CN104616280A - Image registration method based on maximum stable extreme region and phase coherence - Google Patents

Abstract

The invention discloses an image registration method based on a maximum stable extreme region and phase coherence, and aims at solving the defects of low repeating rate of extracted characteristic points and large operation complexity in the prior art. The method comprises the steps of 1, inputting two images with affine transformation, and respectively performing detection and matching for the maximum stable extreme region; 2, fitting the matching areas of the two images, and amplifying and normalizing; 3, performing band-pass decomposition for two normalized areas; 4, detecting the characteristics points based on the maximum phase coherence matrix, and constructing the probability distribution of the detected characteristics points; 5, estimating the accurate affine transformation matrix between two point sets; 6, estimating the transformation matrix of the two images according to the two normalized areas; 7, calculating the accurate affine transformation matrix between the two images, and finishing the image registration. According to the method, the characteristics points with relatively high repeating rate and accurate matching rate can be extracted, the calculation efficiency can be increased, and the image fusion, image splicing and three-dimensional reconstruction can be performed.

Description

Based on the method for registering images of maximum stable extremal region and phase equalization

Technical field

The invention belongs to technical field of image processing, be specifically related to a kind of affined transformation method for registering images, can be applicable to image co-registration, the field such as image mosaic and three-dimensional reconstruction.

Background technology

In image co-registration, the field such as image mosaic and three-dimensional reconstruction, needs first to carry out registration process to several views of Same Scene.Generally, the method for registering images of feature based can be adopted to carry out image registration, this mainly considers that some characteristics of image have unchangeability for the yardstick of image and rotation, and only has the high advantage of counting yield by the geometric relationship that characteristic information is found between image.But, when there is larger affined transformation when between two width images, be often difficult to extract in them there is the accurate feature of higher repetition rate or position, thus cause registration accuracy not even cannot realize the problem of registration.

At present, characteristic information conventional in the method for registering images of feature based has scale invariant feature SIFT, the complete affine invariants ASIFT of maximum stable extremal region MSER characteristic sum, such as Lowe D, " Distinctive image features from scale-invariant keypoints. " International Journal of Computer Vision, vol.60, no.2, pp.91-110.Matas J, Chum O, et al., " Robust wide-baseline stereo from maximally stable extremal regions. " Image and Vision Computing, vol.22, no.10, and Morel J M pp.761-767., Yu G, " ASIFT:A new framework for fully affine invariant image comparison. " SIAM Journal on Imaging Sciences, vol.2, no.2, pp.438-469. disclosed in these three sections of documents, technology is feature extraction and matching process, and then the geometric transformation parameter that the feature of coupling can be utilized to come between computed image realizes image registration.Wherein, the method for registering images based on scale invariant feature SIFT can the larger image of registration yardstick, and obtains good registration effect.But, when there is larger affined transformation when between image, characteristic detection method based on scale invariant feature SIFT often seldom can obtain the enough and matching double points that accuracy is high of number, and the method for registering images therefore based on scale invariant feature SIFT can not have the image of larger affined transformation by registration.Based on the method for registering images of maximum stable extremal region MSER, the barycenter of maximum stable extremal region MSER is adopted to mate as unique point, and then the affine transformation parameter between estimated image, because maximum stable extremal region MSER has higher affined transformation unchangeability, therefore, it is possible to realize the image registration that there is larger affined transformation, but due to the difference of imaging sensor and imaging circumstances, the barycenter of employing often accurately can not reflect the position of feature, thus causes registration accuracy not high.First complete affine invariants ASIFT algorithm carries out the affine space sampling of artificial simulation to original image, obtain several views; Then scale invariant feature SIFT method is utilized to carry out feature extraction and characteristic matching to several views obtained, can obtain so more to mate than scale invariant feature SIFT method and count, can there is the image of larger affined transformation by registration in the method for registering images therefore based on complete affine invariants ASIFT feature.The deficiency that the method exists is, because the method is simulated on affine space image, form the image at each visual angle, consume a large amount of internal memories, also introduced a large amount of Mismatching points when extracting a large amount of correct match points simultaneously, and higher image registration accuracy will be obtained, just need the optimizing process of more complicated to delete Mismatching point, this turn increases computation complexity undoubtedly.

Summary of the invention

The object of the invention is to the shortcoming improving above-mentioned prior art, a kind of affined transformation method for registering images based on maximum stable extremal region and phase equalization is proposed, to obtain better affined transformation image registration effect, and reduce computational complexity, improve counting yield.

