CN104616280B - Method for registering images based on maximum stable extremal region and phase equalization - Google Patents

Abstract

The invention discloses a kind of method for registering images based on maximum stable extremal region and phase equalization, mainly solve that characteristic point repetitive rate that prior art extracts is low and the high defect of computational complexity.Implementation step is：1st, there are the two images of affine transformation and carry out maximum stable extremal region detection and matching respectively in input；2nd, the matching area of two images is fitted, and is expanded and normalized；3rd, band logical decomposition is made to two regions after normalization；4th, characteristic point of the detection based on the maximum square of phase equalization, sets up the characteristic point probability distribution detected；5. estimate the accurate affine transformation matrix between two point sets；6th, the transformation matrix between two images is estimated according to two regions after normalization；7th, calculate the accurate affine transformation matrix between two images and complete image registration.The present invention can extract the characteristic point with higher repetition rate and correct matching rate, improve computational efficiency, available for image co-registration, image mosaic and three-dimensional reconstruction.

Description

Method for registering images based on maximum stable extremal region and phase equalization

Technical field

The invention belongs to technical field of image processing, and in particular to a kind of affine transformation method for registering images, can be applied to The fields such as image co-registration, image mosaic and three-dimensional reconstruction.

Background technology

In image co-registration, the field such as image mosaic and three-dimensional reconstruction to several views of Same Scene, it is necessary to first match somebody with somebody Quasi- processing.Generally, image registration can be carried out using the method for registering images of feature based, this mainly considers one A little characteristics of image have consistency for the yardstick of image and rotation, and only find the geometrical relationship between image with characteristic information Have the advantages that computational efficiency is high.But, it is difficult often to carry in them when there is larger affine transformation between two images Get with the accurate feature of higher repetition rate or position, so as to cause registration accuracy not enough or even asking for registration can not be realized Topic.

At present, the characteristic information commonly used in the method for registering images of feature based has scale invariant feature SIFT, maximum steady Determine extremal region MSER features and complete affine invariants ASIFT, 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, pp.761-767. and Morel J M, Yu G, " ASIFT:A new framework for fully affine invariant image comparison.”SIAM Journal on Imaging Sciences,vol.2, Technology disclosed in this three documents of no.2, pp.438-469. is feature extraction and matching process, and then can utilize matching Feature realizes image registration come the geometric transformation parameter calculated between image.Wherein, the image based on scale invariant feature SIFT is matched somebody with somebody Quasi- method can the larger image of registering yardstick, and obtain preferable registration effect.But, it is larger imitative when existing between image When penetrating conversion, the characteristic detection method based on scale invariant feature SIFT is often less able to obtain number enough and accuracy is high Matching double points, therefore the method for registering images based on scale invariant feature SIFT can not registration have larger affine transformation figure Picture.Based on maximum stable extremal region MSER method for registering images, using maximum stable extremal region MSER barycenter conduct Characteristic point is matched, and then estimates the affine transformation parameter between image, because maximum stable extremal region MSER is with higher Affine transformation consistency, therefore, it is possible to realize the image registration that there is larger affine transformation, but be due to imaging sensor and The difference of imaging circumstances, the barycenter of use tends not to the position of accurate reflection feature, so as to cause registration accuracy not high.Completely The affine space sampling that affine invariants ASIFT algorithms are artificially simulated to original image first, obtains several views；Then Feature extraction and characteristic matching are carried out to several views of acquisition using scale invariant feature SIFT methods, ratio is so resulted in Scale invariant feature SIFT methods more match points, therefore the image based on complete affine invariants ASIFT features is matched somebody with somebody Quasi- method registering can have the image of larger affine transformation.The deficiency that this method is present is, due to this method to image imitative Penetrate and spatially simulated, form the image at each visual angle, consume substantial amounts of internal memory, while extracting substantial amounts of correct With when also introduced substantial amounts of Mismatching point, and to obtain higher image registration accuracy, it is necessary to more complicated optimization Process is deleted Mismatching point, and this undoubtedly adds computation complexity again.

