CN103646399A - Registering method for optical and radar images - Google Patents

Abstract

Disclosed is a registering method for optical and radar images: (1) a radar image is used as a reference image and an optical image is used as a to-be-registered image and downsampling is performed on the reference image and the to-be-registered image respectively so that images of no less than three layers and different resolutions are generated; (2) starting from the image of the first layer and low resolution, image transformation is performed on the image of each layer through use of mutual negative information values; (3) characteristic point sets of coordinates at which gradient magnitudes in the optical and the radar images exceed the gradient magnitude of a preset threshold value, are extracted respectively; (4) translational transformation parameters used in transformation of the image of the last layer in step (2) are used to transfer optical-image coordinate characteristic point sets extracted in the step (3) to radar-image coordinate characteristic point sets; (5) within the range of the transferred point sets, a target function is optimized and a translational transformation parameter corresponding to the maximum of the target function is selected as a fine registration parameter; (6) the fine registration parameter is used to perform transformation and resampling on the to-be-registered image so that a registered image is obtained.

Description

A kind of optics and radar image method for registering

Technical field

The present invention relates to the disposal route of a kind of optics and radar image registration.

Background technology

Remote sensing technology turns to practical application in society and scientific and technological aspect at present, and these application comprise the processing, climate change assessment, natural resource management, environmental protection of disaster etc., and all these relates to long term monitoring earth surface.What in recent years, image registration became in remote sensing application is extremely important.Image registration is the basic task during image is processed, and refers to coupling two width or several same objects from different time, different remote sensor, different visual angles or the image of scene.Image registration is for two width or several digital pictures are accurately aimed to analyze and relatively, relate to a plurality of domain knowledges such as physiology, computer vision, pattern-recognition, image understanding for Digital Image Processing.Accurate registration Algorithm for supporting to inlay remote sensing satellite image, follow the trail of earth surface environmental change, basic scientific research is very important.

Image registration, aims at two width images by calculating one group of transformation parameter, and it is simple and clear that this problem seems definition, seems to have clearly, general method, and is in fact far from so.Because the application corresponding to various different pieces of informations is numerous, image registration developed into one complicated, there is challenging, the task of comprising many methods and strategies very by force.Along with the continuous enhancing that remote sensing, medical science and other field obtain image ability, cause in 20 years, image registration techniques having been carried out to large quantity research in the past.But up to the present, also do not have a kind of method for registering can solve all registration problems, can only and should be used for studying corresponding algorithm according to concrete data type.Image registration algorithm is often divided into based on region and two kinds of methods based on feature, but is generally only applicable to the image that gray scale difference is less, and the linear feature between reflection data, is not suitable for the registration between the image that gray scale difference is larger.

Summary of the invention

Technology of the present invention is dealt with problems and is: because different sensors image-forming principle is different, between the image obtaining, there is larger gray difference (as optical imagery and radar image), the feature that same scenery presents in dissimilar image is also different, common feature is difficult to extract, this just makes the various method for registering of and characteristics of image relevant based on gray scale no longer applicable, effect is bad, so this technology solves an above-mentioned difficult problem and is applied to optics and the registration of radar image proposing new method.

Technical solution of the present invention is: a kind of optics and radar image method for registering, and step is as follows:

(1) take radar image as reference picture, optical imagery, as image subject to registration, carries out to reference picture and image subject to registration the image that down-sampled generation is no less than the different resolution of 3 layers respectively;

(2) from ground floor low-resolution image layer, every one deck is all handled as follows:

(2.1) utilize marginal probability distribution and the joint probability distribution of cuclear density function computing reference image and image subject to registration, and the negative mutual information value between computing reference image and image subject to registration; Take and bear mutual information and carry out iterative optimization as objective function, the translation transformation parameter when asking for minimum similarity measure value or reaching the iterative times of regulation;

(2.2) utilize translation transformation parameter to convert the image subject to registration of lower one deck, to image subject to registration and the reference picture repeating step (2.1) after conversion, until be that resolution maximum layer image conversion subject to registration completes to last one deck;

(3) extract respectively the gradient magnitude place coordinate feature point set that gradient magnitude in optical imagery, radar image surpasses predetermined threshold value;

(4) utilize the translation transformation parameter that in step (2), last tomographic image conversion is used, the optical imagery coordinate feature point set extracting in step (3) is moved on radar image coordinate feature point set, obtain point set S ₁(P);

(5) within the scope of the point set after step (4) migration, objective function is optimized, translation transformation parameter corresponding while choosing objective function maximal value is as smart registration parameter;

(6) utilizing above-mentioned smart registration parameter to treat registering images converts and resamples and obtain the image after registration.

