CN107085607A - A kind of image characteristic point matching method - Google Patents
A kind of image characteristic point matching method Download PDFInfo
- Publication number
- CN107085607A CN107085607A CN201710258205.3A CN201710258205A CN107085607A CN 107085607 A CN107085607 A CN 107085607A CN 201710258205 A CN201710258205 A CN 201710258205A CN 107085607 A CN107085607 A CN 107085607A
- Authority
- CN
- China
- Prior art keywords
- vector
- matched
- cluster centre
- characteristic
- storage
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/51—Indexing; Data structures therefor; Storage structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Library & Information Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Probability & Statistics with Applications (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of image characteristic point matching method, comprise the following steps, storage picture feature point is extracted：Extract the feature of storage image and constitute storage characteristic vector, dimensionality reduction is carried out to its dimension；Vector storage：Split the storage characteristic vector after dimensionality reduction, and quantify to do vector quantization again to first doing product after segmentation per part, constitute product quantizer, vector quantizer, and set up trie tree and Hash table；Picture feature point to be matched is extracted：Extract the feature of image to be matched and constitute characteristic vector to be matched, dimensionality reduction is carried out to its dimension；Vectors matching：Split the characteristic vector to be matched after dimensionality reduction, characteristic vector to be matched is found out with product quantizer, the cluster centre of vector quantizer apart from forward multiple cluster centres, picture according to corresponding to trie tree and Hash table find multiple cluster centres simultaneously constitutes Candidate Set, and the closest picture in Candidate Set with characteristic vector to be matched is calculated using floating point vector；Its speed is fast and precision is high.
Description
Technical field
The present invention relates to picture searching technical field, and in particular to a kind of image characteristic point matching method.
Background technology
In picture search field, characteristic matching is a very important link, and the matching efficiency and precision of feature are determined
Final search speed and precision.During existing picture searching, it uses following steps：The first step is trained by great amount of samples data
One transition matrix, binary code is turned by hash function, and binary code segmentation generates multiple Hash tables, obtained segmentation two
Entrance of the ary codes directly as Hash table.Second step, when vector to be checked reaches, switchs to binary system by same mode
Code, is mapped to corresponding Hash table entry and its in other entrances for r, all pictures in entrance are used as Candidate Set.
3rd step, does all in Candidate Set picture features are vectorial complete Hamming distances with vector to be checked and calculates, reset distance
.When floating point features vector turns binary code, due to the presence of hash function, the precision of vector is caused to be lost, and
Among last rearrangement process, still use the Hamming distances based on binary code to calculate, although speed quickly, but by
It is not so good as floating point vector in the expression precision of binary code, therefore calling rate has and reduced to a certain degree.
With the rapid development of Internet, the picture on internet has reached 10,000,000,000 rank even more highs at present.Adopt
Existing characteristic point matching method is used, it has not adapted to the picture library pattern of existing rapid growth.In mass picture retrieval
Efficient, high-precision search how is carried out in tree becomes focus.
The content of the invention
In order to solve the above-mentioned technical problem the present invention provides a kind of image characteristic point matching method, and its matching precision is high and fast
Degree is fast.
The present invention is achieved through the following technical solutions：
A kind of image characteristic point matching method, comprises the following steps,
Picture feature point is put in storage to extract：Extract the feature of storage image and constitute storage characteristic vector, its dimension is carried out
Dimensionality reduction；
Vector storage：Split the storage characteristic vector after dimensionality reduction, and quantify to do again to first doing product after segmentation per part
Vector quantization, constitutes product quantizer, vector quantizer, and set up trie tree and Hash table；
Picture feature point to be matched is extracted：Extract the feature of image to be matched and constitute characteristic vector to be matched, it is tieed up
Degree carries out dimensionality reduction；
Vectors matching：Split the characteristic vector to be matched after dimensionality reduction, find out characteristic vector to be matched and product quantizer, to
The cluster centre of quantizer is measured apart from forward multiple cluster centres, multiple cluster centre institutes are found according to trie tree and Hash table
Corresponding picture simultaneously constitutes Candidate Set, and the closest figure in Candidate Set with characteristic vector to be matched is calculated using floating point vector
Piece.
