CN101377776B - Method for searching interactive image - Google Patents
Method for searching interactive image Download PDFInfo
- Publication number
- CN101377776B CN101377776B CN2007101210793A CN200710121079A CN101377776B CN 101377776 B CN101377776 B CN 101377776B CN 2007101210793 A CN2007101210793 A CN 2007101210793A CN 200710121079 A CN200710121079 A CN 200710121079A CN 101377776 B CN101377776 B CN 101377776B
- Authority
- CN
- China
- Prior art keywords
- data point
- image data
- image
- mark
- point
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 42
- 230000002452 interceptive effect Effects 0.000 title claims abstract description 15
- 238000012549 training Methods 0.000 claims description 15
- 239000004744 fabric Substances 0.000 description 14
- 238000005516 engineering process Methods 0.000 description 7
- 230000000875 corresponding effect Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 238000002474 experimental method Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 3
- 239000000284 extract Substances 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 244000132059 Carica parviflora Species 0.000 description 1
- 235000014653 Carica parviflora Nutrition 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000003203 everyday effect Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012797 qualification Methods 0.000 description 1
- 230000026676 system process Effects 0.000 description 1
- 239000013598 vector Substances 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The invention provides an interactive image retrieval method based on mobile virtual classified faces, which adopts mobile virtual classified faces for selecting image data points; the image data points are marked one by one; the previous marked image data point is used for selecting the next image data point; the marked image data points are used for searching for images to be accessed, thus completing interactive image retrieval. During the feedback relating to retrieval of images, active learning is often used for reducing the data quantity to be marked manually, which is just to select the data with the maximum information quantity. The relationships between data points are omitted in the traditional method for marking in batches, and the traditional method is not so efficient. As in the point selection policy based on mobile virtual classified faces provided by the invention, the previous marked data are used for providing directions for selecting the next data, thus improving the total information quantity of the marked data without increasing the marked data quantity. The point selection policy based on the mobile virtual classified faces significantly improves the performance of active learning algorithm.
Description
Technical field
The invention belongs to technical field of image processing, relate to interactive image retrieval method based on the mobile virtual classifying face.
Background technology
Along with the raising of Development of Multimedia Technology and computer data processing power, all kinds of resources on the network become increasingly abundant, and especially the digital picture of visual pattern increases especially with surprising rapidity.In worldwide, all can produce the image of enormous amount every day, this just require a kind of can be fast and search the technology of access images, just so-called image retrieval technologies exactly.
CBIR (Content-Based Image Retrieval is called for short CBIR) is current a kind of effective image search method comparatively commonly used, and its main difficult point is how to cross over the wide gap of image high-level semantic and low-level image feature.Relevant feedback (Relevance Feedback) has been proved to be a kind of technology that can handle the problems referred to above preferably, its main thought is to introduce man-machine interaction in the image retrieval process, return some images at every turn and mark, the accuracy that utilizes these information that marked image to remove to improve image retrieval then to the user.Take turns in the relevant feedback every, the image that returns to user's mark is considerably less for huge image data base, how effectively utilizing limited mark image and a large amount of do not mark image to improve the performance of image indexing system as much as possible, is a key issue in the relevant feedback technology.
Initiatively study (Active Learning) is a kind of effective learning algorithm at the situation proposition of " labeled data is few; unlabeled data is many " in the machine learning, its main thought is: the data of only choosing " information is arranged most " mark, and make to obtain big as far as possible raising by its performance of new sorter that these data trained.Study is initiatively applied in the relevant feedback to improve the effect of image retrieval, is a trend of present image retrieval, and wherein the most representative is the SVM that S.Tong and E.Chang propose
Active: they think that the nearest data point of the current training gained classification of distance lineoid is the data of " information is arranged most ", therefore take turns in the relevant feedback every, they return the nearest image of a collection of distance classification face and mark to the user, and train a new sorter with these mark images.It is considered herein that the reconnaissance strategy of above-mentioned traditional " mark in batch " has been ignored the mutual relationship between the unlabeled data point, so has reduced the total quantity of information of institute's labeled data, and then has influenced the performance of entire image searching system.
Summary of the invention
The traditional reconnaissance method often nearest batch data of the current classifying face of selected distance is carried out disposable mark, this method of mark has in batch been ignored the relation between the data point, make that system's retrieval is efficient inadequately, in order to solve prior art problems, the objective of the invention is to adopt efficient reconnaissance strategy, increase the total quantity of information of institute's labeled data, improve the performance of entire image searching system, for this reason, the present invention proposes a kind of based on the novel mobile virtual classifying face strategy interactive image retrieval method of (Moving VirtualBoundary strategy is called for short the MVB strategy).
