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

Abstract

The invention belongs to the field of image processing. Specifically, it relates to a method and device for rapid registration of SAR and panchromatic images. The method includes step 1: performing data block on a reference image and a transformed floating image to obtain multiple data sub-blocks; step 2, calculate the joint probability distribution matrix and cumulative function value of each data sub-block; step 3, merge the joint probability distribution matrix and cumulative function value of each data sub-block, iteratively calculate the reference image and CCRE for floating images. Through this method, the present invention significantly improves the registration speed while ensuring the registration accuracy.

Description

A method and device for rapid registration of SAR and panchromatic images

technical field

The invention belongs to the field of image processing, and in particular relates to a method and device for rapid registration of SAR and panchromatic images.

Background technique

Multi-source image registration refers to determining the transformation parameters between images based on some similarity measures, so that two or more images of the same scene acquired from different sensors, different perspectives, and different times are transformed into the same coordinate system, and the pixel The process of getting the best match on the layer.

Remote sensing image registration methods are generally divided into image registration methods based on feature extraction and image registration methods based on gray information. In the remote sensing image registration method based on feature extraction, the most commonly used features are Harris corner points, Moravec corner points, SIFT feature points, etc. The above method has the following disadvantages: different scenes need to select different features, there is no general model to follow, and there are strong application limitations; it is easily affected by noise. Due to the imaging mechanism of SAR itself, it is very difficult to extract features accurately, and it is also difficult to design a feature extraction algorithm that can be applied to SAR images with different noise intensities, different resolutions, and different scene structures at the same time, which will lead to a series of problems. This registration algorithm is only suitable for one type of SAR image. Therefore, commonly used feature-based methods have great limitations in SAR image registration applications.

The image registration method based on gray information directly uses the gray information of the image for registration, and achieves registration through the global optimization of a certain similarity measure between pixels, and the similarity measure method is mainly considered in gray matching. This method does not need special preprocessing work on the image, is not easily affected by noise, is simple and easy to implement, and is more versatile than the method of feature extraction.

The registration method based on gray statistical properties does not need image segmentation and image feature extraction, so it can avoid the loss of precision caused by these preprocessing. However, this method generally requires a large amount of calculation, the speed of registration is slow, and it is very sensitive to changes in grayscale information. Every time the similarity measure is calculated, all pixels in the image need to be traversed, which obviously increases the amount of calculation. In the process of optimization search, the current mutual information function local extremum problem mainly obtains a smoother function curve by introducing more complex interpolation functions or combining prior information, which will increase the entire calculation process to varying degrees. Mutual information (Mutual information, MI) and normalized mutual information (Normalized mutual information, NMI) are commonly used functions widely used in the field of registration, for image registration, using mutual information and normalized mutual information is often very It is easy to fall into local extremum, which seriously affects the accuracy of registration.

Contents of the invention

The object of the present invention is to provide a method and device for rapid registration of SAR and panchromatic images, so as to solve the above problems.

In an embodiment of the present invention, a method for rapid registration of SAR and panchromatic images is provided, including:

Step 1, performing data block on the reference image and the transformed floating image to obtain multiple data sub-blocks;

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

Step 3, combining the joint probability distribution matrix and cumulative function value of each data sub-block, iteratively calculating the CCRE of the reference image and the floating image.

Further, before step 1, it also includes:

The floating image is divided into image blocks, and parallel difference is performed on the block images to obtain the above-mentioned transformed floating image.

Further, after step 3, it also includes:

Calculate the difference between the CCRE and the CCRE obtained in the previous iteration. If the difference is less than or equal to the preset error value, the iterative calculation ends; if the difference is greater than the preset error value, use the Powell algorithm to adjust the transformation parameters and repeat. Step 1 to Step 3.

Embodiments of the present invention also provide a device for rapid registration of SAR and panchromatic images, including:

The data block module is used to carry out data block to the reference image and the transformed floating image to obtain a plurality of data sub-blocks;

Calculation module, used to calculate the joint probability distribution matrix and cumulative function value of each data sub-block;

The CCRE calculation module is used to combine the joint probability distribution matrix and the cumulative function value of each data sub-block, and iteratively calculate the CCRE of the reference image and the floating image.

Further, the device also includes:

The floating image transformation module is used for performing image block on the floating image, and performing parallel difference on the block image to obtain the above-mentioned transformed floating image.

Further, the device also includes:

The difference calculation module is used to calculate the difference between the CCRE and the CCRE obtained by the previous iteration, if the difference is less than or equal to the preset error value, then the iterative calculation is terminated; if the difference is greater than the preset error value, then use After the Powell algorithm adjusts the transformation parameters, the action steps of the data block module, the calculation module and the CCRE calculation module are repeated.

Compared with the prior art, a method and device for rapid registration of SAR and panchromatic images provided by the embodiment of the present invention:

(1) Aiming at the problem that traditional mutual information image registration is easy to generate local extremum, in SAR and panchromatic image registration algorithm, the present invention adopts cross-cumulative residual entropy as similarity measure, and cross-cumulative residual entropy is stronger than Shannon entropy In general, compared with the traditional registration method using mutual information, the cross-cumulative residual entropy can effectively avoid local extremum, remove noise, and speed up the registration to a certain extent;

(2) Aiming at the problem of slow registration speed in SAR and panchromatic image registration algorithms based on cross accumulated residual entropy, the present invention proposes a parallel optimization strategy. The image registration algorithm is improved and optimized, and the registration speed is significantly improved under the premise of ensuring the registration accuracy.

Description of drawings

Fig. 1 shows a flow chart of an embodiment of a method for rapid registration of a SAR and a panchromatic image in the present invention;

Fig. 2 shows a flow chart of another embodiment of a method for fast registration of a SAR and a panchromatic image in the present invention;

Fig. 3 shows a structural block diagram of an embodiment of a device for rapid registration of a SAR and a panchromatic image in the present invention;

Fig. 4 shows a structural block diagram of another embodiment of a device for rapid registration of a SAR and a panchromatic image in the present invention;

Fig. 5 shows a schematic diagram of joint probability distribution calculation of a method for rapid registration of SAR and panchromatic images according to the present invention;

Fig. 6 shows a schematic diagram of data block parallel computing of a method for rapid registration of SAR and panchromatic images according to the present invention;

FIG. 7 shows an analysis diagram of interpolation test results of a method for rapid registration of SAR and panchromatic images according to the present invention;

Fig. 8 shows a CCRE test result analysis diagram of a method for rapid registration of SAR and panchromatic images in the present invention;

FIG. 9 shows an analysis diagram of a registration result of a method for rapid registration of a SAR and a panchromatic image according to the present invention.

Detailed ways

The present invention will be described in further detail below through specific implementation examples and in conjunction with the accompanying drawings.

Referring to FIG. 1 , FIG. 1 shows a flowchart of an embodiment of a method for rapid registration of a SAR and a panchromatic image according to the present invention.

This embodiment provides a method for rapid registration of SAR and panchromatic images, including:

Step S102, performing data block on the reference image and the transformed floating image to obtain multiple data sub-blocks;

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

Step S104, merging the joint probability distribution matrix and cumulative function value of each data sub-block, iteratively calculating the CCRE of the reference image and the floating image.

This embodiment uses cross-cumulative residual entropy as a similarity measure, and cross-cumulative residual entropy is more general than Shannon entropy. Compared with the traditional registration method using mutual information, cross-cumulative residual entropy can effectively avoid local extremum, Noise is removed, and the registration speed is accelerated to a certain extent; this embodiment adopts the method of parallel multi-threading to improve and optimize the serial image registration algorithm based on cross-accumulated residual entropy. Improved registration speed.

Referring to FIG. 2 , FIG. 2 shows a flow chart of another embodiment of a method for rapid registration of a SAR and a panchromatic image according to the present invention.

In this embodiment, before step S102, it also includes:

In step S101 , image blocks are performed on the floating image, and parallel difference is performed on the block images to obtain the transformed floating image.

In this embodiment, after step S104, it also includes:

Step S105, calculate the difference between the CCRE and the CCRE obtained in the last iteration, if the difference is less than or equal to the preset error value, then end the iterative calculation; if the difference is greater than the preset error value, then use the Powell algorithm to adjust the transformation parameters Then repeat the above step S102 to step S104.

Referring to FIG. 3 , FIG. 3 shows a structural block diagram of an embodiment of a device for rapid registration of a SAR and a panchromatic image according to the present invention.

This embodiment also provides a device for rapid registration of SAR and panchromatic images, including:

Data block module 22, for carrying out data block to reference image and transformed floating image, obtain a plurality of data sub-blocks;

Calculation module 23, used to calculate the joint probability distribution matrix and cumulative function value of each data sub-block;

The CCRE calculation module 24 is used to combine the joint probability distribution matrix and the cumulative function value of each data sub-block, and iteratively calculate the CCRE of the reference image and the floating image.

In this embodiment, by setting the data block module 22, the calculation module 23 and the CCRE calculation module 24, the cross-cumulative residual entropy is used as the similarity measure. The cross-cumulative residual entropy is more general than the Shannon entropy, and it is different from the traditional mutual information method. Compared with the registration method, the cross-accumulated residual entropy can effectively avoid local extremums, remove noise, and accelerate the registration speed to a certain extent; The image registration algorithm is improved and optimized, and the registration speed is significantly improved under the premise of ensuring the registration accuracy.

Referring to FIG. 4 , FIG. 4 shows a structural block diagram of another embodiment of a device for rapid registration of a SAR and a panchromatic image according to the present invention.

In this embodiment, the device also includes:

The floating image conversion module 21 is configured to divide the floating image into image blocks, and perform parallel difference on the block images to obtain the transformed floating image.

In this embodiment, the device also includes:

The difference calculation module 25 is used to calculate the difference between the CCRE and the CCRE obtained by the last iteration, if the difference is less than or equal to the preset error value, then end the iterative calculation; if the difference is greater than the preset error value, then The action steps of the data block module 22 , the calculation module 23 and the CCRE calculation module 24 are repeated after the transformation parameters are adjusted by the Powell algorithm.

The registration method provided by this embodiment will be further described in detail below.

In this embodiment, an image registration method based on grayscale information is adopted, and CCRE (Cross cumulative residual entropy, cross cumulative residual entropy) is used as a similarity measure. Compared with the traditional mutual information, CCRE has many excellent characteristics, so that the registration can obtain higher accuracy. However, the obvious defect of the method based on gray information is the high time complexity of the algorithm. Aiming at the problem of slow registration speed in the SAR and panchromatic image registration algorithm based on CCRE, this embodiment adopts a parallel optimization strategy, utilizes a parallel algorithm, first divides the data, and then distributes these data blocks to different threads, The data sub-blocks are calculated at the same time to achieve the purpose of reducing the running time.

SAR and panchromatic image registration are different images with displacement, rotation, etc. obtained under different viewing angles using different imaging devices in the same scene. In the process of registration, one image is used as a reference image, and the other is used as a floating image, and the floating image is adjusted to make the two images match in position. Image registration has a mathematical definition. To implement a registration algorithm, factors such as feature space, search space, similarity measure, and search strategy need to be considered.

Image registration model:

Let I _F (x, y) and I _R (x, y) represent the gray value of the floating image I _F and the reference image I _R at point (x, y) respectively, and the registration relationship of the images I _R and I _F Available formula (1) said.

I _R (x,y)=g(I _F (f(x,y))) (1)

Where f represents the geometric transformation of two-dimensional space coordinates, including translation transformation and rotation transformation. g represents the grayscale transformation function. The meaning of the formula (1) can be understood as transforming the image I _F so that the images I _F and I _R have the same gray value at corresponding spatial positions. Let T(·) denote the transformation of the image, then the image registration can be described by formula (2).

Find T ₀ ∈ Ω

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

Where Ω is the feasible domain of transformation T(·), S is the similarity measure, and Find...Max means to find the maximum value of S in the domain Ω. Formula (2) shows that image registration is to find the optimal transformation T ₀ in the feasible space of transformation T(·), so that the transformed floating image and the reference image have the greatest similarity. This embodiment uses cross-cumulative residual entropy as a similarity measure.

Matching measure using cross-cumulated residual entropy:

In multi-source remote sensing images, SAR has different imaging principles from optical images, and their features complement each other to a large extent. The optical sensor can only collect images in the presence of light, reflecting the characteristics of the visible light band; the SAR system works continuously for 24 hours, and at the same time reflects the characteristics of the electromagnetic spectrum band, which just makes up for the shortcomings of the optical system. However, due to the completely different imaging mechanisms of the two images, the information integration of the two images faces great challenges.

The coherent noise in the SAR image seriously affects the feature extraction, resulting in a decrease in the success rate of feature extraction. How the features in the optical image correspond to the same features in the SAR image directly affects the registration success rate.

