CN110428392A - A kind of Method of Medical Image Fusion based on dictionary learning and low-rank representation - Google Patents

Abstract

The present invention proposes a kind of Method of Medical Image Fusion based on dictionary learning and low-rank representation, method includes the following steps: first, classify using window sliding technology segmented image, and according to histogram of gradients feature (HOG) to multi-source image sample；Secondly, converting thereof into vector, dictionary learning is carried out, successive ignition singular value decomposition method (K-SVD) training dictionary is passed through；Then low-rank representation (LRR) method is used, the coefficient of low-rank representation is obtained；Subsequently by 1 norm maximum principle and fusion rule, preliminary blending image is obtained；Finally by image compensation, to obtain final blending image.The present invention reaches good result either in terms of the image subjective vision or in terms of objective indicator, obtains the preferable blending image of quality.

Description

A kind of Method of Medical Image Fusion based on dictionary learning and low-rank representation

Technical field

The present invention relates to field of image processing, in particular to a kind of medical image based on dictionary learning and low-rank representation melts Conjunction method.

Background technique

Medical image plays an important role in clinical diagnosis and surgical navigational.However, due to the difference of image-forming mechanism, no Same medical image is different in terms of organizing the expression with organ information.For example, computed tomography (CT) imaging can be quasi- Really detect the compact texture of such as bone and implantation material etc.Magnetic resonance imaging provides high-resolution dissection letter for soft tissue Breath, but the diagnosis fractured is sensitive not as good as CT.The imaging of same human body organ-tissue is only capable of instead by single mode medical image Reflect limited structure, form and information.In order to obtain enough diagnostic messages, doctor needs to extract from the image of different mode Information.Obviously, this method can make troubles to practical operation.In order to solve this problem, a kind of method is needed to integrate not With the supplemental information of the image of mode, i.e. image co-registration.

In the past few decades, the Image Fusion of various principles has been developed.Earlier picture merges main base In non-representative learning method.Multi-scale transform is most common method.Classical way includes more changeable in image co-registration Change domain, including laplacian pyramid (LP), wavelet transform (DWT), discrete cosine transform (DCT), non-down sampling profile Convert (NSCT) etc..There are certain defects in terms of details preservation for classical method.Another kind of method is based on dictionary learning Sparse representation method.Although the fusion method based on rarefaction representation has many advantages, the ability of global structure is captured It is limited.On the contrary, the global structure of low-rank representation (LRR) capture data, but do not retain partial structurtes.Therefore the present invention proposes A kind of multi-focus image fusing method based on dictionary learning and LRR, to solve the above problems.

Summary of the invention

For above-mentioned defect existing in the prior art or deficiency, the present invention is proposed based on dictionary learning and low-rank representation Method of Medical Image Fusion solves what the global structure in existing Medical image fusion, partial structurtes capture and details saved Problem comprises the following steps that

Step 1, to two registered width source images [I₁,I₂] using sliding window technique it is divided into image block, it calculates point The histograms of oriented gradients (HOG) of image block after cutting, and classified according to HOG feature to image block, it is assumed that I₁And I₂Ruler Very little is M × N；

Step 2, sorted image is subjected to rarefaction representation, uses the multiple dictionaries of K-SVD method training；

Step 3, multiple dictionaries of acquisition are synthesized into a dictionary, indicates source images using low-rank representation (LRR) method, obtains Obtain the coefficient of low-rank representation；

Step 4, the selection of low-rank coefficient is carried out by 1 norm maximum principle, and according to the improvement La Pula based on neighborhood The weighted sum WSEML of this operator calculates weight and is merged, and obtains preliminary blending image；

Step 5, image compensation is carried out to the preliminary blending image of acquisition, obtains final blending image.

Fusion method dictionary-based learning can carry out rarefaction representation to image, and low-rank representation (LRR) captures image Global structure, the weighted sum of introducing activity magnitude WSEML calculate weight and are merged, be can solve thin in Medical image fusion Save extraction problem.

