CN101833565A - Method for actively selecting related feedbacks of representative image - Google Patents

Method for actively selecting related feedbacks of representative image Download PDF

Info

Publication number
CN101833565A
CN101833565A CN201010137511.XA CN201010137511A CN101833565A CN 101833565 A CN101833565 A CN 101833565A CN 201010137511 A CN201010137511 A CN 201010137511A CN 101833565 A CN101833565 A CN 101833565A
Authority
CN
China
Prior art keywords
image
width
cloth
row
formula
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
Application number
CN201010137511.XA
Other languages
Chinese (zh)
Other versions
CN101833565B (en
Inventor
周志华
黄圣君
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University
Original Assignee
Nanjing University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Nanjing University filed Critical Nanjing University
Priority to CN201010137511XA priority Critical patent/CN101833565B/en
Publication of CN101833565A publication Critical patent/CN101833565A/en
Application granted granted Critical
Publication of CN101833565B publication Critical patent/CN101833565B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Landscapes

  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a method for actively selecting related feedbacks of a representative image, comprising the following steps of: (1) calculating a score of each image in a database; (2) sequencing in a descending order according to scores, and selecting an image having a high score for a user who judges whether the image is relative or not; (3) recalculating degree of correlation of images in a database after obtaining a feedback of the user; (4) sequencing the images according to the newly generated degree of correlation, and updating an image retrieval result; and (5) returning to step (1) or finishing. The method makes use of active learning thoughts in the machine learning, and gives consideration to the information that a retrieval device has mastered currently and the representation to the whole image database when selecting candidate images of the related feedbacks, so as to solve the problem of obtaining a candidate image randomly or taking partial factors into account only of a traditional related feedback mechanism, and markedly improve the efficiency of related feedbacks.