In order to realize described purpose, the technical scheme of interactive image retrieval method of the present invention is as described below:
Step 1: adopt the mobile virtual classifying face to choose the data point of image;
Step 2: image data point is marked one by one;
Step 3: utilize the image data point of previous mark to choose next image data point;
Step 4: utilize the image data point of mark as training sample sorter to be upgraded, the result for retrieval after being improved is finished interactive image retrieval.
According to embodiments of the invention, described image data point, described image data point, be with the distance metric image data point quantity of information of image data point to classifying face, the distance of promptly utilizing image data point to arrive classifying face is judged the uncertainty of class categories, thereby as the obtainable quantity of information of this image data point of mark.Distance is near more, and uncertainty is big more, thereby it is big more to mark this obtainable quantity of information of image data point institute.
According to embodiments of the invention, the annotation step of described image data point:
Step 31: mark an image data point;
Step 32: the quantity of information of in good time adjusting other each image data point with the virtual class face that moves;
Step 33: the image data point of choosing near true classifying face marks.
According to embodiments of the invention, described virtual class step is as follows:
Step 41: with all images data point according to its to the distance map of current classifying face on the one dimension coordinate axis, the image data point that is divided into positive class is mapped to positive axis, the image data point that is divided into negative class is mapped to negative semiaxis, initial point is corresponding to current classifying face;
Step 42: each moving chooses that promptly nearest apart from the initial point image data point of coordinate absolute value minimum marks on the number axis;
Step 43:, then keep origin position constant if promptly the true classification of this image data point is consistent with the classification results of current classifying face for annotation results; Otherwise the classification results of this image data point was consistent with its true classification after mobile origin position made and moves.
According to embodiments of the invention, described at each mobile virtual classifying face, the mark point that was originally divided by mistake is correctly classified by the virtual class face.
According to embodiments of the invention, described interactive image retrieval method is characterized in that: it is that initial point is moved to the midpoint that the mark point is adjacent data point that the step-length of described each mobile virtual classifying face is chosen.
According to embodiments of the invention, described initial point moves step:
Step 71: initial point moves to a plurality of image data point and is marked;
Step 72: these image data point that mark are added training set;
Step 73: utilize training set to train a new sorter.
According to embodiments of the invention, the quantity of information of unlabeled data is taken turns in the relevant feedback with the mark dynamic change every in the described quantity of information.
Good effect of the present invention:
In the relevant feedback process of image retrieval, initiatively study usually is used to alleviate the data volume of artificial mark, and its main thought is that the data of at every turn only choosing the quantity of information maximum mark.The traditional reconnaissance method often nearest batch data of the current classifying face of selected distance is carried out disposable mark, and the method for this " mark in batch " has been ignored the relation between the data point, and is therefore efficient inadequately.The present invention proposes a kind of novel reconnaissance strategy: " mobile virtual classifying face " strategy (MovingVirtual Boundary strategy is called for short the MVB strategy).Adopted the strategy of " mark one by one " to replace the method for traditional " mark in batch ", utilize the data of previous mark to go, thereby under the condition that does not increase the labeled data amount, improved the total quantity of information of institute's labeled data to choosing of next data provides guidance.Among the present invention, the quantity of information of unlabeled data is taken turns in the relevant feedback with the mark dynamic change every, but not changeless constant." the virtual class face " that move that the present invention adopts removes to approach the true classifying face of data, and chooses in this process and have the information data point to mark most.Experiment showed, and adopted the active learning algorithm performance of MVB reconnaissance strategy to be significantly improved.Experiment showed, that this strategy has important theory and practical significance for the performance of the efficient of improving active learning algorithm in the relevant feedback, raising image indexing system.
Description of drawings
Figure 1 shows that the nearest data point of the current classifying face of distance often is not the data point of " information is arranged most ";
Figure 2 shows that the comparison of MVB reconnaissance strategy and traditional reconnaissance strategy;
Figure 3 shows that initial point moves synoptic diagram;
Figure 4 shows that the image indexing system process flow diagram of band relevant feedback
Fig. 5 adopts the relevant feedback process flow diagram of MVB reconnaissance strategy
Be respectively the degree of accuracy curve after three-wheel and five is taken turns relevant feedback shown in Fig. 6 a and Fig. 6 b;
Fig. 7 a and Fig. 7 b are depicted as the relation curve of 30 and 50 width of cloth pattern accuracy and relevant feedback wheel number.