For achieving the above object, technical scheme of the present invention is: by the partial fitting region of the thick coupling acquisition two width input picture based on maximum stable extremal region MSER feature; Affine region method for normalizing is adopted to overcome the change of the picture structure that Affine distortion brings; Utilize Gabor bandpass filter to carry out band reduction of fractions to a common denominator solution to normalization region, and then in each sub-band images, carry out the feature point detection based on the maximum square of phase equalization; Adopt the method for probability distribution to carry out accuracy registration to the unique point set detected, and calculate the affine transformation matrix between two width input pictures.Its concrete steps comprise as follows:

(1) there are two width image A and B of affined transformation in input respectively, and wherein A is reference picture, and B is image subject to registration;

(2) maximum stable extremal region MSER detection and coupling are carried out to reference picture A and image B subject to registration;

(3) respectively matching is carried out to the maximum stable extremal region that reference picture A and image B subject to registration matches, and obtain reference picture A expand after ellipse fitting region and image B subject to registration expand after ellipse fitting region;

(4) to above-mentioned two ellipse fitting region normalization:

4a) treat normalized point in computing reference image A and image B subject to registration respectively:

$z_A=\frac{1}{{\left(\min \right[\left|\det \left(M_A\right)\right|,\left|\det \left(M_B\right)\right|\left]\right)}^{1/4}}{H_A}^{-1}{M}_A^{1/2}(x_A^{\′}-{\mathrm{\μ}}_A)$

$z_B=\frac{1}{{\left(\min \right[\left|\det \left(M_A\right)\right|,\left|\det \left(M_B\right)\right|\left]\right)}^{1/4}}{H_B}^{-1}{M}_B^{1/2}(x_B^{\′}-{\mathrm{\μ}}_B)$

Wherein, z _aand z _brepresent in reference picture A and image B subject to registration respectively and treat normalized point, M _aand M _brepresent the second-order moments matrix of the barycenter of all maximum stable extremal region MSER in reference picture A and image B subject to registration respectively, H _aand H _brepresent second-order moments matrix M respectively _aand M _bthe symmetrical unitary matrix of the reality that svd obtains, x ' _awith x ' _brepresent the point of the elliptic region after expanding in image A and B respectively, μ _aand μ _brepresent the average of the barycenter of all maximum extremal region MSER in reference picture A and image B subject to registration respectively;

4b) with the normalization point z that needs in reference picture A _aform the normalization region P of reference picture, need normalization point z with image B subject to registration _bform the normalization region Q of image subject to registration;

(5) respectively the band reduction of fractions to a common denominator solution based on Gabor filter is carried out to the normalization region P of reference picture A and the normalization region Q of image B subject to registration, obtain the sub-band images that this two width image comprises different frequency composition;

(6) feature point detection based on the maximum square of phase equalization is carried out to the sub-band images of above-mentioned two width images, and the point set registration based on probability distribution is carried out to the unique point detected, obtain the transformation matrix T between point set ₁;

(7) according to the normalization region P of reference picture and the normalization region Q of image subject to registration, the transformation matrix T between reference picture A and image B subject to registration is estimated _c1, T _c2;

(8) according to the transformation matrix T between point set ₁and the transformation matrix T between reference picture A and image B subject to registration _c1, T _c2affine transformation matrix T between computing reference image A and image B subject to registration:

T＝T _c1 ^-1T ₁T _c2；

(9) treat registering images B according to affine transformation matrix T to convert, then carry out bilinear interpolation to converting the image obtained, complete image registration.

The present invention has the following advantages compared with prior art:

First, the present invention is owing to having carried out the registration based on maximum stable extremal region MSER to the reference picture inputted and image subject to registration, and the band reduction of fractions to a common denominator solution of carrying out in the ellipse fitting region obtained based on Gabor filter and the maximum moment characteristics point detection of phase equalization, improving prior art carries out in the process of feature extraction to the image that there is larger affined transformation, be difficult to the defect obtaining higher feature point repetition rate and correct matching rate, unique point the present invention being designed extract in these cases, has higher unique point repetition rate and correct matching rate.

Second, the present invention is owing to have employed the point set registration strategies based on probability distribution, improve prior art and will set up the defect of higher-dimension descriptor for unique point in the matching process of unique point, make the present invention compared with prior art take less storage space, and there is higher counting yield.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is the registration simulated effect figure of the present invention to large scale modified-image;

Fig. 3 is that the present invention is to the registration simulated effect figure that there is larger affined transformation image.

Embodiment

Below in conjunction with accompanying drawing, the present invention will be further described.

With reference to accompanying drawing 1, concrete steps of the present invention are as follows:

Step 1, input picture: input exists two width images of affined transformation respectively, a width is as reference image A, and another width is as image B subject to registration.

Step 2, carries out maximum stable extremal region MSER detection and coupling to reference picture A and image B subject to registration.

2a) maximum stable extremal region MSER detection is carried out respectively to reference picture A and image B subject to registration, obtain multiple irregular extremal region with affine-invariant features;

2b) by multiple irregular extremal region one_to_one corresponding with affine-invariant features, obtain initial maximum stable extremal region MSER coupling right.

Step 3, carries out matching respectively to the maximum stable extremal region that reference picture A and image B subject to registration matches, and obtains the ellipse fitting region after the ellipse fitting region after reference picture A expansion and image B subject to registration expansion.