Description

A kind of related feedback method of initiatively selecting presentation graphics
Technical field
The present invention relates to a kind of indexing unit of digital picture, particularly a kind of relevant feedback mechanism of initiatively selecting presentation graphics.
Background technology
Digital picture is more and more frequent in daily life and practical application is playing the part of important role.Along with on the network and the quick growth of personal user's digital picture, how effectively these digital pictures of management and use become the work of a very meaningful and challenge, and image retrieval wherein gordian technique especially.Yet, because each user's searching preferences is all different, present retrieval technique can't point-devicely satisfy all users' requirement, therefore, a kind of technology that is called relevant feedback is introduced into, allow the user participate in retrieving, undertaken mutual and feedback, make result for retrieval adapt to personalized requirement more by user and indexing unit.Concrete, when indexing unit after the user presents result for retrieval, the user can be judged as some images relevant or uncorrelated, and these information are submitted to indexing unit.Indexing unit utilizes the information of these acquisitions to improve retrieval quality, makes result for retrieval meet user's requirement more.Such process just is called relevant feedback, and can repeat up to the user satisfied.
At present, often provide less candidate image to allow the user judge whether to be correlated with at random in the relevant feedback process, thereby limited the quantity of information of feedback image by indexing unit.The few relevant feedback mechanism based on initiatively study that exists now can be picked out the higher candidate image of quantity of information automatically, but often only pay close attention to the image that those indexing units are least determined, and do not consider the representativeness of image, thereby the usefulness of raising relevant feedback mechanism that can not be very big.
Summary of the invention
Goal of the invention: the problem that obtains or only consider unilateral factor at candidate image in the existing relevant feedback mechanism at random of the present invention, a kind of related feedback method of initiatively selecting presentation graphics is provided, thereby can when select closing the candidate image of feedback, take into account the current information of having grasped of indexing unit simultaneously and, significantly improve the efficient of relevant feedback the representativeness in whole image data storehouse.
Technical scheme: for achieving the above object, a kind of related feedback method of initiatively selecting presentation graphics of the present invention mainly comprises following series of steps: one, the scoring of every image in the computational data storehouse; Two,, select the maximum image of scoring to judge whether to be correlated with for the user according to scoring ordering from big to small; Three, behind the acquisition user feedback result, recomputate the degree of correlation of image in the database; Four, image is sorted the update image result for retrieval according to newly-generated image correlativity; Five, return one or finish.This process can repeat up to the user satisfied to result for retrieval.
Image in the Computer Database is made up of image data base L and query image U two parts, and wherein image data base L partly passes through user feedback, has had the information of whether being correlated with, and query image U partly is the image without user feedback.Set total N among the image data base L 1Width of cloth image, total N among the query image U 2Width of cloth image, related feedback method comprises following steps:
(1) feature vector, X of acquisition all images L and U 1And X 2, and the whether relevant information Y of L 1If the first round, then L is an empty set;
(2) to each width of cloth image calculation scoring among the U;
(3) scoring that step 2 is calculated is by from big to small rank order;
(4) select the maximum N of scoring 3Open image, represent with S, and obtain its feature vector, X 3
(5) whether the result that step 4 is obtained gives the user and judges relevantly, and represents feedback result with Y3;
(6) S is removed from U, be integrated among the L, corresponding X 3Join X 1In, Y 3Join Y 1, and N correspondingly 1Add N 3, N 2Subtract N 3
(7) according to the X after upgrading 1, X 2And Y 1, calculate the similarity of every image, according to new similarity by ordering back output from big to small as result for retrieval;
(8) if the user has been satisfied with to result for retrieval, then retrieving finishes, otherwise enters the relevant feedback of a new round.
