CN104599233A - Method and device for quickly registering SAR and full-color image - Google Patents

Abstract

The invention belongs to the field of image processing, and particularly relates to a method and a device for quickly registering an SAR and a panchromatic image, wherein the method comprises the following steps of 1, carrying out data blocking on a reference image and a transformed floating image to obtain a plurality of data sub-blocks; step 2, calculating a joint probability distribution matrix and an accumulated function value of each data subblock; and 3, combining the joint probability distribution matrix and the accumulated function value of each data subblock, and iteratively calculating the CCRE of the reference image and the floating image. By the method, the registration speed is obviously improved on the premise of ensuring the registration accuracy.

Description

A kind of method of SAR and full-colour image rapid registering and device

Technical field

The invention belongs to image processing field, in particular to method and the device of a kind of SAR and full-colour image rapid registering.

Background technology

Multi-source image registration refers to the conversion parameter determining between image according to some similarity measurements, make two width or the multiple image of the Same Scene obtained from different sensors, different visual angles, different time, under transforming to the same coordinate system, pixel layer obtains the process of optimum matching.

The method for registering images that remote sensing image registration method is generally divided into feature based to extract and the method for registering images based on half-tone information.In the remote sensing image registration method that feature based extracts, the feature the most often used is Harris angle point, Moravec angle point, SIFT feature point etc.Said method has the following disadvantages: different scenes needs to select different features, does not have general model to follow, has very strong application limitation; Easily be subject to the impact of noise.Due to the imaging mechanism of SAR itself, accurate extraction feature is very difficult, also be difficult to design a kind of energy simultaneously for the feature extraction algorithm that different noise intensity, different resolution and the SAR image containing different scene structure are all suitable for, by causing, this occurs that a kind of registration Algorithm is only applicable to the situation of a class SAR image.Therefore, the method for conventional feature based has significant limitation in SAR image registration application.

Method for registering images based on half-tone information directly utilizes the half-tone information of image to carry out registration, by realizing registration to the global optimum of certain similarity measurement between its pixel, and main its method for measuring similarity of consideration in Gray-scale Matching.The method does not need to carry out special pretreatment work to image, not easily affected by noise, simple, stronger compared with its versatility of method of feature extraction.

Method for registering based on gray-scale statistical characteristic does not need to carry out Iamge Segmentation and image characteristics extraction, thus can avoid the loss of significance caused by these pre-service.But this method generally needs larger calculated amount, the speed of registration is slow, very responsive to half-tone information change.Often calculate a similarity measure, need all pixels in traversing graph picture, this obviously makes calculated amount increase.In the process of Optimizing Search, current mutual information function local extremum problem, mainly through introducing more complicated interpolating function or in conjunction with prior imformation, obtains more smooth function curve, and whole computation process can be made to have increase in various degree.Mutual information (Mutual information, and normalized mutual information (Normalized mutualinformation MI), NMI) be the conventional function being widely used in registration field, for image registration, adopt mutual information and normalized mutual information to be often easy to be absorbed in local extremum, have a strong impact on the precision of registration.

Summary of the invention

The object of the present invention is to provide method and the device of a kind of SAR and full-colour image rapid registering, to solve the above problems.

Provide a kind of method of SAR and full-colour image rapid registering in an embodiment of the present invention, comprising:

Step 1, carries out deblocking to the floating image after reference picture and conversion, obtains multiple data sub-block;

Step 2, calculates joint probability distribution matrix and the cumulative function value of each data sub-block;

Step 3, merges joint probability distribution matrix and the cumulative function value of each data sub-block, the CCRE of this reference picture of iterative computation and floating image.

Further, also comprise before step 1:

Image block is carried out to this floating image, and parallel difference is carried out to block image, obtain the floating image after above-mentioned conversion.

Further, also comprise after step 3:

Calculate the difference of the CCRE that this CCRE and last iteration obtain, if this difference is less than or equal to preset error value, then finishing iteration calculates; If this difference is greater than preset error value, then repeat step 1 to step 3 after utilizing Powell algorithm to adjust conversion parameter.