In this embodiment, CCRE (Cross cumulative residual entropy, cross cumulative residual entropy) is used to describe the similarity between two groups of features, thereby solving the difficulty of extracting more precise and accurately matched features in SAR images and optical images.

The main characteristics of the cross-cumulative residual entropy are: it is suitable for both discrete and continuous environments; the distribution range is used instead of the distribution number; the calculation amount is small, and the results calculated for the sampled data are closer to the actual results. When applied to image registration, it has two obvious advantages compared with mutual information method registration: it is more immune to noise and has a wider range in the parameter transformation domain.

The traditional mutual information method counts the histogram of each gray value, which reflects a discrete feature, and independent noise will have a non-negligible impact on the matching results. The existing discrete Shannon entropy can no longer meet the requirements of overcoming a large amount of noise on SAR images, and an entropy that can reflect the cumulative change in a certain interval is needed to weaken the influence of noise on the whole. This embodiment utilizes CCRE to reflect the cumulative change in a certain interval, so as to achieve the purpose of weakening the impact of noise on the whole. The CCRE described in this embodiment is defined on the basis of cumulative residual entropy, and the definition of cumulative residual entropy is as formula ( 3) as shown.

$\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>{<span\; class="notranslate">\; \&Integral;</span>}_{{\mathrm{<span\; class="notranslate">\; R</span>}}_{<span\; class="notranslate">\; +</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; log</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; d\&lambda;</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; \&Integral;</span>}_{<span\; class="notranslate">\; -</span><span\; class="notranslate">\; \&infin;</span>}^{<span\; class="notranslate">\; +</span><span\; class="notranslate">\; \&infin;</span>}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; log</span>}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; dx</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Where x is a variable belonging to R, F(λ)=P(|x|>λ), R+=(x∈R; x≥0). Compared with Shannon entropy, the probability density function p(x) is replaced by the cumulative residual function F(λ) in formula (3). Cumulative residual functions are more general than probability density functions. Therefore, the CCRE obtained by the cumulative function has more general applicability than the mutual information obtained by the probability density function. The Shannon entropy is represented by h(x), and its definition is shown in formula (4). For variables X and Y, their mutual information MI is defined as formula (5).

MI(X,Y)＝h(X)+h(Y)-h(X,Y) (5)

Among them, h(X) and h(Y) are the entropy of X and Y respectively, and h(X,Y) is the joint entropy of X and Y. From formula (3), define CCRE as formula (6).

c(X,Y)＝ε(X)-E[ε(X/Y)] (6)

For the discrete reference image I _R and the transformed floating image I _T =T(I _F ), the calculation of CCRE needs to obtain the joint histogram, and then calculate the edge probability density and the joint probability density. The CCRE of the discrete reference image I _R and the transformed floating image _IT can be expressed by Equation (7).

$\begin{array}{c}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; R</span>}}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; R</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>\\ <span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; \&Integral;</span>}_{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}^{<span\; class="notranslate">\; \&infin;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; dl</span>}\mathrm{<span\; class="notranslate">\; log</span>}<span\; class="notranslate">\; [</span>{<span\; class="notranslate">\; \&Integral;</span>}_{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}^{<span\; class="notranslate">\; \&infin;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; dl</span>}<span\; class="notranslate">\; ]</span>\\ <span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; R</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\underset{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; T</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; \&Integral;</span>}_{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}^{<span\; class="notranslate">\; \&infin;</span>}\frac{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; R</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; log</span>}<span\; class="notranslate">\; [</span>{<span\; class="notranslate">\; \&Integral;</span>}_{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}^{<span\; class="notranslate">\; \&infin;</span>}\frac{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; R</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; dl</span>}<span\; class="notranslate">\; ]</span>\end{array}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

where p(l,k) represents the joint probability density of images _IR and _IT , and _pT (l) and _pR (k) represent the edge probability densities of the transformed floating image and the reference image, respectively. _LT and _LR represent the sets of discrete pixels of the two images respectively.

CCRE replaces the probability distribution function in the mutual information with the cumulative distribution function, which makes the CCRE change curve of the transformation parameter more gentle, and improves the local extremum problem caused by the fluctuation in the mutual information curve.

In this embodiment, CCRE is used as the similarity measure, which can improve the registration accuracy, is not affected by SAR image noise, has faster convergence speed and lower calculation amount than other registration methods.

Powell search algorithm:

Geometric transformations in image registration come in many forms, such as translation, rotation, scaling, and distortion. Different transformation spaces can be obtained by selecting different transformation forms. Consider the search space generated by translation and rotation, which can be described by three parameters. Horizontal displacement x, vertical displacement y and rotation angle θ, where x and y both belong to the integer set N, -180°≤θ<180°. All combinations of x, y, θ constitute the search space for registration. If for every combination of transformations (x, y, θ), the CCRE of two images is calculated, it will be a redundant and time-consuming process. Therefore, an optimized search algorithm needs to be used in the registration process to reduce the search space.

CCRE is not easy to calculate the derivative of the transformation parameters (x, y, θ). Compared with other optimization algorithms, such as Newton method and gradient descent method, the Powell algorithm does not need to calculate the derivative, and can simplify the extreme value problem of CCRE to a one-dimensional extreme value problem, the convergence speed is fast, and there is an obvious effect of accelerating registration. The specific steps of the algorithm are as follows:

① Given the initial point x ⁰ of the algorithm and n linearly independent directions: d ^(1,1) , d ^(1,2) , ..., d ^{(1, n)} , the allowable error is ε>0, let k= 1;

②Let x ^(k,0) ＝x ^k-1 , starting from x ^(k,0) , search along the directions d ^(k,1) , d ^(k,2) ,...,d ^(k,n) in sequence , even if

$\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; d</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>}\\ {\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; :</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; d</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}{\mathrm{<span\; class="notranslate">\; min</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; )</span>\end{array}\right.$

Get a point in each direction, x ^(k,1) , x ^(k,2) , ..., x ^(k,n) , let d ^(k,n+1) = x ^(k,n) –x ^{( k,0)} , start from x ^(k,n) and perform a one-dimensional search along the direction d ^(k,n+1) to get x ^k ;

③If ||x ^k -x ^k-1 ||<ε, then stop searching, point x ^k is the optimal solution, otherwise, set

d ^(k+1,j) =x ^(k,j+1) (j=1,2,...,n), k=k+1, return to step ②.

Although the Powell algorithm has a fast convergence speed and greatly reduces the search space, it is still a very time-consuming task in the actual registration operation, especially for large-scale image registration. In order to further shorten the registration time, the present invention adopts a parallel strategy to accelerate the registration algorithm.

Parallel strategy:

The realization of parallel computing is inseparable from the parallel computer architecture. According to Flynn's taxonomy, parallel computers are divided into single instruction stream single data stream computer (Single Instruction Single Data, SISD), single instruction stream multiple data stream computer (Single Instruction Multiple Data, SIMD), multiple instruction stream single data stream computer (Multiple Instruction Single Data, MISD) and multiple instruction stream multiple data stream computer (Multiple Instruction Multiple Data, MIMD). Among them, the MIMD system can execute multiple subtasks in parallel to reduce the execution time of the entire program, and is the most widely used parallel computer structure.

According to the structural characteristics of the MIMD system, when designing a parallel algorithm, a large calculation problem is decomposed into several relatively independent sub-problems. Assuming that the number of processors that can be used at the same time is m, for the problem P defined on the feasible domain Ω, the feasible domain is decomposed into Ω ₁ , Ω ₂ ,..., Ω _m , and the corresponding sub-problems are P ₁ , P ₂ ,..., P _m . Assign a subproblem to each processor so that the subproblems can be solved simultaneously. After all the sub-problems are solved, the solutions of the sub-problems are combined to obtain the solution of problem P.

In order for the subproblems to be solved simultaneously, there should not be strong dependencies between the subproblems. If the sub-problem P _s depends on the solution of some other sub-problem P _t , then P _s needs to wait for the solution of P _t to be completed before starting to solve it. When many sub-problems meet the above conditions, the above scheme will only bring time overhead caused by task decomposition, and will not reduce the solution time of the problem.

In the registration algorithm of this embodiment, the CCRE needs to be calculated multiple times. The calculation of CCRE relies on the new transformation parameters obtained after each iteration of the Powell algorithm, and calculating CCRE as a sub-problem of the parallel algorithm will not achieve the effect of optimization and speed-up. In the registration algorithm of this embodiment, in the process of using Powell to find the optimal solution, each round of iterative calculation of CCRE needs to calculate the joint probability distribution and cumulative function of the reference image and the floating image, and then solve the CCRE, so the calculation can be CCRE does the decomposition.