The present invention is mainly based on following thought: the amount of images of user feedback is different, perhaps Fan Kui image is different, the obtainable quantity of information of indexing unit is also just different, if completely random provide candidate image to allow the user judge to the user, on the one hand, the not necessarily the most useful image of the image of picked at random may not help indexing unit to improve retrieval quality like this; On the other hand, offering user's candidate image quantity can be very limited.And the active learning art in the machine learning field can solve such problem, the thought that the indexing unit utilization is initiatively learnt can be automatically from database the image that is suitable for feeding back most of picking out of the overall situation allow the user judge.And whether an image is suitable for feedback, whether can help indexing unit to improve the result for retrieval quality, generally can weigh: on the one hand from two aspects, if whether indexing unit is relevant very uncertain to current this image, the user provides after the judgement so, and indexing unit can obtain bigger quantity of information; On the other hand, if this image has good representativeness, the good a certain class image in the representation database, whether relevant indexing unit just can make judgement to the great amount of images in the database after obtaining its information so, so also be of great use.The present invention just is based on above two aspects, weighs image with the scoring of a quantification and whether is suitable for feedback, picks out the uncertain while of those indexing units and has good representational image again.
Beneficial effect: the present invention compared with prior art, its beneficial effect is: the present invention utilizes and initiatively learns thought in the machine learning, when the candidate image of selecting relevant feedback, take into account the current information of having grasped of indexing unit simultaneously and to the representativeness in whole image data storehouse, before having solved in the relevant feedback mechanism candidate image obtain or only consider the problem of unilateral factor at random, thereby significantly improve the efficient of relevant feedback.
Description of drawings
Fig. 1 is a digital image search device workflow diagram.
Fig. 2 is the process flow diagram of mechanism of the present invention.
Fig. 3 is the process flow diagram that calculates the scoring of every width of cloth image.
Fig. 4 is the process flow diagram of computed image similarity.
Embodiment
Below in conjunction with accompanying drawing, most preferred embodiment is elaborated, but protection scope of the present invention is not limited to described embodiment.
Figure 1 shows that digital image search device workflow diagram.Image data base is made up of two parts, and image data base partly is to pass through user feedback, has had the information of whether being correlated with, and supposes total N 1Width of cloth image is represented with L; Query image partly is the image without user feedback, supposes always to have N 2Width of cloth image is represented with U.Device extracts feature to the image in the database, and each width of cloth image is represented by a proper vector.Feature extraction can use the classical way in the Digital Image Processing textbook to generate the characteristics of image that is suitable for, for example features such as color, texture, shape.
(suppose to use X according to the proper vector of image among L and the U at every turn 1And X 2Expression) whether relevant the information of image (suppose to use Y and among the L 1Expression, value are to represent to be correlated with in 1 o'clock, are worth to represent uncorrelated for-1 o'clock), device carries out image retrieval, calculates the similarity of every width of cloth image, and exports as result for retrieval by similarity order from big to small.If the user has been satisfied with to result for retrieval, then retrieving finishes, otherwise device will select some images to offer the user to judge whether to be correlated with, thereby enters the relevant feedback of a new round.
Figure 2 shows that relevant feedback mechanism of the present invention.Step 0 beginning, step 1 obtains the feature vector, X of all images (comprising L and U) 1, X 2, and the information Y that whether is correlated with of L 1If the first round, then L is an empty set.Each width of cloth figure among the step 2 couple U calculates its scoring, and detailed process is seen Fig. 3.The scoring ordering that step 3 is calculated step 2 is by rank order from big to small.Step 4 is selected the maximum N of scoring 3Open image, represent with S, and suppose that its proper vector is X 3, in step 5, give user's judgement then and whether be correlated with, and use Y3 TableShowing feedback result, represent that promptly it is whether relevant, is that 1 expression is relevant with sample value, be worth for-1 expression uncorrelated.In step 6, S is removed from U, be integrated among the L corresponding X 3Join X 1In, Y 3Join Y 1, and N correspondingly 1Add N 3, N 2Subtract N 3In the step 7, according to the X after upgrading 1, X 2And Y 1, calculate the similarity of every image, detailed process is seen Fig. 4, according to new similarity by ordering back output from big to small as result for retrieval.Arrive like this till the step 8, a relevant feedback process of taking turns finishes.