Embodiments of the invention additionally provide the device of a kind of SAR and full-colour image rapid registering, comprising:

Deblocking module, for carrying out deblocking to the floating image after reference picture and conversion, obtains multiple data sub-block;

Computing module, for calculating joint probability distribution matrix and the cumulative function value of each data sub-block;

CCRE asks for module, for merging joint probability distribution matrix and the cumulative function value of each data sub-block, the CCRE of this reference picture of iterative computation and floating image.

Further, this device also comprises:

Floating image conversion module, for carrying out image block to this floating image, and carrying out parallel difference to block image, obtaining the floating image after above-mentioned conversion.

Further, this device also comprises:

Difference calculating module, for calculating the difference of the CCRE that this CCRE and last iteration obtain, if this difference is less than or equal to preset error value, then finishing iteration calculates; If this difference is greater than preset error value, then repeat the action step that this deblocking module, computing module and CCRE ask for module after utilizing Powell algorithm to adjust conversion parameter.

The method of a kind of SAR that the embodiment of the present invention provides and full-colour image rapid registering and device be compared with prior art:

(1) problem of local extremum is easily produced for traditional mutual information image registration, in SAR and full-colour image registration Algorithm, the present invention adopts and intersects cumulative remainder entropy as similarity measure, the cumulative remainder entropy that intersects has more generality than Shannon entropy, compared with the method for registering of traditional employing mutual information, the cumulative remainder entropy that intersects can avoid local extremum effectively, removes noise, and accelerates registration speed to a certain extent;

(2) for the slow-footed problem of registration that SAR and the full-colour image registration Algorithm based on intersection cumulative remainder entropy exists, the present invention proposes a kind of parallel optimization strategy, the method of parallel multithread is used to improve and optimizate the serial image registration algorithm based on intersection cumulative remainder entropy, under the prerequisite ensureing registration accuracy, significantly improve registration speed.

Accompanying drawing explanation

Fig. 1 shows the process flow diagram of a kind of embodiment of method of a kind of SAR of the present invention and full-colour image rapid registering;

Fig. 2 shows the process flow diagram of the another kind of embodiment of method of a kind of SAR of the present invention and full-colour image rapid registering;

Fig. 3 shows the structured flowchart of a kind of embodiment of device of a kind of SAR of the present invention and full-colour image rapid registering;

Fig. 4 shows the structured flowchart of the another kind of embodiment of device of a kind of SAR of the present invention and full-colour image rapid registering;

The joint probability distribution that Fig. 5 shows the method for a kind of SAR of the present invention and full-colour image rapid registering calculates schematic diagram;

Fig. 6 shows the deblocking parallel computation schematic diagram of the method for a kind of SAR of the present invention and full-colour image rapid registering;

Fig. 7 shows the interpolation test result analysis figure of the method for a kind of SAR of the present invention and full-colour image rapid registering;

Fig. 8 shows the CCRE test result analysis figure of the method for a kind of SAR of the present invention and full-colour image rapid registering;

Fig. 9 shows the registration result analysis chart of the method for a kind of SAR of the present invention and full-colour image rapid registering.

Embodiment

Also by reference to the accompanying drawings the present invention is described in further detail below by specific embodiment.

Shown in ginseng Fig. 1, Fig. 1 shows the process flow diagram of a kind of embodiment of method of a kind of SAR of the present invention and full-colour image rapid registering.

Present embodiments provide a kind of method of SAR and full-colour image rapid registering, comprising:

Step S102, carries out deblocking to the floating image after reference picture and conversion, obtains multiple data sub-block;

Step S103, calculates joint probability distribution matrix and the cumulative function value of each data sub-block;

Step S104, merges joint probability distribution matrix and the cumulative function value of each data sub-block, the CCRE of iterative computation reference picture and floating image.