The content of the invention

It is an object of the invention to improve the shortcoming of above-mentioned prior art, propose a kind of based on maximum stable extremal region With the affine transformation method for registering images of phase equalization, to obtain more preferable affine transformation image registration effect, and fortune is reduced Complexity is calculated, computational efficiency is improved.

To achieve the above object, the technical scheme is that：By based on maximum stable extremal region MSER features Thick matching obtains the partial fitting region of two width input pictures；Overcome what Affine distortion was brought using affine region method for normalizing The change of picture structure；Band logical decomposition is carried out to normalization region using Gabor bandpass filters, and then in each sub-band images It is middle to carry out the feature point detection based on the maximum square of phase equalization；The set of characteristic points of detection is entered using the method for probability distribution Row accuracy registration, and calculate the affine transformation matrix between two width input pictures.Its specific steps includes as follows：

(1) respectively there is the two images A and B of affine transformation in input, and wherein A is reference picture, and B is image subject to registration；

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

(3) reference picture A and image the B subject to registration maximum stable extremal region matched are fitted respectively, and The ellipse fitting region behind ellipse fitting region and image B subject to registration expansions after expanding to reference picture A；

(4) above-mentioned two ellipse fitting region is normalized：

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

Wherein, z_AAnd z_BRepresent to treat normalized point, M in reference picture A and image B subject to registration respectively_AAnd M_BRepresent respectively The second-order moments matrix of all maximum stable extremal region MSER barycenter, H in reference picture A and image B subject to registration_AAnd H_BRespectively Represent second-order moments matrix M_AAnd M_BThe real symmetrical unitary matrice that singular value decomposition is obtained, x '_AWith x '_BRepresent to expand in image A and B respectively The point of elliptic region after big, μ_AAnd μ_BAll maximum extremal region MSER in reference picture A and image B subject to registration are represented respectively Barycenter average；

4b) with needing normalization point z in reference picture A_AThe normalization region P of reference picture is constituted, figure subject to registration is used As B needs normalization point z_BConstitute the normalization region Q of image subject to registration；

(5) reference picture A normalization region P and image B subject to registration normalization region Q are based on respectively The band logical of Gabor filter is decomposed, and obtains the sub-band images that this two images includes different frequency composition；

(6) sub-band images to above-mentioned two images carry out the feature point detection based on the maximum square of phase equalization, and right The characteristic point detected carries out the point set registration based on probability distribution, obtains the transformation matrix T between point set₁；

(7) according to the normalization region P and the normalization region Q of image subject to registration of reference picture, estimation reference picture A with Transformation matrix T between image B subject to registration_c1, T_c2；

(8) according to the transformation matrix T between point set₁Transformation matrix T between reference picture A and image B subject to registration_c1, T_c2 Calculate the affine transformation matrix T between reference picture A and image B subject to registration：

T=T_c1 ^-1T₁T_c2；

(9) enter line translation to image B subject to registration according to affine transformation matrix T, then two-wire is carried out to the image that conversion is obtained Property interpolation, complete image registration.

The present invention has advantages below compared with prior art：

First, the present invention has carried out being based on maximum stable extremal region due to the reference picture to input and image subject to registration MSER registration, and carry out in the ellipse fitting region of acquisition the band logical based on Gabor filter and decompose and phase equalization Maximum moment characteristics point detection, during improving prior art to there is the progress feature extraction of the image of larger affine transformation, It is difficult to obtain higher feature point repetitive rate and the defect of correct matching rate so that present invention design is extracted in these cases Characteristic point, with higher characteristic point repetitive rate and correct matching rate.

Second, the present invention improves prior art in feature as a result of the point set registration strategies based on probability distribution The defect for a little setting up higher-dimension descriptor is characterized in the matching process of point so that the present invention takes less compared with prior art Memory space, and with higher computational efficiency.