The method of this programme calculates binary code without using the algorithm of similar iterative quantization, and retrieval is built using dimensionality reduction cluster
Tree and Hash table, in the cluster of the first level, the data perfectly correlated to its are not clustered, but it is split,
Each several part data are completely independent, and can be accelerated using multi-threading parallel process mode, the training time of quantizer is subtracted significantly
It is few.The whole process of matching retrieval is divided to two sections of progress, and first paragraph is chooses Candidate Set, and second segment is integrally carried out using floating point vector
Distance is calculated, on the premise of Candidate Set scope is larger, then carries out floating-point distance calculating, the recall rate of retrieval result and violence
The gap very little matched somebody with somebody, no more than 1 percentage point.Sorted than Hamming distances more accurate.
If there is N bars record in database, violence matching needs to do n times distance calculating, and using the method for this programme, root
According to the difference of selected parameter, the record number in Candidate Set is N/100~N/10, greatly reduces and calculates, and is substantially increased
With speed.And during trie tree is built, the first stage data of cluster are divided into some, the cluster of some
Process is completely self-contained, therefore multithreading can be used to be clustered, and improves cluster speed.
Preferably, the method for vector storage is：
Storage characteristic vector is divided into disjoint P part；
It is k to carry out cluster centre number in each partial interior1K-means cluster；
For each cluster centre, all data for being assigned to the cluster centre are done into vector quantization, cluster centre number
For k2；
Record the ID for all features for being mapped to correspondence cluster centre or the title of correspondence picture respectively with P Hash table.
Further, the specific method of Vectors matching is：
Characteristic vector to be matched is divided into disjoint P part；
In each partial interior, characteristic vector to be matched and k are calculated1The distance of individual cluster centre, and chosen distance is minimum
W cluster centre；
For W cluster centre of selection, by characteristic vector to be matched k corresponding with the cluster centre2Individual second strata
Row distance calculating is entered at class center one by one, obtains k2Individual distance；
To W*k2Individual distance is ranked up, and takes m closest distance, wherein, m is the natural number more than 1；
M are taken out apart from corresponding cluster centre, corresponding Hash table entry is found, by the picture name in these entrances
Or
ID constitutes Candidate Set；
By the corresponding picture feature of Image ID in Candidate Set it is vectorial with characteristic vector to be matched using floating point vector one by one
Calculate into distance, finally obtain the minimum as target of distance.
Further, the k-means clusters use parallel processing manner.
Preferably, using dimension dimensionality reduction of the principal component analytical method to feature.
The present invention compared with prior art, has the following advantages and advantages：
The present invention builds trie tree and Hash table using dimensionality reduction cluster, and in the cluster of the first level, it is split,
Each several part data are completely independent, and can be accelerated using multi-threading parallel process mode, the training time of quantizer is subtracted significantly
It is few；The process of matching retrieval first chooses Candidate Set, reuses floating point vector and integrally enters row distance calculating, larger in Candidate Set scope
On the premise of, carry out floating-point distance and calculate, the gap very little of the recall rate of retrieval result and violence matching, no more than 1 percentage
Point, its retrieval precision is high and efficient.
Embodiment
For the object, technical solutions and advantages of the present invention are more clearly understood, with reference to embodiment, to present invention work
Further to describe in detail, exemplary embodiment and its explanation of the invention is only used for explaining the present invention, is not intended as to this
The restriction of invention.