The present embodiment adopts and intersects cumulative remainder entropy as similarity measure, the cumulative remainder entropy that intersects has more generality than Shannon entropy, compared with the method for registering of traditional employing mutual information, the cumulative remainder entropy that intersects can avoid local extremum effectively, remove noise, and accelerate registration speed to a certain extent; The present embodiment adopts the method for parallel multithread to improve and optimizate the serial image registration algorithm based on intersection cumulative remainder entropy, under the prerequisite ensureing registration accuracy, significantly improves registration speed.

Shown in ginseng Fig. 2, Fig. 2 shows the process flow diagram of the another kind of embodiment of method of a kind of SAR of the present invention and full-colour image rapid registering.

In the present embodiment, also comprise before step S102:

Step S101, carries out image block to described floating image, and carries out parallel difference to block image, obtains the floating image after above-mentioned conversion.

In the present embodiment, also comprise after step S104:

The difference of step S105, the CCRE that calculating CCRE and last iteration obtain, if this difference is less than or equal to preset error value, then finishing iteration calculates; If this difference is greater than preset error value, then repeat above-mentioned steps S102 to step S104 after utilizing Powell algorithm to adjust conversion parameter.

Shown in ginseng Fig. 3, Fig. 3 shows the structured flowchart of a kind of embodiment of device of a kind of SAR of the present invention and full-colour image rapid registering.

The present embodiment additionally provides the device of a kind of SAR and full-colour image rapid registering, comprising:

Deblocking module 22, for carrying out deblocking to the floating image after reference picture and conversion, obtains multiple data sub-block;

Computing module 23, for calculating joint probability distribution matrix and the cumulative function value of each data sub-block;

CCRE asks for module 24, for merging joint probability distribution matrix and the cumulative function value of each data sub-block, the CCRE of this reference picture of iterative computation and floating image.

The present embodiment asks for module 24 by setting data piecemeal module 22, computing module 23 and CCRE, adopt and intersect cumulative remainder entropy as similarity measure, the cumulative remainder entropy that intersects has more generality than Shannon entropy, compared with the method for registering of traditional employing mutual information, the cumulative remainder entropy that intersects can avoid local extremum effectively, remove noise, and accelerate registration speed to a certain extent; The present embodiment adopts the method for parallel multithread to improve and optimizate the serial image registration algorithm based on intersection cumulative remainder entropy, under the prerequisite ensureing registration accuracy, significantly improves registration speed.

Shown in ginseng Fig. 4, Fig. 4 shows the structured flowchart of the another kind of embodiment of device of a kind of SAR of the present invention and full-colour image rapid registering.

In the present embodiment, this device also comprises:

Floating image conversion module 21, for carrying out image block to this floating image, and carrying out parallel difference to block image, obtaining the floating image after above-mentioned conversion.

In the present embodiment, this device also comprises:

Difference calculating module 25, for calculating the difference of the CCRE that this CCRE and last iteration obtain, if this difference is less than or equal to preset error value, then finishing iteration calculates; If this difference is greater than preset error value, then after utilizing Powell algorithm to adjust conversion parameter, repeating data piecemeal module 22, computing module 23 and CCRE ask for the action step of module 24.

Below the method for registering that the present embodiment provides is explained in further detail.

The present embodiment adopts the method for registering images based on half-tone information, utilizes CCRE (Cross cumulative residual entropy, intersection cumulative remainder entropy) as similarity measure.CCRE, compared with traditional mutual information, has many excellent characteristics, makes registration can obtain more high precision.But be that the time complexity of algorithm is high based on the obvious defect of half-tone information method.For the slow-footed problem of registration that SAR and the full-colour image registration Algorithm based on CCRE exists, the present embodiment adopts parallel optimization strategy, utilize parallel algorithm, first data are divided, then these data blocks are given different threads, data sub-block is calculated simultaneously, to reach the object reducing working time.

SAR and full-colour image registration are under same scene, use different imaging devices, the differential image that there is displacement, rotation etc. obtained under different visual angles.Using piece image as reference image in the process of registration, a width is as floating image, and adjustment floating image, makes two width images mate in position.Image registration has mathematical definition, and realizing a registration Algorithm needs to consider the factors such as feature space, search volume, similarity measure, search strategy.