Objective function in described step (5) is:

$F\left(S_1\left(p\right)\right)=\underset{(x_g,y_g)\∈S_1\left(p\right)}{\mathrm{\Σ}}{|\▿U_1(x_g,y_g)|}^2$

Wherein, $\begin{array}{c}|\▿U_1(x_g,y_g)|=\sqrt{{U_x}^2(x_g,y_g)+{U_y}^2(x_g,y_g)}\\ U_x(x_g,y_g)=0.5(U(x_g,y_g+1)-U(x_g,y_g-1))\\ U_y(x_g,y_g)=0.5(U(x_g+1,y_g)-U(x_g-1,y_g))\end{array}$

U(x _g, y _g+ 1) representative is at (x _g, y _g+ 1) gray-scale value of locating.

The present invention compared with prior art beneficial effect is:

(1) the present invention is directed to the feature of RS data, from the importance similarity measure of registration, break through, adopt new similarity measure criterion, the i.e. method based on image statistics distributed intelligence and the method based on image inherent structure feature, effectively solved similar, the linear limitation of traditional images method for registering similarity measure requirement gradation of image, the pre-service such as the method does not need image to do and cuts apart, feature extraction, applicable surface is wider, has higher precision and good robustness.

(2) the present invention adopts the probability density of cuclear density Function Estimation variable. the probability density of utilizing the approximate variable of histogram the earliest, there is larger evaluated error, and store the required space of histogram along with the characteristic variable number exponentially growth of sample simultaneously.

(3) the present invention adopts Multi-Resolution Registration strategy, by by slightly solving registration problems to smart mode, can avoid mutual information local extremum to improve registration accuracy, has improved speed and the robustness of registration Algorithm simultaneously.

(4) due to the intrinsic similarity structure similarity measure criterion of utilizing in same scene image, having focused on objective function optimization of whole algorithm, characteristic extraction procedure is simplified greatly, only need in optical imagery, extract high gradient amplitude point set relatively preferably at picture structure, effectively avoid all will accurately extracting and mate this difficult point of feature of the same name in two width images.

(5) noiseproof feature is strong, and algorithm robustness is good.For have certain local feature in a width figure, and in another width figure, do not have character pair situation, this method only can affect criterion function optimal value size, but can not have influence on+advantage position, therefore still can obtain registration solution.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is multi-resolution pyramid layering schematic diagram.

Embodiment

Below in conjunction with accompanying drawing, the present invention is elaborated, as shown in Figure 1, a kind of optics and radar image method for registering, step is as follows:

(1) take radar image as reference picture, optical imagery, as image subject to registration, carries out to reference picture and image subject to registration the image that down-sampled generation is no less than the different resolution of 3 layers respectively; This example take 3 layers describe as example.

The above-mentioned employing of the layering to image is set up Gaussian image pyramid mode and is processed as shown in Figure 2, and gaussian pyramid computing formula is as follows:

$g_L(i,j)=\underset{m=-2}{\overset{2}{\mathrm{\&Sigma;}}}\underset{n=-2}{\overset{2}{\mathrm{\&Sigma;}}}w(m,n)g_{L-1}(2i+m,2j+n),0<L\&le;N,0\&le;i<C_L,0\&le;j<R_L$

Gaussian pyramid is an image sequence, and each tomographic image in sequence is all duplicating images of last tomographic image low-pass filtering, the g in above formula _l(i, j) represents the image of L layer, C _lrepresent the columns of L tomographic image, R _lrepresent the line number of L tomographic image, N represents total number of plies, and w (m, n) is window function, often gets 5*5 Gauss template.

(2) at the 1st layer of low-resolution image layer, utilize marginal probability distribution and the joint probability distribution of cuclear density function computing reference image and image subject to registration, and the negative mutual information value between computing reference image and image subject to registration; Take and bear mutual information and carry out iterative optimization as objective function, the translation transformation parameter when asking for minimum similarity measure value or reaching the iterative times of regulation; At middle layer image in different resolution layer, utilize the translation transformation parameter that last layer obtains to convert the image subject to registration in this layer, and then utilize marginal probability distribution and the joint probability distribution of cuclear density function computing reference image and image subject to registration the negative mutual information value between computing reference image and image subject to registration; Take and bear mutual information and carry out iterative optimization as objective function, the translation transformation parameter when asking for minimum similarity measure value or reaching the iterative times of regulation, this translation transformation parameter is designated as thick registration parameter.At high-definition picture layer, utilize the translation transformation parameter that last layer obtains to convert the image subject to registration in this layer.