Embodiment 1
A kind of image characteristic point matching method, comprises the following steps,
Picture feature point is put in storage to extract：Extract the feature of storage image and constitute storage characteristic vector, its dimension is carried out
Dimensionality reduction；
Vector storage：Split the storage characteristic vector after dimensionality reduction, and quantify to do again to first doing product after segmentation per part
Vector quantization, constitutes product quantizer, vector quantizer, and set up trie tree and Hash table；
Picture feature point to be matched is extracted：Extract the feature of image to be matched and constitute characteristic vector to be matched, it is tieed up
Degree carries out dimensionality reduction；
Vectors matching：Split the characteristic vector to be matched after dimensionality reduction, find out characteristic vector to be matched and product quantizer, to
The cluster centre of quantizer is measured apart from forward multiple cluster centres, multiple cluster centre institutes are found according to trie tree and Hash table
Corresponding picture simultaneously constitutes Candidate Set, and the closest figure in Candidate Set with characteristic vector to be matched is calculated using floating point vector
Piece.
Using this method, in retrieving, not only allow for the Hash entrance hit completely, it is also contemplated that near it
Several entrances, therefore, increase the probability in Candidate Set with the inquiry most like picture of picture, improve recall rate.Carrying out
When reordering, original floating-point characteristic vector has been used to enter row distance calculating, it is complete that this mode remains original floating-point characteristic vector
The information in portion, and passing through the mode for quantifying to turn into binary features code has a certain degree of loss of significance.
Embodiment 2
Thought based on above-described embodiment, the present embodiment is refined to each step.
The picture be put in storage as trie tree is either directed to also in picture to be matched, is required to extract characteristic point, it is special
Levy extracting method a little a lot, for example convolutional neural networks, the dimension of its output characteristic vector be one than larger value, can set
It is set to n, n may be 128,256 or 512 etc..Its dimension is larger, can increase the amount of calculation of matching process, it would be desirable to it
Dimensionality reduction is carried out, the dimension of the characteristic vector of output can be reduced to d using principal component analytical method and tieed up by dimensionality reduction, wherein, d is less than or equal to
N, d desirable 128 or 64.The influence of noise not only can remove using the method for dimensionality reduction, and amount of calculation can be reduced, when reducing calculating
Between.
Dimensionality reduction detailed process is as follows：
Assuming that there is S datas, in luv space, S bar n dimensional feature vectors can be by matrix M={ D1,D2,…DnRepresent,
Wherein DnFor S*1 column vector.First, matrix M asks covariance to obtain matrix V ar (M), Var (M)=MTM*1/n, is recycled existing
There is method to ask covariance matrix Var (M) n characteristic value and corresponding characteristic vector, choose maximum d characteristic value and its
Corresponding vector, as the matrix R of dimensionality reduction, calculating MR just can obtain L*d matrix, complete the process of dimensionality reduction.
It is that vector storage structure trie tree is prepared after the feature and dimensionality reduction of extraction storage picture, the specific side of vector storage
Method is：
Storage characteristic vector is divided into disjoint P part, P=4 portions for example are cut into the d=128 data tieed up
Point, using the 1st~32 floating number of characteristic vector as Part I, the 33rd~64 floating number is used as Part II, the 65th
~96 floating numbers are used as Part IV as Part III, the 97th~128 floating number.
It is k to carry out cluster centre number in each partial interior1K-means clusters, be that first layer quantifies herein, with the
Illustrate its detailed process exemplified by a part：
1-1. first chooses k for 32 floating number features of S bars of Part I from S bar features1Bar is as in cluster
The heart；
1-2. is by S bar characteristics respectively at k1Individual cluster centre carries out Euclidean distance calculating, and which characteristic gather with
Class center it is closest, the data is just belonged into the cluster centre；
1-3. will belong to the characteristic of the cluster centre for each cluster centre in previous step, corresponding each
Floating number summation takes average, regard the mean vector finally obtained as new k1Individual cluster centre；
1-4. terminates the k obtained in cluster, step 3 if cluster number of times or cluster error are reduced to certain limit1It is individual
Cluster centre is required, otherwise, returns to step 2.
1-5. records k1The value of individual cluster centre, needs during retrieval.
