CN104167003B - Method for fast registering remote-sensing image - Google Patents

Abstract

The invention relates to a method for fast registering a remote-sensing image. The method comprises the following steps that S1, ORB feature points are extracted from the remote-sensing image to be registered and a reference remote-sensing image; S2, initial matching is carried out on the extracted ORB feature points, and wrongly-matched feature points are removed from the initially-matched feature points; S3, parameter solving is carried out on the remote-sensing image to be registered; S4, resampling is carried out on the remote-sensing image to be registered, and image registration is completed. The method is beneficial to improving the speed and precision of image registration.

Description

A kind of rapid registering method of remote sensing image

Technical field

The present invention relates to technical field of image processing, particularly to a kind of to different time, different visual angles and not simultaneous interpretation The method that remote sensing images more than two width of sensor carries out rapid registering.

Background technology

Image registration is an important topic in image processing techniquess, and current image registration is widely used in each neck Domain, such as remote Sensing Image Analysis, medical image analysis, image co-registration, machine vision and other field.

The basic problem that image registration is processed as remote sensing image, is image data fusion, Monitoring on Dynamic Change etc. The premise of remote sensing image integrated analysis and application and basis.For timely, accurate monitoring tested area change, need acquisition Remote sensing image carries out real-time registration.Conventional Image registration method is generally divided into two classes: the method for registering based on gray scale, Yi Jiji Method for registering in feature (as characteristic point, characteristic curve).Method for registering based on gray scale utilizes gradation of image value metric image Between similarity, such method is realized simple, but speed is slow, and because remote sensing image is by different sensor, different Obtain when visual angle and different weather it may appear that the change such as yardstick, rotation, illumination, will lead to such algorithm cannot be just Really registering image, and be not easily susceptible to affect based on the method for local invariant feature.The method for registering of feature based is joined as needed Between quasi- image, the geometrical relationship of key character determines registration parameter, and this kind of method can reduce the data volume of process, and for The distortion of image, noise etc. have certain robustness.The quality of therefore matching performance depends greatly on feature and retouches The method stated and the quality of feature extraction.

There is size scaling currently for image, rotation, the situation of translation, mainly be used sift extraction characteristic point as joining Quasi- feature.And sift feature needs the dimension setting up characteristic point vector high, computationally intensive, therefore cannot meet the reality of remote sensing image Apply registration.

Image intensity value around fast angle point distinguished point based, detects the pixel value making a circle in candidate feature point week, such as There are enough pixels enough big with the gray value difference of this candidate point then it is assumed that this candidate point is in field around fruit candidate point One characteristic point it may be assumed that

Wherein i (x) is the gray scale of any point on circumference, and i (p) is the gray scale in the center of circle,ε _d For the threshold value of gray value differences, such as Fruit n is more than given threshold value, and generally 3/4ths of surrounding circle points are then it is assumed that p is a characteristic point.

It is that size is in 31 × 31 block of pixels after image smoothing that brief describes son, chooses and obeys n(n=128,256, 512) organize Gauss distribution random point pixel pair, by the size of compared pixels pair, it then follows greatly 1, little be 0 criterion, form two System string.

Binary detection is defined as:

Wherein,p(x ₁)、p(x ₂) be respectively block of pixels p inx ₁Withx ₂The grey scale pixel value of position.

Then brief characterizing definition for n dimension binary string vector it may be assumed that

Orb algorithm has good rotational invariance, and with fast angle point grid characteristic point, fast angle point is that one kind is very fast The Angular Point Extracting Method of speed, because fast does not possess rotational invariance, therefore orb centroid method is the angle point interpolation direction of orb Information is so that feature possesses rotational invariance.In addition, in order that feature is applied to the situation of size scaling, being schemed by foundation As pyramidal method, feature extraction can be carried out in the case of image cause not of uniform size.In addition, setting up feature point description The period of the day from 11 p.m. to 1 a.m, describes son using brief and is described, and it is a kind of binary descriptor that brief describes son, can be rapidly performed by very much Coupling.Therefore real-time registration is applied to based on the image registration of orb feature.

Content of the invention

It is an object of the invention to provide a kind of rapid registering method of remote sensing image, the method is conducive to improving image joins Accurate speed and precision.