This example adopts but is not limited to utilize the matching area of matching algorithm to two described width images based on maximum stable extremal region MSER to carry out matching, and its step is as follows:

3a) detect the barycenter of the maximum stable extremal region MSER of reference picture A and image B subject to registration respectively;

3b) according to the barycenter of the maximum stable extremal region MSER obtained in described two width images, calculate the point of this two width image fitted area according to the following formula:

(x _A-μ _A) ^TU _A ^-1(x _A-μ _A)＝(x _A-μ _A) ^TM _A(x _A-μ _A)＝1

(x _B-μ _B) ^TU _B ^-1(x _B-μ _B)＝(x _B-μ _B) ^TM _B(x _B-μ _B)＝1

Wherein, x _aand x _brepresent the point of fitted area in reference picture A and image B subject to registration respectively, μ _aand μ _brepresent the average of the barycenter of all maximum stable extremal region MSER in reference picture A and image B subject to registration respectively, T represents transposition, U _aand U _brepresent the variance of the barycenter of all maximum stable extremal region MSER in reference picture A and image B subject to registration respectively, M _aand M _brepresent the second-order moments matrix of the barycenter of all maximum stable extremal region MSER in reference picture A and image B subject to registration respectively;

3c) in reference picture A and image B subject to registration, form respective initial ellipse fitting region with the respective fitted area point obtained respectively;

3d) initial the maximum of ellipse fitting region allows exaggerated scale in computing reference image A according to the following formula:

$k_A=\min [\frac{r_A-u_a}{\max \left(x_a\right)-u_a},\frac{c_A-v_a}{\max \left(y_a\right)-v_a},\frac{u_a-1}{u_a-\min \left(x_a\right)},\frac{v_a-1}{v_a-\min \left(y_a\right)}],$

Wherein, k _arepresent that comprising the maximum of the ellipse long and short shaft of initial fitted area in reference picture A allows expansion multiple, r _aand c _arepresent line number and the columns of reference picture A respectively, u _aand v _arepresent the HCCI combustion of the barycenter of all maximum stable extremal region MSER in reference picture A respectively, x _aand y _arepresent row-coordinate and the row coordinate of fitted area point in reference picture A respectively;

3e) calculate initial the maximum of ellipse fitting region in image B subject to registration according to the following formula and allow exaggerated scale:

$k_B=\min [\frac{r_B-u_a}{\max \left(x_b\right)-u_b},\frac{c_B-v_a}{\max \left(y_b\right)-v_b},\frac{u_b-1}{u_b-\min \left(x_b\right)},\frac{v_b-1}{v_b-\min \left(y_b\right)}]$

Wherein, k _brepresent in image B subject to registration that comprising the maximum of the ellipse long and short shaft of initial fitted area allows expansion multiple, r _band c _brepresent line number and the columns of image B subject to registration respectively, u _band v _brepresent the HCCI combustion of the barycenter of the maximum extremal region of all couplings in image B subject to registration respectively, x _band y _brepresent fitted area point row-coordinate and row coordinate in image B subject to registration respectively;

3f) compare initial the maximum of ellipse fitting region in reference picture A and allow exaggerated scale k _awith the exaggerated scale k in ellipse fitting region initial in image B subject to registration _b, using the exaggerated scale k as initial ellipse fitting region less for numerical value in both, i.e. k=min (k _a, k _b);

3g) according to the elliptic region point after expansion in exaggerated scale k computing reference image A and image B subject to registration:

(x′ _A-μ _A) ^TM _A(x′ _A-μ _A)＝k ²

(x′ _B-μ _B) ^TM _B(x′ _B-μ _B)＝k ²

Wherein, x ' _awith x ' _brepresent the point of the elliptic region after expanding in reference picture A and image B subject to registration respectively, μ _aand μ _brepresent the average of the barycenter of all coupling maximum stable extremal regions in reference picture A and image B subject to registration respectively, M _aand M _brepresent the second-order moments matrix of the barycenter of all coupling maximum stable extremal regions in reference picture A and image B subject to registration respectively;

The point x ' of the elliptic region after 3h) expanding with reference picture A respectively _athe point x ' of the elliptic region after expanding with image B subject to registration _b, form the ellipse fitting region of this two width image, i.e. the ellipse fitting region of reference picture A and the ellipse fitting region of image B subject to registration.

Step 4, to above-mentioned two ellipse fitting region normalization:

4a) treat normalized some z in computing reference image A respectively according to the following formula _anormalized some z is treated with in image B subject to registration _b:

$z_A=\frac{1}{{\left(\min \right[\left|\det \left(M_A\right)\right|,\left|\det \left(M_B\right)\right|\left]\right)}^{1/4}}{H_A}^{-1}{M}_A^{1/2}(x_A^{\′}-{\mathrm{\μ}}_A)$

$z_B=\frac{1}{{\left(\min \right[\left|\det \left(M_A\right)\right|,\left|\det \left(M_B\right)\right|\left]\right)}^{1/4}}{H_B}^{-1}{M}_B^{1/2}(x_B^{\′}-{\mathrm{\μ}}_B)$

Wherein, M _aand M _brepresent the second-order moments matrix of the barycenter of all coupling maximum stable extremal region MSER in reference picture A and image B subject to registration respectively, H _aand H _brepresent second-order moments matrix M respectively _aand M _bthe symmetrical unitary matrix of the reality that svd obtains, x ' _awith x ' _brepresent the point of the elliptic region after expanding in reference picture A and image B subject to registration respectively, μ _aand μ _brepresent the average of the barycenter of all coupling maximum stable extremal region MSER in reference picture A and image B subject to registration respectively;

4b) with the normalization point z that needs in reference picture A _aform the normalization region P of reference picture, with the normalization point z that needs in image B subject to registration _bform the normalization region Q of image subject to registration.

Step 5, carries out the band reduction of fractions to a common denominator solution based on Gabor filter to the normalization region P of reference picture A and the normalization region Q of image B subject to registration respectively, obtains the sub-band images that this two width image comprises different frequency composition.

Have the wave filter of the carrying out band reduction of fractions to a common denominator solution in image normalization region: Gaussian-Laplace bandpass filter, DOG wave filter etc.This example adopts Gabor filter to carry out band reduction of fractions to a common denominator solution to the normalization region M of reference picture A and the normalization region N of image B subject to registration respectively, and its step is as follows:

5a) design has the Gabor band-pass filter group G (u, v, λ) of 5 bandpass filter:

$G(u,v,\mathrm{\λ})=\frac{\mathrm{\π}}{K}\underset{i=-K}{\overset{K}{\mathrm{\Σ}}}{e}^{-2{\mathrm{\π}}^2{\mathrm{\λ}}^2({(u-\frac{\cos {\mathrm{\θ}}_i}{2\mathrm{\λ}})}^2+{(v-\frac{\sin {\mathrm{\θ}}_i}{2\mathrm{\λ}})}^2)},$

Wherein, u and v represents the frequency domain coordinates of bandpass filter, and K represents the direction number of each bandpass filter, and the value of K is 6, θ _irepresent bandpass filter towards, λ is the scale factor of bandpass filter, and the λ value of each bandpass filter is different, and namely first bandpass filter value is second bandpass filter value be the 2, three bandpass filter value be 4th bandpass filter value be the 4, five bandpass filter value be

5b) utilize the band-pass filter group of above-mentioned design, according to the following formula band reduction of fractions to a common denominator solution carried out to the normalization region obtained from reference picture A and image B subject to registration:

$I_A^{\mathrm{\λ}}(x,y)=F^{-1}[G(u,v,\mathrm{\λ})\×F[I_A(x,y)\left]\right]$

$I_B^{\mathrm{\λ}}(x,y)=F^{-1}[G(u,v,\mathrm{\λ})\×F[I_B(x,y)\left]\right]$

Wherein, I _a(x, y) and I _b(x, y) represents the normalization region obtained from reference picture A and image B subject to registration respectively, and F [] represents Fourier transform, F ^-1[] represents inverse Fourier transform, represent the sub-band images corresponding to reference picture A, represent the sub-band images corresponding to image B subject to registration.

Step 6, carries out the feature point detection based on the maximum square of phase equalization to the sub-band images of above-mentioned two width images.

The method of the sub-band images of image being carried out to feature point detection has: scale invariant feature SIFT method, complete affine invariants ASIFT method etc.This example adopts the carry image of feature point detecting method to above-mentioned two width images based on the maximum square of phase equalization to carry out feature point detection, and its step is as follows:

6a) respectively to each sub-band images of reference picture A, carry out the feature point detection based on the maximum square of phase equalization, then select unique point and the feature point detection result that it can be used as reference picture A from comprising the maximum sub-band images of unique point number;

6b) form a point set with reference to the unique point detected in image A according to the following formula:

X＝[x ₁x ₂…x _n… x _N]100≤N≤500，

Wherein, X represents the point set that the unique point detected in reference picture A is formed, x _nrepresent the n-th unique point detected in reference picture A, n=1,2 ..., N, N represent the number of the unique point detected in reference picture A;

6c) treat each sub-band images of registering images B respectively, carry out the feature point detection based on the maximum square of phase equalization, then select unique point as the feature point detection result of image B subject to registration from comprising the maximum sub-band images of unique point number;

6d) form a point set by the unique point detected in image B subject to registration according to the following formula:

Y＝[y ₁y ₂…y _m… y _M]100≤M≤500

Wherein, Y represents the point set that the unique point detected in image B subject to registration is formed, y _mrepresent m the unique point detected in image B subject to registration, m=1,2 ..., M, M represent the number of the unique point detected in image B subject to registration.

Step 7, carries out the point set registration based on probability distribution to the above-mentioned unique point detected, obtains the transformation matrix T between point set ₁.

Have the method detecting the unique point that obtains and carry out point set registration: scale invariant feature SIFT method, complete affine invariants ASIFT method etc.This example adopts and carries out point set registration based on the point set method for registering based on probability distribution to the above-mentioned unique point detected, its step is as follows:

7a) using image subject to registration detect in point set Y a little as the center of fiqure of gauss hybrid models GMM;

7b) detect any point x in point set X according to reference picture _nthe corresponding relation of the point in point set Y is detected, formation condition probability density function with image subject to registration:

$P\left(x_n\right|\mathrm{\σ},T_1)=\frac{1}{M{\left({2\mathrm{\π\σ}}^2\right)}^{3/2/}}\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{e}^{(-\frac{{\left|\right|x_n-T_1{y}_m\left|\right|}^2}{{2\mathrm{\σ}}^2})},$

Wherein, σ represents the standard deviation of single Gaussian function in gauss hybrid models, and M represents that image subject to registration detects the number of element in point set Y;

Institute 7c) detected in point set X according to reference picture a little detect with image subject to registration in point set Y corresponding relation a little, generation log-likelihood estimation function:

$l(\mathrm{\σ},T_1)=\log \underset{n=1}{\overset{N}{\mathrm{\Π}}}P\left(x_n\right|\mathrm{\σ},T_1);$

7d) utilize expectation maximization EM algorithm, calculate send as an envoy to log-likelihood estimation function l (σ, T ₁) T when obtaining extreme value ₁matrix.

Step 8, according to the normalization region M of reference picture and the normalization region N of image subject to registration, estimates the transformation matrix T between reference picture A and image B subject to registration _c1, T _c2.

$T_{c1}=\left[\begin{array}{cc}{M}_A^{1/2}{H}_B& -{\mathrm{\μ}}_A\\ 0& 1\end{array}\right],T_{c2}=\left[\begin{array}{cc}R{M}_B^{1/2}{H}_B& -{\mathrm{\μ}}_B\\ 0& 1\end{array}\right]$

Wherein, $R=\left[\begin{array}{cc}\cos \mathrm{\θ}& -\sin \mathrm{\θ}\\ \sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right],$ M _aand M _brepresent the second-order moments matrix of the barycenter of all coupling maximum stable extremal regions in reference picture A and image B subject to registration respectively, H _aand H _brepresent second-order moments matrix M respectively _aand M _bthe symmetrical unitary matrix of the reality that svd obtains, μ _aand μ _brepresent the average of the barycenter of the maximum extremal region of all couplings in reference picture A and image B subject to registration respectively, θ represents the anglec of rotation of image B subject to registration relative to reference picture A.

Step 9, according to the transformation matrix T between point set ₁and the transformation matrix T between reference picture A and image B subject to registration _c1, T _c2, the affine transformation matrix T between computing reference image A and image B subject to registration:

T＝T _c1 ^-1T ₁T _c2；

Step 10, treats registering images B according to affine transformation matrix T and converts, then carries out bilinear interpolation to converting the image obtained, and completes image registration.

Effect of the present invention can be further illustrated by following emulation:

1. simulated conditions: all emulation experiments are all adopt Matlab R2009a software simulating under Windows XP operating system.

2. emulate content:

Emulation 1

The present invention is carried out the experimental result of registration to one group of large scale modified-image and existingly to compare the experimental result of this group image based on scale invariant feature SIFT, result is as Fig. 2.

Wherein:

Fig. 2 (a) is the reference picture of input,

Fig. 2 (b) is the image subject to registration of input,

Fig. 2 (c) is the result of with the method for registering images based on scale invariant feature SIFT, two width input pictures being carried out to registration,

Fig. 2 (d) is the local repressentation to the image registration results based on scale invariant feature SIFT,

The result of Fig. 2 (e) for adopting the present invention two width input pictures to be carried out to registration,

Fig. 2 (f) is the local repressentation to image registration results of the present invention.

As can be seen from Figure 2, there is obvious registration error in regional area corresponding to the registration result adopting the method for registering images based on scale invariant feature SIFT to obtain, and the registration error of regional area corresponding to the registration result adopting the present invention to obtain is less.

Add up based on the method for registering images of scale invariant feature SIFT and method for registering images of the present invention always to mate in unique point count, correctly coupling count, correct matching rate and unique point repetition rate four kinds of objective evaluation indexs, as shown in table 1.

Table 1 based on SIFT method and the inventive method to the Comparative result of four kinds of algorithm evaluation indexes

Algorithm	Total coupling is counted	Correct coupling is counted	Correct matching rate	Unique point repetition rate
					Based on SIFT method	156	123	0.7885	0.1692
The inventive method	340	350	0.9714	0.4920

Data as can be seen from table 1, method for registering images proposed by the invention, when registration has the image compared with large scale conversion, its four kinds of objective evaluation indexs are all better than the method for registering images based on scale invariant feature SIFT.

Emulation 2

With the present invention and the existing method for registering images based on maximum stable extremal region MSER and there is larger affined transformation image based on the method for registering images of complete affine invariants ASIFT to one group and carry out registration and compare, result is as Fig. 3.

Wherein:

Fig. 3 (a) is the reference picture of input,

Fig. 3 (b) is the image subject to registration of input,

Fig. 3 (c) for adopting based on the method for registering images of maximum stable extremal region MSER to the result of two width input picture registrations,

Fig. 3 (d) for adopting based on the method for registering images of complete affine invariants ASIFT to the result of two width input picture registrations,

Fig. 3 (e) adopts the inventive method to the result of two width input picture registrations,

As can be seen from Figure 3, adopt the registration result that the method for registering images based on maximum stable extremal region MSER obtains, the registration result figure that the method for registering images based on complete affine invariants ASIFT obtains and the registration result adopting the inventive method to obtain all have good visual effect.