Embodiment
Describe each related detailed problem in the technical solution of the present invention in detail below in conjunction with accompanying drawing.Be to be noted that described embodiment only is intended to be convenient to the understanding of the present invention, and it is not played any qualification effect.
In the relevant feedback process of image retrieval, initiatively study usually is used to alleviate the data volume of artificial mark, and its main thought is that the data of at every turn only choosing the quantity of information maximum mark.The traditional reconnaissance method often nearest batch data of the current classifying face of selected distance is carried out disposable mark, and the method for this " mark in batch " has been ignored the relation between the data point, and is therefore efficient inadequately.The present invention proposes a kind of novel reconnaissance strategy: " mobile virtual classifying face " strategy (MovingVirtual Boundary strategy is called for short the MVB strategy).The present invention has adopted the method for " mark one by one ", utilizes the data of previous mark to go to choosing of next data provides guidance, thereby improved the total quantity of information of institute's labeled data under the condition that does not increase the labeled data amount.Experiment showed, and adopted the active learning algorithm performance of MVB reconnaissance strategy to be significantly improved.
Initiatively the key issue of study is how to choose the data of " information is arranged most ", and the reconnaissance strategy of current popular is chosen the nearest image of the current training gained of a collection of distance classifying face usually and marked to the user in image retrieval.But, being illustrated in figure 1 as the nearest data point of the current classifying face of distance often is not the data point of " information is arranged most ", because shown in the current training gained classifying face solid line often is not shown in the real classifying face dotted line of data point, and in fact the true nearest image of classifying face of distance is only the data of " information is arranged most ", so its quantity of information of data point of choosing according to traditional reconnaissance strategy often is not very big, thereby very limited for the improvement of sorter performance.
As shown in Figure 1: the negative class data of ▲ expression; ■ represents positive class data; ● expression has information data most; Zero expression unlabeled data;
The present invention proposes " mobile virtual classifying face " the reconnaissance strategy of (Moving Virtual Boundary is called for short MVB)." mark in batch " different from the past, the present invention has adopted the method for " mark one by one ", utilize the data of previous mark to go to provide guidance for choosing of next data, be illustrated in figure 2 as the comparison of MVB reconnaissance strategy and traditional reconnaissance strategy, thereby under the condition that does not increase the labeled data amount, improved the total quantity of information of institute's labeled data.
As shown in Figure 2, the comparison of MVB reconnaissance strategy and traditional reconnaissance strategy, the same with traditional method, the present invention is a kind of tolerance of the distance of data point distance classification face as its quantity of information: data point distance classification face is near more, the uncertainty of its classification is just big more, thereby it is just big more to mark this obtainable quantity of information of some institute.Difference is: take turns in the relevant feedback one, classic method is regarded the quantity of information of data point as constant constant, a collection of data point mark that each current classifying face of selected distance is nearest; The present invention then after data point of every mark, in time adjusts the quantity of information of other each point, the feasible as far as possible more close true classifying face of selecting of data point, but not current classifying face.
Introduce MVB reconnaissance strategy of the present invention below in detail:
At first, the present invention with all data points according to its to the distance map of current classifying face on the one dimension coordinate axis, the data point that is divided into positive class is mapped to positive axis, the data point that is divided into negative class is mapped to negative semiaxis, initial point is corresponding to current classifying face.
At every turn, the present invention's data point of choosing coordinate absolute value minimum (promptly nearest apart from initial point) on the number axis marks.If annotation results (i.e. this true classification) is consistent with the classification results of current classifying face, then keep origin position constant; Otherwise the classification results of this data point was consistent with its true classification after the present invention moved origin position and make to move.Moving synoptic diagram with initial point shown in Figure 3 is example, before mark, the data point nearest apart from initial point is an A, it is divided into positive class by current classifying face, and the result after the mark is that an A belongs to negative class, therefore initial point need be moved to the number axis positive dirction, make invocation point A for the initial point after moving, belong to negative class.Why the process that is equivalent in higher dimensional space mobile " virtual class face " on the process nature of above-mentioned mobile initial point (is referred to as " virtual class face ", be because it is not the classifying face that is obtained by training, but a kind of of true classifying face possible position more reasonably supposed), the origin of this policy name of MVB reconnaissance just.After each moving, guarantee that all the mark point that was originally divided by mistake can correctly be classified by the virtual class face.For each step-length that moves, different choosing methods is arranged, simple, the most stable a kind of method is that initial point is moved to the midpoint that the mark point is adjacent data point.