How explanation shown in Figure 3 calculates the scoring of every width of cloth image among the U.Step 20 is for beginning action.In the step 21, based on X 1And X 2Calculate a nuclear matrix, size is (N 1+ N 2) * (N 1+ N 2), represent with K.Here the computing method of nuclear matrix can be with the most frequently used method in machine learning or the pattern-recognition textbook, such as linear kernel.Each width of cloth image is all corresponding certain delegation and a certain row in K.Step 22, the inverse matrix of calculating (K+I), the result represents that with M wherein I is (N 1+ N 2) * (N 1+ N 2) big or small unit matrix.Same, each width of cloth image is corresponding certain delegation and a certain row in Metzler matrix all.The initial i of step 23 is 1, begins the picture count among the U.Step 24 judges that whether i is greater than N 2If, greater than then image among the U traveled through, jump to step 29, finish, otherwise enter next step.In the step 25, suppose that present image i is correlated with, introduce an intermediate variable z here, z=1 represents to be correlated with, and z=-1 represents uncorrelated, calculates scoring according to following formula (1) then, obtains v 1In the step 26, suppose that image i is incoherent, promptly z=-1 calculates scoring according to formula (1) then equally, obtains v 2Code name M in the formula (1) 1To M 6Represent certain part in the Metzler matrix respectively, form by different row and columns respectively, specific as follows: M 1Row and column by image correspondence among the L is formed, and size is N 1* N 1, M 2Form by removing beyond the i width of cloth row and column of image correspondence among the U, size is (N 2-1) * (N 2-1), M 3Row and column by i width of cloth image correspondence is formed, and size is 1 * 1, M 4Row by image correspondence among the row of i width of cloth image correspondence among the U and the L are formed, and size is 1 * N 1, M 5 byRemoving among the row of image correspondence beyond the i width of cloth and the L row of image correspondence among the U forms size and is (N 2-1) * N 1, M 6Row by i width of cloth image correspondence among the row of removing image correspondence beyond the i width of cloth among the U and the U are formed, and size is (N 2-1) * 1.Step 27 is v relatively 1And v 2, get wherein the smaller as the scoring v of image i.Step 28 will be counted i and add 1, forward step 24 again to.Note being used for calculating the formula (1) of scoring, both used the information of the image correspondence among the L and helped judge the current uncertainty of indexing unit image, used simultaneously among the U information of image correspondence again and weighed the representativeness of image other all images, thereby guaranteed that scoring weighed present image from uncertain and representative two aspects simultaneously and whether be suitable for feedback, just guaranteed that also selected image can farthest improve the performance of indexing unit.What deserves to be explained is, the scoring of more why selecting smaller value in the step 27 as image i, consider that mainly whether image i is correlated with is to know, select smaller value to be equivalent to guarantee under the conservative situation no matter image i is relevant or incoherent, can both choose the image that is suitable for feeding back most.
v = - ( zM 3 z + Y 1 T M 1 Y 1 + 2 z M 4 Y 1 - ( M 5 Y 1 + M 6 z ) T M 2 - 1 ( M 5 Y 1 + M 6 z ) ) - - - ( 3 )
The process flow diagram of computed image similarity shown in Figure 4, the similarity of calculating every width of cloth picture based on L can have several different methods, realizes adopting in the example following computing method the typical case.Suppose that at first hypothesis has n among the L 1The image that the width of cloth is relevant has n 2Incoherent image, always total n width of cloth image need calculate similarity, and supposes that two length are n 1Interim array A, B store intermediate value respectively.Step 70 is an origination action.Step 71 initialization i is 1, begins counting.Step 72 judges whether that n width of cloth image all finishes as calculated, finishes if i＞n then jumps to step 75, otherwise enters next step.Beginning two branches of branch from step 73 carries out.The initial j=1 of step 7310, step 7311 judges that whether j is greater than n 1, be then to jump to step 7316, otherwise enter step 7312 that k is initially 1, and step 7313 judges that whether k is greater than n 1, be then to jump to step 7315, otherwise enter step 7314.Step 7314 is according to formula (2) value of calculating e 1Save as k the element of A afterwards, and k is added 1.X in the formula (2) i, x j, and x kRepresent that respectively the i width of cloth will calculate the image of similarity, j opens the proper vector that k opens uncorrelated image among associated picture and the L among the L, and function d is the distance that is used for calculating two vectors, definition is suc as formula (3), a and b are two vectors in the formula (3), and m is the dimension of vectorial a and b, a tThe value of representing the t dimension of vectorial a, e 0Be a predefined very little constant, being used to avoid denominator is 0, realizes adopting 0.0001 in the example the typical case.Step 7315 is selected a maximal value from A, save as j the element of B, and j is added 1.Step 7316 is calculated the mean value of B, and the value that obtains is designated as s 1Another branch is similar from step 7320 to 7326 and step 7310 to 7316, and different is to calculate according to formula (4) in step 7324, and the value that step 7326 obtains is designated as s 2Identical in the implication of the variable in the attention formula (4) and the formula (2).Step 74 is with s 1And s 2Unification is sued for peace after same dimension, obtains s, and the similarity as i opens image adds 1 with i then, turns back to step 72.
e 1 = d ( x i , x k ) × d ( x i , x j ) + d ( x k , x j ) 2 - - - ( 4 )
d ( a , b ) = 1 Σ t = 1 m | a t - b t | 2 + e 0 - - - ( 5 )
e 2 = 2 d ( x i , x k ) d ( x i , x j ) + d ( x k , x j ) - - - ( 6 )
The personage who knows this area will understand, though described specific embodiment for the ease of explaining here, can make various changes under the situation that does not deviate from spirit and scope of the invention.Therefore, except claims, can not be used to limit the present invention.

