CN103914557B - Mobile visual retrieval method based on key feature descriptor selection - Google Patents
Mobile visual retrieval method based on key feature descriptor selection Download PDFInfo
- Publication number
- CN103914557B CN103914557B CN201410151518.5A CN201410151518A CN103914557B CN 103914557 B CN103914557 B CN 103914557B CN 201410151518 A CN201410151518 A CN 201410151518A CN 103914557 B CN103914557 B CN 103914557B
- Authority
- CN
- China
- Prior art keywords
- feature descriptor
- key feature
- image
- descriptor
- key
- 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.)
- Active
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
Abstract
The invention discloses a mobile visual retrieval system based on key feature descriptor selection, and belongs to the field of mobile visual retrieval. The mobile visual retrieval system based on key feature descriptor selection is characterized by utilizing an affinity propagation clustering method to perform key feature descriptor selection on the feature descriptors of images to be queried, using a method of transmitting and then weighing and matching the key feature descriptors, reducing data transfer volume through key feature descriptor selection to reduce network transmission time delay and further achieving the aim of improving user experience. Meanwhile, the method can effectively eliminate noise point data of the images to be queried, thereby reducing the data transfer volume and effectively improving the retrieval accuracy at the same time.
Description
Technical field
The invention belongs to moving-vision field, it is related to a kind of moving-vision retrieval side choosing based on key feature descriptor
Method.It is related to, using affinity propagation (affinity propagation, ap) clustering method, characteristics of image is carried out with crucial spy
Levy descriptor to choose, and method key feature descriptor transmitted and carries out weighted image coupling.
Background technology
With the high speed development of software and hardware technology, moving-vision retrieval application has obtained rapid development.Growing number of
Number of users and user improve to the inspection of traditional moving-vision based on customer end/server mode for quality of service requirement
Rope service proposes challenge.Under traditional customer end/server mode, the image that the mobile clients such as mobile phone shoot is led to by user
Cross wireless network transmissions, to server, then by server, line retrieval is entered to image, to return the relevant information of user's query image.
Bandwidth due to wireless network environment limits, and the transmission delay of high quality graphic substantially increases the waiting time of user, fall
Low Consumer's Experience.The response time of this application is substantially made up of three aspects: when the process time of client, network transmission
Between, the server process time.
Existing work shows, in moving-vision searching system, compared to the response time of client and server, net
Network transmission delay is the bottleneck of system.Therefore, existing popular system is all absorbed in a kind of low bit feature descriptor of design
Reach the purpose reducing network transmission time delay, to reduce retrieval response time, improve Consumer's Experience.In order to solve due to wireless network
The service bottleneck that network bandwidth limits, low bit feature descriptor is introduced in moving-vision retrieval application.Master in current application
In stream moving-vision retrieval application, employ the low bit feature descriptor of several new propositions, decrease certain data transfer
Amount.Despite the use of low bit feature descriptor, but still suffer from some drawbacks.On the one hand, substantial amounts of when existing in query image
Characteristic point, has also corresponded to substantial amounts of feature descriptor, and the transmitted data amount of this mode is still very high, and consequent network prolongs
When, still have to be reduced, and on the other hand, in the characteristic point of query image, the importance of each characteristic point is different, wherein even more there is portion
Divide noise spot, the accuracy of image retrieval can be reduced, how to avoid transmitting useless noise spot feature descriptor is another value
The problem that must study.
Content of the invention
According to the deficiency in above-mentioned background technology, the invention provides a kind of key feature based on affinity propagation clustering
The moving-vision search method of descriptor algorithm, the main object of the present invention is to be used as pass by the central point that clustering algorithm is chosen
Key characteristic point, carries out image retrieval by choosing and only transmitting key feature descriptor, reaches the purpose reducing transmitted data amount,
Thus reducing network transmission time delay；Meanwhile, using this key descriptor extraction algorithm, can effectively exclude in former query image
Some noise spots；Moreover it is possible to effectively improve image retrieval precision while reducing volume of transmitted data.
The present invention includes query image feature extraction, key based on the moving-vision search method of key feature descriptor
Feature descriptor is chosen, weighting key feature descriptors match；It carries out the technical scheme of image retrieval:
(1) adopt affinity propagation clustering algorithm (affinity propagation) to the query image feature extracted
Carry out key feature descriptor selection: the local feature of query image is extracted in mobile terminal, and the local feature of query image is as sky
Between point set carry out affinity propagation clustering, all central feature points being formed through affinity propagation clustering are chosen for crucial spy
Levy descriptor；
(2) the key feature descriptor obtaining in step (1) is weighted mating, specifically comprises the following steps that
A. as selection class cluster (cluster) ckCentral point feature descriptor dkFor key feature descriptor, this key feature
Descriptor dkRepresent the feature descriptor that each belongs to its class cluster (cluster), according in affinity propagation clustering algorithm
Message parameter (message) r (i, k) and message parameter a (i, k), define dkOne is belonged to the feature of its class cluster (cluster)
Descriptor diRepresentative degreeAs i ≠ k, 0 ＜ pi,k＜ 1；As i=k, represent journey
Degree pi,k=1；
B. define dkImportance parameter be dkTo cluster ckThe representative degree sum of all of feature descriptor, calculates dk's
Importance parameter pkFormula be:
C. carry out characteristic matching in server end, to each d above-mentionedkImage feature base from server end finds
One corresponding arest neighbors characteristic point d'k, between the two apart from diskRepresented with Euclidean distance；Retouched according to key feature
State the distance parameter d of symboliskWith importance parameter pkCalculate weighted value, the weighted value of each key feature descriptor is:Then by weighted value normalization:Wherein nddRepresent the individual of key feature descriptor
Number；
D. for the similarity s equation below meter between the arbitrary image i in image feature base and query image
Calculate:Wherein work as dkArest neighbors characteristic point when belonging to image i, sk=w* k, otherwise sk=0；By in image library
Image presses similarity sequence, you can find matching result.
The invention has the beneficial effects as follows key feature is extracted to query image using the clustering method propagated based on affinity
Descriptor, with respect to other clustering methods, affinity propagation clustering algorithm can in the point to be clustered giving core out point,
Rather than produce a new central point；Choose key characteristics descriptor using the method, decrease the same of inquiry data volume
When moreover it is possible to exclusive segment noise spot, thus reducing transmitted data on network amount, appreciably reducing network transmission time delay, accelerating
The response speed of moving-vision retrieval application, improves Consumer's Experience；On the other hand, affinity propagation clustering algorithm can exclude source
The point of partial noise present in image characteristic point, simultaneously the present invention improve the retrieval of images match using the method for weighted registration
Precision, further increases the performance of system, meets the demand of user.
Brief description
Accompanying drawing is the system framework figure of the present invention.
Specific embodiment
With reference to technical scheme and brief description specific embodiment.
The invention provides a kind of moving-vision search method chosen based on key feature descriptor, concrete steps are such as
Under:
(1) adopt affinity propagation clustering algorithm (affinity propagation) to the query image feature extracted
Carry out key feature descriptor selection: the local feature of query image is extracted in mobile terminal, and the local feature of query image is as sky
Between point set carry out affinity propagation clustering, all central feature points being formed through affinity propagation clustering are chosen for crucial spy
Levy descriptor；
The local feature of query image is extracted in mobile terminal first, and we regard query image feature description vector set as space
Point set carries out affinity propagation clustering, relies on two kinds of message reliabilities (responsibility) between points and can
The transmission of usability (availability), will select a series of class central point.With respect to other clustering methods, affinity passes
Broadcast clustering algorithm can in the point to be clustered being given core out point, rather than produce a new central point, such as k-means gathers
Class；The central feature point that therefore cluster can be formed by we is chosen for key feature descriptor；Chosen crucial using the method
Property feature descriptor, moreover it is possible to exclusive segment noise spot while decreasing inquiry data volume；Simultaneously because these central points pair
Critical characteristic point for query image, only carried out with key feature descriptor images match can also reach considerable
Retrieval precision, by choosing key feature descriptor, to reach the purpose that minimizing system needs the data volume of network transmission, thus
Reduce system response time.The following is retrieval precision and transmission data contrast table:
Which depict two kinds of feature descriptor: surf and chog, method proposed by the present invention with respect to existing method, no
Only the mean size of query image is reduced to 5.2kb and 16.5kb from 34.5kb and 28.2kb respectively, simultaneously to inquiry precision
Also it has been respectively increased 0.9% and 4.9%.
(2) consider that different key feature descriptors represent different number of feature approximate therewith as cluster centre
Descriptor, the present invention proposes the method that key feature descriptor is weighted with mate, and this weighted registration method considers simultaneously
Arrived this key feature descriptor criticality and in characteristic matching with the distance of its arest neighbors feature；
Specifically comprise the following steps that a. works as feature descriptor dkIt is chosen for key feature descriptor, this key feature descriptor dk
Represent the feature descriptor that each belongs to its class cluster (cluster), according to the message parameter in affinity propagation clustering algorithm
(message) r (i, k) and message parameter a (i, k), defines key feature descriptor dkOne is belonged to its class cluster (cluster)
Feature descriptor diRepresentative degreeAs i ≠ k, represent degree pi,k, 0 ＜ pi,k
＜ 1；As i=k, represent degree pi,k=1；
B. as selection class cluster (cluster) ckCentral point feature descriptor dkFor key feature descriptor, define dkWeight
The property wanted parameter is characterized descriptor dkTo cluster ckRepresentative degree p of all of feature descriptori,kSum, calculates dkImportance ginseng
Number pkFormula be:
C. carry out characteristic matching in server end, to above-mentioned each key feature descriptor dkImage from server end
Property data base finds corresponding arest neighbors characteristic point d'k, between the two apart from disk, diskUse Euclidean distance table
Show；Distance parameter d according to key feature descriptoriskWith importance parameter pkCalculate weighted value, each key feature descriptor
Weighted value be:Then by weighted value normalization:Wherein nddRepresent key feature
The number of descriptor；
D. for the similarity s equation below meter between the arbitrary image i in image feature base and query image
Calculate:Wherein work as dkArest neighbors characteristic point when belonging to image i, sk=w* k, otherwise sk=0；By in image library
Image presses similarity sequence, you can find matching result.
Claims (1)
1. a kind of based on key feature descriptor choose moving-vision search method it is characterised in that
(1) key feature descriptor selection is carried out to the query image feature extracted using affinity propagation clustering algorithm: move
The local feature of query image is extracted in moved end, and the local feature of query image carries out affinity propagation clustering as space point set,
The all central feature points being formed through affinity propagation clustering are chosen for key feature descriptor；
(2) the key feature descriptor obtaining in step (1) is weighted mating, specifically comprises the following steps that
A. as selection class cluster ckCentral point feature descriptor dkFor key feature descriptor, this key feature descriptor dkRepresent every
One feature descriptor belonging to its class cluster, according to message parameter r (i, k) in affinity propagation clustering algorithm and message parameter
A (i, k), defines dkOne is belonged to the feature descriptor d of its class clusteriRepresentative degree
As i ≠ k, 0 ＜ pi,k＜ 1；As i=k, represent degree pi,k=1；
B. define dkImportance parameter be dkTo cluster ckThe representative degree sum of all of feature descriptor, calculates dkImportance
Parameter pkFormula be:
C. carry out characteristic matching in server end, to each d above-mentionedkImage feature base from server end finds one
Corresponding arest neighbors characteristic point d'k, between the two apart from diskRepresented with Euclidean distance；According to key feature descriptor
Distance parameter diskWith importance parameter pkCalculate weighted value, the weighted value of each key feature descriptor is:Then by weighted value normalization:Wherein nddRepresent the individual of key feature descriptor
Number；
D. the similarity s equation below between the arbitrary image i in image feature base and query image is calculated:Wherein work as dkArest neighbors characteristic point when belonging to image i, sk=w* k, otherwise sk=0；By the figure in image library
As by similarity sequence, that is, finding matching result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410151518.5A CN103914557B (en) | 2014-04-15 | 2014-04-15 | Mobile visual retrieval method based on key feature descriptor selection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410151518.5A CN103914557B (en) | 2014-04-15 | 2014-04-15 | Mobile visual retrieval method based on key feature descriptor selection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103914557A CN103914557A (en) | 2014-07-09 |
| CN103914557B true CN103914557B (en) | 2017-01-25 |
Family
ID=51040237
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410151518.5A Active CN103914557B (en) | 2014-04-15 | 2014-04-15 | Mobile visual retrieval method based on key feature descriptor selection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103914557B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104951562B (en) * | 2015-07-08 | 2018-06-05 | 天云融创数据科技（北京）有限公司 | A kind of image search method based on VLAD dual adaptions |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8086616B1 (en) * | 2008-03-17 | 2011-12-27 | Google Inc. | Systems and methods for selecting interest point descriptors for object recognition |
| CN102682048A (en) * | 2011-10-28 | 2012-09-19 | 刘晓勇 | Self-adaption attraction transmission clustering method and device |
| CN103605765A (en) * | 2013-11-26 | 2014-02-26 | 电子科技大学 | Mass image retrieval system based on cluster compactness |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9229956B2 (en) * | 2011-01-10 | 2016-01-05 | Microsoft Technology Licensing, Llc | Image retrieval using discriminative visual features |
-
2014
- 2014-04-15 CN CN201410151518.5A patent/CN103914557B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8086616B1 (en) * | 2008-03-17 | 2011-12-27 | Google Inc. | Systems and methods for selecting interest point descriptors for object recognition |
| CN102682048A (en) * | 2011-10-28 | 2012-09-19 | 刘晓勇 | Self-adaption attraction transmission clustering method and device |
| CN103605765A (en) * | 2013-11-26 | 2014-02-26 | 电子科技大学 | Mass image retrieval system based on cluster compactness |
Non-Patent Citations (3)
| Title |
|---|
| Clustering by Passing Messages Between Data Points;Brendan J. Frey et al;《Science》;20070216;第315卷(第5814期);第972-976页 * |
| Compressed Histogram of Gradients: A Low-Bitrate Descriptor;Vijay Chandrasekhar et al;《International Journal of Computer Vision》;20120229;第96卷(第03期);第384-399页 * |
| PCA-SIFT: A More Distinctive Representation for Local Image Descriptors;Yan Ke et al;《Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition》;20040702;第02卷;第506-513页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103914557A (en) | 2014-07-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10102227B2 (en) | Image-based faceted system and method | |
| CN104169946B (en) | Extensible queries for visual search | |
| KR102019336B1 (en) | Matching users in a location-based service | |
| US20150271280A1 (en) | Multi-objective server placement determination | |
| US8718378B2 (en) | Image topological coding for visual search | |
| US9659214B1 (en) | Locally optimized feature space encoding of digital data and retrieval using such encoding | |
| CN105760468A (en) | Large-scale image querying system based on inverted position-sensitive Hash indexing in mobile environment | |
| CN105760469A (en) | High-dimensional approximate image retrieval method based on inverted LSH in cloud computing environment | |
| US8755605B2 (en) | System and method for compact descriptor for visual search | |
| CN103914557B (en) | Mobile visual retrieval method based on key feature descriptor selection | |
| CN104809166B (en) | A kind of determination method and apparatus of business demand | |
| CN104994062B (en) | A kind of information-pushing method and device | |
| US10409857B2 (en) | Method and system of generating and using a geographical hierarchy model | |
| CN111626044A (en) | Text generation method and device, electronic equipment and computer readable storage medium | |
| CN107203578B (en) | Method and device for establishing association of user identifiers | |
| CN103269336A (en) | Method and system for data request and presentation | |
| CN116957095A (en) | Estimation method and estimation device for model application times | |
| TWI499921B (en) | Near duplicate images computer for a method and apparatus | |
| CN105357173A (en) | Data resource processing method and system, electronic equipment and terminal adopted thereby | |
| CN114168763A (en) | Multimedia resource recall method, device, equipment and storage medium | |
| CN116737999A (en) | Data processing method and device | |
| CN113918836A (en) | Space query verification method, device, equipment and storage medium | |
| JP2009140004A (en) | Distributed recommendation system | |
| CN106341311A (en) | Route and real-time location-based data sharing method, terminal and server |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |