CN106776896A - A kind of quick figure fused images search method - Google Patents
A kind of quick figure fused images search method Download PDFInfo
- Publication number
- CN106776896A CN106776896A CN201611078791.5A CN201611078791A CN106776896A CN 106776896 A CN106776896 A CN 106776896A CN 201611078791 A CN201611078791 A CN 201611078791A CN 106776896 A CN106776896 A CN 106776896A
- Authority
- CN
- China
- Prior art keywords
- image
- picture
- retrieval
- weight map
- feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
-
- 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)
- Physics & Mathematics (AREA)
- Library & Information Science (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of quick figure fused images search method.Comprise the following steps：First, by image retrieval data set, the ranking results of feature are obtained；Secondly, weight map is set up, weight map interior joint is the picture in data acquisition system, and the similarity between node between weight and image is directly proportional, for the method that the weight map of different characteristic is merged using figure；Then, in weight map, centered on image to be retrieved according to set forth herein clustering method carry out image sets division；Finally, the order rearrangement candidate image of image sets is added according to candidate image.The present invention builds weight map using jaccard；Thought according to spectral clustering proposes that quick spectral clustering greedy algorithm divides image sets, retrieval result is optimized using the result of cluster draws more excellent retrieval effectiveness；Existing image searching result is resequenced with the thought of spectral clustering and greedy algorithm, the degree of accuracy of image retrieval is substantially increased.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of quick figure fused images search method.
Background technology
Recently as the sharp increase of internet and mobile terminal epigraph data, how fast and effeciently from database
The image that middle retrieval user needs has turned into the hot issue of computer vision field.Therefore I wishes to lifting conventional images inspection
The effect of Suo Fangfa.
Current main flow image retrieval can be divided three classes：Image retrieval based on text, picture material and image, semantic.Base
It is in text and still undesirable based on semantic retrieval effectiveness, and CBIR recent years are achieved preferably
Effect, the main way as image retrieval.CBIR mainly studies image characteristics extraction and image sequence
Both sides works, and the effect that the method and image of image characteristics extraction sort all to image retrieval has a great impact, good
Characteristics of image cooperation image ranking method adaptable therewith can just obtain preferable image retrieval effect.
In terms of image sequence, existing many outstanding achievements.Such as Liu has found that norm order sorts than Euclidean distance and imitates
Fruit is more preferably.But there is ranking mode independence in norm order.Therefore, in follow-up research, graphics is practised in image inspection
Vital effect is played in rope.Manifold ranking (Manifold Ranking, MR) is applied to image retrieval by He etc., is obtained
Better effects.He etc. further provides the MR of broad sense, in the case of data distribution is uncomplicated, can obtain more excellent than MR
Result.K-NN in graph structure is improved to K-RNN, lifting figure by Wang etc. using the degree of sample node and the minimum of weights
As the effect of retrieval.In terms of graph structure, Huang et al. realizes MR using hypergraph, and the weights between node are determined by probability, by
In different node to there is multiple weights, so relatively good ranking results can be obtained in image retrieval.Except hypergraph, scheme more
It is also a kind of realization rate of multiple weighing value between sample point.Zhao etc. using the different characteristic of image build many figures strengthen sample points it
Between relation, realize image retrieval.Zhang etc. is carried out many figures using Graph PageRank and Graph Density methods
Fusion, realizes the sequence of image.
Emphasis is different when extracting feature due to single feature image retrieval method, and such as HSV is primarily upon the face of image
Color information, LBP is primarily upon texture information of image etc., therefore causes single feature image retrieval method for different retrievals
The effect quality of image is different.Single feature image retrieval method can not obtain promising result to all of retrieval image, very unstable
It is fixed.If different single characteristic key methods can be merged, then can just improve the accuracy of image searching result.Graph
Density methods are exactly one of figure fusion learning method of current main flow.The method by merge the big feature group of otherness come
Improve the accuracy of retrieval result.
The content of the invention
The purpose of the present invention is all of retrieval image can not to be obtained for single feature image retrieval method and be satisfied with effect
Really, a kind of very unstable technical problem, it is proposed that quick figure fused images search method.
A kind of quick figure fused images search method that the present invention is provided, comprises the following steps：
Step 100：By image retrieval data set, the ranking results of feature are obtained.
Step 200：Weight map is set up, weight map interior joint is the picture in data acquisition system, weight and image between node
Between similarity be directly proportional, for different characteristic weight map using figure fusion method.