Claims (3)

1. a related feedback method of initiatively selecting presentation graphics is characterized in that the image in the Computer Database is made up of image data base L and query image U two parts, sets total N among the image data base L 1Width of cloth image, total N among the query image U 2Width of cloth image, related feedback method comprises following steps:
(1) feature vector, X of acquisition all images L and U 1And X 2, and the whether relevant information Y of L 1If the first round, then L is an empty set;
(2) to each width of cloth image calculation scoring among the U;
(3) scoring that step 2 is calculated is by from big to small rank order;
(4) select the maximum N of scoring 3Open image, represent with S, and obtain its feature vector, X 3
(5) whether the result that step 4 is obtained gives the user and judges relevantly, and uses Y 3The expression feedback result;
(6) S is removed from U, be integrated among the L, corresponding X 3Join X 1In, Y 3Join Y 1, and N correspondingly 1Add N 3, N 2Subtract N 3
(7) according to the X after upgrading 1, X 2And Y 1, calculate the similarity of every image, according to new similarity by ordering back output from big to small as result for retrieval;
(8) if the user has been satisfied with to result for retrieval, then retrieving finishes, otherwise enters the relevant feedback of a new round.
2. the related feedback method of initiatively selecting presentation graphics according to claim 1 is characterized in that step (2) specifically comprises following steps:
(21) based on X 1And X 2Calculate a nuclear matrix, size is (N 1+ N 2) * (N 1+ N 2), represent with K;
(22) inverse matrix of calculating (K+I), the result represents that with M wherein I is (N 1+ N 2) * (N 1+ N 2) big or small unit matrix;
(23) picture count parameter i is changed to 1, begins the picture count among the U;
(24) judge that whether i is greater than N 2If, greater than then image among the U traveled through, jump to step 29, otherwise enter next step;
(25), suppose that present image i is correlated with, introduce an intermediate variable z here, z=1 represents to be correlated with, and z=-1 represents uncorrelated, calculates scoring according to formula (1) then, obtains υ 1
(26) suppose that image i is incoherent, promptly z=-1 calculates scoring according to formula (1) then equally, obtains υ 2
(27) compare υ 1And υ 2, get wherein the smaller as the scoring υ of image i;
(28) picture count parameter i is added 1, forward step 24 again to;
(29) finish;
υ = - ( z M 3 z + Y 1 T M 1 Y 1 + 2 z M 4 Y 1 - ( M 5 Y 1 + M 6 z ) T M 2 - 1 ( M 5 Y 1 + M 6 z ) ) - - - ( 1 )
Code name M in the formula (1) 1To M 6Represent certain part in the Metzler matrix respectively, form by different row and columns respectively, specific as follows: M 1Row and column by image correspondence among the L is formed, and size is N 1* N 1, M 2Form by removing beyond the i width of cloth row and column of image correspondence among the U, size is (N 2-1) * (N 2-1), M 3Row and column by i width of cloth image correspondence is formed, and size is 1 * 1, M 4Row by image correspondence among the row of i width of cloth image correspondence among the U and the L are formed, and size is 1 * N 1, M 5Forming size by the row of image correspondence among the row of removing image correspondence beyond the i width of cloth among the U and the L is (N 2-1) * N 1, M 6Row by i width of cloth image correspondence among the row of removing image correspondence beyond the i width of cloth among the U and the U are formed, and size is (N 2-1) * 1.
3. the related feedback method of initiatively selecting presentation graphics according to claim 1 is characterized in that in the step (7) that the method for calculating the similarity of every width of cloth picture based on L is: suppose that at first hypothesis has n among the L 1The image that the width of cloth is relevant has n 2Incoherent image, always total n width of cloth image need calculate similarity, and supposes that two length are n 1Interim array A, B store intermediate value respectively, concrete grammar is as follows:
(71) picture count parameter i is changed to 1, begins counting;
(72) judge whether that n width of cloth image all finishes as calculated, finish if i＞n then jumps to step 75, otherwise enter next step;
(73) divide two branches to carry out:
First branches into:
(7310) establish picture count parameter j=1;
(7311) judge that whether j is greater than n 1, be then to jump to step 7316, otherwise enter next step;
(7312) picture count parameter k initially is changed to 1;
(7313) judge that whether k is greater than n 1, be then to jump to step 7315, otherwise enter next step;
(7314) according to formula (2) value of calculating e 1Save as k the element of A afterwards, and k is added 1;
(7315) from A, select a maximal value, save as j the element of B, and j is added 1;
(7316) mean value of calculating B, the value that obtains is designated as s 1
Second branches into:
(7320) establish picture count parameter j=1;
(7321) judge that whether j is greater than n 1, be then to jump to step 7316, otherwise enter next step;
(7322), picture count parameter k initially is changed to 1;
(7323) judge that whether k is greater than n 1, be then to jump to step 7315, otherwise enter next step;
(7324) according to formula (4) value of calculating e 1Save as k the element of A afterwards, and k is added 1;
(7325) from A, select a maximal value, save as j the element of B, and j is added 1;
(7326) mean value of calculating B, the value that obtains is designated as S 2
e 1 = d ( x i , x k ) × d ( x i , x j ) + d ( x k , x j ) 2 - - - ( 2 )
d ( a , b ) = 1 Σ t = 1 m | a t - b t | 2 + e 0 - - - ( 1 )
e 2 = 2 d ( x i , x k ) d ( x i , x j ) + d ( x k , x j ) - - - ( 2 )
(74) with s 1And s 2Unification is sued for peace after same dimension, obtains s, and the similarity as i opens image adds 1 with picture count parameter i then, turns back to step 72;
(75) finish;
X in the formula (2) i, x j, and x kRepresent that respectively the i width of cloth will calculate the image of similarity, j opens the proper vector that k opens uncorrelated image among associated picture and the L among the L, and function d is the distance that is used for calculating two vectors, and definition is suc as formula (3); A and b are two vectors in the formula (3), and m is the dimension of vectorial a and b, a tThe value of representing the t dimension of vectorial a, e 0For predefined greater than 0 very little constant; Identical in the implication of the variable in the formula (4) and the formula (2).
CN201010137511XA 2010-03-31 2010-03-31 Method for actively selecting related feedbacks of representative image Active CN101833565B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201010137511XA CN101833565B (en) 2010-03-31 2010-03-31 Method for actively selecting related feedbacks of representative image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201010137511XA CN101833565B (en) 2010-03-31 2010-03-31 Method for actively selecting related feedbacks of representative image