For achieving the above object, the technical scheme is that a kind of rapid registering method of remote sensing image, including following Step:

Step s1: extract orb characteristic point to remote sensing images subject to registration with reference to remote sensing images respectively；

Step s2: initial matching is carried out to the orb characteristic point extracted, erroneous matching is rejected to the characteristic point of initial matching Characteristic point；

Step s3: parametric solution is carried out to described remote sensing images subject to registration；

Step s4: resampling is carried out to described remote sensing images subject to registration, completes image registration.

Further, in step s1, set up image pyramid to remote sensing images subject to registration with reference to remote sensing images respectively, Orb characteristic point is extracted to every tomographic image pyramid.

Further, in step s1, the method extracting orb characteristic point is: carries out fast Corner Detection first, carries out Harris Corner Detection, n best point before selection, then verify angle point with non-maxima suppression, reject pseudo-edge point, with Characteristic point is made to be evenly distributed.

Further, in step s2, to the orb characteristic point extracted using the Hamming distance characteristic point based on dual threshold Method of completing the square carries out initial matching, adopts angle point angular separation for constraints the characteristic point of initial matching, rejects erroneous matching Characteristic point, method particularly includes:

The angle point direction of characteristic point is tried to achieve by gray scale centroid method, that is, pass through to calculate the matter of angle point circle shaped neighborhood region pixel grey scale The heart, characterizes angle point direction by the vector direction that angle point and barycenter are formed；

Defining angle point circle shaped neighborhood region square is:, the matter of described angle point circle shaped neighborhood region The heart is:, then the angle point direction of the vector direction as characteristic point that angle point is formed with barycenter:；

Wherein, m_pqRepresent p+q rank square, i (x, y) represents the gray value of pixel (x, y) in angle point circle shaped neighborhood region, (x, Y) the pixel point coordinates in angle point circle shaped neighborhood region, m are represented₀₀Represent zeroth order square, m₁₀And m₀₁All represent first moment；

δθ _i For the angle point direction of a characteristic point in remote sensing images subject to registration, δθ _i ’It is with reference to corresponding special in remote sensing images Levy angle point direction a little, then angle point angular separation is δθ _i =δθ _i -δθ _i ’；

Then as follows initial matching is carried out to the orb characteristic point extracted, and rejects the characteristic point of erroneous matching:

A, the angle point that remote sensing images subject to registration are extracted with reference remote sensing images set up rbrief description, if subject to registration distant Sense image feature point set be combined into { a1, a2 ..., an1 }, with reference to remote sensing images feature point set be combined into b1, b2 ..., bn2}；

B, respectively by characteristic point a1 of remote sensing images subject to registration description son with reference to remote sensing images all characteristic points b1, Description of b2 ..., bn2 is compared, and calculates description and the b1 of a1, the Hamming distance of description of b2 ..., bn2 From selecting Hamming distance point the shortest from b1, b2 ..., bn2, and calculate the shortest Hamming distance and time short Hamming distance Ratio, if described ratio is less than the larger threshold value setting, point the shortest for Hamming distance and a1 are left initial pairing Point, and calculate its angle point angular separation δθ ₁, otherwise give up；

C, according to the method described above, obtain characteristic point a2 ... of remote sensing images subject to registration successively, an1 is with reference to remote sensing figure Initial match point in picture, and calculate corresponding angle point angular separation δθ ₂, …, δθ _n ；

D, initial match point is ranked up from big to small according to the ratio of the shortest Hamming distance and time short Hamming distance, that is, Feature Points Matching quality is ranked up from getting well to differing from, and the ratio extracting the shortest Hamming distance with time short Hamming distance is less than setting Small threshold initial match point；

E, the initial match point extracting step d obtain optimal angle point angular separation δ by method of least squareθ _m ；

F, the angle point angular separation of all initial match point obtaining step b and c, with optimal angle point angular separation δθ _m For constraints, will deviate from δθ _m A range of initial match point is rejected.

Further, in step s3, the method carrying out parametric solution to described remote sensing images subject to registration is:

If the pixel with reference to remote sensing images isf(x,y), the pixel of image subject to registration isg(x’,y’) it is assumed that ginseng The coordinate examining on remote sensing images point be (x _i ,y _i ), on corresponding remote sensing images subject to registration point coordinate be (x _i ’,y _i ’), then (x _i ,y _i ) and (x _i ’,y _i ’) between affine transformation be expressed as:

In formula,sFor scale factor,θFor the anglec of rotation, δxAnd δyThe two translation of axes amounts of being respectively；

After obtaining m characteristic point, optimal mapping matrix is obtained according to ransac algorithm, that is, determines registration parameters、θ、δx, δy；

Resampling is carried out to remote sensing images subject to registration according to registration parameter, that is, completes image registration.