In order to compare the performance of each algorithm further, give four kinds of objective evaluation indexs that above-mentioned three kinds of method statistics are obtained: unique point always mate count, correctly coupling count, correct matching rate and unique point repetition rate, as shown in table 2.

Table 2 based on MSER, ASIFT method and the present invention to the Comparative result of four kinds of objective evaluation indexs

Algorithm	Total coupling is counted	Correct coupling is counted	Correct matching rate	Unique point repetition rate
					Based on MSER method	106	90	0.8491	0.1466
Based on ASIFT method	1420	1375	0.9683	0.0448
					The inventive method	137	137	1.0	0.7874

Data as can be seen from table 2, the present invention, compared with the method based on maximum extremal region MSER, can obtain more coupling and count; The present invention compares with the method based on complete affine invariants ASIFT, although based on the method for complete affine invariants ASIFT in correct matching double points number higher than the inventive method, but the unique point repetition rate based on the method for complete affine invariants ASIFT is extremely low, which results in the waste of a large amount of storage space and higher computation complexity.Therefore, the inventive method can not only obtain higher proper characteristics matching rate and unique point repetition rate, and also has certain advantage in raising operation efficiency.

Claims (7)

1., based on a method for registering images for maximum stable extremal region and phase equalization, comprise the steps:

(1) there are two width image A and B of affined transformation in input respectively, and wherein A is reference picture, and B is image subject to registration;

(2) maximum stable extremal region MSER detection and coupling are carried out to reference picture A and image B subject to registration;

(3) respectively matching is carried out to the maximum stable extremal region that reference picture A and image B subject to registration matches, and obtain reference picture A expand after ellipse fitting region and image B subject to registration expand after ellipse fitting region;

(4) to above-mentioned two ellipse fitting region normalization:

4a) treat normalized point in computing reference image A and image B subject to registration respectively:

$z_A=\frac{1}{{\left(\min \right[\left|\det \left(M_A\right)\right|,\left|\det \left(M_B\right)\right|\left]\right)}^{1/4}}{H_A}^{-1}{M}_A^{1/2}(x_A^{\′}-{\mathrm{\μ}}_A)$

$z_B=\frac{1}{{\left(\min \right[\left|\det \left(M_A\right)\right|,\left|\det \left(M_B\right)\right|\left]\right)}^{1/4}}{H_B}^{-1}{M}_B^{1/2}(x_B^{\′}-{\mathrm{\μ}}_B)$

Wherein, z _aand z _brepresent in reference picture A and image B subject to registration respectively and treat normalized point, M _aand M _brepresent the second-order moments matrix of the barycenter of all maximum stable extremal region MSER in reference picture A and image B subject to registration respectively, H _aand H _brepresent second-order moments matrix M respectively _aand M _bthe symmetrical unitary matrix of the reality that svd obtains, x ' _awith x ' _brepresent the point of the elliptic region after expanding in image A and B respectively, μ _aand μ _brepresent the average of the barycenter of all maximum extremal region MSER in reference picture A and image B subject to registration respectively;

4b) with the normalization point z that needs in reference picture A _aform the normalization region P of reference picture, need normalization point z with image B subject to registration _bform the normalization region Q of image subject to registration;

(5) respectively the band reduction of fractions to a common denominator solution based on Gabor filter is carried out to the normalization region P of reference picture A and the normalization region Q of image B subject to registration, obtain the sub-band images that this two width image comprises different frequency composition;

(6) feature point detection based on the maximum square of phase equalization is carried out to the sub-band images of above-mentioned two width images, and the point set registration based on probability distribution is carried out to the unique point detected, obtain the transformation matrix T between point set ₁;

(7) according to the normalization region P of reference picture and the normalization region Q of image subject to registration, the transformation matrix T between reference picture A and image B subject to registration is estimated _c1, T _c2;

(8) according to the transformation matrix T between point set ₁and the transformation matrix T between reference picture A and image B subject to registration _c1, T _c2affine transformation matrix T between computing reference image A and image B subject to registration:

T＝T _c1- ¹T ₁T _c2；

(9) treat registering images B according to affine transformation matrix T to convert, then carry out bilinear interpolation to converting the image obtained, complete image registration.

2. the method for registering images based on maximum extremal region and phase equalization according to claim 1, wherein described in step (2), maximum stable extremal region MSER detection and coupling are carried out to reference picture A and image B subject to registration, carry out as follows:

2a) maximum stable extremal region MSER detection is carried out respectively to reference picture A and image B subject to registration, obtain multiple irregular extremal region with affine-invariant features;

2b) by multiple irregular extremal region one_to_one corresponding with affine-invariant features, obtain initial maximum stable extremal region MSER coupling right.

3. the affined transformation method for registering images based on maximum stable extremal region and phase equalization according to claim 1, wherein described in step (3), respectively matching is carried out to the maximum stable extremal region that reference picture A and image B subject to registration matches, carries out as follows:

3a) detect the barycenter of the maximum stable extremal region MSER of reference picture A and image B subject to registration respectively;

3b) according to the barycenter of the maximum stable extremal region MSER obtained in described two width images, calculate the point of this two width image fitted area according to the following formula:

(x _A-μ _A) ^TU _A ^-1(x _A-μ _A)＝(x _A-μ _A) ^TM _A(x _A-μ _A)＝1

(x _B-μ _B) ^TU _B ^-1(x _B-μ _B)＝(x _B-μ _B) ^TM _B(x _B-μ _B)＝1

Wherein, x _aand x _brepresent the point of fitted area in reference picture A and image B subject to registration respectively, μ _aand μ _brepresent the average of the barycenter of all maximum stable extremal region MSER in reference picture A and image B subject to registration respectively, T represents transposition, U _aand U _brepresent the variance of the barycenter of all maximum stable extremal region MSER in reference picture A and image B subject to registration respectively, M _aand M _brepresent the second-order moments matrix of the barycenter of all maximum stable extremal region MSER in reference picture A and image B subject to registration respectively;

3c) in reference picture A and image B subject to registration, form initial ellipse fitting region with the respective fitted area point obtained respectively;

3d) initial the maximum of ellipse fitting region allows exaggerated scale in computing reference image A according to the following formula:

$k_A=\min [\frac{r_A-u_a}{\max \left(x_a\right)-u_a},\frac{c_A-v_a}{\max \left(y_a\right)-v_a},\frac{u_a-1}{u_a-\min \left(x_a\right)},\frac{v_a-1}{v_a-\min \left(y_a\right)}],$

Wherein, k _arepresent that comprising the maximum of the ellipse long and short shaft of initial fitted area in reference picture A allows expansion multiple, r _aand c _arepresent line number and the columns of reference picture A respectively, u _aand v _arepresent the HCCI combustion of the barycenter of all maximum stable extremal region MSER in reference picture A respectively, x _aand y _arepresent row-coordinate and the row coordinate of fitted area point in reference picture A respectively;

3e) calculate initial the maximum of ellipse fitting region in image B subject to registration according to the following formula and allow exaggerated scale:

$k_B=\min [\frac{r_B-u_b}{\max \left(x_b\right)-u_b},\frac{c_B-v_b}{\max \left(y_b\right)-v_b},\frac{u_b-1}{u_b-\min \left(x_b\right)},\frac{v_b-1}{v_b-\min \left(y_b\right)}]$

Wherein, k _brepresent in image B subject to registration that comprising the maximum of the ellipse long and short shaft of initial fitted area allows expansion multiple, r _band c _brepresent line number and the columns of image B subject to registration respectively, u _band v _brepresent the HCCI combustion of the barycenter of all coupling maximum stable extremal regions in image B subject to registration respectively, x _band y _brepresent fitted area point row-coordinate and row coordinate in image B subject to registration respectively;

3f) allow exaggerated scale k with reference to the maximum of ellipse fitting region initial in image A _awith the exaggerated scale k in ellipse fitting region initial in image B subject to registration _bin the less exaggerated scale k as initial ellipse fitting region, k=min (k _a, k _b);

3g) according to the point expanding rear elliptic region in exaggerated scale k computing reference image A and image B subject to registration:

(x′ _A-μ _A) ^TM _A(x′ _A-μ _A)＝k ²

(x′ _B-μ _B) ^TM _B(x′ _B-μ _B)＝k ²

Wherein, x ' _awith x ' _brepresent the point of the elliptic region after expanding in reference picture A and image B subject to registration respectively, μ _aand μ _brepresent the average of the barycenter of all coupling maximum stable extremal regions in reference picture A and image B subject to registration respectively, M _aand M _brepresent the second-order moments matrix of the barycenter of all coupling maximum stable extremal regions in reference picture A and image B subject to registration respectively;

The point x ' of the elliptic region after 3h) expanding with reference picture A respectively _athe point x ' of the elliptic region after expanding with image B subject to registration _bform the ellipse fitting region of this two width image.

4. the affined transformation method for registering images based on maximum stable extremal region and phase equalization according to claim 1, respectively the band reduction of fractions to a common denominator solution based on Gabor filter is carried out to the normalization region obtained in reference picture A and image B subject to registration wherein described in step (5), carries out as follows:

5a) design has the Gabor band-pass filter group G (u, v, λ) of 5 bandpass filter:

$G(u,v,\mathrm{\λ})=\frac{\mathrm{\π}}{K}\underset{i=-K}{\overset{K}{\mathrm{\Σ}}}{e}^{-2{\mathrm{\π}}^2{\mathrm{\λ}}^2({(u-\frac{\cos {\mathrm{\θ}}_i}{2\mathrm{\λ}})}^2+{(v-\frac{\sin {\mathrm{\θ}}_i}{2\mathrm{\λ}})}^2)},$

Wherein, u and v represents the frequency domain coordinates of bandpass filter, and K represents the direction number of each bandpass filter, and the value of K is 6, θ _irepresent bandpass filter towards, i=-6 ,-5 ,-4 ..., 4,5,6, λ is the scale factor of bandpass filter, and the λ value of each bandpass filter is different, and namely first bandpass filter value is second bandpass filter value be the 2, three bandpass filter value be 4th bandpass filter value be the 4, five bandpass filter value be

5b) utilize the band-pass filter group of design, according to the following formula band reduction of fractions to a common denominator solution carried out to the normalization region obtained from reference picture A and image B subject to registration:

$I_A^{\mathrm{\λ}}(x,y)=F^{-1}[G(u,v,\mathrm{\λ})\×F[I_A(x,y)\left]\right]$

$I_B^{\mathrm{\λ}}(x,y)=F^{-1}[G(u,v,\mathrm{\λ})\×F[I_B(x,y)\left]\right]$

Wherein, I _a(x, y) and I _b(x, y) represents the normalization region obtained from reference picture A and image B subject to registration respectively, and F [] represents Fourier transform, F ^-1[] represents inverse Fourier transform, represent the sub-band images corresponding to reference picture A, represent the sub-band images corresponding to image B subject to registration.

5. the affined transformation method for registering images based on maximum stable extremal region and phase equalization according to claim 1, the feature point detection of carrying out based on the maximum square of phase equalization to the sub-band images of above-mentioned two width images wherein described in step (6), carry out as follows:

6a) respectively to each sub-band images of reference picture A, carrying out the feature point detection based on the maximum square of phase equalization, selecting unique point and the feature point detection result that it can be used as reference picture A from comprising the maximum sub-band images of unique point number;

6b) form a point set with reference to the unique point detected in image A according to the following formula:

X＝[x ₁x ₂…x _n… x _N]100≤N≤500

Wherein, X represents the point set that the unique point detected in reference picture A is formed, x _nrepresent the n-th unique point detected in reference picture A, n=1,2 ..., N, N represent the number of the unique point detected in reference picture A;

6c) treating each sub-band images of registering images B respectively, carry out the feature point detection based on the maximum square of phase equalization, selecting unique point and the feature point detection result that it can be used as image B subject to registration from comprising the maximum sub-band images of unique point number;

6d) according to the following formula the unique point detected in image B subject to registration is formed a point set:

Y＝[y ₁y ₂…y _m… y _M]100≤M≤500

Wherein, Y represents the point set that the unique point detected in image B subject to registration is formed, y _mrepresent m the unique point detected in image B subject to registration, m=1,2 ..., M, M represent the number of the unique point detected in image B subject to registration.

6. the affined transformation method for registering images based on maximum stable extremal region and phase equalization according to claim 1, the point set registration carrying out based on probability distribution to the unique point detected wherein described in step (6), obtains the transformation matrix T between point set ₁, carry out as follows:

6e) using image subject to registration detect in point set Y a little as the center of fiqure of gauss hybrid models GMM;

6f) detect any point x in point set X according to reference picture _nthe corresponding relation of the point in point set Y is detected, formation condition probability density function with image subject to registration:

$P\left(x_n\right|\mathrm{\σ},T_1)=\frac{1}{M{\left(2\mathrm{\π}{\mathrm{\σ}}^2\right)}^{3/2}}\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{e}^{(-\frac{{\left|\right|x_n-T_1{y}_m\left|\right|}^2}{2{\mathrm{\σ}}^2})},$

Wherein, σ represents the standard deviation of single Gaussian function in gauss hybrid models, and M represents that image subject to registration detects the number of element in point set Y;

Institute 6g) detected in point set X according to reference picture a little detect with image subject to registration in point set Y corresponding relation a little, generation log-likelihood estimation function:

$l(\mathrm{\σ},T_1)=\log \underset{n=1}{\overset{N}{\mathrm{\Π}}}P\left(x_n\right|\mathrm{\σ},T_1);$

6 _h) utilize expectation maximization EM algorithm, calculate send as an envoy to log-likelihood estimation function l (σ, T ₁) T when obtaining extreme value ₁matrix.

7. the affined transformation method for registering images based on maximum extremal region MSER and phase equalization according to claim 1, the estimation reference picture A wherein described in step (7) and the transformation matrix T between image B subject to registration _c1, T _c2, calculate according to the following formula:

$T_{c1}=\left[\begin{array}{cc}{M}_A^{1/2}{H}_B& -{\mathrm{\μ}}_A\\ 0& 1\end{array}\right],T_{c2}=\left[\begin{array}{cc}{\mathrm{RM}}_B^{1/2}{H}_B& -{\mathrm{\μ}}_B\\ 0& 1\end{array}\right]$

Wherein, $R=\left[\begin{array}{cc}\cos \mathrm{\θ}& -\sin \mathrm{\θ}\\ \sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right],$ M _aand M _brepresent the second-order moments matrix of the barycenter of all coupling maximum stable extremal regions in reference picture A and image B subject to registration respectively, H _aand H _brepresent second-order moments matrix M respectively _aand M _bthe symmetrical unitary matrix of the reality that svd obtains, μ _aand μ _brepresent the average of the barycenter of all coupling maximum stable extremal regions in reference picture A and image B subject to registration respectively, θ represents the anglec of rotation of image B subject to registration relative to reference picture A.