Publications (2)

Publication Number Publication Date
CN101833565A true CN101833565A (en) 2010-09-15
CN101833565B CN101833565B (en) 2011-10-19

Family

ID=42717635

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201010137511XA Active CN101833565B (en) 2010-03-31 2010-03-31 Method for actively selecting related feedbacks of representative image

Country Status (1)

Country Link
CN (1) CN101833565B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102542014A (en) * 2011-12-16 2012-07-04 华中科技大学 Image searching feedback method based on contents
CN102799635A (en) * 2012-06-27 2012-11-28 天津大学 Image set ordering method driven by user
CN107851097A (en) * 2015-03-31 2018-03-27 株式会社Ｆｒｏｎｔｅｏ Data analysis system, data analysing method, data analysis program and storage medium
CN108764068A (en) * 2018-05-08 2018-11-06 北京大米科技有限公司 A kind of image-recognizing method and device

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1851703A (en) * 2006-05-10 2006-10-25 南京大学 Active semi-monitoring-related feedback method for digital image search
CN1967526A (en) * 2006-09-14 2007-05-23 浙江大学 Image search method based on equity search
CN101021849A (en) * 2006-09-14 2007-08-22 浙江大学 Transmedia searching method based on content correlation
CN101539930A (en) * 2009-04-21 2009-09-23 武汉大学 Search method of related feedback images

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1851703A (en) * 2006-05-10 2006-10-25 南京大学 Active semi-monitoring-related feedback method for digital image search
CN1967526A (en) * 2006-09-14 2007-05-23 浙江大学 Image search method based on equity search
CN101021849A (en) * 2006-09-14 2007-08-22 浙江大学 Transmedia searching method based on content correlation
CN101539930A (en) * 2009-04-21 2009-09-23 武汉大学 Search method of related feedback images

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
《模式识别与人工智能》 20050831 陈可佳等 基于主动相关反馈的图像检索方法 第480-485页 1-3 第18卷, 第4期 2 *
《模式识别与人工智能》 20060430 戴宏斌等 一种基于多示例学习的图像检索方法 第179-185页 1-3 第19卷, 第2期 2 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102542014A (en) * 2011-12-16 2012-07-04 华中科技大学 Image searching feedback method based on contents
CN102542014B (en) * 2011-12-16 2013-09-18 华中科技大学 Image searching feedback method based on contents
CN102799635A (en) * 2012-06-27 2012-11-28 天津大学 Image set ordering method driven by user
CN102799635B (en) * 2012-06-27 2015-10-28 天津大学 The image collection sort method that a kind of user drives
CN107851097A (en) * 2015-03-31 2018-03-27 株式会社Ｆｒｏｎｔｅｏ Data analysis system, data analysing method, data analysis program and storage medium
CN108764068A (en) * 2018-05-08 2018-11-06 北京大米科技有限公司 A kind of image-recognizing method and device

Also Published As

Publication number Publication date
CN101833565B (en) 2011-10-19

Similar Documents

Publication Publication Date Title
Min et al. A discrete hard EM approach for weakly supervised question answering
CN105701216B (en) A kind of information-pushing method and device
CN109190043A (en) Recommended method and device, storage medium, electronic equipment and recommender system
CN103870505B (en) One kind inquiry words recommending method and query word commending system
CN108959603A (en) Personalized recommendation system and method based on deep neural network
CN103793721B (en) Pedestrian repeat recognition method and system based on area related feedback
San Pedro et al. Leveraging user comments for aesthetic aware image search reranking
CN110442781A (en) It is a kind of based on generate confrontation network to grade ranked items recommended method
CN105426529B (en) Image retrieval method and system based on user search intention positioning
CN107220365A (en) Accurate commending system and method based on collaborative filtering and correlation rule parallel processing
CN110674407A (en) Hybrid recommendation method based on graph convolution neural network
CN105824911B (en) Video recommendation method based on LDA user's topic model
CN106777282B (en) The sort method and device of relevant search
CN107103000A (en) It is a kind of based on correlation rule and the integrated recommended technology of Bayesian network
CN106021374A (en) Underlay recall method and device for query result
CN103778227A (en) Method for screening useful images from retrieved images
US8346766B2 (en) Efficient calculation of node proximity on graphs with side information
CN106844518A (en) A kind of imperfect cross-module state search method based on sub-space learning
CN101833565B (en) Method for actively selecting related feedbacks of representative image
CN109165847A (en) A kind of item recommendation method based on recommender system, device and equipment
CN106789338B (en) Method for discovering key people in dynamic large-scale social network
US20160098437A1 (en) Information retrieval method and apparatus
Sheng Simple multiple noisy label utilization strategies
CN108984555A (en) User Status is excavated and information recommendation method, device and equipment
CN107562966A (en) The optimization system and method based on intelligence learning for web page interlinkage retrieval ordering

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C14 Grant of patent or utility model
GR01 Patent grant