The image X and the Y that for two width, need mutual registration, select X as with reference to image, and Y is as image subject to registration, and ideal situation is to utilize mutual information to mate its corresponding pixel position in image X each pixel in image Y.

Image X, the mutual information of Y is defined as:

I(X;Y)=H(X)+H(Y)-H(X,Y)

Wherein: H (X), H (Y) is image X, the edge entropy of Y, H (X, Y) is X, the combination entropy of Y.

$H\left(X\right)=\underset{x}{\mathrm{\&Sigma;}}-P_x\left(x\right)\log P_x\left(x\right)$

$H\left(Y\right)=\underset{y}{\mathrm{\&Sigma;}}-P_y\left(y\right)\log P_y\left(y\right)$

$H(X,Y)=\underset{x,y}{\mathrm{\&Sigma;}}-P_{x,y}(x,y)\log P_{x,y}(x,y)$

P _x(x), P _y(y) be respectively the marginal probability distribution density of image X and Y, P _x,y(x, y) is image X, the joint probability distribution density of Y:

For estimated probability density function P (x), the sample close with x, role seemingly should be than the sample away from X more greatly.Cuclear density method is exactly from measuring a kind of accurate Nonparametric Estimation of sample X (number of samples n) direct estimation stochastic variable probability density.N sample, the probability density that the x estimating from measurement sample X is ordered is defined as:

$P\left(x\right)=\frac{1}{n}\underset{x_l\&Element;X}{\mathrm{\&Sigma;}}W\left(\frac{x-x_l}{h}\right)$

Wherein W (x) is window function, x _lrepresent that choose at random and some x annex; H is window width parameter.Window function must meet two conditions below:

W(x)>=0;

∫W(x)dx=1

Image registration is exactly the problem of a function optimization, obtains and makes one group of transformation parameter μ of similarity measure value S maximum be designated as μ _opt.

μ _opt=arg _μmaxS(μ)

The present invention is usingd the negative value of Minimum mutual information as similarity measure function S, with X and Y, represents reference picture and image subject to registration, and the function that the negative mutual information value representation between reference picture and image subject to registration is transformation parameter μ is:

$S\left(\mathrm{\&mu;}\right)=-\underset{y\&Element;Y}{\mathrm{\&Sigma;}}\underset{x\&Element;X}{\mathrm{\&Sigma;}}{P}_{x,y}(x,y;\mathrm{\&mu;}){\log }_2\frac{P_{x,y}(x,y;\mathrm{\&mu;})}{P_x(x;\mathrm{\&mu;})P_y(y;\mathrm{\&mu;})}$

P _x,y(x, y; μ) the joint probability distribution density of transformation parameter μ is considered in representative; P _y(y; μ) the marginal probability distribution density of the image subject to registration of transformation parameter μ is considered in representative; P _x(x; μ) the marginal probability distribution density of the reference picture of transformation parameter μ is considered in representative.

Mutual information gradient

$\&dtri;S={\left[\begin{array}{cccc}\frac{\&PartialD;S}{\&PartialD;{\mathrm{\&mu;}}_1}& \frac{\&PartialD;S}{\&PartialD;{\mathrm{\&mu;}}_2}& ...& \frac{\&PartialD;S}{\&PartialD;{\mathrm{\&mu;}}_k}\end{array}\right]}^T$

$\frac{\&PartialD;S}{\&PartialD;{\mathrm{\&mu;}}_k}=-\underset{y\&Element;Y}{\mathrm{\&Sigma;}}\underset{x\&Element;X}{\mathrm{\&Sigma;}}\frac{\&PartialD;p(x,y;\mathrm{\&mu;})}{\&PartialD;{\mathrm{\&mu;}}_k}\log \frac{p(x,y;\mathrm{\&mu;})}{p_y(y;\mathrm{\&mu;})}$

be k partial derivative of joint probability distribution, k ∈ Z is that k is integer.

(3) extract respectively the gradient magnitude place coordinate feature point set that gradient magnitude in optical imagery, radar image surpasses predetermined threshold value;

The extracting method of two class images is identical, take optical imagery as example: the gradient magnitude of computed image first:

$|\&dtri;U_1(i,j)|=\sqrt{{U_x}^2(i,j)+{U_y}^2(i,j)}$

U _x(i,j)=0.5(U(i,j+1)-U(i,j-1))

U _y(i,j)=0.5(U(i+1,j)-U(i-1,j))

Get the point of front 25% gradient magnitude as feature point set S ₁(p), put concentrated point and be designated as (x _g, y _g).