Step 300：In weight map, centered on image to be retrieved according to set forth herein clustering method carry out image sets
Divide.
Step 400：The order rearrangement candidate image of image sets is added according to candidate image.
Preferably, step 200 detailed process is：Each picture that the picture and data that will be retrieved are concentrated is considered as not
Same node, and a line is set up between each node according to the similarity of picture after extraction feature, side right is similar by picture
The size of degree determines that similarity picture higher, the side right between node is bigger, otherwise then smaller, it is also possible to be zero；According to
Rule above sets up a characteristic pattern, you can represent per pictures similarity relation.
Preferably, the step 300 includes following sub-step：
Step 310：Using formula：
G1=(V1, E1, w1)+G2=(V2, E2, w2(V, E w) carry out figure fusion to)=G=；Wherein, G1：It is ith feature
Figure, V1：It is the set of the node in figure, E：It is the set on side in figure, w：It is side right function in figure.
Step 320：Using formula：Clustered,
Wherein, q is retrieving image,Preceding k pictures are expressed asDuring v is the figure asked
Node, it is side right function in figure that w is,To the side right size of v；
When finding+1 candidate's picture of kth every time, as long as the larger picture of k image correlation before finding.
Beneficial effect：The present invention builds weight map using jaccard；Thought according to spectral clustering proposes that quick spectral clustering is coveted
Center algorithm divides image sets, retrieval result is optimized using the result of cluster draws more excellent retrieval effectiveness；To existing
Image searching result is resequenced with the thought of spectral clustering and greedy algorithm, substantially increases the accurate of image retrieval
Degree.
Brief description of the drawings
Fig. 1 is the convolutional neural networks that a kind of quick figure fused images search method provided in an embodiment of the present invention is used
Structure.
Fig. 2 is a kind of two kinds of different G of quick figure fused images search method provided in an embodiment of the present inventionm=
(Vm, Em, wm) be fused into G=(V, E, w).
Specific embodiment
For make present invention solves the technical problem that, the technical scheme that uses and the technique effect that reaches it is clearer, below
The present invention is described in further detail in conjunction with the accompanying drawings and embodiments.It is understood that specific implementation described herein
Example is used only for explaining the present invention, rather than limitation of the invention.It also should be noted that, for the ease of description, accompanying drawing
In illustrate only part rather than full content related to the present invention.
First, embodiment
A kind of quick figure fused images search method of the present embodiment, comprises the following steps：
Step 100：By image retrieval data set, the ranking results of feature are obtained；Image characteristics extraction is largely determined
The effect of image retrieval is determined, has been also the important step of multi-characteristic fusion method.
Global image feature：Global characteristics are mainly by picture overall permanence to describe the information entrained by picture.Example
Such as the color histogram (HSV) of statistical color information, the local binary patterns (LBP) of statistic texture information etc..
HSV is statistical color information histogram, and it can describe the global distribution of color in piece image, i.e. different color
The shared ratio in entire image, those are difficult to the image of automatic segmentation and without the concern for object sky to be particularly well-suited to description
Between position image, but it cannot describe the locus residing for the local distribution and every kind of color of color in image, i.e., without
Method describes a certain specific object or object in image.
LBP is the local binary patterns of statistic texture information, and it also illustrates scenery corresponding to image or image-region
Surface nature, but it cannot completely reflect high-level picture material.
In addition there are many global image methods, such as based on local sensitivity degree Hash (LSH), and on LSH
Improve mahalanobis distance etc..But the experiment of this paper mainly uses HSV features.
Local image characteristics：Local image characteristics describe picture feature, wherein picture region by describing the region of picture
Description in domain generally meets consistency (robustness) and ga s safety degree, such as scale invariability feature (SIFT, SURF).
SIFT is widely used local feature, and it becomes to yardstick, rotation and certain visual angle and illumination variation etc. image
Changing all has consistency, and SIFT has very strong ga s safety degree.
SURF is the modified version to SIFT, it using Haar small echos come the gradient operation in approximate SIFT methods, while
Quickly calculated using diagram technology is integrated, the speed of SURF is 3-7 times of SIFT, in most cases its performance with SIFT
Quite.In addition to SIFT feature, convolutional neural networks (CNN) feature application is also relatively broad.
Many binary feature description, such as ORB, BRISK, FREAK etc. are also proposed in the recent period.Local feature description's
The local feature of image is converted into reference to words tree (VOC) or bag of words (BOW).The experiment of this paper mainly uses SIFT feature+BOW
Method and CNN features.
Convolutional neural networks feature extraction mode：Convolutional neural networks are characterized in be trained in advance by Image-Net
Depth network (neural coding obtained by deep convolutional network transmission figure picture trains the class classification of 1000 image-Net).Utilize
Image-Net re -training networks, (can also be using other convolutional Neurals using the convolutional neural networks structure of AlexNet
Network structure, such as ILSVRC CNN).The AlexNet models include five convolutional layers, and every layer includes a convolution, rectification
Linearly (ReLU) conversion (F (X)=MAX (X, 0)), and a maximum pond conversion.There are three at the top of the architecture completely
Articulamentum (layer 6,7,8), wherein being taken as to be input into the output of preceding layer, by a matrix multiple, also, in the feelings of layer 6 and 7
Under condition, apply the linear transformation of rectification.Network be trained so that layer 8 output corresponding to class label a heat coding.
As shown in figure 1, network applies to 224 × 224 image.The size of other sizes image is mark with 224 × 224
Accurate (not cutting).The framework feedforward of CNN, and when giving image I, it produces the activation sequence of layer.We represent and corresponding layer
The activation (output) of L5 (I) L6 (I) and L7 (I).We use the value of L7 (I) as CNN features (L5 (I), L6 of image retrieval
(I) convolutional neural networks feature can equally be done).
Step 200：Weight map is set up, weight map interior joint is the picture in data acquisition system, weight and image between node
Between similarity be directly proportional, for different characteristic weight map using figure fusion method.
The figure fusion of different characteristic is realized in phase sorting, needs to build a good characteristic pattern first.General idea
It between picture is the probability of similar pictures that the weight for being the side for building figure is, but it is incalculable.
Each picture that the picture and data that will be retrieved are concentrated is considered as different nodes, and according to picture after extraction feature
Similarity a line is set up between each node, side right is determined by the size of picture similarity, similarity picture higher,
Side right between node is bigger, otherwise then smaller, it is also possible to be zero.
Rule according to more than sets up a characteristic pattern, you can represent per pictures similarity relation.
Step 300：In weight map, centered on image to be retrieved according to set forth herein clustering method carry out image sets
Divide；The figure fusion method of packet sequencing mainly includes：Multi-characteristic fusion method and packet sequencing method.
Figure fusion：In order to obtain characteristics of image complementary information, the accuracy rate of image retrieval is improved, it is necessary to design various images
Feature fusion.Existing multiple features fusion method mainly has：In the fusion method of feature extraction phases and in phase sorting
Fusion method.Show not good always in terms of efficiency and reliability in the fusion method of feature extraction phases, especially merge
The more efficiency of feature are more difficult to ensure.For requiring efficient image retrieval, exist in the fusion method of feature extraction phases tight
Weight defect.Therefore, in the main direction of studying of the fusion method as multiple features fusion in image retrieval of phase sorting.Exist at present
The fusion method of phase sorting is mainly C Dwork etc.[37]The figure fusion method of proposition, but the method is still present a definite limitation
Limitation.In order to avoid above-mentioned multiple features fusion produced problem, following amalgamation mode is carried out herein.
As shown in Fig. 2 the image candidate set calculated different characteristic can obtain different characteristic pattern Gm=(Vm,
Em,wm), the characteristic pattern after fusion is designated as G=, and (V, E, w), it meets：1) V=∪ Vm；2) E=∪ Em；3) w (i, i ')=∑ wm
(i,i′).From amalgamation mode of this paper algorithms under different characteristic figure：Figure interior joint i and i ' (i ' ∈ N after fusionk
(i)) J (i, i ') more than constant α quantity it is more, then w (q, i) value is bigger.
Clustering algorithm：In order to the characteristic pattern of the fusion more than is ranked up, it is necessary to from suitable to image searching result
Clustering algorithm.Clustering algorithm needs to meet requirement of the image retrieval to efficiency.For retrieval image q, we divide set
It is S={ S1,S2, wherein set S1Size be k, as long as therefore find with k related image of retrieval image, then can be with complete
Into clustering algorithm.Herein using the thought (this paper frameworks can also use other clustering algorithms) of spectral clustering.
The core concept of spectral clustering is G=(V, E, minimal cut problem w), then according to its definitionCan show that the set for this paper divides S={ S1, S2The weights of minimal cut be：
The characteristic vector of Laplce's proof is obtained the need for spectral clustering, its time complexity is O (n3), it is impossible to meet figure
As effectiveness of retrieval requirement, therefore herein using the strategy of greed, approximate solution minimal cut.
For the preceding k pictures that retrieving image q, definition have been retrievedThen according to herein
Greedy algorithm, the pictures of kth+1 are：
WhereinWhen finding+1 candidate's picture of kth every time, we need not travel through whole image database, as long as seeking
The picture for looking for preceding k image correlation larger.
Step 400：The order rearrangement candidate image of image sets is added according to candidate image.
2nd, data analysis
The method that this section illustrates data set and experiment first, then describes the experiment knot on each data acquisition system in detail
Really, the result to each data set is analyzed.
Data set
Assessment herein uses four standard data sets：UK-bench, Corel-1k, Corel-10k and Cifar-10.Its
In, UKbench has picture rotation and dimensional variation, picture classification many and the features such as relatively large data acquisition system；Corel-10k
With every class picture number is more and the features such as relatively large data acquisition system；Corel-1k has and data set many per class picture number
The features such as closing big relatively small；Cifar-10 has the features such as image is smaller and data acquisition system is big.By data above collection, checking
The validity of this paper algorithms.Data set detail parameters are shown in Table 1.
| Database | Image Size | #of Class | #of Each Class | Total Image |
| UKbench | 640x480 | 2550 | 4 | 10200 |
| Corel-1k | 384x256 | 10 | 100 | 1000 |
| Corel-10k | Vary | 100 | 100 | 10000 |
The attribute of the experimental data set of table 1
UK-bench includes 2550 pictures of different data class, and 4 similar pictures are included per class image.
10200 pictures are gathered comprising data acquisition system and inquiry simultaneously.4 candidate's pictures are returned per pictures, is assessed using N-S.
Corel-1k includes 10 pictures of different data class, and 100 pictures are included per class image.1000 pictures
Gather comprising data acquisition system and inquiry.20 candidate's pictures are found per pictures, and assesses the accuracy rate of candidate's picture.
Corel-10k includes 100 pictures of different data class, and 100 pictures are included per class image.10000
Picture is gathered comprising data acquisition system and inquiry.12 candidate's pictures are found per pictures, and assesses the accuracy rate of candidate's picture.
Feature Selection and fusion method
The image retrieval mode on basis includes：CNN, HSV and BOW etc..Herein will be using single characteristic optimization method pair
GCNN、GHSVAnd GBOWOptimize, and application drawing fusion method is to GCNNGHSVAnd GBOWCarry out fusion optimization.
HSV：Herein 11 dimensional vectors are calculated using for retrieving each pixel of image：Black, blue, brown, grey, green, orange,
It is powder, purple, red, white and yellow.For retrieval image, herein description vectors using average vector as color descriptor as picture.
BOW：Herein using the bag of words based on image retrieval, extracted using VLFeat-library per pictures intensive
SIFT feature.Bag of words use 1M vocabulary, carry out kmeans clusters.
CNN：Herein again be based on AlexNet convolutional neural networks structures, and done on imagenet-1000 data sets into
Row pre-training, finally using L7 (I) value as image retrieval CNN features.
UKbench data sets
In order to illustrate effect of the context of methods in multi-characteristic fusion.On UKbench data sets, herein using BOW
With feature based on HSV methods, multi-characteristic fusion is carried out, and with other multi-characteristics fusion (Graph of current main flow
Density etc.) method is compared, and comparative result refers to table 2.
Table 2 on UKbench data sets, the N-S fractions of context of methods and current image retrieval multi-characteristic fusion method
Comparative result
The variety classes picture feature of table 3 is extracted in the performance on UKbench
From table 3 it can be seen that the more current multi-characteristic fusion method of context of methods is showed more on UKbench data sets
It is excellent.Because context of methods obtains more image retrieval information by picture group, context of methods is set to obtain effect substantially excellent
In other algorithms.In optimizations of this paper for single feature, HSV (N-S=3.17) results are relatively low compared with VOC (N-S=3.54).Through
Cross different figure fusion methods and all obtain different raisings.Single characteristic optimization of this paper, makes the N-S of HSV improve 0.12, VOC
Result improve 0.29.Context of methods can also merge other features, it can be deduced that more excellent result, such as table 3, be context of methods
The result of other features is merged, the wherein fusion of CNN, BOW, HSV and MSD can obtain 3.92 (accuracy rate 98.15%) N-S points
Number.
