicate some visual word in positive sample;NijExpression includes or does not include
(i=1 expression include, i=0 expression do not include) visual word t and belong to or be not belonging to (j=1 expression belong to, j=0 indicate not
Belong to) picture number of positive sample.For example, N10The picture number of classification c is indicated comprising visual word t but is not belonging to, with this
Analogize.Ni.=Ni0+Ni1, N.j=N0j+N1j, N=N00+N01+N10+N11。
Then we are from big to small ranked up all visual words according to I (t, c), and (default 100, can also be with by n before choosing
It is other values) feature of a visual word as subsequent this trade mark of identification.Because mutual information is to measure feature and Category Relevance
Important indicator, association relationship is bigger, indicate feature and classification correlation it is bigger, classification (identification) in the process to result
Influence it is also bigger, so using feature selecting, can filter out largely with target trade mark incoherent visual word (corresponding key
Point is also filtered therewith), it on the one hand can reduce their interference to recognition result, after on the other hand capable of greatly shortening
The time of continuous matching and verification, and then improve whole recognition speed.The effect of feature selecting is as shown in figure 3, share four in figure
To image, left figure is characterized the image before selection in each pair of image, and right figure is characterized the image after selection.
(3) key point matches
For testing picture, when identifying whether it includes certain trade mark, need to calculate the institute of it and this trade mark
There is the similarity between positive sample, is then based on maximum similarity to judge whether it includes this trade mark, if it is greater than threshold value
β, then it is assumed that test image includes this trade mark.The process for calculating similarity between test picture and positive sample is as follows:
With (1), identical local feature is extracted to test picture first, then utilizes cluster centre pair obtained in (1)
Feature is quantified, and test picture is expressed as vision set of words.Then utilize n feature obtained in (2) to test picture
Key point filtering is carried out, those key points within the scope of this n visual word are only retained, is remained in test picture and positive sample in this way
Under key point number with n in the same order of magnitude.
Then, key point matching is carried out to this two picture, if two key points are quantized to the same cluster centre,
I (k) i.e. in formula one is equal, then it is assumed that they are matched.Initial matching double points are obtained in this way.Due to single key point
Descriptive power it is limited, be easy to produce error hiding, therefore next reduce using our self-designed topological constraints of one kind
Error hiding.
(4) error hiding is reduced using topological constraints
Key point matched for two, they are called central point by we, are found first spatially nearest from central point
K key point, as it is understood that each key point location information P (k) (as shown in formula one), only need to calculate
The Euclidean distance of central point and other all key points, is then from big to small ranked up key point according to Euclidean distance, and
K key point before choosing all is chosen if the number of key point is less than k.
Then, we add the symmetric points about central point to this k key point respectively, as shown in figure 4, wherein
(a), (b) figure be add symmetric points before key point distribution, (c), (d) figure be add symmetric points after key point be distributed.
In order to distinguish the symmetric points of original point and addition, the cluster centre of symmetric points is set as the opposite of original point cluster centre by us
Number, as shown in formula three:
Formula three: k '={ P (k), S (k) ,-I (k) }
Then, using central point as coordinate origin, any one key point is starting point, along clockwise direction to this 2k key
Point is ranked up, and obtains two sequence S1And S2, as shown in figure 5, here it is considered that S1And S2It is end to end.
Then, we are to S1And S2Longest common subsequence is sought, we can solve using Dynamic Programming, give
The length of two arrays A and B, their longest common subsequence can use the dynamic transfer equation solution in formula four:
Formula four:
Wherein, dp [i] [j] indicates the length of the longest common subsequence of A [1 ... i] and B [1 ... j], it is assumed that the length of A is
The length of m, B are n, then dp [m] [n] is exactly our desired answers.Unlike A and B, S1And S2Be it is end to end,
So we need to enumerate the starting point of one of sequence, longest common subsequence is then sought again, is recorded and is acquired in enumeration process
Maximum length sequence, as S1And S2Longest common subsequence, we are expressed as LCS (S1,S2), as shown in Figure 5.
Next we calculate the matching degree of central point, and calculation method is as shown in formula five, i.e. longest common subsequence
The ratio of length and former sequence length:
Formula five:
Wherein, Length of LCS (S1,S2) indicate S1And S2Longest common subsequence length, min { #S1,#S2}
Indicate S1And S2Key point number smaller value, the matching degree put centered on r, range is between 0~1.If two keys
The matching degree of point is less than some threshold alpha (default 0.6), then it is assumed that they are error hidings, they are given up to fall, and final statistics retains
Similarity of the quantity of different visual words as test picture and positive sample in the matching double points to get off.Fig. 6 is to eliminate error hiding
Effect picture, wherein (a), (b) figure indicate original match effect, effect (c), after (d) expression elimination error hiding.
It is following the experimental results showed that, compared with the conventional method, a kind of trademark image recognition methods of the present invention can obtain
Higher recognition accuracy.
FlickrLogos-32 data set is used in this example, the data set is by document " Scalable Logo
Recognition in Real-World Images " (author S.Romberg, L.G.Pueyo, R.Lienhart and R.Van
Zwol is published in ACM international conference on Multimedia Retrieval in 2011) it proposes,
It wherein include 8240 images, totally 32 trademark class, wherein 320 images are used as training set, 3960 images are as verifying
Collection, 3960 images are as test set.We test following four method as Experimental comparison:
Existing method one: the method SLR of data set author, this method are constituted using the adjacent key point met certain condition
Triangle enhances the descriptive power of single key point;
Existing method two: document " TREE-BASED SHAPE DESCRIPTOR FOR SCALABLE LOGO
(author C.Wan, Z.Zhao, X.Guo and A.Cai are published in 2013 years IEEE Visual to DETECTION "
Communications and Image Processing) in method, this method constructed one kind using four key points
Tree structure describes sub- TSD, enhances the descriptive power of single key point, and have affine-invariant features;
Existing method three: document " Correlation-Based Burstiness for Logo Retrieval " (author
J.Revaud, M.Douze and C.Schmid are published in ACM international conference in 2012
Multimedia the method CBB in), this method can learn automatically from training set out those correctly identify that generation is dry to trade mark
The image-region disturbed, reduces their weight, to improve recognition result.
Experiment is evaluated and tested using the accuracy rate (Precision) and recall rate (Recall) index of information retrieval field use
The accuracy of brand recognition, and the two indexs are comprehensively considered using F1 value, F1 value is higher, illustrates that the result of identification is better.
The contrast and experiment of table 1. and existing method
Control methods | Existing method one | Existing method two | Existing method three | The present invention |
Accuracy rate | 0.980 | 0.980 | 0.980 | 0.994 |
Recall rate | 0.610 | 0.680 | 0.730 | 0.868 |
F1 value | 0.752 | 0.803 | 0.837 | 0.927 |
As it can be seen from table 1 the present invention achieves best brand recognition as a result, comparison existing method one, it is using completely
Three key points of the certain constraint condition of foot construct triangle description, however, the triangle that constructs and not having imitative
Invariance is penetrated, the image of true environment can not be perfectly suitable for;Existing method two is compared, the tree structure that it is constructed describes son
Although having affine-invariant features, this tree structure description is more sensitive to key point missing, constitutes the four of description
The missing of any one key point will lead to total failure in a key point;Existing method three and the present invention are completely not
Same two methods, the two can be complementary in actual application, to further increase recognition result.The present invention is sharp first
Both a large amount of incoherent characteristic points are filtered out with feature selecting, then further remove error hiding using topological constraints, pass through
Combination, the recognition result of trademark image can be greatly improved.
Obviously, various changes and modifications can be made to the invention without departing from essence of the invention by those skilled in the art
Mind and range.In this way, if these modifications and changes of the present invention belongs to the range of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to include these modifications and variations.