Image registration model:

If I _f(x, y), I _r(x, y) represents floating image I respectively _fwith reference picture I _rat the gray-scale value at point (x, y) place, image I _r, I _fthe available formula of registration relation (1) represent.

I _R(x,y)＝g(I _F(f(x,y))) (1)

Wherein f represents that two-dimensional space coordinate geometry converts, and comprises translation transformation and rotational transform.G represents greyscale transformation function.The implication of formula (1) can be understood as, to image I _fconvert, make image I _f, I _rin the locus of correspondence, there is identical gray-scale value.If T () represents the conversion to image, then image registration can use formula (2) to describe.

Find T ₀∈Ω

Max S(T(I _F),I _R) (2)

Wherein Ω is the feasible zone of conversion T (), and S is similarity measure, Find ... Max represents the maximal value finding out S in the Ω of territory.Formula (2) shows that image registration is in the feas ible space of conversion T (), finds optimum conversion T ₀, make the floating image after conversion and reference picture have maximum similarity.The present embodiment uses and intersects cumulative remainder entropy as similarity measure.

Use the match measure of intersection cumulative remainder entropy:

In multi-source Remote Sensing Images, SAR is different from optical imagery image-forming principle, and their feature forms complementation to a great extent.Optical sensor can only gather image when there being illumination, the characteristic of reflection visible light wave range; SAR system 24 continuous throughout the twenty-four hour24, reflects again the characteristic of electromagnetic wave spectrum wave band simultaneously, just in time makes up the deficiency of optical system.But due to two kinds of distinct imaging mechanisms of image, make again that two kinds of image informations are integrated faces very large challenge.

In SAR image, coherent noise has a strong impact on the extraction of feature, causes feature extraction success ratio to reduce.And the feature that in optical imagery, how identical with SAR image feature is is corresponding, directly affect registration success ratio.

The present embodiment is by utilizing CCRE (Cross cumulative residual entropy, intersect cumulative remainder entropy) similarity between two stack features is described, thus solve in SAR image and optical imagery and extract comparatively accurately and the difficulty of the feature of accurate match.

The principal feature of intersection cumulative remainder entropy is: be applicable to discrete and continuous two kinds of environment; Distribution number is replaced by distribution range; Calculated amount is little, and the result calculated sampled data is closer to actual result.Be applied to image registration aspect, compare mutual information method and carry out registration, there are two obvious advantages: be more immune to noise, have scope widely in parameter transformation territory.

Traditional mutual information method statistic be the histogram of each gray-scale value, reflection be a kind of discrete features, self-existent noise can produce very important impact to matching result.Existing discrete Shannon entropy cannot meet the requirement overcoming much noise in SAR image, needs a kind ofly can reflect that the entropy of the accumulated change in certain interval is to weaken the impact of noise on entirety.The present embodiment utilizes CCRE to reflect the accumulated change in certain interval, the object of noise to entire effect is weakened to reach, CCRE described in the present embodiment is defined on the basis of cumulative remainder entropy, and the definition of cumulative remainder entropy is as shown in formula (3).

$\mathrm{\ϵ}\left(x\right)=-{\∫}_{R_{+}}F\left(\mathrm{\λ}\right)\log F\left(\mathrm{\λ}\right)\mathrm{d\λ}---\left(3\right)$

$h\left(x\right)={\∫}_{-\∞}^{+\∞}p\left(x\right)\log p\left(x\right)\mathrm{dx}---\left(4\right)$

Wherein x is the variable belonging to R, F (λ)=P (| x|> λ), R+=(x ∈ R; X >=0).Compared with Shannon entropy, in formula (3), substituted for probability density function p (x) by cumulative remainder function F (λ).Cumulative remainder function ratio probability density function has more generality.Thus the CCRE obtained by cumulative function has more general applicability than the mutual information obtained with probability density function.Represent Shannon entropy with h (x), its definition is as shown in formula (4).For variable X, Y, their mutual information MI is defined as formula (5).