Corel-1k and Corel-10k data sets
In order to illustrate that GRF has good result on different pieces of information collection, herein in Corel-1k and Corel-10k data sets
On tested, and have better effects in Corel-1k and Corel-10k data sets per class figure on the two data sets
Piece includes 100 width pictures.In Corel-1k data sets, the accuracy rate of preceding 20 width picture is compared herein；In Corel-10k data
On collection, the accuracy rate of preceding 12 width picture is compared herein.
The optimization of table 4BOW and HSV and figure fusion and other algorithms are gone forward the standard of 20 width pictures in Corel-1k data sets
True rate
| Performance Type\method | SSH | Ri-HOG | HOG | LBP-MR1 | MSD-MR1 | GRF |
| Precision | 54.88 | 53.13 | 33.29 | 35.84 | 49.65 | 69.08 |
| Recall | 6.58 | 6.25 | 3.94 | 4.3 | 5.96 | 8.29 |
The accuracy rate of 12 width pictures that table 5BOW and HSV are merged and other algorithms are gone forward in Corel-10k data sets
The accuracy rate result explanation of table 4 and 45, context of methods obtains the result of single characterization method using the information of image sets
Larger lifting is arrived.Context of methods for preceding k attach pictures accuracy rate lifting and differ, when k is close to 50 in Corel numbers
It is most obvious according to the effect on collection, on Corel-1k data sets, 15% accuracy rate can be improved, in Corel-10k data sets
On, 9% accuracy rate can be improved.For the method for multiple features fusion, herein can be on Corel-1k data sets, will be accurate
Rate brings up to 88.44%. on Corel-10k data sets, accuracy rate is brought up into 69.08%. it can be found that fusion feature it
Between correlation it is smaller then merge after retrieval effectiveness it is better.
3rd, summarize
Herein using the information of global candidate's picture, the ranking and fusing method of quick figure fusion is proposed.The method is not using
Same characteristic pattern, image retrieval data set is divided into similar pictures group to lift the effect of image retrieval.Can be with from experimental result
Find out, this paper multi-characteristics fusion method is better than existing method；Additionally, the part and global characteristics that pass through fused images, energy
The enough more fully information such as the color of picture engraving, yardstick and rotation, can more former feature (global or local feature) obtain more
Precision high, and the effect of former feature is more excellent, the retrieval effectiveness of multi-characteristic fusion is better.When all features are to the picture retrieval
When result has relatively large deviation, context of methods can still lift the effect of image retrieval.
Finally it should be noted that：Various embodiments above is merely illustrative of the technical solution of the present invention, rather than its limitations；To the greatest extent
Pipe has been described in detail with reference to foregoing embodiments to the present invention, it will be understood by those within the art that：Its is right
Technical scheme described in foregoing embodiments is modified, or which part or all technical characteristic are equally replaced
Change, do not make the scope of the essence disengaging various embodiments of the present invention technical scheme of appropriate technical solution.
Claims (3)
1. a kind of quick figure fused images search method, comprises the following steps：
Step 100：By image retrieval data set, the ranking results of feature are obtained；
Step 200：Weight map is set up, weight map interior joint is the picture in data acquisition system, between node between weight and image
Similarity be directly proportional, for different characteristic weight map using figure fusion method；
Step 300：In weight map, centered on image to be retrieved according to set forth herein clustering method carry out image sets draw
Point；
Step 400：The order rearrangement candidate image of image sets is added according to candidate image.
2. a kind of quick figure fused images search method according to claim 1, it is characterised in that the step 200
Detailed process is：Each picture that the picture and data that will retrieve are concentrated is considered as different nodes, and according to feature is extracted after
The similarity of picture sets up a line between each node, and side right determines that similarity is higher by the size of picture similarity
Picture, the side right between node is bigger, otherwise then smaller, it is also possible to be zero；Rule according to more than sets up a characteristic pattern,
Can represent per pictures similarity relation.