Claims (6)
1. a kind of trademark image recognition methods, comprising the following steps:
(1) using the training sample of trade mark to be identified, the local feature of training sample is extracted, local feature is clustered and is measured
Change;
(2) feature selecting is carried out to training sample using mutual information, filtered out and the incoherent key point of trade mark;
(3) local feature identical with training sample is extracted to test image, and the cluster centre obtained before utilizing is to part
Feature is quantified;
(4) feature remained using feature selecting to test picture key point be filtered, then to test picture and
Trade mark positive sample carries out key point matching, obtains initial matching point pair;
(5) to each pair of match point, by point centered on matched key point, it is crucial to find nearest from central point in image k
Point, then respectively adds the symmetric points about central point to this k key point, and the visual word number of symmetric points is original point view
Feel the opposite number of word number;
(6) using central point as coordinate origin, all 2k key points is ranked up in the direction of the clock, obtain cyclic sequence;
(7) longest common subsequence is asked to two obtained cyclic sequences, it is long with longest common subsequence length and original series
Matching degree of the ratio of degree as two central points, if matching degree be less than threshold value, then it is assumed that they are error hidings, by its from
It is rejected with centering;
(8) number of different visual words is similar to positive sample as test image in the matching double points that statistics finally remains
Degree;
(9) maximum similarity for calculating all positive samples of test image and trade mark to be identified, determines according to similarity size and surveys
Attempt to seem no comprising the trade mark.
2. the method as described in claim 1, which is characterized in that in the step (1), the training samples number of every kind of trade mark is not
Less than 5, the local feature of extraction includes SIFT, SURF, and obtained cluster centre is used for the quantization of follow-up test sample.
3. the method as described in claim 1, which is characterized in that will include the quotient for every kind of trade mark in the step (2)
For target sample as positive sample, then other samples not comprising the trade mark calculate each vision in positive sample as negative sample
The mutual information of word chooses the maximum preceding n visual word of mutual information as the feature for identifying the trade mark.
4. the method as described in claim 1, which is characterized in that in the step (4), for what is extracted in test image
Key point only retains the key point in the characteristic range for including trade mark to be identified, so that key point number to be matched is significantly
It reduces;Then key point matching is carried out to test image and positive sample, is quantized to and if only if two key points same poly-
When class center, it is believed that two key points are matched.
5. the method as described in claim 1, which is characterized in that in the step (7), ask two using the method for Dynamic Programming
The longest common subsequence of sequence needs to enumerate the starting point of one of sequence, then asks longest public again due to being cyclic sequence
Subsequence altogether records the maximum length sequence acquired in enumeration process, as two cyclic sequence longest common subsequences.
6. the method as described in claim 1, which is characterized in that be not direct for utilizing and remaining in the step (8)
Similarity of the number with point as test image and positive sample, but the different visual word number of these matching centerings is counted,
This have the advantage that two different key points are quantized to the same cluster centre when cluster centre number is smaller
Probability will increase, and help to reduce the interference of cluster centre number of variations bring by counting different visual word numbers.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610004214.5A CN105701501B (en) | 2016-01-04 | 2016-01-04 | A kind of trademark image recognition methods |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610004214.5A CN105701501B (en) | 2016-01-04 | 2016-01-04 | A kind of trademark image recognition methods |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105701501A CN105701501A (en) | 2016-06-22 |
CN105701501B true CN105701501B (en) | 2019-01-18 |
Family
ID=56226048
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610004214.5A Active CN105701501B (en) | 2016-01-04 | 2016-01-04 | A kind of trademark image recognition methods |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105701501B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
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 |
CN110084466A (en) * | 2019-03-12 | 2019-08-02 | 中国电力科学研究院有限公司 | A kind of method and system for extracting the critical transient stability boundary characteristic of bulk power grid |
CN111626233B (en) * | 2020-05-29 | 2021-07-13 | 江苏云从曦和人工智能有限公司 | Key point marking method, system, machine readable medium and equipment |
CN112699949B (en) * | 2021-01-05 | 2023-05-26 | 百威投资(中国)有限公司 | Potential user identification method and device based on social platform data |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103745200A (en) * | 2014-01-02 | 2014-04-23 | 哈尔滨工程大学 | Facial image identification method based on word bag model |
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 |
CN104281572A (en) * | 2013-07-01 | 2015-01-14 | 中国科学院计算技术研究所 | Target matching method and system based on mutual information |
CN104462381A (en) * | 2014-12-11 | 2015-03-25 | 北京中细软移动互联科技有限公司 | Trademark image retrieval method |
CN104615676A (en) * | 2015-01-20 | 2015-05-13 | 同济大学 | Picture searching method based on maximum similarity matching |
-
2016
- 2016-01-04 CN CN201610004214.5A patent/CN105701501B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104281572A (en) * | 2013-07-01 | 2015-01-14 | 中国科学院计算技术研究所 | Target matching method and system based on mutual information |
CN103745200A (en) * | 2014-01-02 | 2014-04-23 | 哈尔滨工程大学 | Facial image identification method based on word bag model |
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 |
CN104462381A (en) * | 2014-12-11 | 2015-03-25 | 北京中细软移动互联科技有限公司 | Trademark image retrieval method |
CN104615676A (en) * | 2015-01-20 | 2015-05-13 | 同济大学 | Picture searching method based on maximum similarity matching |
Non-Patent Citations (2)
Title |
---|
INDIVIDUALIZED MATCHING BASED ON LOGO DENSITY FOR SCALABLE LOGO RECOGNITION;Yuan Zhang, Shuwu Zhang, WeiLiang and Qinzhen Guo;《2014IEEE International Conference on Acoustic, Speech and Signal Processing》;20140714;全文 |
基于角点匹配的图像配准算法研究;吴巾一;《中国优秀硕士学位论文全文数据库》;20111215(第2011年S1期);全文 |
Also Published As
Publication number | Publication date |
---|---|
CN105701501A (en) | 2016-06-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wang et al. | The devil of face recognition is in the noise | |
CN107292330B (en) | Iterative label noise identification algorithm based on double information of supervised learning and semi-supervised learning | |
Hu et al. | Learning spatial-temporal features for video copy detection by the combination of CNN and RNN | |
CN110717534B (en) | Target classification and positioning method based on network supervision | |
CN104834693B (en) | Visual pattern search method and system based on deep search | |
CN105718532B (en) | A kind of across media sort methods based on more depth network structures | |
CN105701501B (en) | A kind of trademark image recognition methods | |
CN105389326B (en) | Image labeling method based on weak matching probability typical relevancy models | |
Zhou et al. | Integrating SIFT and CNN feature matching for partial-duplicate image detection | |
Gwo et al. | Plant identification through images: Using feature extraction of key points on leaf contours1 | |
CN106951551A (en) | The cumulative index image search method of joint GIST features | |
CN105930873A (en) | Self-paced cross-modal matching method based on subspace | |
Li et al. | Dating ancient paintings of Mogao Grottoes using deeply learnt visual codes | |
De Boom et al. | Semantics-driven event clustering in Twitter feeds | |
CN103473308B (en) | High-dimensional multimedia data classifying method based on maximum margin tensor study | |
Wu et al. | Plant leaf identification based on shape and convolutional features | |
Al-asadi et al. | Object based image retrieval using enhanced SURF | |
CN112199957A (en) | Character entity alignment method and system based on attribute and relation information joint embedding | |
Liu et al. | Self-supervised transformer-based pre-training method using latent semantic masking auto-encoder for pest and disease classification | |
Wang et al. | Attention-based deep metric learning for near-duplicate video retrieval | |
CN112182275A (en) | Trademark approximate retrieval system and method based on multi-dimensional feature fusion | |
CN105701227B (en) | A kind of across media method for measuring similarity and search method based on local association figure | |
CN111863135B (en) | False positive structure variation filtering method, storage medium and computing device | |
CN110442736B (en) | Semantic enhancer spatial cross-media retrieval method based on secondary discriminant analysis | |
CN112580351A (en) | Machine-generated text detection method based on self-information loss compensation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |