CN106558072A - A kind of method based on SIFT feature registration on remote sensing images is improved - Google Patents
A kind of method based on SIFT feature registration on remote sensing images is improved Download PDFInfo
- Publication number
- CN106558072A CN106558072A CN201611031136.4A CN201611031136A CN106558072A CN 106558072 A CN106558072 A CN 106558072A CN 201611031136 A CN201611031136 A CN 201611031136A CN 106558072 A CN106558072 A CN 106558072A
- Authority
- CN
- China
- Prior art keywords
- feature
- point
- remote sensing
- histogram
- sift
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30181—Earth observation
Abstract
The present invention is claimed a kind of being based on and improves SIFT feature in remote sensing image registration research.For remote sensing image data amount it is big, Duplication is high the characteristics of, the deficiency on the matching error that causes on the complicated factor such as illumination and disparity, matching efficiency, the improved method that have extensively studied SIFT feature descriptor:Using characteristic point border circular areas come structural features descriptor, it is used for the division of regional area and the calculating of histogram of gradients using adaptive quantizing strategy, and principal direction is established again using a kind of interpolation method to each description, improve the feature descriptor of SIFT algorithms;Then a kind of new SIFT feature matching algorithm based on divided-fit surface is adopted, the time loss of feature extraction and matching is reduced by rejecting Non-overlapping Domain, while and ensureing the unique and robustness of feature point description symbol.
Description
Technical field
The present invention relates to field of remote sensing image processing, specifically a kind of to be matched somebody with somebody on remote sensing images based on improvement SIFT feature
Accurate method.
Background technology
It is many that current remote sensing image processing has been widely used in environmental monitoring, topographic(al) reconnaissance, military investigation, disaster alarm etc.
Individual field, wherein image registration are a committed steps, and the accuracy of registration and real-time, there is very big shadow to these applications
Ring.Therefore, many method for registering have been suggested and have solved this problem, and it is divided into two classifications:Method based on region
With the method for feature based.Due to the method based on region it is quite quick to geometry and Nonlinear Grey distortion in remote sensing images
Sense, and remote sensing images generally need substantial amounts of data processing, therefore side of the method for registering images ratio of feature based based on region
Method is more suitable for.The method that its midpoint is characterized as basis, Scale invariant features transform (SIFT) have been widely used in remote sensing figure
As registration,, in image scaling and rotation, significantly the affine deformation of scope, all achieves stably to enter in terms of the change of 3d viewpoint for which
Exhibition.As remote sensing figure has the complicated factors such as different illumination variations, SIFT algorithms are using different sensor images with punctual meeting
Produce many Mismatching points, and real-time is poor for remote sensing image data amount is big, for the characteristics of Duplication is high.To this
The present invention is targetedly improved to the SIFT registration Algorithms of distinguished point based, is proposed a kind of based on area dividing thought
New SIFT matching methods, are optimized to its flow process, make the algorithm all improve in speed and precision.
The content of the invention
Problem to be solved by this invention is the spy that image data amount is big, degree of overlapping is high obtained for remote sensing platform
Point, when illumination variation, rotationally-varying, dimensional variation, geometry deformation, fuzzy and compression of images, traditional SIFT is calculated
Method characteristic matching often brings that precision is high, the low problem of efficiency, and the present invention proposes a kind of improved SIFT descriptors
Building method, and the new SIFT matching methods of application region section thinking on this basis, while characteristic point dimension is reduced,
Take into full account that the different pixel of distance feature point affects different this feature to feature point description symbol in neighborhood.Not only reduce
The time complexity of algorithm, also enhances the unique and robustness of descriptor.
Technical scheme is as follows:
A kind of method based on SIFT feature registration on remote sensing images is improved, which comprises the following steps:
Step 1:Two remote sensing images to be matched respectively to being input into build metric space, for simulated image data
Analysis On Multi-scale Features;
Step 2:Determine the characteristic point under each metric space, and with the position of the three-dimensional quadratic function accurate feature points of fitting
And yardstick;
Step 3:For each characteristic point assigned direction parameter, principal direction is established again using interpolation method, so as to get operator
With rotational invariance;
Step 4:Neighborhood centered on characteristic point is played a game using circular sampling window, employing adaptive quantizing strategy
Portion region is divided and histogram of gradients is calculated, and generates 96 dimensional feature descriptor Expressive Features points;
Step 5:The correlating transforms matrix of the feature point description vector two width images of estimation determined according to SIFT algorithms, is carried out
The calculating of overlapping region and the cutting of image block, so as to the match block to obtaining two width images carries out SIFT feature Vectors matching.
Further, the step 1) multiscale space of remote sensing images is built, figure is built using Gaussian convolution core
The Analysis On Multi-scale Features of metric space simulated image data.
Further, the step 2) extreme point detection is carried out to the metric space of remote sensing source images is by each sampled point
The all consecutive points of metric space that will be located with it are compared, if a point is in DOG this layer of metric space and up and down
When being maximum or minima in two-layer, it is a characteristic point of the image under the yardstick to be considered as the point.By fitting three-dimensional two
The position of secondary function precise positioning feature point and yardstick, remove the key point and unstable skirt response point of low contrast.It is logical
The position and yardstick for fitting three-dimensional quadratic function precise positioning feature point is crossed, the key point of low contrast and unstable side is removed
Edge response point.
Further, the step 3) Feature Descriptor principal direction formula is redefined based on interpolation method it is:
Bin in formulanewRepresent newly-generated principal direction, bininiRepresent former principal direction, mainl, main and mainrDifference table
Show on the left of histogram peak, the value of the post on peak value and right side.
Further, the employing adaptive quantizing strategy is divided to image local area subject to registration and straight to gradient
Square figure is calculated, and generating feature descriptor includes step:
(1) centered on local feature region, normalized portion region R according to different step-length l={ l1,l2,...lm}
It is separated into m nonoverlapping annulus R (1), R (2) ..., R (m);
(2) using adaptive gradient orientation histogram number K={ k1,k2,...km, one is generated in each regional area
Constant rectangular histogram;
(3) uniform angular quantification is carried out to every sub-regions R (m), quantization number is n, is thus assured that every height
Region R (i, j), wherein i=1,2 ... m }, j=1,2 ... n }.
Further, for every sub-regions R (i), its rectangular histogram H (i, j) is with kiWhat quantized directions were calculated, use
The histogram of gradients number of variable variable pixel size:K={ k1,k2,...,km, combine the histogram of gradients of all of subregion
Value formed descriptor D, be defined as follows:D=H (1,1) ... H (i, j) ... H (m, n) } wherein m is radially to quantify number, n is
Subregion R (i) angular quantification number, H (i, j) are the direction histograms of R (i, j), therefore the dimension calculation of descriptor is as follows:
Further, the step of 96 dimensional feature descriptor of the generation it is:Using descriptor dimension calculation parameter:Radius vector
Change number m=4;Step-length number of rings l={ 3,2,2, } 1;Angular quantification number n=4;Histogram of gradients number k={ 8,6,6,4 }, thus may be used
Design improves SIFT arthmetic statements symbol sampling model figure, SIFT descriptor yardsticks:
Four annulus are divided using the number of rings of different step-lengths, i.e., 3,2,2,1,4 are counted respectively with histogram of gradients
4 seed points are formed in the gradient accumulated value of the different gradient directions in annulus, each annulus, 16 seed points are had;Distance
The different annular of central feature point takes the individual direction vector information of k={ 8,6,6,4 } respectively, counts the feature that each seed point is generated
Vector, is always obtained the description subvector of 96 dimensions, and order is arranged the characteristic vector that every sub-regions are generated from inside to outside, sub-district
The characteristic vector in domain 1 is tieed up as the 1~32 of foremost;The characteristic vector of subregion 2 is used as 33~56 dimensions;The feature of subregion 3
Vector is used as 57~80 dimensions;The characteristic vector of subregion 4 adds up as 81~96 dimensions, the gray scale finally counted in each sub-regions
Value, and be normalized.
Further, the step 5) calculating of overlapping region and the dicing step of image block include:
(1) optimal transform matrix according to two width images, calculates overlapping region;
(2) in overlapping region, selected part seed point will overlap cutting estimating the correlating transforms matrix of two width images
For several pieces, the related blocks in matching figure are found out by transformation matrix;
(3) characteristic point in all of match block is matched.
Advantages of the present invention and have the beneficial effect that:
Step 3) as SIFT is to gamma characteristic sensitive, easily judge by accident in the extreme point that gamma characteristic is changed greatly.
Principal direction is established again using a kind of interpolation method to each description,
Bin in formulanewRepresent newly-generated principal direction, bininiRepresent former principal direction, mainl, main and mainrDifference table
Show on the left of histogram peak, the value of the post on peak value and right side.
The step 4) impact due to the nearer field pixel of distance feature point to feature point description is bigger, it is proposed that one
New adaptive strategy is planted, the sub- building method of description of uniqueness is employed.By being used for partial zones using adaptive quantizing strategy
The division in domain and the calculating of histogram of gradients, adopt and successively decrease number of rings to represent feature neighborhood of a point region;Instead of using single
Histogram of gradients is accorded with as existing partial descriptions, uses various sizes of multiple ladders according to the distance apart from local feature center
Degree rectangular histogram.Specific practice is as follows:
1., centered on local feature region, normalized portion region R is separated according to different step-lengths l={ 1,2,2,3 }
Into 4 nonoverlapping annulus R (1), R (2), R (3), R (4).
2. adaptive gradient orientation histogram number K={ 4,6,6,8 } is adopted, and this is different from other sides based on distribution
Method generates a constant rectangular histogram in each regional area.
3. pair every sub-regions R (m) carry out uniform angular quantification, and it is 4 to quantify number, is thus assured that each sub-district
Domain R (i, j), wherein i=1,2 ... 4 }, j=1,2 ... 4 }.
For every sub-regions R (i), its rectangular histogram H (i, j) is with kiWhat quantized directions were calculated, combine all of son
The value of the histogram of gradients in region forms descriptor D, is defined as follows:
D=H (1,1) ... H (i, j) ... H (m, n) }
Wherein m is radially to quantify number, and n is subregion R (i) angular quantification number, and H (i, j) is the direction histogram of R (i, j),
Therefore the dimension calculation of descriptor is as follows:
The present invention employs a kind of new adaptive strategy structure for the feature that remote sensing image data amount is big, Duplication is high
Feature Descriptor is made, the Feature Descriptor of 96 dimensions is finally established according to the parameter value for adopting, area dividing is based on a kind of
The new SIFT matching methods of thought.While characteristic point dimension is reduced, take into full account that distance feature point is different in neighborhood
Pixel affects different this feature to feature point description symbol.The time complexity of algorithm is not only reduced, descriptor is also enhanced
Unique and robustness.
Description of the drawings
Fig. 1 is that histogram of gradients determines main gradient direction schematic diagram;
Fig. 2 is that the present invention provides preferred embodiment SIFT feature point description son construction;
Fig. 3 is that the present invention provides comparison diagram before and after preferred embodiment improvement SIFT feature descriptor sampling model;
Fig. 4 is that the present invention matches comparison diagram using SIFT feature is improved to unmanned aerial vehicle remote sensing image;
Fig. 5 is that the present invention provides preferred embodiment overlapping region estimation illustraton of model;
Fig. 6 is that the present invention provides preferred embodiment SIFT algorithmic match flow charts.
Specific embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, detailed
Carefully describe.Described embodiment is only a part of embodiment of the present invention.
Technical scheme is as follows:
SIFT descriptors are that, based on distribution, it represents outward appearance or variform feature with the straight figure in direction, such as Fig. 1,
2.For this purpose, local characteristic region is divided into different subregions, and calculates the specific rectangular histogram per sub-regions, most
Whole feature point description symbol is obtained by these histogrammic orderly cascades.This often in quality, quantity, extracts dividing for feature
There are existing problems in cloth, the sensitivity to intensity of illumination and disparity, particularly in remote sensing images.The thinking of the present invention
It is, come structural features descriptor, division and the ladder of regional area to be used for using adaptive quantizing strategy using characteristic point border circular areas
Histogrammic calculating is spent, and principal direction is established again using a kind of interpolation method to each description, improve the description of SIFT algorithms
Symbol, the SIFT feature descriptor for thus constructing are as shown in Figure 3.While characteristic point dimension is reduced, ensure feature point description again
The unique and robustness of symbol.Flow chart as shown in Figure 6, specifically in accordance with the following steps:
(1) detect yardstick spatial extrema
The metric space (DoG) of figure, the Analysis On Multi-scale Features of simulated image data are built using Gaussian convolution core.Each is adopted
The all consecutive points of metric space that sampling point will be located with it are compared, as shown in figure 1, the test point of centre and its same yardstick
8 consecutive points and corresponding 9 × 2 points of neighbouring yardstick totally 26 points compare, to guarantee in metric space and X-Y scheme
Image space all detects extreme point.If a point is most in DOG this layer of metric space and bilevel 26 fields
During big or minima, it is a characteristic point of the image under the yardstick to be considered as the point.
(2) determine the position of characteristic point
Position and the yardstick (reaching sub-pixel precision) of key point are accurately determined by fitting three-dimensional quadratic function, while
The key point and unstable skirt response point (because DoG operators can produce stronger skirt response) of low contrast are removed, with
Strengthen matching stability, improve noise resisting ability.The very asymmetric pixel of DoG local curvatures will substantially be removed.
(3) assigned characteristics point direction
Determine in previous step per the characteristic point in width figure, be that each characteristic point calculates a direction, according to this direction
Further calculated, be each key point assigned direction parameter using the gradient direction distribution characteristic of key point neighborhood territory pixel,
Operator is made to possess rotational invariance.
θ (x, y)=α tan2 ((L (x, y+1)-L (x, y-1))/(L (x+1, y)-L (x-1, y))) (2)
Formula (1) and formula (2) are respectively the modulus value and direction formula of (x, y) place gradient.Yardstick used by wherein L is closed for each
The yardstick that key point is each located.
Principal direction of traditional histogrammic peak value of SIFT algorithms selections as characteristic point, others reach maximum
80% direction can be used as auxiliary direction, and the principal direction of characteristic point is to ensure that the key of invariable rotary feature, if principal direction is
It is inaccurate, then it can be with the result of effect characteristicses Point matching.The present invention proposes to be based on
Interpolation method is used for redefining principal direction:
Bin in formulanewRepresent newly-generated principal direction, bininiRepresent former principal direction, mainl, main and mainrDifference table
Show on the left of histogram peak, the value of the post on peak value and right side, be further ensured that the robustness of characteristic point.(4) generate SIFT feature
Vector
Traditional SIFT algorithm characteristics point uses the feature descriptor of 128 dimensions, big, heavy for remote sensing image data amount
The characteristics of folded rate is high is matched on matching accuracy, matching efficiency all Shortcomings with SIFT algorithms.Therefore it is of the invention
Using a kind of new adaptive strategy, it is proposed that unique sub- building method of description.By being used for using adaptive quantizing strategy
The division of regional area and the calculating of histogram of gradients, adopt and successively decrease number of rings to represent feature neighborhood of a point region;Instead of using
Single histogram of gradients is accorded with as existing partial descriptions, according to the distance apart from local feature center using various sizes of
Multiple histogram of gradients.Specific practice is as follows:
1., centered on local feature region, normalized portion region R is separated according to different step-lengths l={ 1,2,2,3 }
Into 4 nonoverlapping annulus R (1), R (2), R (3), R (4).
2. adaptive gradient orientation histogram number K={ 4,6,6,8 } is adopted, and this is different from other sides based on distribution
Method generates a constant rectangular histogram in each regional area.
3. pair every sub-regions R (m) carry out uniform angular quantification, and it is 4 to quantify number, is thus assured that each sub-district
Domain R (i, j), wherein i=1,2 ... 4 }, j=1,2 ... 4 }.
For every sub-regions R (i), its rectangular histogram H (i, j) is with kiWhat quantized directions were calculated, combine all of son
The value of the histogram of gradients in region forms descriptor D, is defined as follows:
D=H (1,1) ... H (i, j) ... H (m, n) }
Wherein m is radially to quantify number, and n is subregion R (i) angular quantification number, and H (i, j) is the direction histogram of R (i, j),
Therefore the dimension calculation of descriptor is as follows:
Thereby determine that the description subvector of available 96 dimension, it is contemplated that the pixel the closer to central feature point is retouched to generation
State that sub- impact is bigger, order is arranged the characteristic vector that every sub-regions are generated from inside to outside.The characteristic vector conduct of subregion 1
1~32 dimension of foremost;The characteristic vector of subregion 2 is used as 33~56 dimensions;The characteristic vector of subregion 3 is used as 57~80 dimensions;
The characteristic vector of subregion 4 is used as 81~96 dimensions.The gray scale accumulated value in each sub-regions is finally counted, and is carried out normalizing
Change is processed, and further removes the impact of the complicated factors such as illumination, noise.The method that the present invention proposes structural features description,
While reducing characteristic point dimension, take into full account that the different pixel of distance feature point is affected not on feature point description symbol in neighborhood
With this feature.The time complexity of algorithm is not only reduced, the unique and robustness of descriptor is also enhanced.
(5) as Fig. 4 overlapping regions estimate model, it is assumed that a pixel of image be (x, y), the point after conversion
For (x', y'), then the transformation relation between them can be represented with following formula:
Wherein matrix H is called homography matrix, and it is the matrix of 3 × 3, has 8 unknown parameters, and each parameter is distinguished
Represent different transformation relations.Through deriving:
Understand only need to 4 pairs of matching double points can be to obtain 8 variables in matrix H, so that it is determined that between two width images
Geometric transform relation.
Therefore the perspective transform coordinate of 4 angular coordinates on right image on the left image of correspondence can be calculated according to formula (3),
The overlapping region of this two width image with this tetragon of 4 perspective transform coordinates as summit as right image;Can try to achieve in the same manner
The overlapping region of two images on left image.According to the stereogram overlapping region scope tried to achieve, with left overlapping region as base
Standard, carries out piecemeal.For the pixel coordinate of 4 angle points of left each sub-block got, close also according to perspective transform
System, can obtain each match block of left image corresponding overlap matching block on right image.
(5) characteristic matching
After obtaining two width image overlap subregion scopes of left and right, characteristic point is carried out to the subregion match block of the same name
Match somebody with somebody.We are used as the similarity determination tolerance of key point in two width images using the Euclidean distance of key point characteristic vector.Take
Certain key point in source images, and find out its with target image European closest the first two key point, at the two
In key point, if nearest distance is less than certain proportion threshold value R divided by secondary near distance, receive this pair of match points.Drop
Low this proportion threshold value Threshold, SIFT matchings are counted out and can be reduced, but more stable.Traditional SIFT algorithmic match and
Experimental comparison such as Fig. 5 of SIFT algorithmic match after being improved using the present invention, experiment show that the matching efficiency of the present invention can be improved
20%~25%, while and can significantly reduce error hiding rate, it is ensured that the unique and robustness of feature descriptor.
The features such as inclination angle of, image many compared with little, quantity is excessive for remote sensing images film size and incline direction does not have rule, profit
Registration is carried out to low latitude image with traditional SIFT algorithms, it may appear that arithmetic speed is slow, efficiency is low and Mismatching point is more etc.
Problem, so as to result in, splicing effect is undesirable, and this present invention is directed to the SIFT registration Algorithms of distinguished point based
The improvement of property, using a kind of new SIFT feature matching algorithm based on divided-fit surface, is optimized to its flow process, makes the calculation
Method all improves in speed and precision, and method flow proposed by the present invention is as shown in Figure 6.
The above embodiment is interpreted as being merely to illustrate the present invention rather than limits the scope of the invention.
After the content of the record for having read the present invention, technical staff can be made various changes or modifications to the present invention, these equivalent changes
Change and modification equally falls into the scope of the claims in the present invention.
Claims (10)
1. it is a kind of based on improve SIFT feature on remote sensing images registration method, it is characterised in that comprise the following steps:
Step 1:Two remote sensing images to be matched respectively to being input into build metric space, for many chis of simulated image data
Degree feature;
Step 2:Determine the characteristic point under each metric space, and the position with the three-dimensional quadratic function accurate feature points of fitting and chi
Degree;
Step 3:For each characteristic point assigned direction parameter, principal direction is established again using interpolation method, so as to get operator have
Rotational invariance;
Step 4:Neighborhood centered on characteristic point using circular sampling window, using with adaptive quantizing strategy to partial zones
Domain is divided and histogram of gradients is calculated, and generates 96 dimensional feature descriptor Expressive Features points;
Step 5:The correlating transforms matrix of the feature point description vector two width images of estimation determined according to SIFT algorithms, is overlapped
The calculating in region and the cutting of image block, so as to the match block to obtaining two width images carries out SIFT feature Vectors matching.
2. it is according to claim 1 based on improve SIFT feature on remote sensing images registration method, it is characterised in that step
The rapid metric space for 1) building remote sensing images builds many chis of the metric space simulated image data of figure using Gaussian convolution core
Degree feature.
3. it is according to claim 1 based on improve SIFT feature on remote sensing images registration method, it is characterised in that institute
State step 2) metric space of remote sensing images is built using the metric space simulated image data of Gaussian convolution core structure figure
Analysis On Multi-scale Features, carry out extreme point detection to metric space, all phases of metric space that each sampled point will be located with it
Adjoint point is compared, if a point is maximum or minima in DOG this layer of metric space and upper and lower two-layer, is considered as
The point is a characteristic point of the image under the yardstick.
4. the method based on SIFT feature registration on remote sensing images is improved according to claim 3, it is characterised in that special
It is by fitting the position of three-dimensional quadratic function precise positioning feature point and yardstick, removing the key point of low contrast to levy point location
With unstable skirt response point.
5. it is according to claim 3 based on improve SIFT feature on remote sensing images registration method, it is characterised in that institute
State step 3) characteristic point principal direction formula is redefined based on interpolation method it is:
Bin in formulanewRepresent newly-generated principal direction, bininiRepresent former principal direction, mainl, main and mainrRepresent respectively straight
On the left of square figure peak value, the value of the post on peak value and right side.
6. it is according to claim 5 based on improve SIFT feature on remote sensing images registration method, it is characterised in that institute
State step 4) image local area subject to registration is divided using adaptive quantizing strategy and histogram of gradients is calculated,
Generating feature descriptor includes step:
(1) centered on local feature region, normalized portion region R according to different step-length l={ l1,l2,...lmSeparate
Into m nonoverlapping annulus R (1), R (2) ..., R (m);
(2) using adaptive gradient orientation histogram number K={ k1,k2,...km, a constant is generated in each regional area
Rectangular histogram;
(3) uniform angular quantification is carried out to every sub-regions R (m), quantization number is n, is thus assured that every sub-regions R
(i, j), wherein i=1,2 ... m }, j=1,2 ... n }.
7. it is according to claim 6 based on improve SIFT feature on remote sensing images registration method, it is characterised in that:It is right
In every sub-regions R (i), its rectangular histogram H (i, j) is with kiWhat quantized directions were calculated, using variable variable pixel size
Histogram of gradients number:K={ k1,k2,...,km, the value for combining the histogram of gradients of all of subregion forms descriptor D, fixed
Justice is as follows:D=H (1,1) ... H (i, j) ... H (m, n) } wherein m is radially to quantify number, n is subregion R (i) angular quantification
Number, H (i, j) is the direction histogram of R (i, j), therefore the dimension calculation of descriptor is as follows:
8. it is according to claim 6 based on improve SIFT feature on remote sensing images registration method, it is characterised in that:Institute
Stating the step of generating 96 dimensional feature descriptor is:Using descriptor dimension calculation parameter:Radially quantify number m=4;Step-length number of rings l
={ 3,2,2, } 1;Angular quantification number n=4;Histogram of gradients number k={ 8,6,6,4 }, thus can design improvement SIFT algorithms and retouch
State symbol sampling model figure, SIFT descriptor yardsticks:
Four annulus are divided using the number of rings of different step-lengths, i.e., 3,2,2,1,4 annulus are counted respectively with histogram of gradients
4 seed points are formed in the gradient accumulated value of interior different gradient directions, each annulus, 16 seed points are had;Distance center
The different annular of characteristic point takes the individual direction vector information of k={ 8,6,6,4 } respectively, count feature that each seed point generates to
Amount, is always obtained the description subvector of 96 dimensions, and order is arranged the characteristic vector that every sub-regions are generated from inside to outside, subregion
1 characteristic vector is tieed up as the 1~32 of foremost;The characteristic vector of subregion 2 is used as 33~56 dimensions;The feature of subregion 3 to
Amount is used as 57~80 dimensions;The characteristic vector of subregion 4 adds up as 81~96 dimensions, the gray scale finally counted in each sub-regions
Value, and be normalized.
9. it is according to claim 1 based on improve SIFT feature on remote sensing images registration method, it is characterised in that:Institute
State step 5) calculating of overlapping region and the dicing step of image block include:
(1) optimal transform matrix according to two width images, calculates overlapping region;
(2) in overlapping region, selected part seed point will overlap cutting for several estimating the correlating transforms matrix of two width images
Block, finds out the related blocks in matching figure by transformation matrix;
(3) characteristic point in all of match block is matched.
10. it is according to claim 9 based on improve SIFT feature on remote sensing images registration method, it is characterised in that:
The SIFT feature Vectors matching that carries out includes:It is used as in two width images using the Euclidean distance of key point characteristic vector crucial
The similarity determination tolerance of point, takes certain key point in source images, and it is European closest with target image to find out which
The first two key point, in the two key points, if nearest distance is less than certain proportion threshold value R divided by secondary near distance,
Then receive this pair of match points.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611031136.4A CN106558072A (en) | 2016-11-22 | 2016-11-22 | A kind of method based on SIFT feature registration on remote sensing images is improved |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611031136.4A CN106558072A (en) | 2016-11-22 | 2016-11-22 | A kind of method based on SIFT feature registration on remote sensing images is improved |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106558072A true CN106558072A (en) | 2017-04-05 |
Family
ID=58444615
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611031136.4A Pending CN106558072A (en) | 2016-11-22 | 2016-11-22 | A kind of method based on SIFT feature registration on remote sensing images is improved |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106558072A (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106990412A (en) * | 2017-05-11 | 2017-07-28 | 哈尔滨工业大学 | A kind of method for registering images of interference synthetic aperture laser radar system |
CN107301661A (en) * | 2017-07-10 | 2017-10-27 | 中国科学院遥感与数字地球研究所 | High-resolution remote sensing image method for registering based on edge point feature |
CN107665479A (en) * | 2017-09-05 | 2018-02-06 | 平安科技(深圳)有限公司 | A kind of feature extracting method, panorama mosaic method and its device, equipment and computer-readable recording medium |
CN109711444A (en) * | 2018-12-18 | 2019-05-03 | 中国科学院遥感与数字地球研究所 | A kind of new remote sensing image matching method based on deep learning |
CN109978862A (en) * | 2019-03-27 | 2019-07-05 | 北京邮电大学 | A kind of air pollution estimation method based on satellite image |
CN110197184A (en) * | 2019-04-19 | 2019-09-03 | 哈尔滨工业大学 | A kind of rapid image SIFT extracting method based on Fourier transformation |
CN110210491A (en) * | 2019-04-24 | 2019-09-06 | 哈尔滨工业大学 | Rapid image feature extracting method based on FAST and SIFT |
CN111275053A (en) * | 2020-01-16 | 2020-06-12 | 北京联合大学 | Method and system for representing local feature descriptor |
CN111666858A (en) * | 2020-05-29 | 2020-09-15 | 中国科学院地理科学与资源研究所 | Forest remote sensing image registration method and system based on single tree recognition |
CN111695396A (en) * | 2019-12-19 | 2020-09-22 | 珠海大横琴科技发展有限公司 | Remote sensing image matching method and device, electronic equipment and readable storage medium |
CN112149728A (en) * | 2020-09-22 | 2020-12-29 | 成都智遥云图信息技术有限公司 | Rapid multi-modal image template matching method |
CN112784761A (en) * | 2021-01-26 | 2021-05-11 | 哈尔滨理工大学 | Special texture background remote sensing image matching method |
CN113221921A (en) * | 2021-05-21 | 2021-08-06 | 吉林大学 | SIFT feature rapid extraction method based on optimal space |
CN115205562A (en) * | 2022-07-22 | 2022-10-18 | 四川云数赋智教育科技有限公司 | Random test paper registration method based on feature points |
CN116958172A (en) * | 2023-08-01 | 2023-10-27 | 金景(海南)科技发展有限公司 | Urban protection and update evaluation method based on three-dimensional space information |
CN117333368A (en) * | 2023-10-10 | 2024-01-02 | 南京矩视科技有限公司 | Image stitching method, device and storage medium based on local edge analysis |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20070110306A1 (en) * | 2005-11-14 | 2007-05-17 | Haibin Ling | Diffusion distance for histogram comparison |
CN104834931A (en) * | 2015-03-13 | 2015-08-12 | 江南大学 | Improved SIFT algorithm based on wavelet transformation |
-
2016
- 2016-11-22 CN CN201611031136.4A patent/CN106558072A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20070110306A1 (en) * | 2005-11-14 | 2007-05-17 | Haibin Ling | Diffusion distance for histogram comparison |
CN104834931A (en) * | 2015-03-13 | 2015-08-12 | 江南大学 | Improved SIFT algorithm based on wavelet transformation |
Non-Patent Citations (2)
Title |
---|
李校林等: "基于SIFT 算法在遥感图像匹配中的研究", 《电视技术》 * |
袁杰: "基于SIFT的图像配准与拼接技术研究", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (24)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106990412B (en) * | 2017-05-11 | 2019-06-11 | 哈尔滨工业大学 | A kind of method for registering images of interference synthetic aperture laser radar system |
CN106990412A (en) * | 2017-05-11 | 2017-07-28 | 哈尔滨工业大学 | A kind of method for registering images of interference synthetic aperture laser radar system |
CN107301661B (en) * | 2017-07-10 | 2020-09-11 | 中国科学院遥感与数字地球研究所 | High-resolution remote sensing image registration method based on edge point features |
CN107301661A (en) * | 2017-07-10 | 2017-10-27 | 中国科学院遥感与数字地球研究所 | High-resolution remote sensing image method for registering based on edge point feature |
CN107665479A (en) * | 2017-09-05 | 2018-02-06 | 平安科技(深圳)有限公司 | A kind of feature extracting method, panorama mosaic method and its device, equipment and computer-readable recording medium |
WO2019047284A1 (en) * | 2017-09-05 | 2019-03-14 | 平安科技(深圳)有限公司 | Methods for feature extraction and panoramic stitching, and apparatus thereof, device, readable storage medium |
CN109711444A (en) * | 2018-12-18 | 2019-05-03 | 中国科学院遥感与数字地球研究所 | A kind of new remote sensing image matching method based on deep learning |
CN109978862A (en) * | 2019-03-27 | 2019-07-05 | 北京邮电大学 | A kind of air pollution estimation method based on satellite image |
CN110197184A (en) * | 2019-04-19 | 2019-09-03 | 哈尔滨工业大学 | A kind of rapid image SIFT extracting method based on Fourier transformation |
CN110210491A (en) * | 2019-04-24 | 2019-09-06 | 哈尔滨工业大学 | Rapid image feature extracting method based on FAST and SIFT |
CN111695396A (en) * | 2019-12-19 | 2020-09-22 | 珠海大横琴科技发展有限公司 | Remote sensing image matching method and device, electronic equipment and readable storage medium |
CN111275053B (en) * | 2020-01-16 | 2023-11-10 | 北京腾信软创科技股份有限公司 | Method and system for representing local feature descriptors |
CN111275053A (en) * | 2020-01-16 | 2020-06-12 | 北京联合大学 | Method and system for representing local feature descriptor |
CN111666858A (en) * | 2020-05-29 | 2020-09-15 | 中国科学院地理科学与资源研究所 | Forest remote sensing image registration method and system based on single tree recognition |
CN112149728A (en) * | 2020-09-22 | 2020-12-29 | 成都智遥云图信息技术有限公司 | Rapid multi-modal image template matching method |
CN112149728B (en) * | 2020-09-22 | 2023-11-17 | 成都智遥云图信息技术有限公司 | Rapid multi-mode image template matching method |
CN112784761A (en) * | 2021-01-26 | 2021-05-11 | 哈尔滨理工大学 | Special texture background remote sensing image matching method |
CN113221921B (en) * | 2021-05-21 | 2022-05-27 | 吉林大学 | SIFT feature rapid extraction method based on optimal space, computer equipment and storage medium |
CN113221921A (en) * | 2021-05-21 | 2021-08-06 | 吉林大学 | SIFT feature rapid extraction method based on optimal space |
CN115205562A (en) * | 2022-07-22 | 2022-10-18 | 四川云数赋智教育科技有限公司 | Random test paper registration method based on feature points |
CN115205562B (en) * | 2022-07-22 | 2023-03-14 | 四川云数赋智教育科技有限公司 | Random test paper registration method based on feature points |
CN116958172A (en) * | 2023-08-01 | 2023-10-27 | 金景(海南)科技发展有限公司 | Urban protection and update evaluation method based on three-dimensional space information |
CN116958172B (en) * | 2023-08-01 | 2024-01-30 | 金景(海南)科技发展有限公司 | Urban protection and update evaluation method based on three-dimensional space information |
CN117333368A (en) * | 2023-10-10 | 2024-01-02 | 南京矩视科技有限公司 | Image stitching method, device and storage medium based on local edge analysis |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106558072A (en) | A kind of method based on SIFT feature registration on remote sensing images is improved | |
CN104778721B (en) | The distance measurement method of conspicuousness target in a kind of binocular image | |
Drost et al. | 3d object detection and localization using multimodal point pair features | |
Novatnack et al. | Scale-dependent/invariant local 3D shape descriptors for fully automatic registration of multiple sets of range images | |
CN103400151B (en) | The optical remote sensing image of integration and GIS autoregistration and Clean water withdraw method | |
Hui et al. | Road centerline extraction from airborne LiDAR point cloud based on hierarchical fusion and optimization | |
CN107301661A (en) | High-resolution remote sensing image method for registering based on edge point feature | |
CN105335973B (en) | Apply to the visual processing method of strip machining production line | |
CN107248159A (en) | A kind of metal works defect inspection method based on binocular vision | |
CN104933434A (en) | Image matching method combining length between perpendiculars (LBP) feature extraction method and surf feature extraction method | |
Yuan et al. | Learning to count buildings in diverse aerial scenes | |
CN107122737A (en) | A kind of road signs automatic detection recognition methods | |
CN104123529A (en) | Human hand detection method and system thereof | |
Chen et al. | Robust affine-invariant line matching for high resolution remote sensing images | |
CN104881671A (en) | High resolution remote sensing image local feature extraction method based on 2D-Gabor | |
Daixian | SIFT algorithm analysis and optimization | |
CN108960115A (en) | Multi-direction Method for text detection based on angle point | |
CN111242050A (en) | Automatic change detection method for remote sensing image in large-scale complex scene | |
CN108537832A (en) | Method for registering images, image processing system based on local invariant gray feature | |
Li et al. | Ship target detection and recognition method on sea surface based on multi-level hybrid network | |
CN104820992B (en) | A kind of remote sensing images Semantic Similarity measure and device based on hypergraph model | |
CN109344846A (en) | Image characteristic extracting method and device | |
CN110765993B (en) | SEM graph measuring method based on AI algorithm | |
CN104700359A (en) | Super-resolution reconstruction method of image sequence in different polar axis directions of image plane | |
Manandhar et al. | Segmentation based building detection in high resolution satellite images |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170405 |