CN101702235A - Image registration method based on triangulation - Google Patents

Abstract

The invention relates to an image registration method based on the triangulation. The method comprises the following steps of: obtaining a video image and carrying the A/D conversion; outputting and transmitting digital video image signals to a DSP device for carrying out a denoising process based on a multi-resolution method, and then carrying out a angular point detection to extracting characteristic points; carrying out a triangular matching on the characteristic points; and processing for obtaining a registering image output after matching. The method has the advantages of wide application range, high registering precision and the like, and is of far reaching importance and use value to the works, such as image fusion process of subsequent video signals, artificially intelligent process based on the characteristic grade, the decision grade and the like, or human-computer interaction judgment and decision and the like.

Description

Method for registering images based on triangulation

Technical field

The present invention relates to a kind of image processing techniques, particularly a kind of method for registering images based on triangulation.

Background technology

Because multi-source image is by different imaging sensors or the multiple image that obtained by the different imaging modes of imaging sensor of the same race, relative translation, rotation and proportional zoom etc. may appear between image, can not directly merge, and must carry out image registration earlier, to set up the corresponding relation of pixel-pixel between image.The preprocessing process that image registration is merged as multi-source image, its importance is conspicuous, and it is that multi-source image merges the precondition of being carried out, and the precision of image registration also has appreciable impact to fusion precision.

This status of image registration algorithm commonly used has three classes: based on the image registration algorithm of pixel grey scale information, as the registration Algorithm based on mutual information; Based on the image registration algorithm of frequency field, as registration Algorithm based on Fourier transform; Based on the image registration algorithm of feature, as registration Algorithm based on profile.Can be divided into two classes from sensor kind branch: same type of sensor method for registering images and foreign peoples's sensor image method for registering.

The same type of sensor multi-source image is to be obtained by the identical sensor of imaging mechanism, and gray distribution of image difference is little.According to bibliographical information, closing for the coordinate between the registering images not is the rigid body translation same type of sensor multi-source image registration problems of (just only having rotation and translation between the registering images), registration Algorithm precision based on mutual information is the highest, but be based in the image registration algorithm of mutual information and relate to a large amount of floating-point operations, time complexity is big, and probably converges to local extremum in the registration parameter optimizing process.

Foreign peoples's sensor multi-source image is to be obtained by the different sensor of imaging mechanism, and the intensity profile between image is widely different, and the general image registration algorithm based on pixel grey scale information is difficult to be suitable for.Image registration algorithm based on feature extracts some notable features such as profile from input picture, then these notable features are mated, and estimates registration parameter at last, and algorithm robustness height is suitable for the registration problems of foreign peoples's sensor multi-source image.Algorithm commonly used at present has the image registration algorithm based on profile, this algorithm just utilizes the closed contour in the input picture to carry out registration, at first extract the closed contour in the input picture, utilize the method for area invariant moment and chain representation that these closed contours are mated then, ask for registration parameter according to the barycenter of the closed contour after the coupling at last.

No matter be same type of sensor method for registering images or foreign peoples's sensor image method for registering, all there are some problems, as registration Algorithm for same type of sensor, problem relates to a large amount of floating-point operations at present, time complexity is big, and probably converges to local extremum in the registration parameter optimizing process.For foreign peoples's sensor image registration, in the actual registration situation, be difficult to sometimes extract enough closed contours from input picture.Experiment shows that the right quantity of the closed contour of coupling can influence the precision that registration parameter is estimated.Right as the closed contour that from input picture, can not detect coupling, then this registration Algorithm just can not be suitable for, the information of in registration process, only utilizing the closed contour in the input picture to provide, and abandoned the information that profile provides of opening in the input picture, certain limitation arranged in actual applications.

Summary of the invention

The present invention be directed to present multi-source image method for registering and have the limitation problem of practical application, a kind of method for registering images based on triangulation has been proposed, used the embedded system of ARM and DSP combination, this system has the mode of the registration strategies called respectively: the unloading phase, user's manual mode and update mode regularly.No matter be same type of sensor registration or foreign peoples's sensor configuration, unified utilize the Delaunay triangulation as a comparison feature carry out registration.Can effectively drive method for registering images utilization in practice, and effectively improve the performance of image registration.