The invention has the beneficial effects as follows there is situations such as different deformation, illumination, in order to reflect in time between remote sensing image Monitoring section dynamic change it is proposed that a kind of rapid registering method of remote sensing image, the method can effectively reject mismatch a little it is ensured that The precision of image registration, has very strong practicality and a wide application prospect.

Brief description

Fig. 1 is the flowchart of the embodiment of the present invention.

Fig. 2 is the reference remote sensing images in the embodiment of the present invention.

Fig. 3 is the remote sensing images subject to registration in the embodiment of the present invention.

Fig. 4 is the initial characteristicses Point matching figure in the embodiment of the present invention.

Fig. 5 is the Feature Points Matching figure mismatching after rejecting in the embodiment of the present invention.

After Fig. 6 is remote sensing image registration subject to registration in the embodiment of the present invention with reference to the figure after remote sensing image fusion.

Specific embodiment

Below in conjunction with the accompanying drawings and specific embodiment the present invention is described in further detail.

The rapid registering method of remote sensing image of the present invention, as shown in figure 1, comprising the following steps:

Step s1: set up image pyramid to remote sensing images subject to registration (Fig. 3) with reference to remote sensing images (Fig. 2) respectively, right Every tomographic image pyramid extracts orb characteristic point.The method extracting orb characteristic point is: carry out fast Corner Detection first, due to The angle point number that fast extracts is too many, and comprises marginal point and pseudo- angle point, harris Corner Detection Algorithm be one stable Corner Detection device, carries out harris Corner Detection, and before selection, n best point, then verifies angle point with non-maxima suppression, Reject pseudo-edge point, so that characteristic point is evenly distributed.

Step s2: in order to obtain the correct characteristic point of as far as possible many couplings, to the orb characteristic point extracted using based on dual threashold The Hamming distance characteristic point matching method of value carries out initial matching, and the characteristic point to initial matching is about using angle point angular separation Bundle condition, rejects the characteristic point of erroneous matching.Method particularly includes:

The angle point direction of characteristic point is tried to achieve by gray scale centroid method, that is, pass through to calculate the matter of angle point circle shaped neighborhood region pixel grey scale The heart, characterizes angle point direction by the vector direction that angle point and barycenter are formed；

Defining angle point circle shaped neighborhood region square is:, the matter of described angle point circle shaped neighborhood region The heart is:, then the angle point direction of the vector direction as characteristic point that angle point is formed with barycenter:；

Wherein, m_pqRepresent p+q rank square, p, q be respectively a coefficient, i (x, y) represent angle point circle shaped neighborhood region in pixel (x, Y) gray value, (x, y) represents the pixel point coordinates in angle point circle shaped neighborhood region, m₀₀Represent zeroth order square, m₁₀And m₀₁All represent one Rank square；

δθ _i For the angle point direction of a characteristic point in remote sensing images subject to registration, δθ _i ’It is with reference to corresponding special in remote sensing images Levy angle point direction a little, then angle point angular separation is δθ _i =δθ _i -δθ _i ’；

Then as follows initial matching is carried out to the orb characteristic point extracted, and rejects the characteristic point of erroneous matching:

A, the angle point that remote sensing images subject to registration are extracted with reference remote sensing images set up rbrief description, if subject to registration distant Sense image feature point set be combined into { a1, a2 ..., an1 }, with reference to remote sensing images feature point set be combined into b1, b2 ..., bn2}；

B, adopt brute-force algorithm, respectively by the description of characteristic point a1 of remote sensing images subject to registration with reference to distant All characteristic points b1 of sense image, description of b2 ..., bn2 is compared, and calculates description and b1, the b2 ... of a1, The Hamming distance of description of bn2, selects Hamming distance point the shortest from b1, b2 ..., bn2, and calculates the shortest Chinese Prescribed distance and the ratio of time short Hamming distance, set larger threshold value as 0.8, if described ratio is less than the larger threshold value setting, Then point the shortest for Hamming distance is left initial match point with a1, and calculates its angle point angular separation δθ ₁, otherwise give up；

C, according to the method described above, obtain characteristic point a2 ... of remote sensing images subject to registration successively, an1 is with reference to remote sensing figure Initial match point in picture, as shown in figure 4, and calculate corresponding angle point angular separation δθ ₂, …, δθ _n ；

D, initial match point is ranked up from big to small according to the ratio of the shortest Hamming distance and time short Hamming distance, that is, Feature Points Matching quality is ranked up from getting well to differing from, and sets small threshold as 0.5, extracts the shortest Hamming distance and time short Hamming The ratio of distance is less than the initial match point of the small threshold setting；

E, the initial match point extracting step d obtain optimal angle point angular separation δ by method of least squareθ _m ；

F, the angle point angular separation of all initial match point obtaining step b and c, with optimal angle point angular separation δθ _m For constraints, will deviate from δθ _m A range of initial match point is rejected, as shown in Figure 5.

Step s3: parametric solution is carried out to described remote sensing images subject to registration.

Convert in view of remote sensing images subject to registration and with reference to there is rotation, size etc. between remote sensing images, determine image it Between transformation matrix be h, h is expressed as:

Optimized transformation parameters are obtained using ransac algorithm, process is as follows:

1) randomly draw m sample in the feature point pairs of coupling, transformation matrix h is obtained by this m sample, further according to change Change same place in reference to remote sensing image for the characteristic point that matrix h obtains in remote sensing images subject to registration, then obtain by transformation matrix h The same place that obtains and the distance of the match point being drawn by Hamming distance coupling, using distance less than threshold value point as interior point；

2) above-mentioned steps are repeated k time, selection comprise in count out a most point set；

3) recalculate transformation matrix h with the sample that the point chosen is concentrated, thus obtaining meeting most of match points Good transformation model.

If the pixel with reference to remote sensing images isf(x,y), the pixel of image subject to registration isg(x’,y’) it is assumed that ginseng The coordinate examining on remote sensing images point be (x _i ,y _i ), on corresponding remote sensing images subject to registration point coordinate be (x _i ’,y _i ’), then (x _i ,y _i ) and (x _i ’,y _i ’) between affine transformation be expressed as:

In formula,sFor scale factor,θFor the anglec of rotation, δxAnd δyThe two translation of axes amounts of being respectively；

Obtaining m(m >=4) after individual characteristic point, optimal mapping matrix is obtained according to ransac algorithm, that is, determines registration Parameters、θ、δx, δy.

Step s4: resampling is carried out to described remote sensing images subject to registration using bilinear interpolation according to registration parameter, completes Image registration.

The conversion such as zoom in and out, rotate to remote sensing images subject to registration with the transformation matrix h obtaining, and adopting bilinear interpolation Carry out resampling, two width images merge in the way of 0.5 × reference picture+0.5 × image subject to registration, as shown in fig. 6, completing figure As registration.

It is more than presently preferred embodiments of the present invention, all changes made according to technical solution of the present invention, produced function is made With without departing from technical solution of the present invention scope when, belong to protection scope of the present invention.

Claims (4)

1. a kind of rapid registering method of remote sensing image is it is characterised in that comprise the following steps:

Step s1: extract orb characteristic point to remote sensing images subject to registration with reference to remote sensing images respectively；

Step s2: initial matching is carried out to the orb characteristic point extracted, the feature to the characteristic point rejecting erroneous matching of initial matching Point；

Step s3: parametric solution is carried out to described remote sensing images subject to registration；

Step s4: resampling is carried out to described remote sensing images subject to registration, completes image registration；

In step s2, the orb characteristic point extracted is carried out just using based on the Hamming distance characteristic point matching method of dual threshold Begin coupling, adopts angle point angular separation for constraints the characteristic point of initial matching, rejects the characteristic point of erroneous matching, specifically Method is:

The angle point direction of characteristic point is tried to achieve by gray scale centroid method, that is, pass through to calculate the barycenter of angle point circle shaped neighborhood region pixel grey scale, Angle point direction is characterized by the vector direction that angle point and barycenter are formed；

Defining angle point circle shaped neighborhood region square is: m_pq=σ_{X, y}x^py^qI (x, y), the barycenter of described angle point circle shaped neighborhood region is:The vector direction that then angle point is formed with barycenter is the angle point direction of characteristic point: θ=atan2 (m₀₁, m₁₀)；