On the basis of first layer quantization, for each cluster centre, all data for being assigned to the cluster centre are done
Vector quantization, cluster centre number is k2, quantify herein for the second layer, its detailed process is as follows：
2-1. quantifies each cluster centre for first layer, and the feature for belonging to certain cluster centre is SiIt is individual, wherein all Si's
With for S, from SiMiddle selection k2Bar is used as cluster centre；
2-2. is by SiBar characteristic is respectively at k2Individual cluster centre carries out Euclidean distance calculating, and which characteristic gather with
Class center it is closest, the data is just belonged into the cluster centre；
2-3. will belong to the characteristic of the cluster centre for each cluster centre in previous step, corresponding each
Floating number summation takes average, regard the mean vector finally obtained as new k2Individual cluster centre；
2-4. terminates the k obtained in cluster, step 3 if cluster number of times or cluster error are reduced to certain limit2It is individual
Cluster centre is required, otherwise, returns to step 2；
2-5. records k1The value of individual cluster centre, needs during retrieval.
Record the ID for all features for being mapped to correspondence cluster centre or the title of correspondence picture respectively with P Hash table,
The Hash codes of each Hash table are log2 (k1)+log2(k2) position.For example：
1. assume k1=16, k2=16, then Hash codes are 4+4=8,16 cluster centres in preceding 4 expressions first layer
One, rear 4 expressions correspond to one of 16 two layers of cluster centres of first layer cluster centre.
2. correspondence is each mapped to the characteristic of second layer cluster centre, its corresponding picture name or ID are added to
In the corresponding Hash entrance of Hash codes.Hash herein is simply encoded for cluster centre, is a kind of data structure, for depositing
Storage.
The essence of vector storage is laid the foundation for Vectors matching, after trie tree, Hash table are put up, you can enter row vector
Matching.If there is picture input to be matched, extract start Vectors matching step after picture feature point to be matched as stated above
Suddenly, the specific method of Vectors matching is：
Characteristic vector to be matched is divided into disjoint P part；When dividing method is with storage as dividing method.
In each partial interior, characteristic vector to be matched and k are calculated1The distance of individual cluster centre, sorts and chosen distance
W minimum cluster centre, with P=4, k1=16, k2Said exemplified by=16, W=4, first vectorial part to be checked
It is bright,
16 cluster centres of first part during to storage, with first vectorial part to be checked, i.e., 1~32
Floating-point number vector and 16 cluster centres carry out the calculating of Euclidean distance.16 distances are ranked up, minimum W=4 is chosen
Individual cluster centre.
For W cluster centre of selection, by characteristic vector to be matched k corresponding with the cluster centre2Individual second strata
Row distance calculating is entered at class center one by one, obtains k2Individual distance, specifically, to the cluster centre of individual selection, inside it, with to be checked
16 cluster centres of the 32 floating-point number vectors and the second hierarchical clustering of asking vector carry out Euclidean distance calculating.
By W*k2Individual distance is sequentially placed into big top heap, it is ensured that only m distance in big top heap, wherein, m is more than 1
Natural number, m rather than one before why taking are to ensure recall rate.
Take out m in big pushing tow individual apart from corresponding cluster centre, find the figure in corresponding Hash table entry, these entrances
Piece title or ID constitute Candidate Set.For m apart from corresponding cluster centre, once encoded during storage for each cluster centre
Hash codes, the sole inlet of each Hash codes correspondence Hash table, ask union to can obtain Candidate Set the ID in m entrance.
Enter the corresponding picture feature of Image ID in Candidate Set is vectorial row distance with characteristic vector to be matched one by one and calculate,
The minimum as target of distance is finally obtained, is specially：
For each ID in Candidate Set, obtain its 128 complete dimension floating point features vector, with vector sum to be checked these
Floating point features vector enters row distance calculating one by one；
K minimum result in selecting step 2, its corresponding ID is required most like picture.When K is 1, it is defined
Really search, works as K>It is k nearest neighbor search when 1.
Embodiment 3
For embodiment 2, a detailed embodiment is now disclosed.
Storage picture feature point is extracted and picture feature point extraction step to be matched is not repeated in this embodiment.
Vector storage：Using a large amount of characteristic vectors in database, the training of product quantizer and vector quantizer is carried out.
Step is as follows：
D dimension storage characteristic vectors after dimensionality reduction are divided into disjoint P part, by taking D=128, P=4 as an example, by spy
The 1st~32 of vector is levied as first paragraph, the 33rd~64 as second segment, the 65th~96 as the 3rd section, the 97th~
128 are used as the 4th section.
It is k to carry out cluster centre number in each partial interior1K-means cluster, common P*k1Individual cluster centre, owns
Cluster centre is [C1 i]p={ [c1 i]p, i=0,1,2 ..., k1；P=0,1,2 ..., P }, by cluster centre [C1 i]pDeposited
Storage, due to being segmented in previous step to feature, so the space consuming of storage is only D/P*k1, this step is referred to as
PQ, i.e. product quantify, and also quantify as first layer, obtain corresponding PQ quantizers.Also, due to the independence of various pieces,
This process can use multithreading, and multi-process even multinode carries out parallel processing, lifting cluster speed.
On the basis of first layer quantization, cluster centre c is arrived to all clusters1 ijData, clustered again, generate
k2Individual cluster centre, common P*k1*k2Individual cluster centre, all cluster centres are [C2 ij]p={ [c2 ij]p, i=0,1,2,3 ...,
k2；J=0,1,2,3 ... k1；P=0,1,2 ..., P }, cluster centre is stored, this step is referred to as VQ, i.e. vector quantization,
Also turn into the second layer to quantify, corresponding VQ quantizers can be obtained.
P Hash table is set up, the vector set of the P part with being separated in the first step is corresponded, one in each Hash table
AltogetherIndividual, the length of each Hash table correspondence Hash codes is alsoFor each part
In k1*k2Individual cluster centre is encoded, and will be mapped in sample data the characteristic vector of each cluster centre ID or
The name of correspondence picture is stored in the entrance in the corresponding Hash table of cluster centre, obtains the row's of falling rope based on many Hash tables
Draw.
After four steps of the above are completed, the PQ quantizers based on great amount of samples data, VQ quantizers, retrieval can be obtained
Tree and the inverted index structure based on many Hash tables.
After the completion of trie tree and Hash table are set up, if there is picture input to be matched, enter Vectors matching retrieval.
When give a characteristic vector y to be matched when, retrieved from trie tree and inverted index therewith it is closest to
Amount, step is as follows：
Vectorial y is divided into P intersecting part, y=[y1,y2,y3,…,yp]。
For yp, its k obtained with the quantization of pth segments first layer is calculated one by one1The distance of individual cluster centre.Define dist
(yp,[c1 i]p)=| | yp–[c1 i]p||2For in vectorial y to be checked pth part and the ith cluster of sample data pth part
The distance of the heart.
Due to the uncertainty of cluster process, with some closest y feature in sample space, it is more likely that ownership
In other neighbouring cluster centres, so to consider from ypOther cluster centres around nearest cluster centre.For upper one
The y obtained in steppReordered with the distance of each cluster centre, choose wherein w minimum cluster centre of distance, as
The scope for the inquiry to be carried out below.
The w first layer cluster centre minimum to the distance selected, to w cluster centre, has under each cluster centre
k2Individual second layer cluster centre, defines dist (yp,[c1 ij]p)=| | yp–[c1 ij]p||2For vectorial y to be checked pth part with
The distance between j-th of second layer cluster centre in sample space under the ith cluster center of first layer.
The w*k obtained to step 42Individual distance is ranked up, due to this distance simply to sample space pth part away from
From, so closest one can not only be taken, before taking m apart from corresponding cluster centre, found pair according to cluster centre
The Hash table entry answered, seeks union by the Image ID or name that are stored in each Hash table entries of m, finally gives most like picture
ID Candidate Set.