(4) utilize the translation transformation parameter that in step (2), last tomographic image conversion is used, the optical imagery coordinate feature point set extracting in step (3) is moved on radar image coordinate feature point set, obtain point set S ₁(P);

The thick registration translation transformation parameter P that step (2) is tried to achieve ₀=(μ 1, and μ 2), by point set S ₁in (x, y), the coordinate of each point adds translation transformation parameter P ₀, complete migration.

For (X, Y) ∈ S ₁(P) have:

X=x+μ ₁

Y=y+μ ₂

(5) within the scope of the point set after step (4) migration, objective function is optimized, translation transformation parameter corresponding while choosing objective function maximal value is as smart registration parameter;

Optimization aim function F (S ₁(p)), be abbreviated as F:

$F\left(S_1\left(p\right)\right)\stackrel{\mathrm{\&Delta;}}{=}\underset{(x_g,y_g)\&Element;S_1\left(p\right)}{\mathrm{\&Sigma;}}{|\&dtri;U_1(x_g,y_g)|}^2$

for point (x _i, y _i) Grad,

Parameter iteration: $p_{n+1}=p_n+{\left({H_p}^n\right)}^{-1}{\&dtri;}_p{F}_n$

to be n at iterations, F gradient when parameter is p, H _p ⁿfor the Hessian matrix of F, ask method as follows:

${\&dtri;}_p{F}_n={\left[\begin{array}{cccc}\frac{\&PartialD;F}{\&PartialD;P_1}& \frac{\&PartialD;F}{\&PartialD;P_2}& ...& \frac{\&PartialD;F}{\&PartialD;P_m}\end{array}\right]}^T$

(6) utilizing above-mentioned smart registration parameter to treat registering images converts and resamples and obtain the image after registration.

The unspecified part of the present invention belongs to general knowledge as well known to those skilled in the art.

Claims (2)

1. optics and a radar image method for registering, is characterized in that step is as follows:

(1) take radar image as reference picture, optical imagery, as image subject to registration, carries out to reference picture and image subject to registration the image that down-sampled generation is no less than the different resolution of 3 layers respectively;

(2) from ground floor low-resolution image layer, every one deck is all handled as follows:

(2.1) utilize marginal probability distribution and the joint probability distribution of cuclear density function computing reference image and image subject to registration, and the negative mutual information value between computing reference image and image subject to registration; Take and bear mutual information and carry out iterative optimization as objective function, the translation transformation parameter when asking for minimum similarity measure value or reaching the iterative times of regulation;

(2.2) utilize translation transformation parameter to convert the image subject to registration of lower one deck, to image subject to registration and the reference picture repeating step (2.1) after conversion, until be that resolution maximum layer image conversion subject to registration completes to last one deck;

(3) extract respectively the gradient magnitude place coordinate feature point set that gradient magnitude in optical imagery, radar image surpasses predetermined threshold value;

(4) utilize the translation transformation parameter that in step (2), last tomographic image conversion is used, the optical imagery coordinate feature point set extracting in step (3) is moved on radar image coordinate feature point set, obtain point set S ₁(P);

(5) within the scope of the point set after step (4) migration, objective function is optimized, translation transformation parameter corresponding while choosing objective function maximal value is as smart registration parameter;

(6) utilizing above-mentioned smart registration parameter to treat registering images converts and resamples and obtain the image after registration.

2. a kind of optics according to claim 1 and radar image method for registering, is characterized in that: the objective function in described step (5) is:

$F\left(S_1\left(p\right)\right)=\underset{(x_g,y_g)\&Element;S_1\left(p\right)}{\mathrm{\&Sigma;}}{|\&dtri;U_1(x_g,y_g)|}^2$

Wherein, $\begin{array}{c}|\&dtri;U_1(x_g,y_g)|=\sqrt{{U_x}^2(x_g,y_g)+{U_y}^2(x_g,y_g)}\\ U_x(x_g,y_g)=0.5(U(x_g,y_g+1)-U(x_g,y_g-1))\\ U_y(x_g,y_g)=0.5(U(x_g+1,y_g)-U(x_g-1,y_g))\end{array}$

U(x _g, y _g+ 1) representative is at (x _g, y _g+ 1) gray-scale value of locating.

Images