CN103714550B - A kind of image registration automatic optimization method based on match curve feature evaluation - Google Patents

Abstract

Match curve can not only registrate performance from the morphological characteristic directviewing description such as smoothness, sharpness, and can be from kurtosis (kurtosis), peak (peak deviation partially, the i.e. deviation of position and optimal location that global maximum occurs), peak value (maximum, i.e. global maximum) and peak value between the characteristic index qualitative assessment registration result such as root-mean-square error (maxinum_RMSE), when image is perfectly aligned, line smoothing, sharp-pointed, global maximum on each degree of freedom is equal, and be positioned at optimal location, i.e. peak is 0 partially.On this basis, present invention further propose that a kind of image registration automatic optimization method based on match curve feature evaluation, registration parameter is partially adjusted with the peak of match curve, set with the root-mean-square error between peak value and stop iterated conditional, constantly registration parameter is optimized, it is achieved image more accuracy registration.

Description

A kind of image registration automatic optimization method based on match curve feature evaluation

Technical field

The invention belongs to technical field of image processing, be specifically related to a kind of based on match curve feature evaluation The image registration automatic optimization method of (Matching Curve feature evaluation, MCfe).

Background technology

Image registration is an important step in art of image analysis.Existing method for registering is broadly divided into two Class: method based on pixel grey scale and the method for feature based.The method for registering early stage of feature based need by Feature detection operator extraction image feature information, including a feature, line feature, region feature and further feature Son is described；Match curve based on pixel grey scale can be from the morphological character directviewing description such as smoothness, sharpness Registration performance, when smoothness, sharpness are good, it is intended that registrate effective.Deep by match curve Enter research, find that when image is perfectly aligned, match curve global maximum on each degree of freedom is equal, and And being positioned at optimal location, i.e. peak is 0 partially.The present inventor is " a kind of based on coupling in a patent application before The image registration Evaluation Method of curvilinear characteristic " (application number: 201310645358.5) proposes with match curve Kurtosis (kurtosis), peak (position that peak deviation, i.e. global maximum occur and optimal location partially Deviation), root-mean-square error between peak value (maximum, i.e. global maximum) and peak value Etc. (maxinum_RMSE) it is characterized as quantitative target, on this basis, features described above based on match curve, A kind of image registration automatic optimization method is proposed.

Summary of the invention

The present invention proposes a kind of image registration automatic optimization method based on match curve feature evaluation, with coupling The peak of curve adjusts registration parameter partially, sets with the root-mean-square error between peak value and stops iterated conditional, constantly Registration parameter is optimized, it is achieved image more accuracy registration.Owing to the present invention uses the structure of correction similar Degree function (Modified Structural Similarity, MSSIM) is water rogulator, and the method is not only suitable for Single mode image, is also applied for multi-modality image registration.Concretely comprise the following steps:

The first step: read in image, according to image mismatch degree set optimization process iterations K, image it Between mismatch the most serious, method convergence needed for iterations K the biggest；Requirement according to registration accuracy sets joins Quasi-precision e: root-mean-square error maxinum_RMSE < e between each match curve peak value；

Second step: with revise structural similarity function MSSIM for estimating, set according to the requirement of registration accuracy Determining the change step of parameter, change step is the finest, and registration accuracy will be the highest, provides reference picture and floating Match curve on each degree of freedom of image, MSSIM metric is obtained by following formula:

$f_{MSSIM}(X,Y)=\frac{(2{\mathrm{\&mu;}}_X{\mathrm{\&mu;}}_Y+C_1)\left(2\right|{\mathrm{\&sigma;}}_{XY}|+C_2)}{({\mathrm{\&mu;}}_X^2+{\mathrm{\&mu;}}_Y^2+C_1)({\mathrm{\&sigma;}}_X^2+{\mathrm{\&sigma;}}_Y^2+C_2)}---\left(1\right)$

Wherein X represents the subimage of reference picture, Y represents the subimage of floating image, C₁、C₂For little just , zero there is instability, μ to prevent denominator from being in constant_XRepresent the luminance mean value of X, μ_YRepresent Y's Luminance mean value, σ_XThe luminance standard of expression X is poor, σ_YThe luminance standard of expression Y is poor, σ_XYExpression X, The brightness covariance of Y；

3rd step: calculate the most according to the following formula the kurtosis of each curve, peak partially, mean square between peak value and peak value Root error:

1) kurtosis (kurtosis)

$kurtosis=\frac{\mathrm{\&Sigma;}{(X_i-\stackrel{\&OverBar;}{X})}^4/n}{{(\mathrm{\&Sigma;}{\left(X_i-\stackrel{\&OverBar;}{X}\right)}^2/n)}^2}---\left(2\right)$

X_iFor sample measured value,For the meansigma methods of n measured value of sample, quote the concept of statistics kurtosis, but Difference is that sample measurements is MSSIM measure value, can weigh match curve equally maximum in the overall situation Intensity near value or acuity；

2) peak is partially (peak deviation)

F (X, Y) is the measure value of match curve, and max f (X, Y) represents the global maximum that match curve is estimated I.e. peak value, peak is partially the most different from the statistical degree of bias (Skewness), refer to position that global maximum occurs and The deviation of optimal location, translation, the optimal location of rotating curve are 0 position, the optimal location of scaling curve Being 1 position, when image is perfectly aligned, global maximum will be in optimal location, otherwise, the optimum position of deviation Putting, irrelevance is represented partially by peak, and the inclined ideal value in peak is 0；

3) peak value (maximum)

Peak value refers to the global maximum of match curve, i.e. f_max(X, Y), when image is perfectly aligned, each curve Peak value size is completely the same；

4) root-mean-square error (Root Mean Square Error, RMSE) between peak value

${\stackrel{\&OverBar;}{f}}_{max}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_{max}\left(i\right)---\left(4\right)$

$\max imum\_RMSE=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\&lsqb;f_{max}\left(i\right)-{\stackrel{\&OverBar;}{f}}_{max}\&rsqb;}^2}---\left(5\right)$

N is the degree of freedom of registration transformation, and maxinum_RMSE is effectively equivalent to standard error, reflects The consistent degree of each peak of curve, thus also reflect the degree of registration between image, i.e. registration accuracy, image is complete When entirely aliging, maxinum_RMSE ideal value is 0；

4th step: judge whether to meet iteration stopping condition, if be unsatisfactory for, each by descending secondary ordered pair Curve peak angle value is ranked up, and then registration parameter corresponding for the match curve that kurtosis value is maximum is optimized Adjusting, method of adjustment is:

New parameter value=old parameter value+α peak is partially (6)

Dynamic gene 0 < α < 1, carries out spatial alternation with the registration parameter after adjusting to original floating image, obtains new Floating image replace original floating image, repeat second step to the 4th step, wherein repeat four steps time, for The each degree of freedom of traversal, it is ensured that the registration parameter on each degree of freedom can be optimized, to new peak After angle value sequence, the registration parameter that the match curve that kurtosis value time is big is corresponding is optimized adjustment, adjustment side Method ibid, is repeated in above-mentioned steps, until registration parameter corresponding to the minimum match curve of kurtosis value obtains Optimizing and revising, if degree of freedom is N, then iteration n times complete once to circulate, and then move in circles；If Meet iteration stopping condition or reach the iterations upper limit set, then entering the 5th step；

5th step: original floating image is carried out spatial alternation with the registration parameter after optimizing and revising, by merging Image display registration effect, provides match curve simultaneously, and the smoothness of curve and sharpness can directviewing descriptions Registration performance before and after optimizing and revising.

The water rogulator that the present invention uses is MSSIM, or is the convex of other or recessed water rogulator, described Convex or recessed water rogulator includes that similarity measure and diversity are estimated, and described similarity measure has phase relation Count (CC), normalized crosscorrelation (NCC), structural similarity (SSIM), mutual information (MI) and return One changes mutual information (NMI), and described diversity estimates distance measure SSD, SAD, the RC having gray scale, repeatedly For the Euclidean distance that closest approach algorithm (iterative closest point, ICP) is inner, chamfering distance (chamfer Distance), Hausdorff distance and theory of information in Hamming distance (Hamming distance).

The registration transformation being suitable for include the translation of x direction, y direction translate, rotate around z-axis, consistent scaling Deng the affine transformation (N=4) of 4 degree of freedom, and it is generalized to any multivariant affine transformation, projection Conversion and nonlinear transformation.

Other algorithm that registration parameter optimization method uses, including golden section search algorithm, experiment proves with public The effect of optimization of formula (6) is close.

When mismatch is not serious when between image, the present invention is used to be directly registrable image；When image has registrated, Use present invention optimization further to registration parameter, make registration more accurate.

The positive effect of the present invention:

The present invention proposes a kind of image registration automatic optimization method based on match curve feature evaluation, can not only Enough it is directly registrable image, additionally it is possible to the image registrated is optimized further, makes registration more accurate, It is not only suitable for single mode image simultaneously, is also applied for multi-modality image registration.

Accompanying drawing explanation

Fig. 1 is satellite remote sensing images registration result, and size is 150 × 150 pixels, (a) reference picture, and (b) floats Motion video, floating image after the optimization of (c) registration, (d) is with (floating image after 0.5 × reference picture+0.5 × registration) The fusion image of display.

Fig. 2 optimizes front Fig. 1 (a) and the match curve of (b) for registration.

Fig. 3 is Fig. 1 (a) and the match curve of (c) after registration optimization.

Fig. 4 for (a) kurtosis, (b) peak during optimizing partially, between (c) peak value and (d) peak value root-mean-square error along with The convergence curve of iterations.

Fig. 5 is multi-modal medical science brain image MR-T2 and CT registration result, and size is 256 × 256 pixels, A (initial mismatch parameter is for () reference picture MR-T2, (b) floating image CT(c) registration optimize after floating image, (d) with (0.5 × with reference to figure As+0.5 × registration after floating image) shown in fusion image.

Fig. 6 optimizes front Fig. 5 (a) and the match curve of (b) for registration.

Fig. 7 is Fig. 5 (a) and the match curve of (c) after registration optimization.

Fig. 8 for (a) kurtosis, (b) peak during optimizing partially, between (c) peak value and (d) peak value root-mean-square error along with The convergence curve of iterations.

Table 1 in Fig. 9 is multi-modal medical science brain image registration result, and wherein initial mismatch parameter isCorresponding physical coordinates root-mean-square error RMSE=9.8961 picture Element.

Detailed description of the invention:

The method step of the present invention is further illustrated below in conjunction with the accompanying drawings with embodiment:

The first step: read in satellite remote sensing images [Fig. 1 (a) (b)], sets iterations during registration parameter optimizes K=15, concurrently set stop iteration condition: root-mean-square error maxinum_RMSE between each match curve peak value <0.0001；

Second step: with revise structural similarity function MSSIM for estimating, provide reference picture and floating figure As the match curve (Fig. 2) on each degree of freedom, wherein in order to ensure that line smoothing uses spline method, Translation is with " pixel " as unit, and the anglec of rotation is with " spending " as unit, and translating step is 0.2 pixel, rotates Step-length is 0.2 °, and scaling step-length is 0.005；

3rd step: calculate respectively the kurtosis of each curve, peak partially, root-mean-square error between peak value and peak value；

4th step: judge whether to meet iteration stopping condition, if be unsatisfactory for, by descending secondary ordered pair Each curve peak angle value is ranked up, and then registration parameter corresponding for the match curve that kurtosis value is maximum is carried out excellent Changing and adjust, method of adjustment is:

New parameter value=old parameter value+0.618 × peak is inclined

With the registration parameter after adjusting, original floating image is carried out spatial alternation, obtain new floating image and replace former Beginning floating image, repetition second step is to the 4th step, when wherein repeating four steps, in order to travel through each degree of freedom, Ensure that the registration parameter on each degree of freedom can be optimized, after new kurtosis value sequence, to kurtosis The registration parameter that the secondary big match curve of value is corresponding is optimized adjustment, and method of adjustment ibid, is repeated in State step, until registration parameter corresponding to the minimum match curve of kurtosis value is optimized and revised, if freely Degree is N, then iteration n times complete once to circulate, and then move in circles, and draws kurtosis, peak during optimizing Partially, between peak value and peak value root-mean-square error along with the convergence curve (Fig. 4) of iterations；If met repeatedly For stop condition or reach set the iterations upper limit, then enter the 5th step；