MI(X,Y)＝h(X)+h(Y)-h(X,Y) (5)

Wherein h (X), h (Y) are respectively the entropy of X, Y, the combination entropy that h (X, Y) is X, Y.By formula (3), definition CCRE is formula (6).

c(X,Y)＝ε(X)-E[ε(X/Y)] (6)

For discrete reference picture I _rwith the floating image I after conversion _t=T (I _f), calculating CCRE needs to obtain joint histogram, then edge calculation probability density and joint probability density.Discrete reference picture I _rwith the floating image I after conversion _tcCRE can represent with formula (7).

$\begin{array}{c}c(I_T,I_R)\\ =\mathrm{\ϵ}\left(I_T\right)-E\left[\mathrm{\ϵ}(I_T/I_R)\right]\\ =-\underset{\mathrm{\λ}\∈L_T}{\mathrm{\Σ}}{\∫}_{\mathrm{\λ}}^{\∞}{p}_T\mathrm{dl}\log \left[{\∫}_{\mathrm{\λ}}^{\∞}{p}_T\mathrm{dl}\right]\\ +\underset{k\∈L_T}{\mathrm{\Σ}}{p}_R\left(k\right)\underset{\mathrm{\λ}\∈L_T}{\mathrm{\Σ}}{\∫}_{\mathrm{\λ}}^{\∞}\frac{p(l,k)}{p_R\left(k\right)}\×\log \left[{\∫}_{\mathrm{\λ}}^{\∞}\frac{p(l,k)}{p_R\left(k\right)}\mathrm{dl}\right]\end{array}---\left(7\right)$

Wherein p (l, k) representative image I _rand I _tjoint probability density, p _t(l) and p _rk () represents the marginal probability density of the floating image after conversion and reference picture respectively.L _tand L _rrepresent the set of 2 width image discrete pixels points respectively.

CCRE replaces the probability distribution function in mutual information with cumulative distribution function, makes CCRE more mild to conversion Parameters variation curve, improves in mutual information curve the local extremum problem brought that fluctuates.

The present embodiment adopts CCRE as similarity measure, can improve registration accuracy, and not by the impact of SAR image noise, have speed of convergence faster than other method for registering, calculated amount is lower.

Powell searching algorithm:

Geometric transformation in image registration has various ways, as translation, rotation, convergent-divergent and distortion etc.Choose different variations and can obtain different transformation spaces.Consider by translation and rotate the search volume generated, it can be described by three parameters.Horizontal shift x, vertical displacement y and anglec of rotation θ, wherein x, y all belong to set of integers N ,-180 °≤θ <180 °.All combinations of x, y, θ constitute the search volume of registration.If for the combination (x, y, θ) of each conversion, all calculate the CCRE of two width images, this will be the time-consuming again process of a redundancy.Thus need to adopt Optimizing Search algorithm in registration process, reduce search volume.

CCRE not easily calculates conversion parameter (x, y, derivative θ), Powell algorithm relatively other optimized algorithm, as Newton method, gradient descent method, do not need to calculate derivative, the extreme-value problem of CCRE can be reduced to one dimension extreme-value problem again, fast convergence rate, has and significantly accelerates registration effect.These algorithm concrete steps are as follows:

1. given algorithm initial point x ⁰with the direction of n linear independence: d ^(1,1), d ^(1,2)..., d ^{(1, n)}, the error of permission is ε >0, makes k=1;

2. x is made ^{(k, 0)}=x ^k-1, from x ^{(k, 0)}set out, successively along direction d ^{(k, 1)}, d ^{(k, 2)}..., d ^{(k, n)}search for, even

$\left\{\begin{array}{c}{x}^{(x,j)}=x^{(k,j)}+{\mathrm{\&lambda;}}_j{d}^{(k,j)}\\ {\mathrm{\&lambda;}}_j:f(x^{(k,j-1)}+{\mathrm{\&lambda;}}_j{d}^{(k,j)})=\underset{\mathrm{\&lambda;}}{\min }f(x^{(k,j-1)}+\mathrm{\&lambda;}{d}^{(k,j)})\end{array}\right.$

Obtain a bit in each direction, x ^{(k, 1)}, x ^{(k, 2)}..., x ^{(k, n)}, make d ^{(k, n+1)}=x ^{(k, n)}– x ^{(k, 0)}, from x ^{(k, n)}set out along direction d ^{(k, n+1)}carry out linear search and obtain x ^k;

If 3. || x ^k-x ^k-1|| < ε, then stop search, some x ^kbe optimum solution, otherwise, order

D ^{(k+1, j)}=x ^{(k, j+1)}(j=1,2 ..., n), 2. k=k+1, return step.

Although Powell algorithm has speed of convergence faster, greatly reduce search volume, in actual registration operation, especially for the registration of significantly image, remain a job very consuming time.In order to shorten the registration time further, the present invention adopts paralleling tactic to accelerate registration Algorithm.

Paralleling tactic:

The realization that parallel processing calculates be unable to do without concurrent computer architecture.According to Flynn classification, parallel computer is divided into single instruction single data stream computing machine (SingleInstruction Single Data, SISD), single instruction stream multiple data stream computing machine (SingleInstruction Multiple Data, SIMD), Multiple Instruction stream Single Data stream computing machine (Multiple Instruction Single Data, and multiple-instruction multiple-data stream computer (Multiple Instruction Multiple Data, MIMD) MISD).Wherein MIMD system can reduce execution time of whole program in the multiple subtask of executed in parallel, is the parallel computer architecture of most widespread use.

For the design feature of MIMD system, during design parallel algorithm, a larger computational problem is decomposed into several relatively independent subproblems.Suppose that the number of the processor that can simultaneously use is for m, for the problem P be defined on feasible zone Ω, resolves into Ω by feasible zone ₁, Ω ₂..., Ω _m, corresponding subproblem is P ₁, P ₂..., P _m.For each multi-frame system subproblem, subproblem can be resolved simultaneously.After all subproblems solve, the solution merging subproblem obtains the solution of problem P.

In order to make subproblem solve simultaneously, do not have between subproblem very strong dependence.If subproblem P _srely on other certain subproblem P _tsolution, then P _sneed to wait for P _tcould start after having solved to solve.When a lot of subproblem meets above situation, the time overhead that such scheme only can bring Task-decomposing to produce, that can not reduce problem solves the time.

In the registration Algorithm of the present embodiment, need repeatedly to calculate CCRE.The calculating of CCRE relies on the new conversion parameter obtained after Powell algorithm often takes turns iteration, will not reach the effect optimizing speed-raising using calculating the subproblem of CCRE as parallel algorithm.The registration Algorithm of the present embodiment is in the process utilizing Powell searching optimum solution, when each takes turns iterative computation CCRE, all need joint probability distribution and the cumulative function of computing reference image and floating image, and then solve CCRE, thus calculating CCRE can be decomposed.

To calculate joint probability distribution, reference picture I can be established _rnumber of greyscale levels be M, through conversion after floating image I _tgray level be N, then their joint probability distribution is as shown in formula (8).

$p_{R,T}={\left[\begin{array}{cccc}p\left(\mathrm{0,0}\right)& p\left(\mathrm{0,1}\right)& ...& p(0,N-1)\\ p\left(\mathrm{1,0}\right)& p\left(\mathrm{1,1}\right)& ...& p(1,N-1)\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ p(M,0)& p(M,1)& ...& p(M,N-1)\end{array}\right]}_{M\&times;N}---\left(8\right)$

Joint probability distribution is that computation process as shown in Figure 5 by traveling through two width images, obtaining according to the gray-scale value of correspondence position pixel.In joint probability distribution, the calculating of each element is only by the image I shared _rand I _tobtain, do not rely on the calculating of other element, this is that the division of data is provided convenience.

Data can by row or divided by column, and final object makes the calculated amount of each data sub-block suitable, can make full use of the computing power of each thread like this, reach load balancing.Each unit of joint probability density have similar form of calculation, and the computing power of consumption is suitable, and thus two kinds of dividing mode all can realize the uniform distribution of calculation task.According to the available Thread Count of running environment, calculation task can be distributed to each thread fifty-fifty.Under Fig. 6 gives the running environment of p thread, the schematic diagram of piecemeal by row.

In the parallel running stage, after waiting for that each thread computes terminates, all thread computes results are gathered, obtains final result of calculation.In theory, the 1/p before computing time can reduce to by the parallel algorithm realized by this strategy.But the calculating data that each thread obtains there are differences, cause calculation task synchronously not complete, communicate between thread the time overhead existed in addition, often makes actual run time be greater than the theoretical used time.In the operation such as the image interpolation in registration Algorithm, copying image, data all have the feature of above-mentioned weak dependence, can implement this paralleling tactic.According to above-mentioned thinking, registration Algorithm is carried out parallel optimization, can rapid registering be realized.

Parallel registration Algorithm:

1. prepare storage space for changing image and carry out piecemeal, trying to achieve inverse-transform matrix according to conversion parameter, calculate the correspondence position of each pixel of changing image in floating SAR image, obtain changing image by floating image parallel interpolation;

2. paralleling tactic is used, first to associating probability distribution matrix piecemeal, by the floating image after reference picture and conversion, parallel computation obtains joint probability distribution matrix, and obtain cumulative function value, pooled data block, the CCRE of computing reference full-colour image and floating image;

If the difference of CCRE that 3. CCRE and last iteration obtain is greater than the permissible error of setting, use Powell algorithm adjustment running parameter, jump to step 1., otherwise show to meet registration requirement, obtain the floating image of registration, registration process terminates.

From above-mentioned 1. to 3., all use paralleling tactic to be optimized acceleration in the data manipulation that view data participates in, and the judgement of parameter and adjustment all do not adopt paralleling tactic.Consider that adjustment and the judgement of parameter have less operation expense, thus this registration Algorithm can realize good acceleration effect.

Parallel registration Algorithm testing and analyzing:

Hardware test environment is configured to: 16 pieces of Xeon E5649 CPU, dominant frequency 2.40GHz, total amount of memory 16G, disk total volume 3T.Software environment is: Microsoft VisualStudio 2005, OpenMP.Test uses SAR and the full-colour image of many group the same areas, and the present embodiment lists four groups of test data experimental results.

First with the SAR floating image of different size, serial and parallel interpolation algorithm are tested.Conversion parameter (x, y, θ) is got (128,64 ,-64 °).Table 1 gives the result of the interpolation test being of a size of 1024 × 906.

Table 1 interpolation test result

Process number	Serial	2	4	8	16
						Working time (ms)	797	453	282	210	142
Speed-up ratio	-	1.76	2.83	3.80	5.61
						Parallel efficiency	-	0.88	0.71	0.48	0.36

In table 1, speed-up ratio is defined as the ratio of serial working time and parallel runtime, illustrates parallel accelerate effect intuitively.Parallel efficiency is the ratio of speed-up ratio and Thread Count, represents each thread average service rate in the algorithm.Its numerical value is larger, represents in algorithm operational process to have longer parallel time.Other sized image test result as shown in Figure 7.

As shown in Figure 7, along with the increase of Thread Count, the acceleration effect of parallel algorithm constantly improves.The image measurement result less from size, when Thread Count reaches some, speed-up ratio can not continue to increase.And too much Thread Count can increase the burden of process communication, speed-up ratio can be caused on the contrary to decline.

Fig. 8 is the speed-up ratio test result calculating single CCRE.As shown in Figure 8, use paralleling tactic to calculate CCRE and also can improve computing velocity.