Technical scheme of the present invention is: a kind of method for registering images based on triangulation comprises the steps:

1) obtaining of video image: will be converted into 24 bit digital pixel values of LDVS form from the analog video signal that CCD video camera and thermal infrared imager obtain, after the represented pixel of all simulating signals is changed through A/D, obtain the digital video image signal of a frame one frame, send it to the DSP device and carry out denoising based on multiresolution method;

2) based on the decomposition of multiresolution: adopt method, original image and Hi-pass filter h based on wavelet frame transform _n, low-pass filter g _nCarry out filtering respectively, filtered result is respectively the low-pass coefficients of wavelet frame at the high-pass coefficient and the n layer of n layer, continue if desired to decompose, then needing to proceed zero insertion handles, up to finishing set decomposition goal, for the high pass part of wavelet frame, carried out Filtering Processing, and low pass partly is used to carry out the registration strategies based on Delaunay;

3) Corner Detection: utilize the low-pass coefficients after wavelet frame decomposes, finish by each picture element that scans whole low-pass coefficients image, with each picture element is the center of circle, with certain r that configures in advance is the circle that radius is made a graphics, and on the particular location of image, realize locating, getting the picture element that drops on the graphics circle and the picture element of circle centre position compares, if the picture element on all graphics circle is greater than or less than rank t of gray-scale value arrival that the center of circle goes out picture element, think that then this point is unique point;

4) triangle coupling: construct a super triangle, comprise all unique points, put into the triangle chained list; The unique point that point is concentrated is inserted successively, in the triangle chained list, find out the triangle that its circumscribed circle comprises the insertion point, the leg-of-mutton common edge of deletion influence couples together the insertion point, thereby finishes the insertion of a point in Delaunay triangle chained list with the leg-of-mutton whole summits of influence; According to optimizing criterion the new triangle that forms in part is optimized, the triangle that forms is put into Delaunay triangle chained list; Circulation is carried out unique point and is inserted, and inserts up to all points that loose to finish;

5) registering images: the coupling in the Delaunay triangle chained list is carried out two rigidity characteristic invariants and is calculated, and supposes l ₁, l ₂, l ₃Be three sides of a triangle, m ₁, m ₂, m ₃Be leg-of-mutton three angles:

First rigid transformation invariant:

Second rigid transformation invariant: V _m=[∠ m ₁, ∠ m ₂, ∠ m ₃],

Then the ratio between them is compared, when the summation of the result after the comparison is worth less than one, then can select corresponding anglec of rotation parameter and translation parameters as final registration results.

Described step 2) in based on the method for wavelet frame transform: the employed original small echo of setting of the wavelet frame coefficient in the filtering algorithm program predefine part of wavelet frame is the DB3 small echo, and its high-pass filtering coefficient is:

[0.037828；-0.023849；-0.110624；0.377403；0.852699；0.377404；-0.110624；-0.023849；0.037828]

The low-pass filtering coefficient is:

[-0.064539；0.040689；0.418092；-0.788486；0.418092；0.040689；-0.064539；0；0]；

The coefficient that obtains after wavelet coefficient carry out frameworkization is

Its high-pass filtering coefficient is:

$h1=[0.037828;0;-0.023849;0;-0.110624;0;0.377403;0;$

$0.852699;0;0.377404;0;-0.110624;0;-0.023849;0;$

$0.037828]/\sqrt{2};$

The low-pass filtering coefficient is:

$g1=[-0.064539;0;0.040689;0;0.418092;0;-0.788486;0;$

$0.418092;0;0.040689;0;-0.064539;0;0;0;0]/\sqrt{2};$ The above wavelet frame coefficient when once decomposing only, more high-rise if desired decomposition then needs to proceed zero insertion and handles.

Beneficial effect of the present invention is: the method for registering images that the present invention is based on triangulation, have applied widely, registration accuracy advantages of higher, handle, judge the significant and practical values of work such as decision-making based on the processing or the man-machine interaction of artificial intelligence such as feature level and decision level for the image co-registration that vision signal is follow-up.

Description of drawings

Fig. 1 is the system hardware structure block diagram that the present invention is based on the method for registering images of triangulation;

Fig. 2 is the software architecture diagram that the present invention is based on the method for registering images of triangulation;

Fig. 3 is the frame in the actual video in the used ccd image in the embodiment of the invention;

Fig. 4 is the frame in the actual video in the used infrared image in the embodiment of the invention;

Fig. 5 is a frame of the final registration rear video image that obtains in the embodiment of the invention.

Embodiment

As shown in Figure 1 based on the system hardware structure block diagram of the method for registering images of triangulation, two imaging sensors have been adopted, one is AVT industrial CCD camera F-032B/C, another one is that Infrared Solutions company can provide a thermal imaging system based on non-cooling bolometer technology: IR-160 type imager, its exportable 160 * 120 pixel NTSC or PAL video output signals; Sensor kind selector switch and DSP image co-registration chip have all adopted the TMS320C6000 of TI company; A/D multi pass acquisition chip adopts 24 A/D converter CS5381; The ARM940T that the ARM chip has adopted ARM company to produce.

Based on the software architecture diagram of the method for registering images of triangulation, the registration process of arranging in the software architecture realizes calling by three kinds of modes respectively as shown in Figure 2.1) realized the automated graphics registration process of system start-up initial phase, this calls in cold start-up, warm start process and loads automatically; 2) realized regularly calling the function of image registration, this calls is that software mark after overflowing by the clock register that is provided with in the system is automatically realized calling, the initial value of this clock register has determined the interval time that timing diagram calls as registration procedure, generally speaking, this time is set to 300 seconds/time; 3) realized that the calling graph of user's triggering mode is as registration procedure, this calls is to trigger by a button on based on the function interface of the embedded OS of ARM to finish the action that this is compensated when normally the simple fused images after the user finds registration occurs than large deviation.

On the ARM chip, finish the control of startup, on-line operation and the service logic of embedded Control platform program: comprised the display interface of function button such as " being written into ", " image A ", " image B ", " image registration ", " option " and view data in the embedded Control platform program, the operable support platform of this Presentation Function (including but not limited to): OpenGL, DirectX etc.; What " be written into " that button mainly finishes view data is written into procedure for displaying automatically; " image registration " button is realized calling of aforesaid image registration program; What " option " specifically provided is the setting of various image registration parameters and image expanding method for registering, and the present invention has only provided " based on the method for registering images of Delaunay triangulation "; The display interface of view data only is that the api function that has called support platforms such as OpenGL, DirectX has been realized view window, and shows the image behind the registration in this view window; " image A ", " image B " button are respectively the dynamic images that shows two foreign peoples's sensors in view window.

Collection and input, the denoising of video image and the output of video image of video image have been carried out on the DSP module based on the method for registering images of triangulation.Its specific descriptions are as follows:

The collection of image and input: obtaining of video image is by A/D converter, at first will be converted into 24 bit digital pixel values of LDVS form from the analog video signal that CCD video camera (F-032B/C) and thermal infrared imager (IR-160) obtain.After the pixel that all CCD acquisition simulating signals is represented is changed through A/D, obtain the digital video image signal of a frame one frame at last, because the resolution of F-032B/C is 640 * 480, and the resolution of IR-160 is 160 * 120, therefore need be in the DSP side to the ccd signal processing that resamples, make it and 160 * 120 can mating fully that infrared sensor obtains, here need to prove, through manual adjustment, be that the video image that can make two cameras obtain reaches complete registration, according to this angle it carried out mechanical fixation then.After obtaining digital video image, send it to the DSP device and carry out denoising based on multiresolution method.

Each frame of data image signal carries out the denoising based on the data image signal of multiresolution on the DSP device, this processing procedure is divided into the decomposition based on multiresolution, adopts the method based on wavelet frame transform here:

The employed original small echo of setting of the wavelet frame coefficient in the filtering algorithm program predefine part of wavelet frame is the DB3 small echo, and its high-pass filtering coefficient is:

[0.037828；-0.023849；-0.110624；0.377403；0.852699；0.377404；-0.110624；-0.023849；0.037828]

The low-pass filtering coefficient is:

[-0.064539；0.040689；0.418092；-0.788486；0.418092；0.040689；-0.064539；0；0]；

The coefficient that obtains after wavelet coefficient carry out frameworkization is

Its high-pass filtering coefficient is:

$h1=[0.037828;0;-0.023849;0;-0.110624;0;0.377403;0;$

$0.852699;0;0.377404;0;-0.110624;0;-0.023849;0;$

$0.037828]/\sqrt{2};$

The low-pass filtering coefficient is:

$g1=[-0.064539;0;0.040689;0;0.418092;0;-0.788486;0;$

$0.418092;0;0.040689;0;-0.064539;0;0;0;0]/\sqrt{2};$

The above wavelet frame coefficient when once decomposing only, more high-rise if desired decomposition then needs to proceed zero insertion and handles.

Decomposable process is original image (or low-pass pictures of every layer) and Hi-pass filter h _n, low-pass filter g _nCarry out filtering respectively, filtered result is respectively the low-pass coefficients of wavelet frame at the high-pass coefficient and the n layer of n layer.Continue if desired to decompose, then use the low-pass coefficients of n layer to carry out this part computing once more, and the like, up to finishing set decomposition goal.

For the low pass of wavelet frame part and high pass part, follow-uply carried out different processing respectively.For the high pass part, this algorithm has carried out Filtering Processing, and low pass partly is used to carry out the registration strategies based on Delaunay.

Concrete steps based on the filtering algorithm of wavelet frame are: at first the multiresolution that original image is carried out wavelet frame decomposes, and sets up based on mixed Gauss model.

Wherein mixed Gauss model is:

W wherein _iBe hybrid weight, and

g _i() is complex element, i=1 wherein ..., M.The distribution of each element all is a Gaussian distribution, and its average is μ _i, variance is a ∑ _i:

Wherein d is the dimension of sampled data.

Used the coefficient of 5 * 5 windows that mixed Gauss model is estimated in the present embodiment, this estimation procedure is the estimation problem of a fragmentary data, and the present invention adopts the method for twice iteration to ask for two optimal values respectively:

First step design conditions probability density:

$g_{k,i,l}\left[I_i\left(l\right)\right]=\frac{\frac{{\mathrm{\λ}}_{k,i}}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_{k,i}^2}}\exp (-\frac{{(I_i\left(l\right)-{\mathrm{\α}}_{i}s\left(l\right)-{\mathrm{\β}}_i)}^2}{2{\mathrm{\σ}}_{k,i}^2})}{\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\frac{{\mathrm{\λ}}_{p,i}}{\sqrt{2\mathrm{\π}{\mathrm{\σ}}_{p,i}^2}}\exp (-\frac{{(I_i\left(l\right)-{\mathrm{\α}}_{i}s\left(l\right)-{\mathrm{\β}}_i)}^2}{2{\mathrm{\σ}}_{p,i}^2})}---\left(3\right)$

This step is to calculate in parameter and raw data (image) under the stable condition, the distribute conditional probability density of k item of mixed Gaussian.

The second step undated parameter α _i, α _iValue 1,0, select among the 1} so that the value maximum of (5) formula:

$Q=-\frac{1}{2}\underset{i=1}{\overset{q}{\mathrm{\Σ}}}\underset{l=1}{\overset{L}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}[\ln \left({\mathrm{\σ}}_{k,i}^2\right)+\frac{{(I_i\left(l\right)-{{\mathrm{\α}}^{\′}}_{i}S\left(l\right)-{\mathrm{\β}}_i)}^2}{2{\mathrm{\σ}}_{k,i}^2}]\·g_{k,i,l}\left[I_i\left(l\right)\right]---\left(4\right)$

The 3rd step was recomputated conditional probability density distribution g _{K, i, l}, upgrade real scene S (l):

$S^{\′}\left(l\right)=\frac{\underset{i=1}{\overset{q}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}(I_i\left(l\right)-{\mathrm{\β}}_i){{\mathrm{\α}}^{\′}}_i\frac{g_{k.i.l}\left(I_i\left(l\right)\right)}{{\mathrm{\σ}}_{k.i}^2}}{\underset{i=1}{\overset{q}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\α}}_i^{\′2}\frac{g_{k.i.l}\left(I_i\left(l\right)\right)}{{\mathrm{\σ}}_{k.i}^2}}---\left(5\right)$

${\mathrm{\β}}_i=\frac{\underset{l=1}{\overset{L}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}(I_i\left(l\right)-{\mathrm{\α}}_i^{\′}{S}^{\′}\left(l\right))\frac{g_{k.i.l}\left(I_i\left(l\right)\right)}{{\mathrm{\σ}}_{k.i}^2}}{\underset{l=1}{\overset{L}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}\frac{g_{k.i.l}\left(I_i\left(l\right)\right)}{{\mathrm{\σ}}_{k.i}^2}}---\left(6\right)$

The 4th step was recomputated conditional probability density distribution g _{K, i, l}, real scene S (l) and sensor offset amount β _i, the model parameter ω of renewal noise _{K, i}, σ _{K, i} ²And β _i:

${{\mathrm{\λ}}^{\′}}_{k,i}=\frac{1}{L}\underset{l=1}{\overset{L}{\mathrm{\Σ}}}{g}_{k.i.l}\left(I_i\left(l\right)\right),k=1,K,K;i=1,K,q---\left(7\right)$

${\mathrm{\σ}}_{k.i}^{\′2}=\frac{\underset{l=1}{\overset{L}{\mathrm{\Σ}}}{(I_i\left(l\right)-{{\mathrm{\α}}^{\′}}_{i}s\left(l\right)-{\mathrm{\β}}_i)}^2{g}_{k.i.l}\left(I_i\left(l\right)\right)}{\underset{l=1}{\overset{L}{\mathrm{\Σ}}}{g}_{k.i.l}\left(I_i\left(l\right)\right)},k=1,K,K;i=1,K,q---\left(8\right)$

In the 5th step, use new argument to repeat for first to the 4th step.

In the above-mentioned iterative process, when parameter convergence to some specific scopes, determine that fused images is:

$S^{\′}\left(l\right)=\frac{\underset{i=1}{\overset{q}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{I}_i\left(l\right){\mathrm{\α}}_i^{\′2}\frac{g_{k.i.l}\left(I_i\left(l\right)\right)\·C_i\left(l\right)}{{\mathrm{\σ}}_{k.i}^2}}{\underset{i=1}{\overset{q}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\α}}_i^{\′2}\frac{g_{k.i.l}\left(I_i\left(l\right)\right)}{{\mathrm{\σ}}_{k.i}^2}}---\left(9\right)$

C wherein _iBe the edge feature factor of influence,

Iterate the w that obtains in formula (1) and the formula (2) _i, expectation value μ _iAnd variance ∑ _iThe wavelet frame coefficient is by expectation value μ _iReplace.

Cut apart the detected characteristics extractive technique based on carrying out accumulation on the DSP module of the method for registering images of triangulation, its specific descriptions are as follows:

Utilize the low-pass coefficients after wavelet frame decomposes.Detection method is finished by each picture element that scans whole low-pass coefficients image, with each picture element is the center of circle, with certain r that configures in advance is the circle that radius is made a graphics, and on the particular location of image, realize locating, getting the picture element that drops on the graphics circle and the picture element of circle centre position compares, if the picture element on all graphics circle is greater than or less than rank t of gray-scale value arrival that the center of circle goes out picture element, think that then this point is unique point; In this embodiment, the present invention has used 3 as radius, and this means on the graphics circle has 12 pixels.Through a large amount of experimental verification, 9 pixel values should be the best-of-breed technology numbers, that is to say have in these 12 pixels 9 pixels than the gray-scale value of the pixel of circle centre position greater than t, then can think unique point.

Carry out method of addition at random on the DSP module based on the method for registering images of triangulation, its specific descriptions are as follows:

According to the point that at random order insertion point is successively concentrated, in whole process, all to safeguard and upgrade a Delaunay triangulation corresponding with current point set.The present invention adopts shortcut calculation to realize the Delaunay triangulation, and its concrete steps are: 1, initialization: construct a super triangle, comprise all diffusing points, put into the triangle chained list.2, will put concentrated diffusing point inserts successively, in the triangle chained list, find out the triangle (triangle that influences that is called this point) that its circumscribed circle comprises the insertion point, the leg-of-mutton common edge of deletion influence, the insertion point is coupled together with the leg-of-mutton whole summits of influence, thereby finish the insertion of a point in Delaunay triangle chained list.3, according to optimizing criterion the new triangle that forms in part is optimized.The triangle that forms is put into Delaunay triangle chained list.4, above-mentioned the 2nd step is carried out in circulation, inserts up to all points that loose to finish.

Triangle Optimum Matching program has used two indexs that the Delaunay triangle in the iterative process is compared, and when the summation of the result after the comparison is worth less than one, then can select corresponding anglec of rotation parameter and translation parameters as final registration results; The employed optimized Algorithm of matching process has adopted the algorithm of adaptive modeling annealing among the present invention.