5th step: original floating image is carried out spatial alternation with the registration parameter after optimizing and revising, by merging Image display registration effect [Fig. 1 (c) (d)], provides match curve (Fig. 3) after registration optimizes, from curve simultaneously Smoothness and the morphological character directviewing description such as sharpness optimize and revise before and after registration performance；

6th step: read in multi-modal medical science brain image, image initial is alignment, initial floating image along x Direction moves horizontally to the right 3 pixels, and y direction is moved vertically downward 2 pixels, turns clockwise 4 degree, whole Body amplifies 1.05 times, and the image after spatial alternation is as original floating image [Fig. 5 (a) (b)]；

7th step: set iterations K=25 during registration parameter optimizes, concurrently set the bar stopping iteration Part: root-mean-square error maxinum_RMSE between each match curve peak value < 0.0001, repeat above-mentioned second step and arrive 5th step, obtains image [Fig. 5 (c) (d)], registration after match curve (Fig. 6) before registration optimizes, registration optimization Match curve (Fig. 7) after optimization, and during optimizing kurtosis, peak partially, root-mean-square between peak value and peak value Error is along with the convergence curve (Fig. 8) of iterations；

8th step: read in other multi-modal brain image, repeats above-mentioned 6th step and the 7th step, is given more than many groups Modality image co-registration result (table 1 in Fig. 9).

Claims (4)

1. based on match curve feature evaluation (a Matching Curve feature evaluation, MCfe) Image registration automatic optimization method, it is characterised in that: propose with the kurtosis of match curve, peak partially, peak value And the root-mean-square error amount of being characterized between each peak of curve, partially adjust registration parameter with peak, with between peak value all Square error sets and stops iterated conditional, the registration parameter on each degree of freedom of Automatic Optimal, it is achieved more accurate Registration, concretely comprise the following steps:

The first step: read in image, according to image mismatch degree set optimization process iterations K, image it Between mismatch the most serious, method convergence needed for iterations K the biggest；Requirement according to registration accuracy sets joins Quasi-precision e: root-mean-square error maxinum_RMSE < e between each match curve peak value；

Second step: with revise structural similarity function MSSIM for estimating, set according to the requirement of registration accuracy Determining the change step of parameter, change step is the finest, and registration accuracy will be the highest, provides reference picture and floating Match curve on each degree of freedom of image, MSSIM metric is obtained by following formula:

$f_{MSSIM}(X,Y)=\frac{(2{\mathrm{\&mu;}}_X{\mathrm{\&mu;}}_Y+C_1)\left(2\right|{\mathrm{\&sigma;}}_{XY}|+C_2)}{({\mathrm{\&mu;}}_X^2+{\mathrm{\&mu;}}_Y^2+C_1)({\mathrm{\&sigma;}}_X^2+{\mathrm{\&sigma;}}_Y^2+C_2)}---\left(1\right)$

Wherein X represents the subimage of reference picture, Y represents the subimage of floating image, C₁、C₂For little just , zero there is instability, μ to prevent denominator from being in constant_XRepresent the luminance mean value of X, μ_YRepresent Y's Luminance mean value, σ_XThe luminance standard of expression X is poor, σ_YThe luminance standard of expression Y is poor, σ_XYExpression X, The brightness covariance of Y；

3rd step: calculate the most according to the following formula the kurtosis of each curve, peak partially, mean square between peak value and peak value Root error:

1) kurtosis (kurtosis)

$kurtosis=\frac{\mathrm{\&Sigma;}{(X_i-\stackrel{\&OverBar;}{X})}^4/n}{{(\mathrm{\&Sigma;}{\left(X_i-\stackrel{\&OverBar;}{X}\right)}^2/n)}^2}---\left(2\right)$

X_iFor sample measured value,For the meansigma methods of n measured value of sample, quote the concept of statistics kurtosis, but Difference is that sample measurements is MSSIM measure value, can weigh match curve equally maximum in the overall situation Intensity near value or acuity；