Taking the calculation of the joint probability distribution as an example, the gray level of the reference image I _R can be set as M, and the gray level of the transformed floating image _IT is N, then their joint probability distribution is shown in formula (8) .

${\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; =</span>{\left[\begin{array}{cccc}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 0,0</span>}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 0,1</span>}<span\; class="notranslate">\; )</span>& <span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>& \mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1,0</span>}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1,1</span>}<span\; class="notranslate">\; )</span>& <span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>& \mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>\\ <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>\\ <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>\\ \mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>& <span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>& \mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\end{array}\right]}_{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

The joint probability distribution is obtained by traversing the two images and according to the gray value of the pixel at the corresponding position. The calculation process is shown in Figure 5. The calculation of each element in the joint probability distribution is only obtained from the shared images I _R and _IT , without relying on the calculation of other elements, which facilitates the division of data.

Data can be divided by row or column, and the ultimate goal is to make the amount of calculation of each data sub-block equal, so that the computing power of each thread can be fully utilized to achieve load balancing. Each element of the joint probability density has a similar calculation form and consumes the same computing power, so the two division methods can achieve a uniform distribution of computing tasks. According to the number of threads available in the operating environment, computing tasks can be evenly distributed to each thread. Fig. 6 shows a schematic diagram of partitioning by row in an operating environment with p threads.

In the parallel running stage, after waiting for the calculation of each thread to finish, the calculation results of all threads are aggregated to obtain the final calculation result. Theoretically, a parallel algorithm implemented by this strategy can reduce the calculation time to 1/p before. However, there are differences in the calculation data obtained by each thread, which makes it impossible to complete the calculation tasks synchronously. In addition, the time overhead of communication between threads often makes the actual running time longer than the theoretical time. In the operations of image interpolation and image copying in the registration algorithm, the data all have the characteristics of the above-mentioned weak correlation, so this parallel strategy can be implemented. According to the above ideas, the registration algorithm is optimized in parallel to achieve fast registration.

Parallel registration algorithm:

① Prepare the storage space for the transformed image and divide it into blocks, obtain the inverse transformation matrix according to the transformation parameters, calculate the corresponding position of each pixel of the transformed image on the floating SAR image, and obtain the transformed image by parallel interpolation of the floating image;

②Using the parallel strategy, first divide the joint probability distribution matrix into blocks, calculate the joint probability distribution matrix from the reference image and the transformed floating image in parallel, and obtain the cumulative function value, merge the data blocks, and calculate the reference panchromatic image and floating image CCRE;

③If the difference between the CCRE and the CCRE obtained in the last iteration is greater than the set allowable error, use the Powell algorithm to adjust the change parameters, and jump to step ①, otherwise it means that the registration requirements have been met, and the registered floating image is obtained. The registration process Finish.

From the above ① to ③, it can be seen that the parallel strategy is used for optimization and acceleration in the data operation involving image data, but the parallel strategy is not used in the judgment and adjustment of parameters. Considering that the adjustment and judgment of parameters have a small running cost, the registration algorithm can achieve a better acceleration effect.

Parallel registration algorithm testing and analysis:

The hardware test environment configuration is: 16 Xeon E5649 CPUs, the main frequency is 2.40GHz, the total memory capacity is 16G, and the total disk capacity is 3T. The software environment is: Microsoft VisualStudio 2005, OpenMP. The test uses multiple sets of SAR and panchromatic images of the same area, and this embodiment lists the experimental results of four sets of test data.

The serial and parallel interpolation algorithms are first tested with SAR floating images of different sizes. Transformation parameters (x, y, θ) take (128, 64, -64°). Table 1 presents the results of interpolation tests with size 1024×906.

Table 1 Interpolation test results

number of processes serial 2 4 8 16 running time (ms) 797 453 282 210 142 Speedup ratio - 1.76 2.83 3.80 5.61 parallel efficiency - 0.88 0.71 0.48 0.36

The speedup ratio in Table 1 is defined as the ratio of the serial running time to the parallel running time, which intuitively shows the effect of parallel speedup. Parallel efficiency is the ratio of the speedup ratio to the number of threads, indicating the average utilization rate of each thread in the algorithm. The larger its value, the longer the parallelism time during the running of the algorithm. The test results of images of other sizes are shown in Figure 7.