Wherein, m_pqRepresent p+q rank square, i (x, y) represents the gray value of pixel (x, y) in angle point circle shaped neighborhood region, (x, y) represents Pixel point coordinates in angle point circle shaped neighborhood region, m₀₀Represent zeroth order square, m₁₀And m₀₁All represent first moment；

δθ_iFor the angle point direction of a characteristic point in remote sensing images subject to registration, δ θ_i' it is with reference to character pair point in remote sensing images Angle point direction, then angle point angular separation is δ θ_i=δ θ_i-δθ_i’；

Then as follows initial matching is carried out to the orb characteristic point extracted, and rejects the characteristic point of erroneous matching:

A, the angle point that remote sensing images subject to registration are extracted with reference remote sensing images set up rbrief description, if remote sensing figure subject to registration The feature point set of picture is combined into { a1, a2 ..., an1 }, and the feature point set with reference to remote sensing images is combined into { b1, b2 ..., bn2 }；

B, respectively by characteristic point a1 of remote sensing images subject to registration description son with reference to remote sensing images all characteristic points b1, Description of b2 ..., bn2 is compared, and calculates description and the b1 of a1, b2 ..., bn2 describe sub Hamming distance, from B1, b2 ..., select Hamming distance point the shortest in bn2, and calculate the shortest Hamming distance and the ratio of time short Hamming distance, If described ratio is less than the larger threshold value setting, point the shortest for Hamming distance is left initial match point with a1, and counts Calculate its angle point angular separation δ θ₁, otherwise give up；

C, according to the method described above, obtain characteristic point a2 ... of remote sensing images subject to registration successively, an1 is in reference to remote sensing images Initial match point, and calculate corresponding angle point angular separation δ θ₂,…,δθ_n；

D, initial match point is ranked up from big to small according to the ratio of the shortest Hamming distance and time short Hamming distance, i.e. feature Point matching quality is ranked up from getting well to differing from, and the ratio extracting the shortest Hamming distance with time short Hamming distance is less than setting relatively The initial match point of little threshold value；

E, the initial match point extracting step d obtain optimal angle point angular separation δ θ by method of least square_m；

F, the angle point angular separation of all initial match point obtaining step b and c, with optimal angle point angular separation δ θ_mIt is about Bundle condition, will deviate from δ θ_mA range of initial match point is rejected.

2. a kind of rapid registering method of remote sensing image according to claim 1 is it is characterised in that in step s1, divide Other set up image pyramid to remote sensing images subject to registration with reference to remote sensing images, orb characteristic point is extracted to every tomographic image pyramid.

3. a kind of rapid registering method of remote sensing image according to claim 1 is it is characterised in that in step s1, carry The method taking orb characteristic point is: carry out fast Corner Detection first, carry out harris Corner Detection, before selection n best Point, then verifies angle point with non-maxima suppression, rejects pseudo-edge point, so that characteristic point is evenly distributed.

4. a kind of rapid registering method of remote sensing image according to claim 1 is it is characterised in that in step s3, right The method that described remote sensing images subject to registration carry out parametric solution is:

If the pixel with reference to remote sensing images is f (x, y), the pixel of image subject to registration is g (x ', y ') it is assumed that with reference to remote sensing On image, the coordinate of point is (x_i,y_i), on corresponding remote sensing images subject to registration, the coordinate of point is (x_i’,y_i'), then (x_i, y_i) and (x_i’,y_i') between affine transformation be expressed as:

$\left[\begin{array}{c}{x_i}^{\′}\\ {y_i}^{\′}\end{array}\right]=s\left[\begin{array}{cc}\cos \mathrm{\θ}& \sin \mathrm{\θ}\\ -\sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right]\left[\begin{array}{c}{x}_i\\ y_i\end{array}\right]+\left[\begin{array}{c}\mathrm{\δ}x\\ \mathrm{\δ}y\end{array}\right]$

In formula, s is scale factor, and θ is the anglec of rotation, and δ x and δ y is respectively two translation of axes amounts；

After obtaining m characteristic point, optimal mapping matrix is obtained according to ransac algorithm, that is, determines registration parameter s, θ, δ X, δ y；

Resampling is carried out to remote sensing images subject to registration according to registration parameter, that is, completes image registration.