These two indexs have been represented two feature invariants respectively, suppose l ₁, l ₂, l ₃Be three sides of a triangle, m ₁, m ₂, m ₃Be leg-of-mutton three angles:

First rigid transformation invariant:

$V_t=[\frac{l_1}{l_3},\frac{l_2}{l_3},\cos \left(m_3\right)]$ (formula 1)

Second rigid transformation invariant:

V _m=[∠ m ₁, ∠ m ₂, ∠ m ₃] (formula 2)

There are 24 Delaunay triangles among the present invention in the wavelet frame low-pass coefficients of infrared image, in the low-pass coefficients of visible images, have 18 Delaunay triangles.Such one has 24 * 18 (432) individual matching results, and these matching results all carry out two rigidity characteristic invariants and calculate, and then the ratio between them are compared, and chooses that group coupling of ratio maximum.Image after the coupling carries out simple pixel level and merges in the wavelet frame field, the present invention has adopted the fusion method based on traditional small echo, with the effect behind the expression registration.

The input that needs to use two width of cloth video images in the entire image fusion process of the inventive method is ccd image and infrared image as Fig. 3 and Fig. 4, and Fig. 5 is final registration results.

Claims (2)

1. the method for registering images based on triangulation is characterized in that method comprises the steps:

1) obtaining of video image: will be converted into 24 bit digital pixel values of LDVS form from the analog video signal that CCD video camera and thermal infrared imager obtain, after the represented pixel of all simulating signals is changed through A/D, obtain the digital video image signal of a frame one frame, send it to the DSP device and carry out denoising based on multiresolution method;

2) based on the decomposition of multiresolution: adopt method, original image and Hi-pass filter h based on wavelet frame transform _n, low-pass filter g _nCarry out filtering respectively, filtered result is respectively the low-pass coefficients of wavelet frame at the high-pass coefficient and the n layer of n layer, continue if desired to decompose, then needing to proceed zero insertion handles, up to finishing set decomposition goal, for the high pass part of wavelet frame, carried out Filtering Processing, and low pass partly is used to carry out the registration strategies based on Delaunay;

3) Corner Detection: utilize the low-pass coefficients after wavelet frame decomposes, finish by each picture element that scans whole low-pass coefficients image, with each picture element is the center of circle, with certain r that configures in advance is the circle that radius is made a graphics, and on the particular location of image, realize locating, getting the picture element that drops on the graphics circle and the picture element of circle centre position compares, if the picture element on all graphics circle is greater than or less than rank t of gray-scale value arrival that the center of circle goes out picture element, think that then this point is unique point;

4) triangle coupling: construct a super triangle, comprise all unique points, put into the triangle chained list; The unique point that point is concentrated is inserted successively, in the triangle chained list, find out the triangle that its circumscribed circle comprises the insertion point, the leg-of-mutton common edge of deletion influence couples together the insertion point, thereby finishes the insertion of a point in Delaunay triangle chained list with the leg-of-mutton whole summits of influence; According to optimizing criterion the new triangle that forms in part is optimized, the triangle that forms is put into Delaunay triangle chained list; Circulation is carried out unique point and is inserted, and inserts up to all points that loose to finish;

5) registering images: the coupling in the Delaunay triangle chained list is carried out two rigidity characteristic invariants and is calculated, and supposes l ₁, l ₂, l ₃Be three sides of a triangle, m ₁, m ₂, m ₃Be leg-of-mutton three angles:

First rigid transformation invariant:

Second rigid transformation invariant: V _m=[∠ m ₁, ∠ m ₂, ∠ m ₃], then the ratio between them is compared, when the summation of the result after the comparison is worth less than one, then can select corresponding anglec of rotation parameter and translation parameters as final registration results.

2. the method for registering images based on triangulation according to claim 1, it is characterized in that, described step 2) in based on the method for wavelet frame transform: the employed original small echo of setting of the wavelet frame coefficient in the filtering algorithm program predefine part of wavelet frame is the DB3 small echo, and its high-pass filtering coefficient is:

[0.037828；-0.023849；-0.110624；0.377403；0.852699；0.377404；-0.110624；-0.023849；0.037828]

The low-pass filtering coefficient is:

[-0.064539；0.040689；0.418092；-0.788486；0.418092；0.040689；-0.064539；0；0]；

The coefficient that obtains after wavelet coefficient carry out frameworkization is

Its high-pass filtering coefficient is:

$h1=[0.037828;0;-0.023849;0;-0.110624;0;0.37703;0;$

$0.852699;0;0.377404;0;-0.110624;0;-0.023849;0;$

$0.037828]/\sqrt{2};$

The low-pass filtering coefficient is:

The above wavelet frame coefficient when once decomposing only, more high-rise if desired decomposition then needs to proceed zero insertion and handles.

Images