The corresponding characteristic vector of Image ID is taken out, row distance calculating is once entered with complete vectorial y to be checked, using small
This data structure of heap is pushed up, no matter last data set has much, can be worked well, and committed memory amount is also pole
Small.Finally obtain the topK of K as search result before distance-taxis, as K=1, as most like picture.
The present embodiment is quantified using product, and vector quantization, many hash indexs solve the problems, such as nearest neighbor search, utilize cluster
Parallel computation process and retrieving two-part divide raising and retrieved recall rate.
Above-described embodiment, has been carried out further to the purpose of the present invention, technical scheme and beneficial effect
Describe in detail, should be understood that the embodiment that the foregoing is only the present invention, be not intended to limit the present invention
Protection domain, within the spirit and principles of the invention, any modification, equivalent substitution and improvements done etc. all should be included
Within protection scope of the present invention.
Claims (6)
1. a kind of image characteristic point matching method, it is characterised in that comprise the following steps,
Picture feature point is put in storage to extract：Extract the feature of storage image and constitute storage characteristic vector, dimensionality reduction is carried out to its dimension；
Vector storage：Split the storage characteristic vector after dimensionality reduction, and quantify to do vector quantization again to first doing product after segmentation per part,
Product quantizer, vector quantizer are constituted, and sets up trie tree and Hash table；
Picture feature point to be matched is extracted：Extract the feature of image to be matched and constitute characteristic vector to be matched, its dimension is entered
Row dimensionality reduction；
Vectors matching：Split the characteristic vector to be matched after dimensionality reduction, find out characteristic vector to be matched and product quantizer, vector quantity
Change the cluster centre of device apart from forward multiple cluster centres, according to corresponding to trie tree and Hash table find multiple cluster centres
Picture and constitute Candidate Set, closest picture with characteristic vector to be matched in Candidate Set is calculated using floating point vector.
2. a kind of image characteristic point matching method according to claim 1, it is characterised in that：It is specific that the vector is put in storage
Method is：
Storage characteristic vector is divided into disjoint P part；
It is k to carry out cluster centre number in each partial interior1K-means cluster；
For each cluster centre, all data for being assigned to the cluster centre are done into vector quantization, cluster centre number is k2；
Record the ID for all features for being mapped to correspondence cluster centre or the title of correspondence picture respectively with P Hash table.
3. a kind of image characteristic point matching method according to claim 2, it is characterised in that：The Vectors matching it is specific
Method is：
Characteristic vector to be matched is divided into disjoint P part；
In each partial interior, characteristic vector to be matched and k are calculated1The distance of individual cluster centre, and chosen distance minimum W
Cluster centre；
For W cluster centre of selection, by characteristic vector to be matched k corresponding with the cluster centre2Individual second layer cluster centre
Enter row distance calculating one by one, obtain k2Individual distance；
To W*k2Individual distance is ranked up, and takes m closest distance, wherein, m is the natural number more than 1；
M is taken out apart from corresponding cluster centre, corresponding Hash table entry is found, by the picture name in these entrances or
ID constitutes Candidate Set；
Calculated the corresponding picture feature of Image ID in Candidate Set is vectorial one by one using floating point vector with characteristic vector to be matched
Enter distance, finally obtain the minimum as target of distance.
4. a kind of image characteristic point matching method according to claim 2, it is characterised in that：The k-means clusters are adopted
Use parallel processing manner.
5. a kind of image characteristic point matching method according to claim 1, it is characterised in that：Using principal component analytical method
To the dimension dimensionality reduction of feature.