3. a kind of quick figure fused images search method according to claim 1 and 2, it is characterised in that the step
300 include following sub-step：
Step 310：Using formula：G1=(V1, E1, w1)+G2=(V2, E2, w2(V, E w) carry out figure fusion to)=G=；Wherein,
G1：It is ith feature figure, V1：It is the set of the node in figure, E：It is the set on side in figure, w：It is side right function in figure；
Step 320：Using formula：Clustered,
Wherein, q is retrieving image,Preceding k pictures are expressed asV is the section in the figure asked
Point, it is side right function in figure that w ' is,To the side right size of v；
When finding+1 candidate's picture of kth every time, as long as the larger picture of k image correlation before finding.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611078791.5A CN106776896A (en) | 2016-11-30 | 2016-11-30 | A kind of quick figure fused images search method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611078791.5A CN106776896A (en) | 2016-11-30 | 2016-11-30 | A kind of quick figure fused images search method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106776896A true CN106776896A (en) | 2017-05-31 |
Family
ID=58898913
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611078791.5A Pending CN106776896A (en) | 2016-11-30 | 2016-11-30 | A kind of quick figure fused images search method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106776896A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107506401A (en) * | 2017-08-02 | 2017-12-22 | 大连理工大学 | A kind of image retrieval rearrangement method based on drawing method |
| CN108009191A (en) * | 2017-09-24 | 2018-05-08 | 贵州师范学院 | A kind of image search method |
| CN108764245A (en) * | 2018-04-03 | 2018-11-06 | 南昌奇眸科技有限公司 | A method of it improving brand logo similarity and judges accuracy |
| CN108932950A (en) * | 2018-05-18 | 2018-12-04 | 华南师范大学 | It is a kind of based on the tag amplified sound scenery recognition methods merged with multifrequency spectrogram |
| CN109215091A (en) * | 2018-08-02 | 2019-01-15 | 浙江理工大学 | A kind of Fashionable Colors of Garment extraction method indicated based on figure |
| CN109284411A (en) * | 2017-07-19 | 2019-01-29 | 哈尔滨工业大学深圳研究生院 | One kind being based on having supervision hypergraph discretized image binary-coding method |
| CN109711352A (en) * | 2018-12-28 | 2019-05-03 | 中国地质大学（武汉） | Vehicle front road environment based on geometry convolutional neural networks has an X-rayed cognitive method |
| CN110134811A (en) * | 2019-05-20 | 2019-08-16 | 清华大学深圳研究生院 | Single picture geographic positioning based on broad sense weight limit figure |
| CN110162654A (en) * | 2018-01-30 | 2019-08-23 | 西安邮电大学 | It is a kind of that image retrieval algorithm is surveyed based on fusion feature and showing for search result optimization |
| WO2020102881A1 (en) * | 2018-11-20 | 2020-05-28 | The Toronto-Dominion Bank | Automated image retrieval with image graph |
| CN111950545A (en) * | 2020-07-23 | 2020-11-17 | 南京大学 | Scene text detection method based on MSNDET and space division |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103678480A (en) * | 2013-10-11 | 2014-03-26 | 北京工业大学 | Personalized image retrieval method with privacy controlled in grading mode |
| CN104156423A (en) * | 2014-08-06 | 2014-11-19 | 聂秀山 | Multiscale video key frame extraction method based on integer programming |
-
2016
- 2016-11-30 CN CN201611078791.5A patent/CN106776896A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103678480A (en) * | 2013-10-11 | 2014-03-26 | 北京工业大学 | Personalized image retrieval method with privacy controlled in grading mode |
| CN104156423A (en) * | 2014-08-06 | 2014-11-19 | 聂秀山 | Multiscale video key frame extraction method based on integer programming |
Non-Patent Citations (2)
| Title |
|---|
| ZIQIONG LIU等: "（"Visual reranking with improved image graph"", 《2014 IEEE INTERNATIONAL CONFERENCE ON ACOUSTIC。SPEECH AND SIGNAL PROCESSING（ICASSP）》 * |
| 刘胜蓝 等: "分组排序多特征融合的图像检索方法", 《计算机研究与发展》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109284411A (en) * | 2017-07-19 | 2019-01-29 | 哈尔滨工业大学深圳研究生院 | One kind being based on having supervision hypergraph discretized image binary-coding method |
| CN107506401A (en) * | 2017-08-02 | 2017-12-22 | 大连理工大学 | A kind of image retrieval rearrangement method based on drawing method |
| CN108009191A (en) * | 2017-09-24 | 2018-05-08 | 贵州师范学院 | A kind of image search method |
