CN106203449A - The approximation space clustering system of mobile cloud environment - Google Patents
The approximation space clustering system of mobile cloud environment Download PDFInfo
- Publication number
- CN106203449A CN106203449A CN201610539136.9A CN201610539136A CN106203449A CN 106203449 A CN106203449 A CN 106203449A CN 201610539136 A CN201610539136 A CN 201610539136A CN 106203449 A CN106203449 A CN 106203449A
- Authority
- CN
- China
- Prior art keywords
- point
- image
- key point
- cloud environment
- clustering system
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- 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
- G06F18/23211—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with adaptive number of clusters
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The approximation space clustering system of mobile cloud environment, belong to image processing field, for the problem solving image clustering, have technical point that and be mainly made up of cloud center service system, cloud center service system is set up and is performed SIFT feature and extracts and matching algorithm, extracts the correlated characteristic data of image, and mates, and cloud center service system also sets up and perform over-sampling correction algorithm to carry out image correction, feedback cluster approximate image information is to client.Effect is: this image clustering system can be uploaded by the corresponding information that correlation technique gets picture at user side and be saved in cloud server, and then cloud server processes, and obtains optimal image clustering scheme and feeds back to user.
Description
Technical field
Patent of the present invention belongs to field of mobile communication, is the approximation space clustering system of a kind of mobile cloud environment, this system
Relate to the mass data processing under large-scale data analysis, cloud computing environment, relate to data intelligence processing and application and development.
Background technology
Along with Internet technology develops rapidly with digitized, and electronic digital product is universal, and people can obtain
Digital Image Data got more and more.Image shows directly perceived, abundant in content multimedia messages as a kind of, each at each row
Industry is the most increasingly widely applied, as numeral multi-media book shop, medical image applications management, satellite remote sensing images and
GIS-Geographic Information System, authentication identification system, ecommerce, the supervision etc. of trade mark copyright.But, the image of explosive growth
Data are own through considerably beyond the receptible degree of people, for the view data of magnanimity, how can fast and effeciently enter
Line pipe reason and retrieval, then depending from obtaining potential valuable information becomes problem of concern.And retrieval time is figure
As one of most critical issue of retrieval, traditional search method is the sample image to be checked provided according to user, system according to
Specific similarity measurement rule, all images in ergodic data storehouse, and select most like some width to return as Query Result
Back to user but due in real world image resource the abundantest, view data storage capacity is huge, if still employing order inspection
Rope, amount of calculation will be considerable, causes the recall precision can be the lowest.If can first the image in image library be entered before retrieval
Row cluster, sets up image index to all kinds of, and retrieving is carried out in specific a certain class then, thus can be at very great Cheng
The retrieval matching range of downscaled images on degree, to reach accurately and the purpose of rapidly retrieving images.
The algorithm being presently used for image clustering mainly has K-Means clustering algorithm, scalable k-means++ to cluster calculation
Method etc., but the characteristic of these sequence of algorithms limits its extensibility, and it needs when reconnaissance to carry out substantial amounts of iteration behaviour
Make so that efficiency comparison when it processes mass data under parallel processing environment is low.Therefore, exploitation better image cluster is calculated
Method just seems necessary.
In recent years, cloud computing was developing progressively an important branch into cloud computing.Any intelligent terminal such as intelligence
Mobile phone and panel computer the most on-demand from wireless network environment can obtain service, and are not only restricted to limited hardware money
Source, computing capability and bandwidth etc..It will be apparent that in cloud computing, efficient analysis and process massive spatio-temporal data, and gather with image
Class application combines, it is simply that an emerging practical technique, under cloud computing environment, effective Indexing Techniques for Spatial Databases is to carrying
High spatial database lookup efficiency is most important with application Consumer's Experience, and based on this starting point, we have designed and Implemented this
Bright patent.
Summary of the invention
According to defect and deficiency present in above-mentioned background technology, by use, patent of the present invention includes that SIFT feature describes
Method, over-sampling correction algorithm, at interior multiple image processing algorithms, devise a kind of new approximate image clustering software, more preferably
Solve image clustering problem.
To achieve these goals, the technical solution adopted in the present invention is:
The approximation space clustering system of a kind of mobile cloud environment, is mainly made up of cloud center service system, cloud center service
System is set up and is performed SIFT feature and extracts and matching algorithm, extracts the correlated characteristic data of image, and mates, and in cloud
Central server system also sets up and performs over-sampling correction algorithm to carry out image correction, and feedback cluster approximate image information is to client
End.
Further, the image that needs are carried out clustering by client as required sends to cloud center service system, and connects
Receive the request of cloud center service system.
Further, described foundation the method performing SIFT feature extraction and matching algorithm, including:
S1. detection yardstick spatial extrema；
S2. key point location；
S3. key point direction coupling；
S4. Feature Descriptor is generated；
S5. characteristic matching.
Further, the method for step S1. detection yardstick spatial extrema is: each sampled point will be with the 8 of present image
Individual consecutive points and 9 × 2 points corresponding to neighbouring scalogram picture compare, if this sampled point is both greater than or all
Less than other 26 consecutive points, this point is then key point.
Further, the method for step S2. key point location is: key point matching three-dimensional quadratic function is to position key point
Yardstick and position.
Further, the method for step S3. key point direction distribution is: adopt in the neighborhood window centered by key point
Sample, and with the gradient direction of gradient orientation histogram statistics neighborhood territory pixel, it is whole that histogrammic peak value represents at this key point
The principal direction of neighborhood gradient, this principal direction is as the direction of this key point.
Further, in gradient orientation histogram, when there is another peak value of main peak value 80% energy above, then
It is the auxiliary direction of this key point by the direction so that main and auxiliary direction is combined.
Further,
Step S4. generates the method for Feature Descriptor:
S4.1., after obtaining the principal direction of key point, centered by key point, the window of 16 × 16 is taken, and by image coordinate
Axle rotates to be the principal direction of key point；
S4.2. calculate in the window of 16 × 16 pixels centered by key point each pixel gradient direction and
Amplitude；
S4.3. Gauss weighting is carried out；
S4.4. window is divided into 4 × 4 wickets, each wicket calculates the gradient direction Nogata in 8 directions
Figure, and draw the accumulated value of each gradient direction, form a seed points；
Wherein, key point by 4 × 4 totally 16 seed points form, each seed points has 8 direction vector information, one
Individual characteristic point forms the SIFT feature vector of 128 dimensions.
Further, the method for step S5. characteristic matching is: take certain feature in piece image in two images
Point, find out its with another piece image in European closest the first two characteristic point, in the two characteristic point, if nearest away from
The ratio of the distance close to homogeneous is less than certain threshold value, then accept this pair match point, otherwise abandon.
Further,
The iterative process of described over-sampling correction algorithm is as follows:
Step one. calculate global clustering error ψ；
Step 2. each Map processing procedure, task uses scalable k-means++ initialization algorithm to select
Point, each some probability selected for x ∈ X is px=l*d2(x,U0)/φX(c1)；
Wherein: U0For initial all central point set, c1Uniformly random first central point selected, d2(x,U0) it is every
One some x ∈ XiWith U0Squared-distance；
Step 3. an all central point from Map task of Reduce task merging, the set U of output center point1；
Step 4. during iteration, OnR utilizes the global error that-2 iteration gained central points of jth are corresponding to select
Point；OnR utilizes another oversample factor o to expand each point further and is chosen as the probability of central point；OnR appoints at Reduce
Business use a process revised remove the central point of multiselect.
Beneficial effect: this image clustering system can be uploaded by the corresponding information that correlation technique gets picture at user side
And it is saved in cloud server, then cloud server processes, and obtains optimal image clustering scheme and feeds back to user.
Accompanying drawing explanation
The functional block diagram of Fig. 1 present invention；
The schematic diagram of the extensive approximate image clustering system framework of Fig. 2 present invention；
Fig. 3 inventive feature extracts process；
Fig. 4 inventive feature vector generates process；
The code of the over-sampling correction algorithm of Fig. 5 present invention；
The over-sampling correction algorithm flow process of Fig. 6 present invention；
The structured flowchart of the described system of Fig. 7 present invention.
Detailed description of the invention
Embodiment 1:With reference to Fig. 1, the approximation space clustering system of a kind of mobile cloud environment, described system is by a Ge Yun center
Service system and a cell phone intelligent mobile client software system composition.Wherein, cloud service system is responsible for carrying out SIFT feature
Extraction algorithm sets up the execution with over-sampling correction algorithm, and cluster result is fed back to user side；As required will in mobile terminal
Need the image carrying out image clustering to send to cloud center service system, and receive high in the clouds request.
With reference to Fig. 2, as an embodiment, the execution flow process of this approximate image clustering system is, when cluster user sends
After associated picture cluster request, cloud system obtain best according to SIFT feature extraction algorithm and over-sampling correction algorithm
Clustering schemes also returns final result to user, carries out business confirmation by mobile intercommunion platform.
This large-scale image clustering system uses a kind of data processing method based on cloud computing, gathers when user sends image
During class request, data center can extract rapidly the correlated characteristic data of image by the SIFT feature extraction algorithm set up.SIFT
The process step of feature extraction algorithm particularly as follows:
With reference to Fig. 3,4, its detailed step is: in order to detect the local maximum minimal point of metric space, each sampled point
9 × 2 points totally 26 points that will be corresponding with neighbouring scalogram picture with the 8 of present image consecutive points compare, if
This sampled point will be more than or will be less than other 26 consecutive points, and this point is then key point.These key points also need to intend
Close three-dimensional quadratic function to be accurately positioned yardstick and the position of key point.Sample in the neighborhood window centered by key point,
And the gradient direction of neighborhood territory pixel is added up with gradient orientation histogram.The angular range of this gradient orientation histogram is 0-360 degree,
The most every 10 degree as a post, totally 36 pillars, histogrammic peak value represents the master of the whole neighborhood gradient at this key point
Direction, the direction is then as the direction of this key point.In gradient orientation histogram, a key point may be designated to be had
Multiple directions, a principal direction and more than one auxiliary direction.When there is another peak value of main peak value 80% energy, then will
Individual direction is considered the auxiliary direction of this key point, can be strengthened the robustness of algorithmic match by the associating of primary and secondary direction.Obtaining
After the principal direction of key point, centered by key point, then take the window of 16 × 16, and image coordinate axle is rotated to be key point
Principal direction, to guarantee rotational invariance.Then calculate each in the window of 16 × 16 pixels centered by key point
The gradient direction of individual pixel and amplitude, then carry out Gauss weighting.Finally window is divided into 4 × 4 wickets, each little
The gradient orientation histogram in 8 directions of calculating on window, and draw the accumulated value of each gradient direction, a kind can be formed
Sub-point.One key point by 4 × 4 totally 16 seed points form, each seed points has 8 direction vector information, such a spy
Levy the SIFT feature vector that a little just can form 128 dimensions.As in Figure 2-4.The method of this associating neighborhood directivity information is not
Only enhance the noiseproof feature of SIFT algorithm, and the characteristic matching that there is error when location is also had reasonable fault-tolerant
Property.After the SIFT feature vector of two width images generates, it next it is exactly SIFT feature matching stage.SIFT feature Vectors matching
Using the Euclidean distance between characteristic vector as the similarity measures of characteristic point in two width images.Take one in two images
Certain characteristic point in width image, and find out itself and European closest the first two characteristic point in another piece image.This two
In individual characteristic point, if the ratio of the near distance of nearest distance homogeneous is less than certain threshold value, then accept this pair match point, otherwise
Abandon.When reducing this threshold value when, SIFT match point number can reduce, but match point more accurately and is stablized.
With reference to Fig. 5, as another embodiment, the definition of over-sampling correction algorithm is: in each iteration, over-sampling
Revise (Oversampling and Refining, referred to as OnR) use a MapReduce operation select l central point and
Calculate the error of the overall situation.OnR method is inspired by scalable k-means++ method, and except oversample factor, it uses
Another oversample factor o increases the number of the central point that the Map stage is selected further.
With reference to Fig. 6, Job1 is (from P1To P4) remain responsible for calculate ψ, ψ be global clustering error；Job2 is (from P5To P8) every
One Map (Map is a processing procedure, and this process naming is Map by we) task uses scalable k-means++
Initialization algorithm carries out reconnaissance, and each some probability selected for x ∈ X is px=l*d2(x,U0)/φX(c1), wherein: U0For initially
All central point set, c1Uniformly random first central point selected, d2(x,U0) it is that each puts x ∈ XiWith U0Square
Distance (P5Stage), all central point (P from Map task of right later Reduce task merging7Stage), output center point
Set U1(P8Stage).The process of OnR iteration is from P9To P11, the process of iteration j is as follows, and wherein, j is iterations, and 1
≤ j≤n, n are maximum iteration time, data point number determine: because each Map task cannot obtain-1 iteration of jth
The global error that gained central point is corresponding, i.e. cannot obtain φX(Uj-1), our method OnR utilizes-2 iteration gained of jth
The global error that central point is corresponding carries out reconnaissance, i.e. uses φX(Uj-2)；Yet with the central point obtained after-2 iteration of jth
Number fewer than the number of the central point of gained after j-1 iteration, so-2 corresponding global errors φ of jthX(Uj-2) ratio
-1 corresponding global error of jth is big, result in each some x and is chosen as the Probability p of central pointx=l*d2(x,Uj-1)/φX
(Uj-2) diminish, and then the number that result in the central point that iteration j selects tails off.In order to solve the problems referred to above, OnR utilizes
Another oversample factor o expands each point further and is chosen as the probability of central point, i.e. x is chosen as the probability of central point and is
px=o*l*d2(x,Uj-1)/φX(Uj-2).In this case, the number of the central point of iteration Map task choosing each time
Expected value can be more than l (l is the central point number that the expectation of scalable k-means++ initialization algorithm each iteration selects), because of
This OnR needs the central point using a process revised to remove multiselect in Reduce task.
The iterative process that OnR algorithm is detailed is as follows: at P9In the stage, each Map utilizes the Probability p of each some xx=o*l*
d2(x,Uj-1)/φX(Uj-2) select new central point and calculate center point set Uj-1Corresponding local errorAnd institute
Some local error valuesEach random probability value corresponding for some x chosen(The parameter arranged voluntarily, with p'x
=l*d2(x,Uj-1)/φX(Uj-1) compare to select central point) and l*d2(x,Uj-1) it is transferred to Reduce (P10Stage).Repair
Positive operation performs (P in a single Reduce task11Stage), the central point that its merging Map stage is selectedTo all of local errorCarry out sum operation and obtain-1 iteration rear center's point correspondence of jth
Global error φX(Uj-1), real probit corresponding to selected central point x, i.e. p' also can be obtained at thisx=l*d2(x,
Uj-1)/φX(Uj-1)；IfThen some x is correct central point, otherwise x is removed from C, this time the output of iteration
For UjAnd φX(Uj-1), they become the input data of next iteration.
By above analysis, show that the transmission volume of OnR iteration each time includes 4 parts: this time iteration is selected
Center point setAll of local errorThe random number that selected central point is correspondingl*d2(x,
Uj-1).With the transmission quantity of MRSKMI algorithm (With) compare, the major advantage of OnR is not have at network overhead
On the premise of having growth too many, it greatly reduces I/O expense.
Embodiment 2:The approximation space clustering system of a kind of mobile cloud environment is by a cloud center service system and a hands
Quick-witted energy mobile client software system composition.Wherein cloud service system is responsible for carrying out the foundation of SIFT feature extraction algorithm and excessively adopting
The execution of sample correction algorithm, and cluster result is fed back to user side；Mobile terminal carries out image clustering by needing as required
Image sends to cloud center service system, and receives high in the clouds request.
Supplementing as technical scheme, the execution method of this image clustering system is: when cluster user sends associated picture
After cluster request, cloud system obtain best clustering schemes also according to SIFT feature extraction algorithm and over-sampling correction algorithm
Return final result, to user, carries out business confirmation by mobile intercommunion platform.
Supplementing further as technical scheme, 1. the process step of SIFT feature extraction algorithm particularly as follows: detect yardstick
Spatial extrema；2. key point location；3. key point direction distribution；4. Feature Descriptor is generated；5. characteristic matching.
Its detailed step is: in order to detect the local maximum minimal point of metric space, and each sampled point will be with current figure
9x2 point totally 26 points that 8 consecutive points of picture are corresponding with neighbouring scalogram picture compare, if this sampled point will
Being more than or will be less than other 26 consecutive points, this point is then key point.These key points also need to matching three-dimensional secondary letter
Number is with the yardstick and the position that are accurately positioned key point.Sample in the neighborhood window centered by key point, and use gradient direction
The gradient direction of statistics with histogram neighborhood territory pixel.The angular range of this gradient orientation histogram is 0-360 degree, the most every 10 degree of works
Being a post, totally 36 pillars, histogrammic peak value represents the principal direction of the whole neighborhood gradient at this key point, and the direction is then
Direction as this key point.In gradient orientation histogram, a key point may be designated has multiple directions, one
Principal direction and more than one auxiliary direction.When there is another peak value of main peak value 80% energy, then individual direction is considered
The auxiliary direction of this key point, can strengthen the robustness of algorithmic match by the associating of primary and secondary direction.Obtaining the main formula of key point
Backward, centered by key point, then take the window of 16 × 16, and image coordinate axle is rotated to be the principal direction of key point, with really
Protect rotational invariance.Then in the window of 16 × 16 pixels centered by key point, calculate the gradient side of each pixel
To and amplitude, then carry out Gauss weighting.Finally window is divided into 4 × 4 wickets, each wicket calculates 8
The gradient orientation histogram in direction, and draw the accumulated value of each gradient direction, a seed points can be formed.One key point
By 4 × 4 totally 16 seed points form, each seed points has 8 direction vector information, and such a characteristic point just can be formed
The SIFT feature vector of 128 dimensions.As in Figure 2-4.The method of this associating neighborhood directivity information not only increases SIFT and calculates
The noiseproof feature of method, and the characteristic matching that there is error when location is also had reasonable fault-tolerance.Display in Fig. 2-4
It is the window of 4 × 4 pixels, but classical SIFT uses the window of 16 × 16 pixels in Practical Calculation.When two width figures
After the SIFT feature vector of picture generates, it next it is exactly SIFT feature matching stage.SIFT feature Vectors matching is with characteristic vector
Between Euclidean distance as the similarity measures of characteristic point in two width images.Take in two images in piece image
Certain characteristic point, and find out itself and European closest the first two characteristic point in another piece image.In the two characteristic point,
If the ratio of the distance that nearest distance homogeneous is near is less than certain threshold value, then accepts this pair match point, otherwise abandon.Work as reduction
The when of this threshold value, SIFT match point number can reduce, but match point more accurately and is stablized.
Supplementing as technical scheme, the definition of over-sampling correction algorithm is: in each iteration, and OnR uses one
MapReduce operation is selected l central point and calculates the error of the overall situation.OnR method is by scalable k-means++ method
Inspiration, except oversample factor l, it uses another oversample factor o to increase the number of the central point that the Map stage is selected further
Mesh.
The most supplementary as technical scheme, the method for approximate image based on over-sampling correction cluster is:
Job1 is (from P1To P4) remain responsible for calculating ψ；Job2 is (from P5To P8) each Map task use scalable k-means+
+ initialization algorithm carries out reconnaissance, and each some probability selected for x ∈ X is px=l*d2(x,U0)/φX(c1)(P5Stage), then
One all central point (P from Map task of Reduce task merging7Stage), the set U of output center point1(P8Stage).
The process of OnR iteration is from P9To P11, the process of iteration j is as follows: because each Map task cannot obtain jth and change for-1 time
For the global error that gained central point is corresponding, φ i.e. cannot be obtainedX(Uj-1), our method OnR utilizes-2 iteration institutes of jth
The global error obtaining central point corresponding carries out reconnaissance, i.e. uses φX(Uj-2)；Yet with the center obtained after-2 iteration of jth
The number of point is fewer than the number of the central point of gained after j-1 iteration, so-2 corresponding global errors φ of jthX
(Uj-2) bigger than-1 corresponding global error of jth, result in each some x and be chosen as the Probability p of central pointx=l*d2(x,
Uj-1)/φX(Uj-2) diminish, and then the number that result in the central point that iteration j selects tails off.In order to solve the problems referred to above,
OnR utilizes another oversample factor o to expand each point further and is chosen as the probability of central point, i.e. x is chosen as central point
Probability be px=o*l*d2(x,Uj-1)/φX(Uj-2).In this case, the central point of iteration Map task choosing each time
The expected value of number can be more than l (l is the central point that the expectation of scalable k-means++ initialization algorithm each iteration selects
Number), therefore OnR needs the central point using a process revised to remove multiselect in Reduce task.
The iterative process that OnR algorithm is detailed is as follows: at P9In the stage, each Map utilizes the Probability p of each some xx=o*l*
d2(x,Uj-1)/φX(Uj-2) select new central point and calculate center point set Uj-1Corresponding local errorAnd institute
Some local error valuesEach random probability value corresponding for some x chosenAnd l*d2(x,Uj-1) be transferred to
Reduce(P10Stage).The operation revised performs (P in a single Reduce task11Stage), it merges the Map stage
The central point selectedTo all of local errorCarry out during sum operation obtains after-1 iteration of jth
Global error φ that heart point is correspondingX(Uj-1), real probit corresponding to selected central point x, i.e. p' also can be obtained at thisx
=l*d2(x,Uj-1)/φX(Uj-1)；IfThen some x is correct central point, otherwise x is removed from C, the most repeatedly
In generation, is output as UjAnd φX(Uj-1), they become the input data of next iteration.
Embodiment 3:In the various embodiments described above, at clustering phase, this image clustering system can be by correlation technique user
End gets the corresponding information of picture and uploads and be saved in cloud server, and then cloud server processes, and obtains optimal
Image clustering scheme and feed back to user.
Image clustering system described in the present embodiment has following structure and a benefit:
(1) design of single terminal is used.User side is the software being arranged in Android smartphone, enters for user
Use during row image clustering.
User passes through system and the base station of mobile phone operators of embedded in mobile phone, relies on 2G/3G network, and wifi etc. obtains and needs
Image to be clustered also sends to high in the clouds.
(2) cloud computing is a kind of calculation based on the Internet, in this way, and the software and hardware resources shared and letter
Breath on-demand can be supplied to computer and other equipment.The cloud service that in our invention, the image clustering system of design is used
Device is the webserver by multiple cloud data centers or fictitious host computer is constituted, and uses this parallelization of cloud computing to calculate
Process large-scale data and tackle the image clustering user on line, in such a mode, it is ensured that addressing during high access is stable
Property, also accelerate response speed when user searches for, enhance extensibility simultaneously.
(3) when user sends cluster request, data center can pass through SIFT feature extracting method, over-sampling correction algorithm
Carry out large-scale image cluster, and provide the image clustering scheme that applicable user is best.
The above, only the invention preferably detailed description of the invention, but the protection domain of the invention is not
Being confined to this, any those familiar with the art is in the technical scope that the invention discloses, according to the present invention
The technical scheme created and inventive concept thereof in addition equivalent or change, all should contain the invention protection domain it
In.
Claims (10)
1. the approximation space clustering system moving cloud environment, it is characterised in that: mainly it is made up of cloud center service system, cloud
Center service system is set up and is performed SIFT feature and extracts and matching algorithm, extracts the correlated characteristic data of image, and carries out
Join, and cloud center service system also sets up and perform over-sampling correction algorithm to carry out image correction, feedback cluster approximate image
Information is to client.
2. the approximation space clustering system of mobile cloud environment as claimed in claim 1, it is characterised in that client is as required
The image carrying out needs clustering sends to cloud center service system, and receives the request of cloud center service system.
3. the approximation space clustering system of mobile cloud environment as claimed in claim 1, it is characterised in that described foundation also performs
SIFT feature extracts the method with matching algorithm, including:
S1. detection yardstick spatial extrema；
S2. key point location；
S3. key point direction coupling；
S4. Feature Descriptor is generated；
S5. characteristic matching.
4. the approximation space clustering system of mobile cloud environment as claimed in claim 3, it is characterised in that step S1. detection ruler
The method of degree spatial extrema is: each sampled point will be with the 8 of present image consecutive points and neighbouring scalogram picture pair
9 × 2 points answered compare, if this sampled point both greater than or is both less than other 26 consecutive points, this point is then crucial
Point.
5. the approximation space clustering system of mobile cloud environment as claimed in claim 3, it is characterised in that step S2. key point
The method of location is: key point matching three-dimensional quadratic function is to position yardstick and the position of key point.
6. the approximation space clustering system of mobile cloud environment as claimed in claim 3, it is characterised in that step S3. key point
The method of direction distribution is: samples in the neighborhood window centered by key point, and adds up neighborhood with gradient orientation histogram
The gradient direction of pixel, histogrammic peak value represents the principal direction of the whole neighborhood gradient at this key point, this principal direction conduct
The direction of this key point.
7. the approximation space clustering system of mobile cloud environment as claimed in claim 6, it is characterised in that at gradient direction Nogata
In figure, when there is another peak value of main peak value 80% energy above, then it is the auxiliary direction of this key point by the direction so that
Main and auxiliary direction is combined.
8. the approximation space clustering system of mobile cloud environment as claimed in claim 3, it is characterised in that step S4. generates spy
The method levying description is:
S4.1., after obtaining the principal direction of key point, centered by key point, take the window of 16 × 16, and image coordinate axle is revolved
Transfer the principal direction of key point to；
S4.2. in the window of 16 × 16 pixels centered by key point, calculate gradient direction and the width of each pixel
Value；
S4.3. Gauss weighting is carried out；
S4.4. window is divided into 4 × 4 wickets, each wicket calculates the gradient orientation histogram in 8 directions,
And draw the accumulated value of each gradient direction, form a seed points；
Wherein, key point by 4 × 4 totally 16 seed points form, each seed points has 8 direction vector information, a spy
Levy a SIFT feature vector forming 128 dimensions.
9. the approximation space clustering system of mobile cloud environment as claimed in claim 3, it is characterised in that step S5. feature
The method joined is: take certain characteristic point in piece image in two images, finds out itself and Euclidean distance in another piece image
Nearest the first two characteristic point, in the two characteristic point, if the ratio of the near distance of nearest distance homogeneous is less than certain threshold
Value, then accept this pair match point, otherwise abandon.
10. the approximation space clustering system of mobile cloud environment as claimed in claim 1, it is characterised in that described over-sampling is repaiied
The iterative process of normal operation method is as follows:
Step one. calculate global clustering error ψ；
Step 2. each Map processing procedure, task uses scalable k-means++ initialization algorithm to carry out reconnaissance, often
Individual some probability selected for x ∈ X is
Wherein: U0For initial all central point set, c1Uniformly random first central point selected, d2(x,U0) be each point
x∈XiWith U0Squared-distance；
Step 3. an all central point from Map task of Reduce task merging, the set U of output center point1；
Step 4. during iteration, OnR utilizes the global error that-2 iteration gained central points of jth are corresponding to carry out reconnaissance；
OnR utilizes another oversample factor o to expand each point further and is chosen as the probability of central point；OnR is in Reduce task
The process that middle use one is revised removes the central point of multiselect.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610539136.9A CN106203449A (en) | 2016-07-08 | 2016-07-08 | The approximation space clustering system of mobile cloud environment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610539136.9A CN106203449A (en) | 2016-07-08 | 2016-07-08 | The approximation space clustering system of mobile cloud environment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106203449A true CN106203449A (en) | 2016-12-07 |
Family
ID=57473110
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610539136.9A Pending CN106203449A (en) | 2016-07-08 | 2016-07-08 | The approximation space clustering system of mobile cloud environment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106203449A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107657175A (en) * | 2017-09-15 | 2018-02-02 | 北京理工大学 | A kind of homologous detection method of malice sample based on image feature descriptor |
| CN107679216A (en) * | 2017-10-19 | 2018-02-09 | 大连大学 | The distributed temporal index method of the row's of falling Thiessen polygon of portable medical and application |
| CN107766495A (en) * | 2017-10-19 | 2018-03-06 | 大连大学 | Distributed temporal index method based on Voronoi diagram |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070183653A1 (en) * | 2006-01-31 | 2007-08-09 | Gerard Medioni | 3D Face Reconstruction from 2D Images |
| CN102184551A (en) * | 2011-05-10 | 2011-09-14 | 东北大学 | Automatic target tracking method and system by combining multi-characteristic matching and particle filtering |
| CN102682091A (en) * | 2012-04-25 | 2012-09-19 | 腾讯科技（深圳）有限公司 | Cloud-service-based visual search method and cloud-service-based visual search system |
| US20140219569A1 (en) * | 2013-02-07 | 2014-08-07 | Raytheon Company | Image recognition system and method for identifying similarities in different images |
| CN104142921A (en) * | 2013-05-06 | 2014-11-12 | 苏州普达新信息技术有限公司 | Image feature optimal-combination method based on cloud computing |
-
2016
- 2016-07-08 CN CN201610539136.9A patent/CN106203449A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070183653A1 (en) * | 2006-01-31 | 2007-08-09 | Gerard Medioni | 3D Face Reconstruction from 2D Images |
| CN102184551A (en) * | 2011-05-10 | 2011-09-14 | 东北大学 | Automatic target tracking method and system by combining multi-characteristic matching and particle filtering |
| CN102682091A (en) * | 2012-04-25 | 2012-09-19 | 腾讯科技（深圳）有限公司 | Cloud-service-based visual search method and cloud-service-based visual search system |
| US20140219569A1 (en) * | 2013-02-07 | 2014-08-07 | Raytheon Company | Image recognition system and method for identifying similarities in different images |
| CN104142921A (en) * | 2013-05-06 | 2014-11-12 | 苏州普达新信息技术有限公司 | Image feature optimal-combination method based on cloud computing |
Non-Patent Citations (1)
| Title |
|---|
| 许玉杰: "云计算环境下海量数据的并行聚类算法研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107657175A (en) * | 2017-09-15 | 2018-02-02 | 北京理工大学 | A kind of homologous detection method of malice sample based on image feature descriptor |
| CN107679216A (en) * | 2017-10-19 | 2018-02-09 | 大连大学 | The distributed temporal index method of the row's of falling Thiessen polygon of portable medical and application |
| CN107766495A (en) * | 2017-10-19 | 2018-03-06 | 大连大学 | Distributed temporal index method based on Voronoi diagram |
| CN107958087A (en) * | 2017-10-19 | 2018-04-24 | 大连大学 | The representative point acquisition methods of the distributed temporal index of the row's of falling Thiessen polygon of portable medical |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chen et al. | An edge traffic flow detection scheme based on deep learning in an intelligent transportation system | |
| Cong et al. | Going from RGB to RGBD saliency: A depth-guided transformation model | |
| US10102227B2 (en) | Image-based faceted system and method | |
| CN108108499B (en) | Face retrieval method, device, storage medium and equipment | |
| CN109711228B (en) | Image processing method and device for realizing image recognition and electronic equipment | |
| CN102831405B (en) | Method and system for outdoor large-scale object identification on basis of distributed and brute-force matching | |
| Faraki et al. | Log‐Euclidean bag of words for human action recognition | |
| CN103927387A (en) | Image retrieval system, method and device | |
| CN102737243A (en) | Method and device for acquiring descriptive information of multiple images and image matching method | |
| CN106250909A (en) | A kind of based on the image classification method improving visual word bag model | |
| Bi et al. | Multiple instance dense connected convolution neural network for aerial image scene classification | |
| CN106203449A (en) | The approximation space clustering system of mobile cloud environment | |
| Petkos et al. | Graph-based multimodal clustering for social event detection in large collections of images | |
| CN108830283B (en) | Image feature point matching method | |
| CN107315984B (en) | Pedestrian retrieval method and device | |
| CN110929081A (en) | Picture query method, computer equipment and storage medium | |
| CN109948624A (en) | Method, apparatus, electronic equipment and the computer storage medium of feature extraction | |
| Yu et al. | Quantization-based clustering algorithm | |
| CN111415196A (en) | Advertisement recall method, device, server and storage medium | |
| CN108875828A (en) | A kind of fast matching method and system of similar image | |
| Zhai et al. | FMT: fusing multi-task convolutional neural network for person search | |
| CN106250907A (en) | Cloud computing environment large-scale image based on over-sampling correction clustering method | |
| CN103955713A (en) | Icon recognition method and device | |
| He et al. | D2former: Jointly learning hierarchical detectors and contextual descriptors via agent-based transformers | |
| Gao et al. | Intelligent query: open another door to 3D object retrieval |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20161207 |