2) peak is partially (peak deviation)

F (X, Y) is the measure value of match curve, and max f (X, Y) represents the global maximum that match curve is estimated I.e. peak value, peak is partially the most different from the statistical degree of bias (Skewness), refer to position that global maximum occurs and The deviation of optimal location, translation, the optimal location of rotating curve are 0 position, the optimal location of scaling curve Being 1 position, when image is perfectly aligned, global maximum will be in optimal location, otherwise, the optimum position of deviation Putting, irrelevance is represented partially by peak, and the inclined ideal value in peak is 0；

3) peak value (maximum)

Peak value refers to the global maximum of match curve, i.e. f_max(X, Y), when image is perfectly aligned, each curve Peak value size is completely the same；

4) root-mean-square error (Root Mean Square Error, RMSE) between peak value

${\stackrel{\&OverBar;}{f}}_{max}=\frac{1}{N}\underset{i=1}{\overset{N}{\&Sigma;}}{f}_{max}\left(i\right)---\left(4\right)$

$\max imum\_RMSE=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\&Sigma;}}{\&lsqb;f_{max}\left(i\right)-{\stackrel{\&OverBar;}{f}}_{max}\&rsqb;}^2}---\left(5\right)$

N is the degree of freedom of registration transformation, and maxinum_RMSE is effectively equivalent to standard error, reflects The consistent degree of each peak of curve, thus also reflect the degree of registration between image, i.e. registration accuracy, image is complete When entirely aliging, maxinum_RMSE ideal value is 0；

4th step: judge whether to meet iteration stopping condition, if be unsatisfactory for, each by descending secondary ordered pair Curve peak angle value is ranked up, and then registration parameter corresponding for the match curve that kurtosis value is maximum is optimized Adjusting, method of adjustment is:

New parameter value=old parameter value+α peak is partially (6)

Dynamic gene 0 < α < 1, carries out spatial alternation with the registration parameter after adjusting to original floating image, obtains new Floating image replace original floating image, repeat second step to the 4th step, wherein repeat four steps time, for The each degree of freedom of traversal, it is ensured that the registration parameter on each degree of freedom can be optimized, to new peak After angle value sequence, the registration parameter that the match curve that kurtosis value time is big is corresponding is optimized adjustment, adjustment side Method ibid, is repeated in above-mentioned steps, until registration parameter corresponding to the minimum match curve of kurtosis value obtains Optimizing and revising, if degree of freedom is N, then iteration n times complete once to circulate, and then move in circles；If Meet iteration stopping condition or reach the iterations upper limit set, then entering the 5th step；

5th step: original floating image is carried out spatial alternation with the registration parameter after optimizing and revising, by merging Image display registration effect, provides match curve simultaneously, and the smoothness of curve and sharpness can directviewing descriptions Registration performance before and after optimizing and revising.

2. image registration automatic optimization method based on match curve feature evaluation as claimed in claim 1, it is special Levy and be: wherein second step water rogulator is MSSIM, or be the convex of other or recessed water rogulator, described Convex or recessed water rogulator include that similarity measure and diversity are estimated, described similarity measure has relevant Coefficient (CC), normalized crosscorrelation (NCC), structural similarity (SSIM), mutual information (MI) and Normalized mutual information (NMI), described diversity estimates distance measure SSD, SAD, the RC having gray scale, The Euclidean distance that iterative closest point algorithm (iterative closest point, ICP) is inner, chamfering distance (chamfer Distance), Hausdorff distance and theory of information in Hamming distance (Hamming distance).

3. image registration automatic optimization method based on match curve feature evaluation as claimed in claim 1, it is special Levy and be: the degree of freedom of registration transformation includes the translation of x direction, the translation of y direction, rotates around z-axis and one Cause scaling, and be generalized to any multivariant affine transformation, projective transformation and nonlinear transformation.

4. image registration automatic optimization method based on match curve feature evaluation as claimed in claim 1, it is special Levy and be: wherein the 4th step registration parameter optimizes and revises other algorithm that method uses, and searches including golden section Rope algorithm, experiment proves close with the effect of optimization of formula (6), but required time increased.