Initial point moves synoptic diagram as shown in Figure 3, repeats above process, is marked up to the data point of some.The data point that then these is marked adds training set, thereby trains a new sorter.
Now the concrete implementation step with MVB reconnaissance strategy in the image retrieval process is summarized as follows:
Known: current classifying face f, training image set L (marking) does not mark image collection U, and each takes turns the picture number N of the required mark of relevant feedback.
Step 1: initialization virtual class face f
vBe current classifying face f; To not mark that each width of cloth does not mark image mapped on the one dimension coordinate axis among the image collection U, image x is f corresponding to coordinate on the coordinate axis
v(x) point; Initial point is corresponding to virtual class face f
v
Step 2: repeat following flow process N time:
(1) the nearest image data point of chosen distance initial point, i.e. coordinate absolute value | f
v(x) | minimum image data point marks, and is assumed to be x
1
(2) if x
1Be positioned at the positive dirction (being the right side) of initial point, but be noted as " uncorrelated " (being negative sample), then initial point is moved to x
1And the midpoint of right adjoint point; If x
1Be positioned at the negative direction (i.e. left side) of initial point, but be noted as " being correlated with " (being positive sample), then initial point is moved to x
1And the midpoint of left adjoint point; Otherwise, keep origin position constant.
(3) with x
1Never mark among the image collection U and reject, and add training image set L.
(4) recomputate the coordinate f that each does not mark image data point x according to new origin position
v(x).
Step 3: L trains a new classifying face f ' with the training image set.
Because the virtual class face moves to the direction of the mark point of can correctly classifying all the time, so can think that it is to level off to real classifying face as much as possible, thus make select at every turn also near true classifying face annex the information data point arranged most apart from the nearest data point of virtual class face constantly.Obviously, the data point of choosing out by the MVB strategy has the more information amount than only being confined to the data point that current classifying face annex chooses, and is therefore more helpful for the raising of sorter performance.
The band relevant feedback the image indexing system flow process as shown in Figure 4, specific explanations is as follows:
1. utilize Feature Extraction Technology, extract characteristics of image (as color, texture, shape etc.) from database images, and be kept in the image feature base, this step can be finished by off-line.
2. the query image of submitting to for the user need be by corresponding feature in feature extraction acquisition and the image feature base, and this step is online to be finished.
3. with characteristics of image input category device, obtain Query Result, promptly in the database image according to the descending sort of query image correlativity.If the user is satisfied to Query Result, then poll-final; Otherwise the user need provide feedback information, so that sorter obtains better result.
The relevant feedback flow process that adopts MVB reconnaissance strategy as shown in Figure 5, specific explanations is as follows:
(1) uses MVB reconnaissance strategy, from image library, choose the image (being the nearest image of distance classification face) that a width of cloth " has information most ", mark by the user, annotation results will determine whether moving of virtual class face to reach moving direction, position, and then instruct choosing of image that next width of cloth " has information most ".This step can circulate N time, has marked image thereby obtain the N width of cloth.
(2) after MVB reconnaissance mark process finished, the N width of cloth image that mark is good was trained sorter as training image, the sorter after upgrading with acquisition.
Implementation result
Be the validity of proof MVB reconnaissance strategy, the present invention uses different reconnaissance strategies to carry out the image retrieval experiment on a subclass of Coral image set.5000 width of cloth images are arranged in image set of the present invention, be divided into 50 and have different semantic classifications, every class 100 width of cloth images.In the experiment, the present invention as query image, calculates average retrieval accuracy with preceding 10 width of cloth images of every class at last.
The present invention get colors and textural characteristics as the level image feature representation.Color characteristic is made up of the color histogram of rgb space 125 dimensions and the color moment of 6 dimensions.Textural characteristics extracts as follows: at first image is carried out 3 layers wavelet transform, form the proper vectors of totally 20 dimensions with separately average and variance on 10 subbands then, as image texture features.
The present invention with MVB reconnaissance application of policies in SVM
ActiveAnd adopt the original SVM of traditional reconnaissance strategy
ActiveCompare.The sorter that adopts is a svm classifier device commonly used in the image retrieval, selects RBF nuclear, and parameter obtains by cross validation.Take turns in the relevant feedback every, two kinds of methods all mark the image (20 width of cloth) of equal number.
Fig. 6 a and Fig. 6 b have provided respectively the curve of taking turns image retrieval degree of accuracy after the relevant feedback through three-wheel and five, and wherein ordinate is a precision value, and horizontal ordinate is illustrated in the degree of accuracy in the preceding x width of cloth image range of returning.
Fig. 7 a and Fig. 7 b are shown is respectively the relation curve of degree of accuracy and relevant feedback wheel number in preceding 30 and 50 width of cloth images that return, and wherein ordinate is a precision value, and horizontal ordinate is the wheel number of relevant feedback.
Adopt the SVM of MVB reconnaissance strategy among Fig. 6 a and Fig. 6 b, Fig. 7 a and Fig. 7 b with zero curve representation
Active, the original SVM of curve representation of band △
ActiveThe present invention adopts after the MVB reconnaissance strategy SVM as can be seen from figure
ActivePerformance obtained obvious improvement, the degree of accuracy of returning image is than original SVM
ActiveHad and significantly improved.For example, after taking turns relevant feedback through 3, the accuracy rate of preceding 30 width of cloth images that return brings up to 43.69% by original 33.97%; And after 5 took turns relevant feedback, the accuracy rate of preceding 50 width of cloth images that return brought up to 49.58% by original 37.45%.This has illustrated that using the MVB strategy to choose out the data point that marks has the more information amount than traditional reconnaissance strategy, thereby can improve the performance of image indexing system more effectively.
The above; only be the embodiment among the present invention; but protection scope of the present invention is not limited thereto; anyly be familiar with the people of this technology in the disclosed technical scope of the present invention; can understand conversion or the replacement expected; all should be encompassed in of the present invention comprising within the scope, therefore, protection scope of the present invention should be as the criterion with the protection domain of claims.
Claims (5)
1. an interactive image retrieval method is characterized in that,
Step 1: adopt the mobile virtual classifying face to choose the data point of image; With all data points according to its to the distance map of current mobile virtual classifying face on the one dimension coordinate axis, the data point that is divided into positive class is mapped to positive axis, the data point that is divided into negative class is mapped to negative semiaxis, initial point is corresponding to current mobile virtual classifying face;
Step 2: image data point is marked one by one, each, choose that promptly nearest apart from the initial point data point of coordinate absolute value minimum marks on the number axis; If annotation results i.e. this true classification is consistent with the classification results of current mobile virtual classifying face, then keep origin position constant; Otherwise the classification results of this data point was consistent with its true classification after mobile origin position made and moves;
Step 3: utilize the image data point of previous mark choose next image data point specifically be according to: after data point of every mark, adjust the quantity of information that other does not mark image data point with the virtual class face that moves in good time, the image data point of choosing near the mobile virtual classifying face marks, the feasible as far as possible more close true classifying face of selecting of data point;
Step 4: utilize the image data point of mark as training sample sorter to be upgraded, the result for retrieval after being improved is finished interactive image retrieval.
2. interactive image retrieval method according to claim 1, it is characterized in that: described image data point, be with the distance metric image data point quantity of information of image data point to classifying face, the distance of promptly utilizing image data point to arrive classifying face is judged the uncertainty of class categories, thereby as the obtainable quantity of information of this image data point of mark; Distance is near more, and uncertainty is big more, thereby it is big more to mark this obtainable quantity of information of image data point institute.
3. interactive image retrieval method according to claim 1 is characterized in that: described mobile virtual classifying face is that the mark point that was originally divided by mistake is correctly classified by the virtual class face.
4. interactive image retrieval method according to claim 1 is characterized in that: it is that initial point is moved to the midpoint that the mark point is adjacent data point that the step-length of described mobile virtual classifying face is chosen.
