CN106204657A - Moving target based on gaussian pyramid and wavelet transformation describes method across yardstick - Google Patents

Abstract

The invention discloses a kind of moving target based on gaussian pyramid and wavelet transformation and describe method across yardstick, it is characterised in that including: obtain image sequence；Build 3 layers or more than 3 layers gaussian pyramids of described image sequence；Each scale layer of described gaussian pyramid is carried out wavelet decomposition, obtains the initial subband coefficient of the subband of each scale layer；According to default convergence strategy, described initial subband coefficient is merged, obtain fused image sequence；It is trained according to described fused image sequence pair dictionary.As can be seen from above, the problems such as the method that the present invention provides dimensional properties dictionary single for yardstick present in prior art, structure is inaccurate, propose and based on gaussian pyramid and wavelet transformation describe method across yardstick, not only allow for the scale invariability between each level under different resolution, but also consider the similarity between same interlayer different sub-band, improve the accuracy that moving target is described.

Description

Moving target based on gaussian pyramid and wavelet transformation describes method across yardstick

Technical field

The present invention relates to image sequence analysis technical field, particularly relate to a kind of based on gaussian pyramid and wavelet transformation Moving target describes method across yardstick.

Background technology

Describe method across yardstick and not only there are the unexistent a lot of advantages of description method of traditional single yardstick, and to depositing Information in different scale feature has preferable expressive ability.The spy that the picture signal processed through wavelet method is had Different property, such as property openness, multiple dimensioned and the similarity etc. with human vision recognition method so that build based on wavelet method Dictionary more has an advantage, and become application most across yardstick, method is described.Wavelet method is utilized to obtain multiresolution information, And draw function and Gaussian function to combine with Dick the scale feature of image, image is carried out denoising or classification and Detection.So And in sampling process, sampling sheet is relatively big, have a strong impact on the speed of the method.

Small echo and deriving method thereof are combined with other character description methods, the information under different scale is extracted Analyze.Obtained the multiple dimensioned model of a dictionary with multiple dimensional properties by the feature of different scale, pass through Bayes Statistical model study has the dictionary of multiple dimensioned characteristic, can improve the speed of the method.But due to Bayesian frame itself Limiting, need the priori of some parameters, therefore the method can not be widely used in detection field.

Traditional moving target describes method and only spatially constructs target characteristic at single scale, and extracts it, The Analysis On Multi-scale Features of target can not be made full use of, be primarily present both sides problem: be on the one hand to utilize gaussian pyramid to obtain When taking the different scale information produced in moving target change procedure, the metric space built due to gaussian pyramid only remains The profile information of image, weakens detailed information, the phenomenon that yardstick is the biggest, details is the fewest occurs；On the other hand it is in little wavelength-division During solution, although remain the detailed information of image to greatest extent, but the profile during dimensional variation can not be utilized Information.Existing yardstick describes method in research process only to one of both accounting for, it is impossible to merge both excellent Pair graph picture is described.How can both consider, while building metric space, the dimensional information produced because of dimensional variation, Details and the profile of image itself can be taken into account again, be the key issue carrying out moving target describing across yardstick.

Summary of the invention

In view of this, it is an object of the invention to propose a kind of moving target based on gaussian pyramid and wavelet transformation across Yardstick describes method.

A kind of based on gaussian pyramid and wavelet transformation moving target based on above-mentioned purpose present invention offer is across yardstick Description method, including:

Obtain image sequence；

Build 3 layers or more than 3 layers gaussian pyramids of described image sequence；

Each scale layer of described gaussian pyramid is carried out wavelet decomposition, obtains the initial subband of the subband of each scale layer Coefficient；

According to default convergence strategy, described initial subband coefficient is merged, obtain fused image sequence；

It is trained according to described fused image sequence pair dictionary.

In some optional embodiments, 3 layers or more than 3 layers gaussian pyramids of the described image sequence of described structure, bag Include:

Described image sequence and Gaussian function are carried out convolution, obtains image after convolution；

Image after described convolution is carried out successively down-sampling, obtains image after the convolution of 3 layers or more than 3 layers, constitute Gauss Pyramid.

In some optional embodiments, described initial subband coefficient is melted by the convergence strategy that described basis is preset Close, obtain merging sub-band coefficients, including:

From the beginning of the 2nd tomographic image, the image of this tomographic image with last layer corresponding subband is entered according to matching degree convergence strategy Row merges.

In some optional embodiments, described from the beginning of the 2nd tomographic image, by this tomographic image and last layer corresponding subband Image merges according to matching degree convergence strategy, including:

Described tomographic image is carried out interpolation calculation, obtains this layer of interpolation image；

Described layer interpolation image low frequency sub-band coefficient is entered with described last layer image low frequency sub-band coefficient weighted average Row calculates, and obtains low-frequency subband fusion coefficient；

Calculate the significance measure of described layer interpolation image high-frequency sub-band；

Calculate the matching degree of described layer interpolation image high-frequency sub-band coefficient and described last layer image high-frequency sub-band coefficient, Judge the magnitude relationship of the two；

If it is determined that described layer interpolation image high-frequency sub-band coefficient less than described last layer image high-frequency sub-band coefficient Degree of joining, determines high-frequency sub-band fusion coefficients according to significance measure；

If it is determined that described layer interpolation image high-frequency sub-band coefficient is not less than described last layer image high-frequency sub-band coefficient Matching degree, determines high-frequency sub-band fusion coefficients jointly according to described significance measure and matching degree.

From the above it can be seen that the method that the present invention provides is single for yardstick present in prior art, structure The problem such as dimensional properties dictionary is inaccurate, it is proposed that based on gaussian pyramid and wavelet transformation across yardstick, method is described, no Only account for the scale invariability between each level under different resolution, but also consider between same interlayer different sub-band Similarity, improves the accuracy described for moving target.

Accompanying drawing explanation

The moving target based on gaussian pyramid and wavelet transformation that Fig. 1 provides for the present invention describes the reality of method across yardstick Execute the schematic flow sheet of example；

The moving target based on gaussian pyramid and wavelet transformation that Fig. 2 provides for the present invention across yardstick method described can Select the schematic flow sheet of embodiment；

The moving target based on gaussian pyramid and wavelet transformation that Fig. 3 provides for the present invention across yardstick method described can Select the schematic flow sheet of embodiment.

Detailed description of the invention

For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and reference Accompanying drawing, the present invention is described in more detail.

It should be noted that the statement of all uses " first " and " second " is for distinguishing two in the embodiment of the present invention The entity of individual same names non-equal or the parameter of non-equal, it is seen that " first " " second ", only for the convenience of statement, should not Being interpreted as the restriction to the embodiment of the present invention, this is illustrated by subsequent embodiment the most one by one.

The moving target based on gaussian pyramid and wavelet transformation that Fig. 1 provides for the present invention describes the reality of method across yardstick Execute the schematic flow sheet of example, as it can be seen, at a kind of motion mesh based on gaussian pyramid and wavelet transformation disclosed by the invention Mark in yardstick describes the embodiment of method, including:

S10, obtains image sequence.

S11, builds 3 layers or more than 3 layers gaussian pyramids of described image sequence.

S12, carries out wavelet decomposition to each scale layer of described gaussian pyramid, obtains subband initial of each scale layer Sub-band coefficients.

S13, merges described initial subband coefficient according to default convergence strategy, obtains fused image sequence.

S14, is trained according to described fused image sequence pair dictionary.

From the above it can be seen that the method for the present embodiment is single for yardstick present in prior art, structure The problems such as dimensional properties dictionary is inaccurate, it is proposed that based on gaussian pyramid and wavelet transformation describe method across yardstick, not only Consider the scale invariability between each level under different resolution, but also consider the phase between same interlayer different sub-band Like property, improve the accuracy that moving target is described.

The moving target based on gaussian pyramid and wavelet transformation that Fig. 2 provides for the present invention across yardstick method described can Select the schematic flow sheet of embodiment, as it can be seen, in some optional embodiments, S11, the described image sequence of described structure 3 layers or more than 3 layers gaussian pyramids, including:

S20, carries out convolution by described image sequence and Gaussian function, obtains image after convolution.

S21, carries out successively down-sampling to image after described convolution, obtains image after the convolution of 3 layers or more than 3 layers, constitutes Gaussian pyramid.