Finally to whole registration test of heuristics.The full-colour image of one width SAR image and same size is as one group of test data.Test respectively based on mutual information and CCRE serial registration Algorithm, and different threads number based on CCRE parallel algorithm.

Table 2 is the test result based on mutual information and CCRE.The maximum pixel number that in table, registration accuracy departs from reference picture by registration result image is weighed.

Table 2 is based on mutual information and CCRE serial registration result

As shown in Table 2, the image registration algorithm registration speed based on CCRE that the present embodiment realizes is faster, and the method for registering registration accuracy compared based on mutual information is higher.

The present embodiment uses multiple series of images to test parallel registration Algorithm.Be of a size of 1024 × 906 one group of image test result as shown in table 3.The accurate test result of other assembly as shown in Figure 9.

Table 3 walks abreast registration test result

Process number	Serial	2	4	8	16
						Working time (s)	388	245	194	174	150
Speed-up ratio	-	1.58	2.00	2.23	2.59
						Parallel efficiency	-	0.79	0.50	0.43	0.16
Precision (pixels)	2.68	2.68	2.68	2.68	2.68

Consolidated statement 3 and Fig. 9 test result can be seen, parallel speedup ratio becomes large with the increase of process number, and the value of all speed-up ratios is greater than 1.This shows that parallel algorithm serves the effect accelerating registration, and the increase of parallel line number of passes can make acceleration effect become more and more obvious.If registering images size is less, parallel thread number should be avoided too much.

The rate of rise of speed-up ratio increases along with number of threads and slows down, and this is owing to there is serial execution part in registration Algorithm.Thread increase the execution time that only can reduce parallel section, serial section accounts for the continuous rising of ratio regular meeting of total execution time, which results in slowing down of speed-up ratio rate of rise.

The foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, for a person skilled in the art, the present invention can have various modifications and variations.Within the spirit and principles in the present invention all, any amendment done, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (6)

1. a method for SAR and full-colour image rapid registering, is characterized in that, comprising:

Step 1, carries out deblocking to the floating image after reference picture and conversion, obtains multiple data sub-block;

Step 2, calculates joint probability distribution matrix and the cumulative function value of each data sub-block;

Step 3, merges joint probability distribution matrix and the cumulative function value of each data sub-block, the CCRE of reference picture described in iterative computation and floating image.

2. the method for a kind of SAR according to claim 1 and full-colour image rapid registering, is characterized in that, also comprises before described step 1:

Image block is carried out to described floating image, and parallel difference is carried out to block image, obtain the floating image after described conversion.

3. the method for a kind of SAR according to claim 2 and full-colour image rapid registering, is characterized in that, also comprises after described step 3:

Calculate the difference of the CCRE that described CCRE and last iteration obtain, if described difference is less than or equal to preset error value, then terminate described iterative computation; If described difference is greater than preset error value, then utilize Powell algorithm adjust conversion parameter after repeating said steps 1 to step 3.

4. a device for SAR and full-colour image rapid registering, is characterized in that, comprising:

Deblocking module, for carrying out deblocking to the floating image after reference picture and conversion, obtains multiple data sub-block;

Computing module, for calculating joint probability distribution matrix and the cumulative function value of each data sub-block;

CCRE asks for module, for merging joint probability distribution matrix and the cumulative function value of each data sub-block, the CCRE of reference picture described in iterative computation and floating image.

5. the device of a kind of SAR according to claim 4 and full-colour image rapid registering, is characterized in that, also comprise:

Floating image conversion module, for carrying out image block to described floating image, and carrying out parallel difference to block image, obtaining the floating image after described conversion.

6. the device of a kind of SAR according to claim 5 and full-colour image rapid registering, is characterized in that, also comprise:

Difference calculating module, for calculating the difference of the CCRE that described CCRE and last iteration obtain, if described difference is less than or equal to preset error value, then terminates described iterative computation; If described difference is greater than preset error value, then repeat the action step that described deblocking module, computing module and CCRE ask for module after utilizing Powell algorithm to adjust conversion parameter.