5. interactive image retrieval method according to claim 1 is characterized in that: it is as follows that described sorter carries out step of updating:
Step 71: initial point moves up to a plurality of image data point and is marked;
Step 72: these image data point that mark are added training set;
Step 73: utilize training set to train a new sorter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2007101210793A CN101377776B (en) | 2007-08-29 | 2007-08-29 | Method for searching interactive image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2007101210793A CN101377776B (en) | 2007-08-29 | 2007-08-29 | Method for searching interactive image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101377776A CN101377776A (en) | 2009-03-04 |
| CN101377776B true CN101377776B (en) | 2010-06-30 |
Family
ID=40421322
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2007101210793A Expired - Fee Related CN101377776B (en) | 2007-08-29 | 2007-08-29 | Method for searching interactive image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101377776B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104077344B (en) * | 2013-12-31 | 2018-07-17 | 河南大学 | Interactive image retrieval method and system based on adaptive learning region importance |
| KR20160087640A (en) * | 2015-01-14 | 2016-07-22 | 엘지전자 주식회사 | Mobile terminal and method for controlling the same |
| KR101981075B1 (en) * | 2015-03-31 | 2019-05-22 | 가부시키가이샤 프론테오 | Data analysis system, data analysis method, data analysis program, and recording medium |
| CN109033372A (en) * | 2018-07-27 | 2018-12-18 | 北京未来媒体科技股份有限公司 | A kind of content information retrieval method and system based on artificial intelligence |
-
2007
- 2007-08-29 CN CN2007101210793A patent/CN101377776B/en not_active Expired - Fee Related
Non-Patent Citations (9)
| Title |
|---|
| Avrim Blum,Tom Mitchell.Combining Labeled and Unlabeled Data with Co-Training.Proceedings of the 11th Annual Conference on Computational Learning Theory.1998,92-100. * |
| Jian Cheng,Kongqiao Wang.Active learning for image retrieval with Co-SVM.pattern recognition 40.2006,(40),330-334. |
| Jian Cheng,Kongqiao Wang.Active learning for image retrieval with Co-SVM.pattern recognition 40.2006,(40),330-334. * |
| Simon Tong,Edward Chang.Suppot Vector Machine Active Learning for Image Retrieval.Proceedings of the 9th ACM International Conference on Multimedia.2001,107-118. |
| Simon Tong,Edward Chang.Suppot Vector Machine Active Learning for Image Retrieval.Proceedings of the 9th ACM International Conference on Multimedia.2001,107-118. * |
| Xu Jie,Shi Pengfei.Active learning based on maximizing information gain forcontent-based image retrieval.Journal of Southeast University(English Edition)20 4.2004,20(4),431-435. |
| Xu Jie,Shi Pengfei.Active learning based on maximizing information gain forcontent-based image retrieval.Journal of Southeast University(English Edition)20 4.2004,20(4),431-435. * |
| 吴洪,卢汉清,马颂德.基于内容图像检索中相关反馈技术的回顾.计算机学报28 12.2005,28(12),1969-1979. |
| 吴洪,卢汉清,马颂德.基于内容图像检索中相关反馈技术的回顾.计算机学报28 12.2005,28(12),1969-1979. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN101377776A (en) | 2009-03-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bui et al. | Scalable sketch-based image retrieval using color gradient features | |
| Isola et al. | Scene collaging: Analysis and synthesis of natural images with semantic layers | |
| Zhou et al. | Sketch-based image retrieval on a large scale database | |
| CN106126581A (en) | Cartographical sketching image search method based on degree of depth study | |
| CN103631928A (en) | LSH (Locality Sensitive Hashing)-based clustering and indexing method and LSH-based clustering and indexing system | |
| CN108897791B (en) | Image retrieval method based on depth convolution characteristics and semantic similarity measurement | |
| CN101377776B (en) | Method for searching interactive image | |
| Yu et al. | A rule based technique for extraction of visual attention regions based on real-time clustering | |
| CN115424059A (en) | Remote sensing land use classification method based on pixel level comparison learning | |
| Wu et al. | Content-based image retrieval using growing hierarchical self-organizing quadtree map | |
| Min et al. | Overview of content-based image retrieval with high-level semantics | |
| Fung et al. | Learning primitive and scene semantics of images for classification and retrieval | |
| Peng et al. | A hierarchical model to learn object proposals and its applications | |
| Laga et al. | A boosting approach to content-based 3D model retrieval | |
| CN112487241A (en) | Video abstraction method based on convolutional neural network | |
| CN104484418B (en) | A kind of characteristic quantification method and system based on dual resolution design | |
| Tanaka et al. | Fisher vector based on full-covariance gaussian mixture model | |
| Shaheen et al. | Efficient Content-Based Image Retrieval System with Two-Tier Hybrid Frameworks | |
| Shah et al. | Object detection using convolutional neural networks and transformer-based models: a review | |
| Memon et al. | A novel technique for region-based features similarity for content-based image retrieval | |
| Ding et al. | A novel few-shot action recognition method: temporal relational crosstransformers based on image difference pyramid | |
| CN109766467A (en) | Remote sensing image retrieval method and system based on image segmentation and improvement VLAD | |
| Qu et al. | iDistance based interactive visual surveillance retrieval algorithm | |
| Maihami et al. | Color Features and Color Spaces Applications to the Automatic Image Annotation | |
| Wang et al. | PCCN‐RE: Point cloud colourisation network based on relevance embedding |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20100630 Termination date: 20210829 |