Brief description of the drawings

Fig. 1 is implementation process figure of the invention；

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

Fig. 3 is the present invention to there is the registering simulated effect figure of larger affine transformation image.

Embodiment

The present invention will be further described below in conjunction with the accompanying drawings.

Referring to the drawings 1, of the invention comprises the following steps that：

Step 1, input picture：Respectively there are the two images of affine transformation in input, and a width is another as reference picture A Width is used as image B subject to registration.

Step 2, maximum stable extremal region MSER detections and matching are carried out to reference picture A and image B subject to registration.

Maximum stable extremal region MSER detections 2a) are carried out respectively to reference picture A and image B subject to registration, obtain multiple Irregular extremal region with affine-invariant features；

2b) multiple irregular extremal regions with affine-invariant features are corresponded, obtain initial maximum steady Determine extremal region MSER matchings pair.

Step 3, reference picture A and image the B subject to registration maximum stable extremal region matched are fitted respectively, obtained The ellipse fitting region behind ellipse fitting region and image B subject to registration expansions after expanding to reference picture A.

This example is used but is not limited to using the matching algorithm based on maximum stable extremal region MSER to described two The matching area of width image is fitted, and its step is as follows：

Reference picture A and image B subject to registration maximum stable extremal region MSER barycenter 3a) is detected respectively；

3b) according to the maximum stable extremal region MSER obtained in the two images barycenter, this is calculated according to the following formula The point of two images fitted area：

(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_BThe point of fitted area in reference picture A and image B subject to registration, μ are represented respectively_AAnd μ_BRepresent respectively The average of all maximum stable extremal region MSER barycenter in reference picture A and image B subject to registration, T represents transposition, U_AAnd U_B The variance of all maximum stable extremal region MSER barycenter in reference picture A and image B subject to registration, M are represented respectively_AAnd M_BPoint Not Biao Shi in reference picture A and image B subject to registration all maximum stable extremal region MSER barycenter second-order moments matrix；

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

The maximum allowable exaggerated scale in initial ellipse fitting region in reference picture A 3d) is calculated according to the following formula：

Wherein, k_AThe maximum allowable expansion multiple of the ellipse long and short shaft comprising initial fitted area in reference picture A is represented, r_AAnd c_AReference picture A line number and columns, u are represented respectively_aAnd v_aAll maximum stable extremals in reference picture A are represented respectively The HCCI combustion of region MSER barycenter, x_aAnd y_aThe row coordinate and row coordinate of fitted area point in reference picture A are represented respectively；

The maximum allowable exaggerated scale in initial ellipse fitting region in image B subject to registration 3e) is calculated according to the following formula：

Wherein, k_BRepresent the maximum allowable expansion times of the ellipse long and short shaft comprising initial fitted area in image B subject to registration Number, r_BAnd c_BImage B subject to registration line number and columns, u are represented respectively_bAnd v_bRepresent that all matchings are most in image B subject to registration respectively The HCCI combustion of the barycenter of big extremal region, x_bAnd y_bRepresent that fitted area point row coordinate and row are sat in image B subject to registration respectively Mark；

3f) in comparison reference image A initial ellipse fitting region maximum allowable exaggerated scale k_AIn image B subject to registration The exaggerated scale k in initial ellipse fitting region_B, using the less exaggerated scale as initial ellipse fitting region of numerical value in both K, i.e. k=min (k_A,k_B)；

The elliptic region point after expanding in reference picture A and image B subject to registration 3g) is calculated according to exaggerated scale k：

(x′_A-μ_A)^TM_A(x′_A-μ_A)=k²

(x′_B-μ_B)^TM_B(x′_B-μ_B)=k²

Wherein, x '_AWith x '_BThe point of the elliptic region after expanding in reference picture A and image B subject to registration, μ are represented respectively_AWith μ_BThe average of the barycenter of all matching maximum stable extremal regions in reference picture A and image B subject to registration, M are represented respectively_AAnd M_B The second-order moments matrix of the barycenter of all matching maximum stable extremal regions in reference picture A and image B subject to registration is represented respectively；

The point x ' of elliptic region after 3h) being expanded respectively with reference picture A_AArea elliptica after expanding with image B subject to registration The point x ' in domain_B, constitute the ellipse fitting region of this two images, i.e. reference picture A ellipse fitting region and image B subject to registration Ellipse fitting region.

Step 4, above-mentioned two ellipse fitting region is normalized：

4a) calculate respectively in reference picture A according to the following formula and treat normalized point z_ATreat normalized with image B subject to registration Point z_B：

Wherein, M_AAnd M_BAll matching maximum stable extremal regions in reference picture A and image B subject to registration are represented respectively The second-order moments matrix of MSER barycenter, H_AAnd H_BSecond-order moments matrix M is represented respectively_AAnd M_BThe real symmetrical tenth of the twelve Earthly Branches that singular value decomposition is obtained Matrix, x '_AWith x '_BThe point of the elliptic region after expanding in reference picture A and image B subject to registration, μ are represented respectively_AAnd μ_BDifference table Show the average of all matching maximum stable extremal region MSER barycenter in reference picture A and image B subject to registration；

4b) with needing normalization point z in reference picture A_AThe normalization region P of reference picture is constituted, figure subject to registration is used As needing normalization point z in B_BConstitute the normalization region Q of image subject to registration.

Step 5, reference picture A normalization region P and image B subject to registration normalization region Q are based on respectively The band logical of Gabor filter is decomposed, and obtains the sub-band images that this two images includes different frequency composition.

Have to the wave filter of the carry out band logical decomposition in image normalization region：Gaussian-Laplace bandpass filters, DOG wave filters etc..This example is using Gabor filter respectively to reference picture A normalization region M and image B subject to registration Normalize region N and carry out band logical decomposition, its step is as follows：

5a) Gabor bandpass filter group G (u, v, λ) of the design with 5 bandpass filters：

Wherein, u and v represent the frequency domain coordinates of bandpass filter, and K represents the direction number of each bandpass filter, K value For 6, θ_iThe direction of bandpass filter is represented,I=-6, -5, -4 ..., 4,5,6, λ be bandpass filter yardstick because Son, the λ values of each bandpass filter are different, i.e., first bandpass filter value isSecond bandpass filter value For 2, the 3rd bandpass filter value is4th bandpass filter value for 4, the 5th bandpass filter value is

5b) using the bandpass filter group of above-mentioned design, according to the following formula to being obtained from reference picture A and image B subject to registration The normalization region obtained carries out band logical decomposition：

Wherein, I_A(x, y) and I_B(x, y) represents the normalization area obtained from reference picture A and image B subject to registration respectively Domain, F [] represents Fourier transformation, F^-1[] represents inverse Fourier transform,Represent the son corresponding to reference picture A Band image,Represent the sub-band images corresponding to image B subject to registration.

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

Method to the sub-band images progress feature point detection of image has：Scale invariant feature SIFT methods, it is completely affine Invariant features ASIFT methods etc..This example uses the feature point detecting method based on the maximum square of phase equalization to above-mentioned two width The image that carries of image carries out feature point detection, and its step is as follows：

6a) each sub-band images respectively to reference picture A, carry out the characteristic point inspection based on the maximum square of phase equalization Survey, then from most sub-band images of being counted out comprising feature select characteristic point and as reference picture A characteristic point examine Survey result；

The characteristic point detected in reference picture A 6b) is constituted into a point set according to the following formula：

X=[x₁ x₂ … x_n … x_N] 100≤N≤500,

Wherein, X represents the point set that the characteristic point detected in reference picture A is constituted, x_nRepresent to detect in reference picture A N-th of characteristic point, n=1,2 ..., N, N represent the number of characteristic point that is detected in reference picture A；

6c) each sub-band images respectively to image B subject to registration, carry out the characteristic point inspection based on the maximum square of phase equalization Survey, then characteristic point is selected as image B subject to registration feature point detection knot from most sub-band images of being counted out comprising feature Really；

6d) constitute a point set with the characteristic point detected in image B subject to registration according to the following formula：

Y=[y₁ y₂ … y_m … y_M] 100≤M≤500

Wherein, Y represents the point set that the characteristic point detected in image B subject to registration is constituted, y_mRepresent to examine in image B subject to registration M-th of the characteristic point measured, m=1,2 ..., M, M represent the number of characteristic point that is detected in image B subject to registration.

Step 7, the point set registration based on probability distribution is carried out to the above-mentioned characteristic point detected, the conversion between point set is obtained Matrix T₁。

Have to the method that the characteristic point that detection is obtained carries out point set registration：Scale invariant feature SIFT methods, it is completely affine Invariant features ASIFT methods etc..This example uses the point set method for registering for being based on probability distribution to the above-mentioned spy detected A progress point set registration is levied, its step is as follows：

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

Any point x in point set X 7b) is detected according to reference picture_nWith pair of the point in image detection point set Y subject to registration It should be related to, formation condition probability density function：

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

7c) according to reference picture detect in point set X a little with image detection point set Y subject to registration pair a little It should be related to, generate log-likelihood estimation function：

Expectation maximization EM algorithms 7d) are utilized, send as an envoy to log-likelihood estimation function l (σ, T is calculated₁) obtain extreme value when T₁ Matrix.

Step 8, according to the normalization region M and the normalization region N of image subject to registration of reference picture, reference picture is estimated Transformation matrix T between A and image B subject to registration_c1, T_c2。

Wherein,M_AAnd M_BAll matchings in reference picture A and image B subject to registration are represented respectively The second-order moments matrix of the barycenter of maximum stable extremal region, H_AAnd H_BSecond-order moments matrix M is represented respectively_AAnd M_BSingular value decomposition is obtained The real symmetrical unitary matrice arrived, μ_AAnd μ_BAll maximum extremal regions of matching in reference picture A and image B subject to registration are represented respectively The average of barycenter, θ represents the anglecs of rotation of the image B subject to registration relative to reference picture A.

Step 9, according to the transformation matrix T between point set₁Transformation matrix T between reference picture A and image B subject to registration_c1, T_c2, calculate the affine transformation matrix T between reference picture A and image B subject to registration：

T=T_c1 ^-1T₁T_c2；

Step 10, enter line translation to image B subject to registration according to affine transformation matrix T, then the image that conversion is obtained is carried out Bilinear interpolation, completes image registration.

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

1. simulated conditions：All emulation experiments are all soft using Matlab R2009a under Windows XP operating systems Part is realized.

2. emulation content：

Emulation 1

The present invention is subjected to registering experimental result to one group of large scale modified-image with existing based on scale invariant feature SIFT is compared to the experimental result of this group of image, as a result such 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 that with the method for registering images based on scale invariant feature SIFT two width input pictures are carried out with registration As a result,

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

Fig. 2 (e) is the result that using the present invention two width input pictures are carried out with registration,

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

Figure it is seen that the registration result pair obtained using the method for registering images based on scale invariant feature SIFT There is obvious registration error in the regional area answered, and using the registration of the registration result corresponding regional area of the invention obtained Error is smaller.

The method for registering images based on scale invariant feature SIFT is counted with method for registering images of the invention in characteristic point Total matching points, correct matching points, four kinds of objective evaluation indexs of correct matching rate and characteristic point repetitive rate, as shown in table 1.

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

Algorithm	Total matching points	Correct matching points	Correct matching rate	Characteristic point repetitive rate
					Based on SIFT methods	156	123	0.7885	0.1692
The inventive method	340	350	0.9714	0.4920

It can be seen from the data in Table 1 that method for registering images proposed by the invention, has large scale conversion in registration Image when, its four kinds of objective evaluation indexs will be better than the method for registering images based on scale invariant feature SIFT.