Preferably, the step 1 specifically includes:

(1) to two registered width source images [I₁,I₂] use sliding window technique progress image segmentation, it is assumed that I₁With I₂Size be M × N, window size be n × n, step-length s, so will be divided intoA fritter P_i (j=1,2 ..., Q)；

(2) in extracting HOG characteristic procedure, it is assumed that have L chest { θ₁,θ₂,…,θ_L, G_i(θ_j) (j=1,2 ..., L) generation Table is in j-th of chest, the gradient value of i-th of fritter, defines J_iAs fritter P_iGrade, the classification method of grade is as follows,

G_iMax=max { G_i(θ_i), J=arg max_J{G_i(θ_i) represent P_iDominant gradient, T be a threshold value be used to determine Whether fritter P_iThere are dominant gradient, J_i=0 represents P_iThere is no dominant gradient, that is to say, that P_iIt is random.

The step 2 specifically includes:

(1) after classifying, all sorted image blocks is reconstructed into column vector, constitute corresponding matrix V_j(j=0,1 ..., L)；

(2) matrix V_jDictionary D_jIt can be obtained by KSVD.Subordinate dictionary D_j(j=0,1 ..., L) it respectively obtains Afterwards, it is combined into Global Dictionary D according to figure.Global Dictionary D will be used to carry out the process of image co-registration, and the word as LRR Allusion quotation input.

The step 3 specifically includes:

(1) during image co-registration, each source images are divided into Q image block first, are then sorted by dictionary All image blocks are converted into vector, by these Vector Groups composograph matrix Vs I_A, to source images I₂Identical operation is carried out, Obtain image array VI_B。

(2) LRR parameter matrix Z_AAnd Z_BIt can be by calculating VI_A, VI_BAnd following formula,

s,t.,VI_C=DZ_C+E_C,

VI_C, C ∈ { A, B } representative representative is by I₁Or I₂The image array of acquisition, D represent Global Dictionary, Z_C, C ∈ { A, B } representative VI_CCorresponding LRR coefficient matrix, | | | |_*Indicating nuclear norm, it is the summation of singular values of a matrix,Referred to as l_2,1- norm, y, x respectively represent E_CY row and xth column, λ > 0 be balance system Number, E_C, C ∈ { A, B } represents VI_CCorresponding error matrix.

The step 4 specifically includes:

(1)Z_CI, C ∈ { A, B } represent Z_CThe i-th column vector i={ 1,2..., Q }, pass through 1 norm and MAXIMUM SELECTION strategy Determining related LRR coefficient vector, fusion LRR related coefficient vector can obtain by following formula,

Z_fRepresent the fusion LRR correlation matrix obtained by formula, Z_fiRepresent Z_fThe i-th column vector i={ 1,2..., Q }, | |·||₁Indicate 1 norm；

(2) blending image block matrix V_fIt can be obtained by formula, D represents Global Dictionary, Z in this formula_fIt represents The fusion LRR correlation matrix obtained by formula,

V_f=DZ_f

Vector V_fI represents matrix V_fThe i-th column vector, i={ 1,2..., Q }, by vector V_fI is reconstructed into the image block of n × n, fixed Justice is block_i, i={ 1,2 ..., Q }；

(3) assume that two image blocks for having lap are respectively g_aAnd g_b, improvement drawing of the calculating activity magnitude based on neighborhood The weighted sum WSEML of general Laplacian operater_s,

S ∈ { A, B }, q and p are respectively represented in g_aOr g_bIn q row pth column, q={ 1,2 ..., n }, p={ 1,2 ..., n }, W are One (2r+1) × (2r+1) weight matrix, r is radius.For each element in W, value 2^2r-d, d is four neighborhoods To the distance at center, it is exactly g that S represents the image to be calculated herein_aOr g_b,

(4) work as W_A<W_B,

Fusion (a, b)=g_A(q,p)×W₁+g_B(q,p)×W₂, a={ 1,2,3..., M }, b={ 1,2,3..., N }

W in formula_A=WSEMLA (q, p), W_B=WSEMLB (q, p), W₁, W₂If representing the weight W calculated_A>W_B, exchange equation In W_AAnd W_BPosition,

Fusion (a, b)=g_B(q,p)×W₁+g_A(q,p)×W₂, a={ 1,2,3..., M }, b={ 1,2,3..., N }

Fusion is defined as preliminary blending image, and Fusion (a, b) represents the pixel in preliminary blending image at a row b column Point, q={ 1,2 ..., M }, p={ 1,2 ..., N }, g_a(q, p) and g_b(q, p) is two image blocks for having lap.Q and p points It Dai Biao not be in g_aOr g_bIn q row pth column, q={ 1,2 ..., n }, p={ 1,2 ..., n }；

The step 5 specifically includes:

(1) image Fusion and I are obtained by formula₁And I₂The difference figure DiffF of mean value,

(2) the difference figure DiffF of acquisition is subjected to morphological dilations, is as a result used as parameter adaptive Pulse Coupled Neural Network (PA-PCNN) input obtains image DiffF_PCNN, then to DiffF_PCNNNormalization；

(3) difference figure DiffF is handled according to formula, is combined with Fusion, obtains final blending image Finish,

Finish (a, b)=Fusion (a, b)+DiffF_dilate (a, b) × DiffF_PCNN(a,b)。

DiffF_dilate is that difference figure DiffF passes through the image that morphological dilations obtain.

Detailed description of the invention

In order to illustrate the embodiments of the present invention more clearly and technical solution in the prior art, below will to embodiment or Attached drawing needed to be used in the description of the prior art is briefly described, it should be apparent that, the accompanying drawings in the following description is only The embodiment of the present invention for those of ordinary skill in the art without creative efforts, can also basis The attached drawing of offer obtains other accompanying drawings.

Fig. 1 is the Medical image fusion side based on dictionary learning and low-rank representation accord to a specific embodiment of that present invention The structural block diagram of method；

Fig. 2 is two width source images accord to a specific embodiment of that present invention；

Fig. 3 is image co-registration result accord to a specific embodiment of that present invention；

Specific embodiment

The embodiment of the present invention is described below in detail, examples of the embodiments are shown in the accompanying drawings, wherein from beginning to end Same or similar label indicates same or similar element or element with the same or similar functions.Below with reference to attached The embodiment of figure description is exemplary, and for explaining only the invention, and is not considered as limiting the invention.

Referring to Figure of description 1, the Method of Medical Image Fusion of the invention based on dictionary learning and low-rank representation includes Following steps:

Step 1, to two registered width source images [I₁,I₂] using sliding window technique it is divided into image block, it calculates point The histograms of oriented gradients (HOG) of image block after cutting, and classified according to HOG feature to image block, it is assumed that I₁And I₂Ruler Very little is M × N；

(1) to two registered width source images [I₁,I₂] use sliding window technique progress image segmentation, it is assumed that I₁With I₂Size be M × N, window size be n × n, step-length s, so will be divided intoA fritter P_i (j=1,2 ..., Q)；

(2) in extracting HOG characteristic procedure, it is assumed that have L chest { θ₁,θ₂,…,θ_L, G_i(θ_j) (j=1,2 ..., L) generation Table is in j-th of chest, the gradient value of i-th of fritter, defines J_iAs fritter P_iGrade, the classification method of grade is as follows,

G_iMax=max { G_i(θ_i), J=arg max_J{G_i(θ_i) represent P_iDominant gradient, T be a threshold value be used to determine Whether fritter P_iThere are dominant gradient, J_i=0 represents P_iThere is no dominant gradient, that is to say, that P_iIt is random.

Step 2, sorted image is subjected to rarefaction representation, uses the multiple dictionaries of K-SVD method training；

(1) after classifying, all sorted image blocks is reconstructed into column vector, constitute corresponding matrix V_j(j=0,1 ..., L)；

(2) matrix V_jDictionary D_jIt can be obtained by KSVD.Subordinate dictionary D_j(j=0,1 ..., L) it respectively obtains Afterwards, it is combined into Global Dictionary D according to figure.Global Dictionary D will be used to carry out the process of image co-registration, and the word as LRR Allusion quotation input.