6. a kind of image characteristic point matching method according to claim 1, it is characterised in that：The specific steps of the dimensionality reduction
For：
Matrix M is constituted using L bar n dimensional feature vectors data, asks matrix M covariance to obtain matrix V ar (M), wherein, matrix M
={ D1,D2,…Dn, n is characterized the dimension of vector, and L is the natural number more than 1；
Ask covariance matrix Var (M) n characteristic value and corresponding characteristic vector, and choose maximum d characteristic value and its
Corresponding vector, is used as the matrix R of dimensionality reduction；
MR is calculated with regard to L*d matrix can be obtained, dimensionality reduction is realized.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710258205.3A CN107085607B (en) | 2017-04-19 | 2017-04-19 | Image feature point matching method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710258205.3A CN107085607B (en) | 2017-04-19 | 2017-04-19 | Image feature point matching method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107085607A true CN107085607A (en) | 2017-08-22 |
| CN107085607B CN107085607B (en) | 2020-06-30 |
Family
ID=59611717
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710258205.3A Active CN107085607B (en) | 2017-04-19 | 2017-04-19 | Image feature point matching method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107085607B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107944046A (en) * | 2017-12-15 | 2018-04-20 | 清华大学 | Extensive high dimensional data method for quickly retrieving and system |
| CN108763481A (en) * | 2018-05-29 | 2018-11-06 | 清华大学深圳研究生院 | A kind of picture geographic positioning and system based on extensive streetscape data |
| CN110019096A (en) * | 2017-12-29 | 2019-07-16 | 上海全土豆文化传播有限公司 | The generation method and device of index file |
| CN110019905A (en) * | 2017-10-13 | 2019-07-16 | 北京京东尚科信息技术有限公司 | Information output method and device |
| CN110019907A (en) * | 2017-12-01 | 2019-07-16 | 北京搜狗科技发展有限公司 | A kind of image search method and device |
| CN110019875A (en) * | 2017-12-29 | 2019-07-16 | 上海全土豆文化传播有限公司 | The generation method and device of index file |
| CN110175546A (en) * | 2019-05-15 | 2019-08-27 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN110795582A (en) * | 2019-10-31 | 2020-02-14 | 京东方科技集团股份有限公司 | Image recommendation method, system, terminal device and server |
| CN110889424A (en) * | 2018-09-11 | 2020-03-17 | 阿里巴巴集团控股有限公司 | Vector index establishing method and device and vector retrieval method and device |
| CN111324760A (en) * | 2020-02-19 | 2020-06-23 | 名创优品（横琴）企业管理有限公司 | Image retrieval method and device |
| CN112418298A (en) * | 2020-11-19 | 2021-02-26 | 北京云从科技有限公司 | Data retrieval method, device and computer readable storage medium |
| CN112988747A (en) * | 2021-03-12 | 2021-06-18 | 山东英信计算机技术有限公司 | Data retrieval method and system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102842032A (en) * | 2012-07-18 | 2012-12-26 | 郑州金惠计算机***工程有限公司 | Method for recognizing pornography images on mobile Internet based on multi-mode combinational strategy |
| US9436758B1 (en) * | 2011-12-27 | 2016-09-06 | Google Inc. | Methods and systems for partitioning documents having customer feedback and support content |
-
2017
- 2017-04-19 CN CN201710258205.3A patent/CN107085607B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9436758B1 (en) * | 2011-12-27 | 2016-09-06 | Google Inc. | Methods and systems for partitioning documents having customer feedback and support content |
| CN102842032A (en) * | 2012-07-18 | 2012-12-26 | 郑州金惠计算机***工程有限公司 | Method for recognizing pornography images on mobile Internet based on multi-mode combinational strategy |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110019905A (en) * | 2017-10-13 | 2019-07-16 | 北京京东尚科信息技术有限公司 | Information output method and device |
| CN110019905B (en) * | 2017-10-13 | 2022-02-01 | 北京京东尚科信息技术有限公司 | Information output method and device |
| CN110019907B (en) * | 2017-12-01 | 2021-07-16 | 北京搜狗科技发展有限公司 | Image retrieval method and device |
| CN110019907A (en) * | 2017-12-01 | 2019-07-16 | 北京搜狗科技发展有限公司 | A kind of image search method and device |
| CN107944046B (en) * | 2017-12-15 | 2019-02-05 | 清华大学 | Extensive high dimensional data method for quickly retrieving and system |