Emulation 2

With of the invention and existing method for registering images based on maximum stable extremal region MSER and based on completely affine There is larger affine transformation image to one group and carry out registering comparison in invariant features ASIFT method for registering images, as a result such 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) is to two width input pictures registration using the method for registering images based on maximum stable extremal region MSER Result,

Fig. 3 (d) is to two width input pictures registration using the method for registering images based on complete affine invariants ASIFT Result,

Fig. 3 (e) is the result to two width input pictures registration using the inventive method,

From figure 3, it can be seen that the registration obtained using the method for registering images based on maximum stable extremal region MSER is tied Really, registration result figure and use the inventive method that the method for registering images based on complete affine invariants ASIFT is obtained Obtained registration result is respectively provided with good visual effect.

In order to further compare the performance of each algorithm, give four kinds obtained to above-mentioned three kinds of method statistics and objective comment Valency index：Characteristic point always matches points, correct matching points, correct matching rate and characteristic point repetitive rate, as shown in table 2.

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

Algorithm	Total matching points	Correct matching points	Correct matching rate	Characteristic point repetitive rate
					Based on MSER methods	106	90	0.8491	0.1466
Based on ASIFT methods	1420	1375	0.9683	0.0448
					The inventive method	137	137	1.0	0.7874

It can be seen from the data in Table 2 that the present invention is resulted in more compared with the method based on maximum extremal region MSER Many matching points；The present invention is compared with the method based on complete affine invariants ASIFT, although based on completely affine Invariant features ASIFT method is higher than the inventive method in terms of correct matching double points number, but based on complete affine constant spy The characteristic point repetitive rate for levying ASIFT method is extremely low, and which results in the waste of a large amount of memory spaces and higher calculating are multiple Miscellaneous degree.Therefore, the inventive method can not only obtain higher proper characteristics matching rate and characteristic point repetitive rate, and improve fortune Also there is certain advantage in terms of calculating efficiency.

Claims (7)

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

(1) respectively there is the two images A and B of affine transformation in input, and wherein A is reference picture, and B is image subject to registration；

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

(3) reference picture A and image the B subject to registration maximum stable extremal region matched are fitted respectively, and joined Examine the ellipse fitting region after image A expands and the ellipse fitting region after image B subject to registration expansions；

(4) above-mentioned two ellipse fitting region is normalized：

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

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

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

Wherein, z_AAnd z_BRepresent to treat normalized point, M in reference picture A and image B subject to registration respectively_AAnd M_BReference is represented respectively The second-order moments matrix of all maximum stable extremal region MSER barycenter, H in image A and image B subject to registration_AAnd H_BRepresent respectively Second-order moments matrix M_AAnd M_BThe real symmetrical unitary matrice that singular value decomposition is obtained, x '_AWith x '_BRepresent respectively after expanding in image A and B Elliptic region point, μ_AAnd μ_BThe matter of all maximum extremal region MSER in reference picture A and image B subject to registration is represented respectively The average of the heart；

4b) with needing normalization point z in reference picture A_AThe normalization region P of reference picture is constituted, with image B institutes subject to registration Need normalization point z_BConstitute the normalization region Q of image subject to registration；

(5) reference picture A normalization region P and image B subject to registration normalization region Q are carried out based on Gabor filters respectively The band logical of ripple device is decomposed, and obtains the sub-band images that this two images includes different frequency composition；

(6) sub-band images to above-mentioned two images carry out the feature point detection based on the maximum square of phase equalization, and to detection The characteristic point arrived carries out the point set registration based on probability distribution, obtains the transformation matrix T between point set₁；

(7) according to the normalization region P and the normalization region Q of image subject to registration of reference picture, estimation reference picture A matches somebody with somebody with waiting Transformation matrix T between quasi- image B_c1, T_c2；

(8) according to the transformation matrix T between point set₁Transformation matrix T between reference picture A and image B subject to registration_c1, T_c2Calculate Affine transformation matrix T between reference picture A and image B subject to registration：

T=T_c1 ^-1T₁T_c2；

(9) enter line translation to image B subject to registration according to affine transformation matrix T, then bilinearity carried out to the image that conversion is obtained to insert Value, completes image registration.