The concrete calculating formula of the present embodiment step is given below.Assume that (x y) represents the image sequence got, passes through I Each image in image sequence is carried out convolution with Gaussian function respectively and obtains corresponding gaussian pyramid, be shown below:

L (x, y, σ)=G (x, y, σ) * I (x, y)

Wherein, use normal distyribution function as Gaussian function, be shown below:

$G(x,y,\mathrm{\σ})=\frac{1}{\sqrt{2\mathrm{\π}\mathrm{\σ}}}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2})$

Gaussian convolution core can be obtained by following formula:

$H_{i,j}=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-\frac{{(i-k-1)}^2+{(j-k-1)}^2}{2{\mathrm{\σ}}^2}}$

Some preferred embodiment in, make σ=1, k=1, then the size of template is 11 × 11, utilizes above formula to obtain To following convolution kernel:

$H=\left[\begin{array}{ccccccccccc}0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000\\ 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0001& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000\\ 0.0000& 0.0000& 0.0000& 0.0002& 0.0011& 0.0018& 0.0011& 0.0002& 0.0000& 0.0000& 0.0000\\ 0.0000& 0.0000& 0.0002& 0.0029& 0.0131& 0.0215& 0.0131& 0.0029& 0.0002& 0.0000& 0.0000\\ 0.0000& 0.0000& 0.0011& 0.0131& 0.0585& 0.0965& 0.0585& 0.0131& 0.0011& 0.0000& 0.0000\\ 0.0000& 0.0001& 0.0018& 0.0215& 0.0965& 0.1592& 0.0965& 0.0215& 0.0018& 0.0001& 0.0000\\ 0.0000& 0.0000& 0.0011& 0.0131& 0.0585& 0.0965& 0.0585& 0.0131& 0.0011& 0.0000& 0.0000\\ 0.0000& 0.0000& 0.0002& 0.0029& 0.0131& 0.0215& 0.0131& 0.0029& 0.0002& 0.0000& 0.0000\\ 0.0000& 0.0000& 0.0000& 0.0002& 0.0011& 0.0018& 0.0011& 0.0002& 0.0000& 0.0000& 0.0000\\ 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0001& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000\\ 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0000\end{array}\right]$

By image after current Gaussian convolution is carried out successively down-sampling, the information that available different scale layer is corresponding, structure Build gaussian pyramid.

The moving target based on gaussian pyramid and wavelet transformation that Fig. 3 provides for the present invention across yardstick method described can Select the schematic flow sheet of embodiment.As it can be seen, in some optional embodiments, S13, according to default convergence strategy to institute State initial subband coefficient to merge, obtain merging sub-band coefficients, including:

S30, from the beginning of the 2nd tomographic image, merges the image of this tomographic image Yu last layer corresponding subband with plan according to matching degree Slightly merge, specifically include:

S31, carries out interpolation calculation to described tomographic image, obtains this layer of interpolation image.

S32, by flat with the weighting of described last layer image low frequency sub-band coefficient for described layer interpolation image low frequency sub-band coefficient All calculate, obtain low-frequency subband fusion coefficient.

S33, calculates the significance measure of described layer interpolation image high-frequency sub-band.

S34, calculate described layer interpolation image high-frequency sub-band coefficient and described last layer image high-frequency sub-band coefficient Degree of joining, it is judged that the magnitude relationship of the two.

S35 is if it is determined that described layer interpolation image high-frequency sub-band coefficient is less than described last layer image high-frequency sub-band coefficient Matching degree, determines high-frequency sub-band fusion coefficients according to significance measure.

S36, if it is determined that described layer interpolation image high-frequency sub-band coefficient is not less than described last layer image high-frequency sub-band system The matching degree of number, determines high-frequency sub-band fusion coefficients jointly according to described significance measure and matching degree.

When realizing the present embodiment, first, wavelet scale decomposition is carried out for the image information on each yardstick, Arrive:

${\{W_i^1,W_i^2,W_i^3\}}_{HL},{\{W_i^1,W_i^2,W_i^3\}}_{LH},{\{W_i^1,W_i^2,W_i^3\}}_{HH},{\{W_i^1,W_i^2,W_i^3\}}_{LL}$

Wherein, i represents i-th image,Represent that i-th image is on yardstick pyramid jth layer, subband MN Number, wherein MN={HL, LH, HH, LL}, LL represent that low frequency sub-band, HL, LH and HH represent high-frequency sub-band.

By the image of jth layer after interpolation, the imagery exploitation matching degree convergence strategy of corresponding subband upper with jth-1 layer is entered Row merges, and obtains new image, until ground floor position, wherein matching degree fusion mainly completes through following steps:

For low frequency sub-band (LL) part, the low frequency part of each source images it is weight averaged and obtains, be shown below:

$cF(m,n)=\frac{1}{2}\[cA(m,n)+cB(m,n)\]$

For high-frequency sub-band (HL or LH or HH) part, it is calculated as follows shown in formula:

${\mathrm{EX}}_j^{\mathrm{\ϵ}}(m,n)=\underset{s=-1}{\overset{1}{\Σ}}\underset{t=-1}{\overset{1}{\Σ}}R(s+2,t+2){\[{\mathrm{dX}}_j^{\mathrm{\ϵ}}(m+s,n+t)\]}^2{\mathrm{dA}}_j^{\mathrm{\ϵ}}={\mathrm{dB}}_j^{\mathrm{\ϵ}}$

Wherein,Represent image X high-frequency wavelet coefficient on jth layer ε direction,Reflect high frequency The significance measure of information significance is comprised with regional area in image.Weight matrix R is shown below:

$R=\left(\begin{array}{ccc}1/16& 1/16& 1/16\\ 1/16& 1/2& 1/16\\ 1/16& 1/16& 1/16\end{array}\right)$

The matching degree of the regional area that image A and image B is corresponding in jth layer ε directional subband is shown below:

${\mathrm{MAB}}_j^{\mathrm{\ϵ}}(m,n)=\frac{2\underset{s=-1}{\overset{1}{\Σ}}\underset{t=-1}{\overset{1}{\Σ}}R(s+2,t+2){\mathrm{dA}}_j^{\mathrm{\ϵ}}(m+s,n+t){\mathrm{dB}}_j^{\mathrm{\ϵ}}(m+s,n+t)}{{\mathrm{EA}}_j^{\mathrm{\ϵ}}(m,n)+{\mathrm{EB}}_j^{\mathrm{\ϵ}}(m,n)}$

If thr is the threshold value of matching degree, when the matching degree of two width source images is less, i.e.By showing Work property tolerance determine to merge after the choosing of wavelet coefficient, be shown below:

${\mathrm{dF}}_j^{\mathrm{\&epsiv;}}(m,n)=\left\{\begin{array}{cc}{\mathrm{dA}}_j^{\mathrm{\&epsiv;}}(m,n),& {\mathrm{EA}}_j^{\mathrm{\&epsiv;}}(m,n)\&GreaterEqual;{\mathrm{EB}}_j^{\mathrm{\&epsiv;}}(m,n)\\ {\mathrm{dB}}_j^{\mathrm{\&epsiv;}}(m,n),& {\mathrm{EA}}_j^{\mathrm{\&epsiv;}}(m,n)<{\mathrm{EB}}_j^{\mathrm{\&epsiv;}}(m,n)\end{array}\right.$

When the matching degree of two width source images is bigger, i.e.Then it is total to by significance measure and matching degree With determining the choosing of wavelet coefficient after merging, it is shown below:

${\mathrm{dF}}_j^{\mathrm{\&epsiv;}}(m,n)=\left\{\begin{array}{cc}{\mathrm{\&omega;}}_L{\mathrm{dA}}_j^{\mathrm{\&epsiv;}}(m,n)+{\mathrm{\&omega;}}_S{\mathrm{dB}}_j^{\mathrm{\&epsiv;}}(m,n),& {\mathrm{EA}}_j^{\mathrm{\&epsiv;}}(m,n)\&GreaterEqual;{\mathrm{EB}}_j^{\mathrm{\&epsiv;}}(m,n)\\ {\mathrm{\&omega;}}_S{\mathrm{dA}}_j^{\mathrm{\&epsiv;}}(m,n)+{\mathrm{\&omega;}}_L{\mathrm{dB}}_j^{\mathrm{\&epsiv;}}(m,n),& {\mathrm{EA}}_j^{\mathrm{\&epsiv;}}(m,n)<{\mathrm{EB}}_j^{\mathrm{\&epsiv;}}(m,n)\end{array}\right.$

ω in above formula_LAnd ω_SAvailable following formula represents respectively:

${\mathrm{\&omega;}}_L=\frac{1}{2}+\frac{1}{2}\left(\frac{1-{\mathrm{MAB}}_j^{\mathrm{\&epsiv;}}(m,n)}{1-thr}\right)$

ω_S=1-ω_L

In some optional embodiments, described step S14, it is trained according to described fusion image sequence pair dictionary, Can be described by following calculation:

Build about described fusion image sequence, dictionary and the object function of sparse matrix；Assume that Y is for through abovementioned steps The fusion image sequence obtained, D is dictionary, and X is corresponding sparse matrix, builds object function such as following formula, merges figure with described As sequence is trained as training data set pair dictionary:

$\begin{array}{ccc}\underset{X,D}{\min }\left|\right|Y-DX|{|}_F^2,& subjectto& \&ForAll;i,\left|\right|x_i|{|}_0\&le;\mathrm{\&epsiv;}\end{array}$

Solving described object function, solve above formula, it is carried out by the method utilizing two steps to solve.If dictionary D is Know, solve sparse matrix X by above formula；Utilize the X that the first step is asked for, ask for D by following formula:

$\hat{D}=\arg \underset{D}{\min }\left|\right|Y-DX|{|}_F^2$

Due toSo (1) formula can be obtained:

$\begin{array}{c}\left|\right|Y-DX|{|}_F^2=tr\left(Y^{T}Y\right)-tr\left(X^T{D}^{T}Y\right)-tr\left(Y^{T}DX\right)+tr\left(X^T{D}^{T}DX\right)\\ =tr\left(Y^{T}Y\right)-tr\left(X^T{D}^{T}Y\right)-tr\left(X^T{D}^{T}Y\right)+tr\left(X^T{D}^{T}DX\right)\\ =tr\left(Y^{T}Y\right)-2tr\left(D^T{\mathrm{YX}}^T\right)+tr\left({\mathrm{DXX}}^T{D}^T\right)\end{array}---\left(1\right)$

To the every D in (1) formula, in dictionary D_ijDerivation, can descend 3 formulas:

$\frac{\&part;tr\left(Y^{T}Y\right)}{\&part;D_{ij}}=0$

$\frac{\&part;tr\left(D^T{\mathrm{YX}}^T\right)}{\&part;D_{ij}}={\&lsqb;{\mathrm{YX}}^T\&rsqb;}_{ij}$

$\frac{\&part;tr\left({\mathrm{DXX}}^T{D}^T\right)}{\&part;D_{ij}}={\&lsqb;{\left({\mathrm{XX}}^T{D}^T\right)}^T\&rsqb;}_{ij}$

From above-mentioned 3 formulas, formula (1) can be written as following formula:

$\begin{array}{c}\frac{\&part;\left|\right|Y-{\mathrm{DX}}_i|{|}_F^2}{\&part;D_{ij}}=\frac{\&part;}{\&part;D_{ij}}\{tr\left(Y^{T}Y\right)-2tr\left(D^T{\mathrm{YX}}^T\right)+tr\left({\mathrm{DXY}}^T{D}^T\right)\}\\ =\frac{\&part;tr\left(Y^{T}Y\right)}{\&part;D_{ij}}-2\frac{\&part;tr\left(D^T{\mathrm{YX}}^T\right)}{\&part;D_{ij}}-\frac{\&part;tr\left({\mathrm{DXX}}^T{D}^T\right)}{\&part;D_{ij}}\\ =2{\&lsqb;{\mathrm{YX}}^T+{\mathrm{DXX}}^T\&rsqb;}_{ij}\end{array}---\left(2\right)$

Described in order to ask forsubject toIn minima, make (2) formula etc. In zero, obtain following formula:

YX^T+DXX^T=0

The analytic solutions of available formula (2), are shown below:

D=YX^T(XX^T)^-1

Above formula comprises (XX^T)^-1If, to matrix XX^TInverting, time complexity is O (n³), amount of calculation is the biggest.Can lead to Cross the every string to dictionary to be updated respectively, reduce amount of calculation.Such as, if kth row are updated, can be by only for kth item Vertical out, thus formula (2) can be written as following formula:

$\left|\right|Y-DX|{|}_F^2=||(Y-\underset{j=1,j\&NotEqual;k}{\overset{m}{\&Sigma;}}{d}_j{x}_j^T)-d_k{x}_k^T|{|}_F^2$

OrderThen can obtain further:

$\left|\right|Y-DX|{|}_F^2=||E_k-d_k{x}_k^T|{|}_F^2$

Wherein, j ≠ k.Utilize singular value method for solving can ask for d_kAnd the sparse vector of correspondence.

In sum, gaussian pyramid and wavelet transformation are combined and build across metric space by the method that the present invention provides, Training obtains the dictionary with different scale characteristic.Build yardstick pyramid, use low pass filter smoothed image, to smooth figure Down-sampled as carrying out, obtain the diminishing image of a series of size；The image of different resolution in each yardstick is entered respectively Row wavelet scale decomposes, and obtains the decomposition result under same resolution by fusion；Utilize these decomposition result that dictionary is carried out Training study, describes effect by evaluation of classification across yardstick.This method, by image carries out Gaussian smoothing, uses down-sampling side Method obtains the information in next scale layer, thus builds yardstick pyramid to obtain the Scale invariant between different resolution Property, the image on each yardstick is carried out one layer of wavelet decomposition, isolates low-frequency approximation component, high frequency horizontal direction approximation point Amount, frequency vertical direction approximation component and diagonal approximation component.Special in order to take into account scale invariability between different layers Seek peace the similarity relationships between same layer difference component, successively the different inter-layer information of each component are merged, melted Four components after conjunction.Therefore, original input picture, through the extraction again of metric space and distribution, is changed into final four Individual component, for follow-up dictionary training and classification.

Compared with prior art, the method for present invention yardstick single for yardstick present in prior art, structure is special The problem such as property dictionary is inaccurate, it is proposed that based on gaussian pyramid and wavelet transformation describe method across yardstick, not only allows for Scale invariability between each level under different resolution, but also consider the similarity between same interlayer different sub-band, Improve the accuracy that moving target is described.

Those of ordinary skill in the field are it is understood that the discussion of any of the above embodiment is exemplary only, not It is intended to imply that the scope of the present disclosure (including claim) is limited to these examples；Under the thinking of the present invention, above example Or can also be combined between the technical characteristic in different embodiments, step can realize with random order, and exists such as Other change of the many of the different aspect of the upper described present invention, in order to concisely they do not provide in details.Therefore, all Within the spirit and principles in the present invention, any omission of being made, amendment, equivalent, improvement etc., should be included in the present invention's Within protection domain.

Claims (4)

1. a moving target based on gaussian pyramid and wavelet transformation describes method across yardstick, it is characterised in that including:

Obtain image sequence；

Build 3 layers or more than 3 layers gaussian pyramids of described image sequence；

Each scale layer of described gaussian pyramid is carried out wavelet decomposition, obtains the initial subband system of the subband of each scale layer Number；

According to default convergence strategy, described initial subband coefficient is merged, obtain fused image sequence；

It is trained according to described fused image sequence pair dictionary.

Method the most according to claim 1, it is characterised in that 3 layers or more than 3 layers of the described image sequence of described structure are high This pyramid, including:

Described image sequence and Gaussian function are carried out convolution, obtains image after convolution；

Image after described convolution is carried out successively down-sampling, obtains image after the convolution of 3 layers or more than 3 layers, constitute Gauss gold word Tower.

Method the most according to claim 1, it is characterised in that the convergence strategy that described basis is preset is to described initial subband Coefficient merges, and obtains merging sub-band coefficients, including:

From the beginning of the 2nd tomographic image, the image of this tomographic image with last layer corresponding subband is melted according to matching degree convergence strategy Close.

Method the most according to claim 3, it is characterised in that described from the beginning of the 2nd tomographic image, by this tomographic image and upper The image of layer corresponding subband merges according to matching degree convergence strategy, including:

Described tomographic image is carried out interpolation calculation, obtains this layer of interpolation image；

Described layer interpolation image low frequency sub-band coefficient is counted with described last layer image low frequency sub-band coefficient weighted average Calculate, obtain low-frequency subband fusion coefficient；

Calculate the significance measure of described layer interpolation image high-frequency sub-band；

Calculate the matching degree of described layer interpolation image high-frequency sub-band coefficient and described last layer image high-frequency sub-band coefficient, it is judged that The magnitude relationship of the two；

If it is determined that described layer interpolation image high-frequency sub-band coefficient is less than the matching degree of described last layer image high-frequency sub-band coefficient, High-frequency sub-band fusion coefficients is determined according to significance measure；

If it is determined that described layer interpolation image high-frequency sub-band coefficient is not less than the coupling of described last layer image high-frequency sub-band coefficient Degree, determines high-frequency sub-band fusion coefficients jointly according to described significance measure and matching degree.