| CN110162654A (en) * | 2018-01-30 | 2019-08-23 | 西安邮电大学 | It is a kind of that image retrieval algorithm is surveyed based on fusion feature and showing for search result optimization |
| CN108764245A (en) * | 2018-04-03 | 2018-11-06 | 南昌奇眸科技有限公司 | A method of it improving brand logo similarity and judges accuracy |
| CN108764245B (en) * | 2018-04-03 | 2022-04-29 | 南昌奇眸科技有限公司 | Method for improving similarity judgment accuracy of trademark graphs |
| CN108932950B (en) * | 2018-05-18 | 2021-07-09 | 华南师范大学 | Sound scene identification method based on label amplification and multi-spectral diagram fusion |
| CN108932950A (en) * | 2018-05-18 | 2018-12-04 | 华南师范大学 | It is a kind of based on the tag amplified sound scenery recognition methods merged with multifrequency spectrogram |
| CN109215091A (en) * | 2018-08-02 | 2019-01-15 | 浙江理工大学 | A kind of Fashionable Colors of Garment extraction method indicated based on figure |
| CN109215091B (en) * | 2018-08-02 | 2021-09-14 | 浙江理工大学 | Clothing fashion color automatic extraction method based on graph representation |
| WO2020102881A1 (en) * | 2018-11-20 | 2020-05-28 | The Toronto-Dominion Bank | Automated image retrieval with image graph |
| US11397765B2 (en) | 2018-11-20 | 2022-07-26 | The Toronto-Dominion Bank | Automated image retrieval with image graph |
| CN109711352A (en) * | 2018-12-28 | 2019-05-03 | 中国地质大学（武汉） | Vehicle front road environment based on geometry convolutional neural networks has an X-rayed cognitive method |
| CN110134811A (en) * | 2019-05-20 | 2019-08-16 | 清华大学深圳研究生院 | Single picture geographic positioning based on broad sense weight limit figure |
| CN110134811B (en) * | 2019-05-20 | 2021-01-15 | 清华大学深圳研究生院 | Single picture geographical positioning method based on generalized maximum weight graph |
| CN111950545A (en) * | 2020-07-23 | 2020-11-17 | 南京大学 | Scene text detection method based on MSNDET and space division |
| CN111950545B (en) * | 2020-07-23 | 2024-02-09 | 南京大学 | Scene text detection method based on MSDNet and space division |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106776896A (en) | A kind of quick figure fused images search method | |
| CN106126581B (en) | Cartographical sketching image search method based on deep learning | |
| Lu et al. | Object detection based on SSD-ResNet | |
| Hu et al. | Recognition of pornographic web pages by classifying texts and images | |
| CN109241317B (en) | Pedestrian Hash retrieval method based on measurement loss in deep learning network | |
| CN106649487B (en) | Image retrieval method based on interest target | |
| Liu | An improved faster R-CNN for object detection | |
| Qi et al. | Making better use of edges via perceptual grouping | |
| CN102054178B (en) | A kind of image of Chinese Painting recognition methods based on local semantic concept | |
| CN108280187B (en) | Hierarchical image retrieval method based on depth features of convolutional neural network | |
| Zamberletti et al. | Text localization based on fast feature pyramids and multi-resolution maximally stable extremal regions | |
| KR101443187B1 (en) | medical image retrieval method based on image clustering | |
| Li et al. | A generative/discriminative learning algorithm for image classification | |
| CN102289522A (en) | Method of intelligently classifying texts | |
| Niu et al. | A novel image retrieval method based on multi-features fusion | |
| CN107766860A (en) | Natural scene image Method for text detection based on concatenated convolutional neutral net | |
| Fu et al. | Chinfood1000: A large benchmark dataset for chinese food recognition | |
| CN106845513A (en) | Staff detector and method based on condition random forest | |
| Zhang et al. | Character proposal network for robust text extraction | |
| CN110414587A (en) | Depth convolutional neural networks training method and system based on progressive learning | |
| CN107577994A (en) | A kind of pedestrian based on deep learning, the identification of vehicle auxiliary product and search method | |
| Shalunts | Architectural style classification of building facade towers | |
| CN104834891A (en) | Method and system for filtering Chinese character image type spam | |
| Akhlaghi et al. | Farsi handwritten phone number recognition using deep learning | |
| Luqman et al. | Subgraph spotting through explicit graph embedding: An application to content spotting in graphic document images |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170531 |