| CN107944046A (en) * | 2017-12-15 | 2018-04-20 | 清华大学 | Extensive high dimensional data method for quickly retrieving and system |
| CN110019096A (en) * | 2017-12-29 | 2019-07-16 | 上海全土豆文化传播有限公司 | The generation method and device of index file |
| CN110019875A (en) * | 2017-12-29 | 2019-07-16 | 上海全土豆文化传播有限公司 | The generation method and device of index file |
| CN108763481A (en) * | 2018-05-29 | 2018-11-06 | 清华大学深圳研究生院 | A kind of picture geographic positioning and system based on extensive streetscape data |
| CN110889424A (en) * | 2018-09-11 | 2020-03-17 | 阿里巴巴集团控股有限公司 | Vector index establishing method and device and vector retrieval method and device |
| CN110889424B (en) * | 2018-09-11 | 2023-06-30 | 阿里巴巴集团控股有限公司 | Vector index establishing method and device and vector retrieving method and device |
| CN110175546A (en) * | 2019-05-15 | 2019-08-27 | 深圳市商汤科技有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN110795582A (en) * | 2019-10-31 | 2020-02-14 | 京东方科技集团股份有限公司 | Image recommendation method, system, terminal device and server |
| CN111324760A (en) * | 2020-02-19 | 2020-06-23 | 名创优品（横琴）企业管理有限公司 | Image retrieval method and device |
| CN111324760B (en) * | 2020-02-19 | 2023-09-26 | 创优数字科技(广东)有限公司 | Image retrieval method and device |
| CN112418298A (en) * | 2020-11-19 | 2021-02-26 | 北京云从科技有限公司 | Data retrieval method, device and computer readable storage medium |
| CN112988747A (en) * | 2021-03-12 | 2021-06-18 | 山东英信计算机技术有限公司 | Data retrieval method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107085607B (en) | 2020-06-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107085607A (en) | A kind of image characteristic point matching method | |
| CN109960737B (en) | Remote sensing image content retrieval method for semi-supervised depth confrontation self-coding Hash learning | |
| Tao et al. | Quality and efficiency in high dimensional nearest neighbor search | |
| CN108491430A (en) | It is a kind of based on the unsupervised Hash search method clustered to characteristic direction | |
| CN107291895B (en) | Quick hierarchical document query method | |
| CN102254015A (en) | Image retrieval method based on visual phrases | |
| CN109886406A (en) | A kind of complex convolution neural network compression method based on depth-compression | |
| CN109886349A (en) | A kind of user classification method based on multi-model fusion | |
| CN109376772A (en) | A kind of Combination power load forecasting method based on neural network model | |
| CN108388656A (en) | A kind of image searching method based on label correlation | |
| CN101324926B (en) | Method for selecting characteristic facing to complicated mode classification | |
| CN108763295A (en) | A kind of video approximate copy searching algorithm based on deep learning | |
| CN114118369B (en) | Image classification convolutional neural network design method based on group intelligent optimization | |
| CN107180079B (en) | Image retrieval method based on convolutional neural network and tree and hash combined index | |
| CN110738053A (en) | News theme recommendation algorithm based on semantic analysis and supervised learning model | |
| CN112434553A (en) | Video identification method and system based on deep dictionary learning | |
| CN113742292B (en) | Multithread data retrieval and access method of retrieved data based on AI technology | |
| CN106919658B (en) | A kind of large-scale image words tree search method and system accelerated based on GPU | |
| CN113222109A (en) | Internet of things edge algorithm based on multi-source heterogeneous data aggregation technology | |
| CN107704872A (en) | A kind of K means based on relatively most discrete dimension segmentation cluster initial center choosing method | |
| CN105653686A (en) | Domain name network address activeness statistics method and system | |
| CN115618039A (en) | Image retrieval method based on multi-label projection online Hash algorithm | |
| CN115329116A (en) | Image retrieval method based on multi-layer feature fusion | |
| CN114610941A (en) | Cultural relic image retrieval system based on comparison learning | |
| KR20110080966A (en) | An associative classification method for detecting useful knowledge from huge multi-attributes dataset |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |