CN104182974B - A speeded up method of executing image matching based on feature points - Google Patents

A speeded up method of executing image matching based on feature points Download PDF

Info

Publication number
CN104182974B
CN104182974B CN201410392413.9A CN201410392413A CN104182974B CN 104182974 B CN104182974 B CN 104182974B CN 201410392413 A CN201410392413 A CN 201410392413A CN 104182974 B CN104182974 B CN 104182974B
Authority
CN
China
Prior art keywords
point
subregion
matching
points
characteristic
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
Application number
CN201410392413.9A
Other languages
Chinese (zh)
Other versions
CN104182974A (en
Inventor
林秋华
曹建超
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian University of Technology
Original Assignee
Dalian University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Dalian University of Technology filed Critical Dalian University of Technology
Priority to CN201410392413.9A priority Critical patent/CN104182974B/en
Publication of CN104182974A publication Critical patent/CN104182974A/en
Application granted granted Critical
Publication of CN104182974B publication Critical patent/CN104182974B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Landscapes

  • Image Analysis (AREA)

Abstract

The invention discloses a speeded up method of executing image matching based on feature points, and belongs to the field of computer vision. The speeded up method of executing image matching based on feature points is characterized by adding a fishing strategy between extraction of the feature points of a target image and a reference image and construction of feature descriptors, evenly dividing all of the feature points of the target image into N*N sub-areas in accordance with locations, randomly selecting a certain proportion of feature points, and constructing feature descriptors and executing matching only for these feature points, wherein, if feature points selected from some sub-area have more matching points in the reference image, the number of feature points selected next from the area is increased, and if feature points selected from some sub-area do not have more matching points in the reference image, the number of feature points selected next from the area is decreased, until the total number of the matching points reaches threshold requirement or the feature points participating in the matching reach a certain proportion. In comparing with the original method of executing image matching based on feature points, the speeded up method of executing image matching based on feature points can increase the speed of the image matching by about 5 times without reducing matching accuracy and with saving memory, and can resolve the problem of gathering of the matching points to a certain extent.

Description

A kind of distinguished point based carries out the accelerated method of images match
Technical field
The present invention relates to computer vision field, more particularly to a kind of accelerated method of images match.
Background technology
Images match is always the study hotspot of computer vision field, and knows in vision guided navigation, target positioning, target It is not widely applied with aspects such as tracking, remote sensing image processing, image retrieval, stereoscopy passive ranging and three-dimensional reconstructions.
Distinguished point based carries out the image matching method that images match is that a class is commonly used.Its process mainly has three, first, point Indescribably take the characteristic point of target image and reference picture, such as speckle or angle point;Then, it is each characteristic point construction feature description; Finally, the distance between target image and each Feature Descriptor of reference picture are measured, when the distance of two characteristic points is less than one When determining threshold value, judge that two characteristic points can be mated, when match point quantity is more than certain threshold value, assert and deposit between two width images In matching relationship., the result of coupling can be shown that taking object recognition task as a example, identifies target image from reference picture.
Below taking famous SIFT (scale invariant feature transform) algorithm as a example, brief description Three key links of images match.(1) target image and reference picture feature point extraction.In this link, SIFT algorithm is first Build gaussian pyramid, then difference of Gaussian pyramid is built according to gaussian pyramid, finally by the extreme point of this metric space Position is defined as characteristic point position.(2) Feature Descriptor builds.Constant for making Feature Descriptor have rotation, affine, illumination etc. Property, SIFT adopts construction feature description with the following method:First, selected respectively by feature neighborhood of a point gradient orientation histogram Principal direction α of characteristic point.Secondly, using principal direction α as reference coordinate direction of principal axis, character pair neighborhood of a point is divided into 16 Subregion, and build the histogram of gradients in 8 directions in every sub-regions.Finally, 16 sub-regions are connected in certain sequence The histogram of gradients in interior 8 directions, generates the feature description subvector that length is 128 dimensions, and builds k dimension search for reference picture Tree.It should be noted that in the application such as target recognition, reference picture is known.At this moment, in order to save match time, can Carry out reference picture feature point extraction in advance, Feature Descriptor builds, and k dimension search tree builds and the work such as preservation, is connecing The characteristic matching link got off, directly invokes this k dimension search tree.(3) characteristic matching.K dimension search based on reference picture Tree, calculates the Euclidean distance between target image and reference picture feature description subvector, as arest neighbors Euclidean distance and time neighbour The ratio of Euclidean distance is less than during given threshold then it is assumed that being matching double points.
SIFT has generally acknowledged superperformance, but there is also following three aspect problems.One is that the speed of service is slow, real-time Difference.Although many scholars give multiple improvement projects, mostly with the matching precision of sacrifice SIFT as cost.For example, it is people Known to SURF (speeded up robust features) algorithm, its speed is 3 times of SIFT, but actual match performance is relatively SIFT declined.In fact, in three links of SIFT images match, the Feature Descriptor of the second link builds and needs Amount of calculation maximum.Through experiment test, the first link feature point extraction spends about 10%~20% about time, three link model Characteristic matching spends for about 20%~30% time, and the second link Feature Descriptor builds and then consumes about 50%~70% about Time.Therefore, if the structure time of Feature Descriptor in the second link can be reduced, the coupling of SIFT algorithm can be obviously improved Speed.
The problem of second aspect is, when the characteristic point that SIFT algorithm extracts is a lot, build the Feature Descriptors of 128 dimensions to Amount needs to take larger internal memory, and this is totally unfavorable to Embedded Application such as smart mobile phone.If only choosing sub-fraction every time Characteristic point carries out Feature Descriptor structure, then can significantly reduce the requirement to internal memory.
The problem of the third aspect is that the characteristic point that SIFT extracts is usually present gathering (differing only by several pixels) phenomenon, enters And lead to the gathering of match point.One of reason is, in order to ensure scale invariability, SIFT algorithm constructs gaussian pyramid. In different groups of gaussian pyramid, same characteristic point may repeat.Due to processing the presence of error, correspond to artwork When in picture, the characteristic point position repeating not exclusively overlaps, but position is extremely near.However, in the application such as image flame detection, people The degree of scatter of match point, and its effectiveness of information provided to subsequent treatment are more provided.Give one example, if There are 10 match points, wherein 5 match points occur in that clustering phenomena (i.e. 5 match points differ only by several pixels each other), So in image flame detection application, the information that this 5 match points provide is equivalent to the information that a match point provides.In other words Say, in this 10 match points, actually active match point only has 6,4 feature point pairs image flame detection in addition almost do not have Contribution, but increased amount of calculation and the requirement to internal memory of coupling.If characteristic point is divided into the subregion of multiple spatial dispersion, and Randomly select Partial Feature point from every sub-regions to be mated, then can improve feature scattering of points, thus to a certain degree The upper rendezvous problem solving match point.
Content of the invention
The present invention carries out a class method of images match for distinguished point based, such as the stains algorithm such as SIFT, SURF, or Harris and FAST isocenter algorithm, provides a kind of speeding scheme, in the case of not reducing images match precision, significantly improves The speed of images match.Meanwhile, reduce the requirement to internal memory, and solve the rendezvous problem of match point to a certain extent.
Target image and reference picture feature point extraction in the first link for the present invention and the Feature Descriptor of the second link Fishing strategy is added, by the characteristic point subregion of target image, circulation are randomly selected the little portion of all subregion one between structure Point characteristic point, and only this fraction characteristic point is sent into Feature Descriptor and build link, realizes the saving time, saves internal memory, excellent First coupling has the effect of the characteristic point of preferable dispersion.The reason be named fishing is, if by the feature in target image Point is compared to bait, the characteristic point in reference picture is compared to fish, then the Feature Points Matching mistake between target image and reference picture Journey is closely similar with fishing operations.Specific technical scheme is as follows:
The first step, application speckle or Robust Algorithm of Image Corner Extraction extract characteristic point from target image and reference picture, set up ginseng Examine the k dimension search tree of image.If reference picture is it is known that carry out the feature point extraction of reference picture, Feature Descriptor structure in advance Build, k dimension search tree builds and preserves work;This step only need to call the k of reference picture to tie up search tree, and need not carry out with reference to figure As feature point extraction and k dimension search tree are set up.
Second step, all characteristic points of target image is evenly dividing as N × N number of subregion according to position, is each sub-district Domain setting characteristic point chooses ratio ai=a0, i=1 ..., N2, a0 is initial selection ratio.
3rd step, randomly selects the characteristic point that ratio is ai, i=1 ..., N in the subregion i having residue character point2.
4th step, for selecting characteristic point construction feature description.
5th step, the k based on reference picture ties up search tree, describes son and with reference to figure to the target image characteristics of above-mentioned structure As Feature Descriptor carries out characteristic matching.
6th step, Mismatching point is rejected in application RANSAC (random sample consensus).
7th step, calculates total coupling points.If always coupling points >=threshold value TH, terminate matching process;Otherwise execute the 8th Step.
8th step, if the characteristic point ratio participating in target image mating is more than setting value P, terminates matching process;Otherwise, Match condition according to all subregion dynamically adjusts its characteristic point and chooses ratio ai, i=1 ..., N2:If taken out from the i of region Characteristic point have more match point in a reference image, just increase ai, that is, increase selected characteristic point from this region next time Quantity;Whereas if only having less match point in a reference image or not having match point, it is reduced by ai, that is, reduces next The quantity of secondary selected characteristic point from this region.The feature already engaged in coupling is rejected from the characteristic point subset of all subregion Point, goes to the 3rd step.
In fishing strategy, when carrying out N × N uniform segmentation to all characteristic points of target image according to position, N can Take 3,4, or 5.Terminate total coupling points threshold value TH needed for matching process and may be configured as 10, terminate the target needed for matching process Characteristic point ratio P participating in image mating may be configured as 50%, or according to actual needs TH and P is adjusted.Each sub-district The characteristic point in domain choose ratio ai (i=1 ..., N2) setting as follows:Initial selection ratio a0=10%, or work as limited memory system When, counted out according to the treatable feature of each coupling institute and a0 is set;Weighed each using last matching rate mi/ni The match condition of subregion, wherein ni are the total characteristic points participating in last coupling in subregion i, and mi is match point therein Number, the method dynamically adjusting ai according to the matching rate mi/ni of all subregion is as follows:If (mi/ni)<10%, then reduce and choose ratio Example, makes ai=0.5a0;If (mi/ni) >=10%, increase selection ratio, make ai=2a0.
The effect that the present invention is reached and benefit be, first, in the case of pre-stored reference image k dimension search tree, and former Have distinguished point based carry out images match method (same pre-stored reference image k dimension search tree, simply not using fishing Strategy) to compare, images match speed can be improved 5 times about by the present invention.In the case that target image characteristics point is more, image Matching speed can improve more than 5 times.Even if the k dimension search tree of non-pre-stored reference image, the present invention also has 2 times about of speed to carry Rise.Secondly as employ mode construction feature description of batch cycle, and only special for the sub-fraction of target image every time Levy construction feature description, the present invention can save a large amount of internal memories.Again, benefit from characteristic point subregion and randomly select characteristic point Strategy, the characteristic point that the present invention makes preferentially to participate in mate has good dispersion, to some extent solves match point Rendezvous problem.Finally, due to the present invention does not carry out any approximate to SIFT algorithm, and simply pick fraction coupling Potentiality are big, the characteristic point of good dispersion degree carries out Feature Descriptor structure and coupling, so images match precision does not reduce.
Brief description
Accompanying drawing is the flow chart that the present invention carries out images match based on SIFT algorithm.
Specific embodiment
An existing width target image, and reference picture known to a width.Target image is had altogether with the scene of reference picture Same part, but the aspect such as the source of the two, size, illumination, scene coverage is all different.Because of reference picture it is known that carrying Before carried out the feature point extraction of reference picture, Feature Descriptor builds, and k ties up the structure of search tree and preserves work.? Carry out target image characteristic point subregion when, make N=4, will all characteristic points be evenly dividing as 4 × 4 sub-districts according to position Domain.Additionally, initial selection ratio a0=10% of setting characteristic point, terminate required total coupling points threshold value TH=10 of coupling, Terminate the required target image characteristics point of coupling and participate in ratio P=50%.Above-mentioned target image and ginseng are realized based on SIFT algorithm Examine the flow chart of images match as shown in drawings.
The first step, application SIFT algorithm extracts the characteristic point (speckle) of target image, calls the k of reference picture to tie up search Tree.
Second step, all characteristic points of target image are evenly dividing as 4 × 4 sub-regions according to position.Make each sub-district Characteristic point selection ratio ai=10% in domain, i=1 ..., 16.
3rd step, randomly selects the characteristic point that ratio is ai, the spy that record is chosen in the subregion i having residue character point Levying a sum is ni, i=1 ..., 16.
4th step, for selecting characteristic point construction feature description.
5th step, the k based on reference picture ties up search tree, describes son and with reference to figure to the target image characteristics of above-mentioned structure As Feature Descriptor carries out characteristic matching.
6th step, application RANSAC rejects Mismatching point, and the coupling of record subregion i is counted as mi.
7th step, calculates total coupling points.If total coupling points >=10, terminate matching process;Otherwise execute the 8th step.
8th step, if the characteristic point ratio participating in target image mating is more than 50%, terminates matching process;Otherwise, root Matching rate according to all subregion dynamically adjusts ai, i=1 ..., 16:If (mi/ni)<10%, make ai=5%;If (mi/ni) >= 10%, make ai=20%.Reject the characteristic point already engaged in coupling from the characteristic point subset of all subregion, go to the 3rd step.

Claims (10)

1. a kind of distinguished point based carries out the accelerated method of images match it is characterised in that following steps:
The first step, application speckle or Robust Algorithm of Image Corner Extraction extract characteristic point from target image and reference picture, set up with reference to figure The k dimension search tree of picture;If reference picture is it is known that carry out the feature point extraction of reference picture in advance, Feature Descriptor builds, k Dimension search tree builds and preserves work;
Second step, all characteristic points of target image is evenly dividing as N × N number of subregion according to position, is that all subregion sets Put characteristic point and choose ratio ai=a0, i=1 ..., N2, a0 is initial selection ratio;
3rd step, randomly selects the characteristic point that ratio is ai, i=1 ..., N in the subregion i having residue character point2
4th step, for selecting characteristic point construction feature description;
5th step, the k based on reference picture ties up search tree, special with reference picture to target image characteristics description of above-mentioned structure Levy description and carry out characteristic matching;
6th step, application RANSAC rejects Mismatching point;
7th step, calculates total coupling points;If always coupling points >=threshold value TH, terminate matching process;Otherwise execute the 8th step;
8th step, if the characteristic point ratio participating in target image mating is more than setting value P, terminates matching process;Otherwise, according to The match condition of all subregion dynamically adjusts its characteristic point and chooses ratio ai, i=1 ..., N2:If the spy taking out from the i of region Levy and a little have more match point in a reference image, just increase ai, increase the quantity of selected characteristic point from this region next time; Whereas if only having less match point in a reference image or not having match point, it is reduced by ai, reduce next time from this region The quantity of middle selected characteristic point;Reject the characteristic point already engaged in coupling from the characteristic point subset of all subregion, go to the 3rd Step;Wherein, the total characteristic participating in last coupling in subregion i is counted as ni, and coupling points are mi.
2. a kind of distinguished point based according to claim 1 carries out the accelerated method of images match, it is characterized in that, to target When all characteristic points of image carry out N × N uniform segmentation according to position, N takes 3,4 or 5.
3. a kind of distinguished point based according to claim 2 carries out the accelerated method of images match, it is characterized in that, end Join total coupling points threshold value TH needed for process and be set to 10.
4. a kind of distinguished point based according to claim 1,2 or 3 carries out the accelerated method of images match, it is characterized in that, Characteristic point ratio P terminating to participate in the target image needed for matching process mating is set to 50%.
5. a kind of distinguished point based according to claim 1,2 or 3 carries out the accelerated method of images match, it is characterized in that, The characteristic point of all subregion initially chooses ratio a0=10%, or when limited memory, treatable according to each coupling institute Feature is counted out and a0 is set.
6. a kind of distinguished point based according to claim 1,2 or 3 carries out the accelerated method of images match, it is characterized in that, Weigh the match condition of all subregion using last matching rate mi/ni, wherein ni is to participate in last coupling in subregion i Total characteristic points, mi be therein coupling points.
7. a kind of distinguished point based according to claim 4 carries out the accelerated method of images match, it is characterized in that, each sub-district The characteristic point in domain initially chooses ratio a0=10%, or when limited memory, according to the treatable characteristic point of each coupling institute Number sets to a0.
8. a kind of distinguished point based according to claim 7 carries out the accelerated method of images match, it is characterized in that, using upper Matching rate mi/ni once weighs the match condition of all subregion, and wherein ni is the total spy participating in last coupling in subregion i Levy points, mi is coupling points therein.
9. a kind of distinguished point based according to claim 1,2,3,7 or 8 carries out the accelerated method of images match, its feature It is that the method dynamically adjusting ai according to the matching rate mi/ni of all subregion is as follows:If (mi/ni)<10%, then reduce and choose ratio Example, makes ai=0.5a0;If (mi/ni) >=10%, increase selection ratio, make ai=2a0.
10. a kind of distinguished point based according to claim 6 carries out the accelerated method of images match, it is characterized in that, according to The method that the matching rate mi/ni of all subregion dynamically adjusts ai is as follows:If (mi/ni)<10%, then reduce selection ratio, make ai =0.5a0;If (mi/ni) >=10%, increase selection ratio, make ai=2a0.
CN201410392413.9A 2014-08-12 2014-08-12 A speeded up method of executing image matching based on feature points Active CN104182974B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201410392413.9A CN104182974B (en) 2014-08-12 2014-08-12 A speeded up method of executing image matching based on feature points

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201410392413.9A CN104182974B (en) 2014-08-12 2014-08-12 A speeded up method of executing image matching based on feature points

Publications (2)

Publication Number Publication Date
CN104182974A CN104182974A (en) 2014-12-03
CN104182974B true CN104182974B (en) 2017-02-15

Family

ID=51963992

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201410392413.9A Active CN104182974B (en) 2014-08-12 2014-08-12 A speeded up method of executing image matching based on feature points

Country Status (1)

Country Link
CN (1) CN104182974B (en)

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105518709B (en) * 2015-03-26 2019-08-09 北京旷视科技有限公司 The method, system and computer program product of face for identification
CN106202130A (en) * 2015-05-08 2016-12-07 无锡天脉聚源传媒科技有限公司 A kind of method and device of shot segmentation
FR3046691B1 (en) * 2016-01-13 2019-03-29 Commissariat A L'energie Atomique Et Aux Energies Alternatives DEVICE FOR SELECTING AND DESCRIBING POINTS OF INTERESTS IN A SEQUENCE OF IMAGES, FOR EXAMPLE FOR THE MATCHING OF POINTS OF INTERESTS
CN107065895A (en) * 2017-01-05 2017-08-18 南京航空航天大学 A kind of plant protection unmanned plane determines high-tech
CN107229935B (en) * 2017-05-16 2020-12-11 大连理工大学 Binary description method of triangle features
CN107194424B (en) * 2017-05-19 2019-08-27 山东财经大学 A kind of image similar block method for fast searching
CN108021921A (en) * 2017-11-23 2018-05-11 塔普翊海(上海)智能科技有限公司 Image characteristic point extraction system and its application
CN108304863B (en) * 2018-01-12 2020-11-20 西北大学 Terra-cotta warriors image matching method using learning invariant feature transformation
CN108764297B (en) * 2018-04-28 2020-10-30 北京猎户星空科技有限公司 Method and device for determining position of movable equipment and electronic equipment
CN110276289B (en) * 2019-06-17 2021-09-07 厦门美图之家科技有限公司 Method for generating matching model and face characteristic point tracking method

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102722731A (en) * 2012-05-28 2012-10-10 南京航空航天大学 Efficient image matching method based on improved scale invariant feature transform (SIFT) algorithm
CN103136751A (en) * 2013-02-05 2013-06-05 电子科技大学 Improved scale invariant feature transform (SIFT) image feature matching algorithm

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8189925B2 (en) * 2009-06-04 2012-05-29 Microsoft Corporation Geocoding by image matching
US9235780B2 (en) * 2013-01-02 2016-01-12 Samsung Electronics Co., Ltd. Robust keypoint feature selection for visual search with self matching score

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102722731A (en) * 2012-05-28 2012-10-10 南京航空航天大学 Efficient image matching method based on improved scale invariant feature transform (SIFT) algorithm
CN103136751A (en) * 2013-02-05 2013-06-05 电子科技大学 Improved scale invariant feature transform (SIFT) image feature matching algorithm

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Li Y et al..Fast SIFT algorithm based on Sobel edge detector.《2nd IEEE International Conference on Consumer Electronics, Communications and Networks (CECNet)》.2012, *
基于SITF特征提取的加速匹配方法;程丹等;《科技创新导报》;20131231(第21期);14-18 *
基于区域分块的SIFT图像匹配技术研究与实现;杜京义等;《光电工程》;20130831;第40卷(第8期);52-58 *

Also Published As

Publication number Publication date
CN104182974A (en) 2014-12-03

Similar Documents

Publication Publication Date Title
CN104182974B (en) A speeded up method of executing image matching based on feature points
CN107633209B (en) Electronic device, the method for dynamic video recognition of face and storage medium
CN110414533A (en) A kind of feature extracting and matching method for improving ORB
CN107301402B (en) Method, device, medium and equipment for determining key frame of real scene
Zhang et al. Multiple-level feature-based measure for retargeted image quality
CN109961399B (en) Optimal suture line searching method based on image distance transformation
CN108769821A (en) Scene of game describes method, apparatus, equipment and storage medium
CN104134200B (en) Mobile scene image splicing method based on improved weighted fusion
CN105427333B (en) Real-time Registration, system and the camera terminal of video sequence image
CN108022262A (en) A kind of point cloud registration method based on neighborhood of a point center of gravity vector characteristics
AU2013237718A1 (en) Method, apparatus and system for selecting a frame
CN103927785B (en) A kind of characteristic point matching method towards up short stereoscopic image data
US11023781B2 (en) Method, apparatus and device for evaluating image tracking effectiveness and readable storage medium
CN105335952B (en) Matching power flow computational methods and device and parallax value calculating method and equipment
CN110825900A (en) Training method of feature reconstruction layer, reconstruction method of image features and related device
CN110956078B (en) Power line detection method and device
CN105913069A (en) Image identification method
CN108229583B (en) Method and device for fast template matching based on main direction difference characteristics
CN109815763A (en) Detection method, device and the storage medium of two dimensional code
CN113610178A (en) Inland ship target detection method and device based on video monitoring image
CN106909552A (en) Image retrieval server, system, coordinate indexing and misarrangement method
CN110135422B (en) Dense target detection method and device
Matusiak et al. Unbiased evaluation of keypoint detectors with respect to rotation invariance
CN116342955A (en) Target detection method and system based on improved feature pyramid network
CN107464257B (en) Wide base line matching method and device

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