CN101989352B - Image registration method based on improved scale invariant feature transform (SIFT) algorithm and Lissajous figure track - Google Patents

Abstract

Technical field: the method belongs to the field of computer image alignment and registration. It can be used for alignment and registration of remote sensing, medical and general images. By combining boundary information that is more stable to multi-modality, the inherent shortcomings of the SIFT (SURF) algorithm are overcome. It effectively improves the correct matching rate of the features of the SIFT (SURF) algorithm in the multi-modal image alignment algorithm, thereby improving the stability of the alignment algorithm. In order to improve the accuracy of image alignment, a similarity measure function based on Lissajous figure trajectory is proposed, which has better stability and higher alignment accuracy. Based on the above improved SIFT (SURF) algorithm and the proposed similarity measure function based on the Lissajous figure trajectory, an image alignment algorithm with better stability and higher alignment accuracy is constructed.

Description

Method for registering images based on improved SIFT algorithm and Li Sa such as figure track

Technical field:

This method belongs to computer picture alignment, registration field.Can be used for carrying out alignment, the registration of remote sensing, medical science, general pattern.

Background technology:

SIFT and SURF algorithm are detected characteristics points from the image that two parafacies close, and the algorithm that carries out matching operation.But because this algorithm is to geometry deformation and the more sensitive shortcoming of variation of image grayscale.Thereby cause when using it and carry out the remote sensing images alignment very unstable, and the extremely low phenomenon of correct matching rate between unique point.This paper sets out for these two problems of solution, has proposed a kind of improved SIFT (SURF) algorithm and similarity measure function.

Summary of the invention:

Carry out remote sensing images when assorting owing to using SIFT algorithm or SURF algorithm, have two very large defectives.These two defectives mainly are because the multimode state property of image causes.This method combines multi-modal more stable edge and the profile information to image, thus above two defectives that effectively overcome.The SIFT algorithm and the SURF algorithm stability in this case that improve greatly.

Simultaneously this algorithm has also proposed a kind of more similarity measurement function of high resolving power and discernment that has.This similarity measurement function is based on Lissajous trajectory and calculates.Have better stability and higher precision and recognition capability.

Description of drawings:

Fig. 1 unique point and near limit thereof

Fig. 2 .TAR image of edges shown in Fig. 1.

Fig. 3. Lee's Sa such as figure track and the point set of selecting at this track thereof indicate with asterisk

Fig. 4. the advantage of Lee's Sa such as figure

Fig. 5. (a) and the remote sensing images that (b) will align, (c) and (d) be from (a) and the parts of images that (b) selects, 2 same geographic position of correspondence that the criss-cross among the figure marks

Fig. 6 by the similarity measure function that this algorithm proposes calculate similar matrix.

Fig. 7. be used for the calculating schematic diagram of the track point set of calculating similarity measure.

Fig. 8, the correct matching rate between the matching double points of former algorithm

Fig. 9, the correct matching rate between the matching double points of improved algorithm

Embodiment:

One, improved SURF and SIFT algorithm

1. from two width of cloth images, detect and coupling with SURF or SIFT algorithm

2. two stack features point, the unique point of coupling sorts from high to low to the similarity degree according to them.

3. detect the profile information of two width of cloth images or the information on limit.

To the unique point of a pair of coupling to and a near opposite side calculate its corresponding TAR figure.The foundation of the calculating of TAR figure is the affine constant TAR of being, it calculates according to leg-of-mutton three apex coordinates.If an Atria summit is respectively: p _B(x _b, y _b), p _M(x _m, y _m), p _E(x _e, y _e), we have so

$\mathrm{TAR}(p_B,p_M,p_E)=\frac{1}{2}\left|\begin{array}{ccc}{x}_b& y_b& 1\\ x_m& y_m& 1\\ x_e& y_e& 0\end{array}\right|=\frac{1}{2}(x_b{y}_m+x_m{y}_e+x_e{y}_b-x_e{y}_m-x_b{y}_e-x_m{y}_b)$

To the unique point pR among Fig. 1 (a) and limit p _r ⁱ(i=0,1,2 ... n).p _r ⁱ(i=0,1,2 ... n) limit E _rOn point set.Fig. 2 (a) is the TAR figure ImR of their correspondences.To the unique point pT among Fig. 1 (b) and limit E _tp _t ⁱ(i=0,1,2 ... m) limit E _tOn point set.Same Fig. 2 (b) is point and TAR figure ImT corresponding to limit among Fig. 1 (b).Their mutually element value computing formula is:

$\mathrm{ImR}\left[i\right]\left[j\right]=\mathrm{TAR}(p_r^i,\mathrm{pR},p_r^j)$

$\mathrm{ImT}\left[i\right]\left[j\right]=\mathrm{TAR}(p_t^i,\mathrm{pT},p_t^j)$

5. TAR figure ImR and ImT are used SURF or SIFT algorithm, and their Feature Descriptor is improved mutually element value (TAR value information) and the isocontour information of having added corresponding point among the TAR figure.Thereby find stack features point, and the descriptor after the application enhancements is found out the corresponding relation between them.The unique point correspondence of a pair of mutual coupling like this one diabolo, such as: shown in Fig. 1 (c) and Fig. 1 (d).

6. calculated near the affined transformation of two parts the unique point by the triangle pair that mutually finds coupling.

7. extracting equably one out from the limit of image simultaneously marks words and phrases for special attention to CP _rAnd CP _s, and allow them take on the role of a part of unique point descriptor.To a pair of unique point to p _rAnd p _s, find out three pairs of points from overlapped border to (pr according to the local affine transformations of finding out in the 5th step ₁, ps ₁), (pr ₂, ps ₂), and (pr ₃, ps ₃) any 3 conllinear not among the .l.Ps ₁ ⁱ, ps ₂ ^j, ps ₃ ^kTo meet the following conditions and the point set on the border:

Quadrilateral pr then, pr ₁, pr ₂, pr ₃With one group of quadrilateral ps, ps ₁ ⁱ, ps ₂ ^j, ps ₃ ^kDetermine one group of geometric transformation F.Wherein l is in borderline hunting zone.Then this part similar value is:

$\mathrm{SME}(p_r,p_s)=\underset{f\∈F}{\max }\{\underset{p\∈{\mathrm{CP}}_r}{\mathrm{\Σ}}(\mathrm{\ξ}-||p,f\left(p\right)|\left|\right)\mathrm{\δ}(\mathrm{\ξ}-||p,f\left(p\right)|\left|\right)+\underset{p\∈C{P}_s}{\mathrm{\Σ}}(\mathrm{\ξ}-||p,\mathrm{ft}\left(p\right)|\left|\right)\mathrm{\δ}(\mathrm{\ξ}-||p,\mathrm{ft}\left(p\right)|\left|\right)\}$

Then improved similar value is:

SM(p _r，p _t)＝SMD(p _r，p _t)*||p _t，f(p _r)||-α*SME(p _r，p _t)

8. according to the geometric transformation parameter of having estimated and improved Feature Descriptor, recomputate the coupling between unique point.And then calculate overall geometric transformation parameter.

9. on this basis, adopt similarity measure function in this paper, adopt alternative manner, realize the accuracy registration of image.

The advantage that SURF after the improvement or SIFT algorithm possess:

Such as Fig. 8 and shown in Figure 9.The unique point that algorithm after the improvement improves significantly to correct matching rate.

Two, the new similarity measure function based on Li Sa such as figure that proposes in this method:

1. Lee's Sa is such as figure

In the mathematics category, Li Sa such as figure are the movement locus that has following parameter system of equations to determine

$\left\{\begin{array}{c}x=A_x\sin ({\mathrm{\ω}}_{x}t+{\mathrm{\φ}}_x)\\ y=A_y\sin ({\mathrm{\ω}}_{y}t+{\mathrm{\φ}}_y)\end{array}\right.$

2. be used for to calculate the selection of the point set on the track of similarity.At first, given parameters

A _x, A _y, ω _x, ω _y, φ _x, φ _y, for the some pS among the pR in figure R and the figure S, the Trajectories Toggle of generation is gR ₁, choose equally spacedly one group of point set according to parameter t and be designated as pR ₁ ⁱ(i=1,2,3 ... n).By track gR ₁The track with respect to a pR that produces is designated as gR ₂, the point set on it is designated as pR ₂ ⁱ(i=1,2,3 ... n).Satisfy following relational expression between them.

PR ₂ ⁱ(i=1,2,3 ... n) be with respect to a pR and the point set on the selected track that is used for calculating similarity measure function, wherein α is given constant.

If the geometric transformation between figure R and figure S is described with f, the track point set selection course with respect to pS in figure S is as follows: to point set pR+pR ₁ ⁱ(i=1,2,3 ... n) carry out the f conversion and obtain point set pS ₁ ⁱ(i=1,2,3 ... n), by point set pS ₁ ⁱ(i=1,2,3 ... n) and the point set pS that determines of some pS ₂ ⁱ(i=1,2,3 ... n).Satisfy following relational expression between them

PS ₂ ⁱ(i=1,2,3 ... n) be with respect to a pS and the point set on the selected track that is used for calculating similarity measure function.

3. suppose: S _lBe labeled as a kind of similarity measure function, based on the point set of above definition, the similarity measure that we propose is: α.

PR wherein ₂Be selected point set pR ₂ ⁱ(i=1,2,3 ... n), pS ₂Be selected point set pS ₂ ⁱ(i=1,2,3 ... n).

The advantage of this similarity measure function:

(a) relative other similarity measure function, this similarity measure function can in the situation that other similarity measure function lost efficacy, still effective, wherein sea area as shown in Figure 5.

(b) possesses the error amplification.

(c) can calculate according to different tracks a plurality of similarity measures, thus more stable.

(d) by the track disturbing phenomenon, this similarity measure has higher alignment accuracy.

Claims (2)

1. a method for registering images is characterized in that, may further comprise the steps:

(1) from two width of cloth images with SURF or SIFT algorithm detected characteristics point and mate;

(2) two stack features points, the unique point of coupling sorts from high to low to the similarity degree according to them;

(3) detect the profile information of two width of cloth images or the information on limit;

(4) to the unique point of a pair of coupling to and a near opposite side calculate its corresponding TAR figure ImR and ImT;

(5) TAR figure ImR and ImT are used SURF or SIFT algorithm, and their Feature Descriptor improved pixel value and the isocontour information of having added corresponding point among the TAR figure, thereby find stack features point, and the descriptor after the application enhancements finds out the corresponding relation between them, the unique point correspondence of so a pair of mutual coupling one diabolo;

(6) calculated near the affined transformation of two parts the unique point by the triangle pair of mutual coupling;

(7) extracting equably one out from the limit of image simultaneously marks words and phrases for special attention to CP _rAnd CP _s, and allow them take on the role of a part of unique point descriptor, to a pair of unique point to p _rAnd p _s, find out three pairs of points from overlapped border to (pr according to the local affine transformations of finding out ₁, ps ₁), (pr ₂, ps ₂), (pr ₃, ps ₃), any 3 conllinear not wherein,

To meet the following conditions and the point set on the border:

Quadrilateral pr then, pr ₁, pr ₂, pr ₃With one group of quadrilateral ps,

Determine one group of geometric transformation F, wherein l is in borderline hunting zone;

(8) according to geometric transformation parameter and the improved Feature Descriptor estimated, recomputate the coupling between unique point, and then calculate overall geometric transformation parameter;

(9) adopt the similarity measure function, adopt alternative manner, realize the accuracy registration of image.

2. method for registering images as claimed in claim 1, wherein similarity measure function MILF is based on Li Sa such as figure track, and Li Sa such as figure track are the curve maps that following system of equations produces,

$\left\{\begin{array}{c}x=A_x\sin ({\mathrm{\ω}}_{x}t+{\mathrm{\φ}}_x)\\ y=A_y\sin ({\mathrm{\ω}}_{y}t+{\mathrm{\φ}}_y)\end{array}\right.$

Given one group of parameter A _x, A _y, ω _x, ω _y, φ _x, φ _y, according to following formula, just can produce a track TR ₁, for benchmark image with treat 2 pr and ps in the accurate figure picture, used similarity measure function MILF is in this method: MILF (pr, ps)=S _l(pR ₂, pS ₂), S _lIdentify a kind of similarity measure function, it is MI (mutual information), wherein pR ₂And pS ₂The point set that produces as follows:

From track TR ₁In the point set of equidistantly choosing by parameter t, given amplification coefficient α is according to equation

Just can generate point set

Given geometric transformation f, just can obtain with $\{{\mathrm{pR}}_1^k:k=\mathrm{1,2},...,n\}$ Another corresponding group point set $\{{\mathrm{pS}}_1^k:k=\mathrm{1,2},...,n\},$ In like manner, according to equation $\frac{\stackrel{\→}{{\mathrm{pS}}_1^k{\mathrm{pS}}_2^k}}{\stackrel{\→}{{\mathrm{pS}}_2^k\mathrm{ps}}}=-\frac{1+\mathrm{\α}}{\mathrm{\α}},$ Just can produce point set $:{\mathrm{pS}}_2=\{{\mathrm{pS}}_2^k:k=\mathrm{1,2},...,n\}.$

Images