2. the method for registering images according to claim 1 based on maximum stable extremal region and phase equalization, wherein Maximum stable extremal region MSER detections and matching are carried out to reference picture A and image B subject to registration described in step (2), by such as Lower step is carried out：

Maximum stable extremal region MSER detections 2a) are carried out respectively to reference picture A and image B subject to registration, obtain multiple having The irregular extremal region of affine-invariant features；

2b) multiple irregular extremal regions with affine-invariant features are corresponded, obtain initial maximum stable pole It is worth region MSER matchings pair.

3. the method for registering images according to claim 1 based on maximum stable extremal region and phase equalization, wherein Being fitted respectively to reference picture A and image the B subject to registration maximum stable extremal region matched described in step (3), presses Following steps are carried out：

Reference picture A and image B subject to registration maximum stable extremal region MSER barycenter 3a) is detected respectively；

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

(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_BThe point of fitted area in reference picture A and image B subject to registration, μ are represented respectively_AAnd μ_BReference is represented respectively The average of all maximum stable extremal region MSER barycenter in image A and image B subject to registration, T represents transposition, U_AAnd U_BRespectively Represent the variance of all maximum stable extremal region MSER barycenter in reference picture A and image B subject to registration, M_AAnd M_BDifference table Show the second-order moments matrix of all maximum stable extremal region MSER barycenter in reference picture A and image B subject to registration；

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

The maximum allowable exaggerated scale in initial ellipse fitting region in reference picture A 3d) is calculated 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 the maximum allowable expansion multiple of the ellipse long and short shaft comprising initial fitted area in reference picture A, r_AWith c_AReference picture A line number and columns, u are represented respectively_aAnd v_aAll maximum stable extremal regions in reference picture A are represented respectively The HCCI combustion of MSER barycenter, x_aAnd y_aThe row coordinate and row coordinate of fitted area point in reference picture A are represented respectively；

The maximum allowable exaggerated scale in initial ellipse fitting region in image B subject to registration 3e) is calculated according to the following formula：

$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 the maximum allowable expansion multiple of the ellipse long and short shaft comprising initial fitted area in image B subject to registration, r_B And c_BImage B subject to registration line number and columns, u are represented respectively_bAnd v_bRepresent that all matchings are maximum steady in image B subject to registration respectively Determine the HCCI combustion of the barycenter of extremal region, x_bAnd y_bRepresent that fitted area point row coordinate and row are sat in image B subject to registration respectively Mark；

3f) by the maximum allowable exaggerated scale k in initial ellipse fitting region in reference picture A_AWith it is initial ellipse in image B subject to registration The exaggerated scale k of circle fitted area_BIn less exaggerated scale k, k=min (k as initial ellipse fitting region_A,k_B)；

The point of elliptic region after expanding in reference picture A and image B subject to registration 3g) is calculated according to exaggerated scale k：

(x′_A-μ_A)^TM_A(x′_A-μ_A)=k²

(x′_B-μ_B)^TM_B(x′_B-μ_B)=k²

Wherein, x '_AWith x '_BThe point of the elliptic region after expanding in reference picture A and image B subject to registration, μ are represented respectively_AAnd μ_BPoint Not Biao Shi in reference picture A and image B subject to registration the barycenter of all matching maximum stable extremal regions average, M_AAnd M_BRespectively Represent the second-order moments matrix of the barycenter of all matching maximum stable extremal regions in reference picture A and image B subject to registration；

The point x ' of elliptic region after 3h) being expanded respectively with reference picture A_AThe point of elliptic region after expanding with image B subject to registration x′_BConstitute the ellipse fitting region of this two images.

4. the method for registering images according to claim 1 based on maximum stable extremal region and phase equalization, wherein Respectively carrying out the normalization region obtained in reference picture A and image B subject to registration based on Gabor filtering described in step (5) The band logical of device is decomposed, and is carried out as follows：

5a) Gabor bandpass filter group G (u, v, λ) of the design with 5 bandpass filters：

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

Wherein, u and v represent the frequency domain coordinates of bandpass filter, and K represents the direction number of each bandpass filter, and K value is 6, θ_iThe direction of bandpass filter is represented,I=-6, -5, -4 ..., 4,5,6, λ be the scale factor of bandpass filter, The λ values of each bandpass filter are different, i.e., first bandpass filter value isSecond bandpass filter value be 2, the 3rd bandpass filter value be4th bandpass filter value for 4, the 5th bandpass filter value is

The bandpass filter group of design 5b) is utilized, according to the following formula the normalizing to being obtained from reference picture A and image B subject to registration Change region and carry out band logical decomposition：

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

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

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, F respectively [] represents Fourier transformation, F^-1[] represents inverse Fourier transform,Represent the subband figure corresponding to reference picture A Picture,Represent the sub-band images corresponding to image B subject to registration.

5. the method for registering images according to claim 1 based on maximum stable extremal region and phase equalization, wherein The sub-band images to above-mentioned two images described in step (6) carry out the feature point detection based on the maximum square of phase equalization, press Following steps are carried out：

6a) each sub-band images respectively to reference picture A, carry out the feature point detection based on the maximum square of phase equalization, from Characteristic point is selected in most sub-band images of being counted out comprising feature and as reference picture A feature point detection result；

The characteristic point detected in reference picture A 6b) is constituted into a point set according to the following formula：

X=[x₁ x₂ …x_n… x_N]100≤N≤500

Wherein, X represents the point set that the characteristic point detected in reference picture A is constituted, x_nRepresent n-th detected in reference picture A Individual characteristic point, n=1,2 ..., N, N represent the number of characteristic point that is detected in reference picture A；

6c) each sub-band images respectively to image B subject to registration, carry out the feature point detection based on the maximum square of phase equalization, Characteristic point is selected from most sub-band images of being counted out comprising feature and as image B subject to registration feature point detection knot Really；

The characteristic point detected in image B subject to registration 6d) is constituted into a point set according to the following formula：

Y=[y₁ y₂ …y_m… y_M]100≤M≤500

Wherein, Y represents the point set that the characteristic point detected in image B subject to registration is constituted, y_mRepresent to detect in image B subject to registration M-th of characteristic point, m=1,2 ..., M, M represent the number of characteristic point that is detected in image B subject to registration.

6. the method for registering images according to claim 1 based on maximum stable extremal region and phase equalization, wherein The point set registration carried out to the characteristic point that detects based on probability distribution described in step (6), obtains the transformation matrix between point set T₁, carry out as follows：

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

Any point x in point set X 6f) is detected according to reference picture_nIt is corresponding with the point in image detection point set Y subject to registration to close System, formation condition probability density function：

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

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

6g) according to reference picture detect institute in point set X a little with image detection point set Y subject to registration a little corresponding close System, generates log-likelihood estimation function：

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

Expectation maximization EM algorithms 6h) are utilized, send as an envoy to log-likelihood estimation function l (σ, T is calculated₁) obtain extreme value when T₁Matrix.

7. the method for registering images according to claim 1 based on maximum stable extremal region and phase equalization, wherein The transformation matrix T between estimation reference picture A and image B subject to registration described in step (7)_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,M_AAnd M_BRepresent that all matchings are maximum steady in reference picture A and image B subject to registration respectively Determine the second-order moments matrix of the barycenter of extremal region, H_AAnd H_BSecond-order moments matrix M is represented respectively_AAnd M_BThe reality that singular value decomposition is obtained Symmetrical unitary matrice, μ_AAnd μ_BThe matter of all matching maximum stable extremal regions in reference picture A and image B subject to registration is represented respectively The average of the heart, θ represents the anglecs of rotation of the image B subject to registration relative to reference picture A.