It can be seen from Figure 7 that as the number of threads increases, the acceleration effect of the parallel algorithm continues to improve. It can be seen from the test results of images with a smaller size that when the number of threads reaches a certain number, the speedup ratio will not continue to increase. Too many threads will increase the burden of process communication, but will lead to a decrease in speed-up ratio.

Figure 8 shows the speedup test results for calculating a single CCRE. It can be seen from Fig. 8 that using the parallel strategy to calculate CCRE can also improve the calculation speed.

Finally, the overall registration algorithm is tested. A SAR image and a panchromatic image of the same size are used as a set of test data. Test the serial registration algorithm based on mutual information and CCRE, and the parallel algorithm based on CCRE with different numbers of threads.

Table 2 shows the test results based on mutual information and CCRE. The registration accuracy in the table is measured by the maximum number of pixels that the registration result image deviates from the reference image.

Table 2 Serial registration results based on mutual information and CCRE

It can be seen from Table 2 that the CCRE-based image registration algorithm implemented in this embodiment has faster registration speed and higher registration accuracy than the mutual information-based registration method.

In this embodiment, multiple groups of images are used to test the parallel registration algorithm. A set of image test results with a size of 1024×906 are shown in Table 3. The registration test results of other groups are shown in Fig. 9 .

Table 3 Parallel registration test results

number of processes serial 2 4 8 16 running time(s) 388 245 194 174 150 Speedup ratio - 1.58 2.00 2.23 2.59 parallel efficiency - 0.79 0.50 0.43 0.16 Accuracy (pixels) 2.68 2.68 2.68 2.68 2.68

Based on the test results in Table 3 and Figure 9, it can be seen that the parallel speedup ratio increases with the increase of the number of processes, and all speedup ratios are greater than 1. This shows that the parallel algorithm plays a role in accelerating the registration, and the increase in the number of parallel threads will make the acceleration effect more and more obvious. If the size of the registered image is small, an excessive number of parallel threads should be avoided.

The growth rate of the speedup slows down as the number of threads increases, which is due to the serial execution part of the registration algorithm. The increase in threads will only reduce the execution time of the parallel part, and the proportion of the serial part to the total execution time will continue to rise, which leads to a slowdown in the growth rate of the speedup ratio.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (6)

1. A method for rapid registration of SAR and panchromatic images, characterized in that, comprising: Step 1, performing data block on the reference image and the transformed floating image to obtain multiple data sub-blocks; Step 2, calculating the joint probability distribution matrix and cumulative function value of each data sub-block; Step 3, combining the joint probability distribution matrix and cumulative function value of each data sub-block, iteratively calculating the CCRE of the reference image and floating image. 2. The method for rapid registration of a SAR and a panchromatic image according to claim 1, characterized in that, before the step 1, it also includes: performing image block on the floating image, and performing parallel difference on the block image to obtain the transformed floating image. 3. The method for rapid registration of a SAR and a panchromatic image according to claim 2, characterized in that, after the step 3, it also includes: Calculate the difference between the CCRE and the CCRE obtained by the last iteration, if the difference is less than or equal to a preset error value, then end the iterative calculation; if the difference is greater than the preset error value, then use the Powell algorithm Repeat steps 1 to 3 after adjusting the transformation parameters. 4. A device for rapid registration of SAR and panchromatic images, characterized in that it comprises: The data block module is used to carry out data block to the reference image and the transformed floating image to obtain a plurality of data sub-blocks; Calculation module, used to calculate the joint probability distribution matrix and cumulative function value of each data sub-block; The CCRE calculation module is used to combine the joint probability distribution matrix and the cumulative function value of each data sub-block, and iteratively calculate the CCRE of the reference image and the floating image. 5. A device for rapid registration of SAR and panchromatic images according to claim 4, further comprising: The floating image conversion module is used to divide the floating image into image blocks, and perform parallel difference on the block images to obtain the transformed floating image. 6. A device for rapid registration of SAR and panchromatic images according to claim 5, further comprising: A difference calculation module, configured to calculate the difference between the CCRE and the CCRE obtained in the last iteration, if the difference is less than or equal to a preset error value, then end the iterative calculation; if the difference is greater than the preset For the error value, the Powell algorithm is used to adjust the transformation parameters and then repeat the action steps of the data block module, calculation module and CCRE calculation module.