Step 3, multiple dictionaries of acquisition are synthesized into a dictionary, indicates source images using low-rank representation (LRR) method, obtains Obtain the coefficient of low-rank representation；

(1) during image co-registration, each source images are divided into Q image block first, are then sorted by dictionary All image blocks are converted into vector, by these Vector Groups composograph matrix Vs I_A, to source images I₂Identical operation is carried out, Obtain image array VI_B；

(2) LRR parameter matrix Z_AAnd Z_BIt can be by calculating VI_A, VI_BAnd following formula,

s,t.,VI_C=DZ_C+E_C,

VI_C, C ∈ { A, B } representative representative is by I₁Or I₂The image array of acquisition, D represent Global Dictionary, Z_C, C ∈ { A, B } representative VI_CCorresponding LRR coefficient matrix, | | | |_*Indicating nuclear norm, it is the summation of singular values of a matrix,Referred to as l_2,1- norm, y, x respectively represent E_CY row and xth column, λ > 0 be balance system Number, E_C, C ∈ { A, B } represents VI_CCorresponding error matrix.

Step 4, the selection of low-rank coefficient is carried out by 1 norm maximum principle, and according to the improvement La Pula based on neighborhood The weighted sum WSEML of this operator calculates weight and is merged, and obtains preliminary blending image；

(1)Z_CI, C ∈ { A, B } represent Z_CThe i-th column vector i={ 1,2..., Q }, pass through 1 norm and MAXIMUM SELECTION strategy Determining related LRR coefficient vector, fusion LRR related coefficient vector can obtain by following formula,

Z_fRepresent the fusion LRR correlation matrix obtained by formula, Z_fiRepresent Z_fThe i-th column vector i={ 1,2..., Q }, | |·||₁Indicate 1 norm；

(2) blending image block matrix V_fIt can be obtained by formula, D represents Global Dictionary, Z in this formula_fIt represents The fusion LRR correlation matrix obtained by formula,

V_f=DZ_f

Vector V_fI represents matrix V_fThe i-th column vector, i={ 1,2..., Q }, by vector V_fI is reconstructed into the image block of n × n, fixed Justice is block_i, i={ 1,2 ..., Q }；

(3) assume that two image blocks for having lap are respectively g_aAnd g_b, improvement drawing of the calculating activity magnitude based on neighborhood The weighted sum WSEML of general Laplacian operater_s,

S ∈ { A, B }, q and p are respectively represented in g_aOr g_bIn q row pth column, q={ 1,2 ..., n }, p={ 1,2 ..., n }, W are One (2r+1) × (2r+1) weight matrix, r is radius.For each element in W, value 2^2r-d, d is four neighborhoods To the distance at center, it is exactly g that S represents the image to be calculated herein_aOr g_b,

(4) work as W_A<W_B,

Fusion (a, b)=g_A(q,p)×W₁+g_B(q,p)×W₂, a={ 1,2,3..., M }, b={ 1,2,3..., N }

W in formula_A=WSEMLA (q, p), W_B=WSEMLB (q, p), W₁, W₂The weight calculated is represented, if W_A>W_B, exchange public affairs W in formula_AAnd W_BPosition,

Fusion (a, b)=g_B(q,p)×W₁+g_A(q,p)×W₂, a={ 1,2,3..., M }, b={ 1,2,3..., N }

Fusion is defined as preliminary blending image, and Fusion (a, b) represents the pixel in preliminary blending image at a row b column Point, q={ 1,2 ..., M }, p={ 1,2 ..., N }, g_a(q, p) and g_b(q, p) is two image blocks for having lap.Q and p points It Dai Biao not be in g_aOr g_bIn q row pth column, q={ 1,2 ..., n }, p={ 1,2 ..., n }；

Step 5, to the preliminary blending image image compensation of acquisition, final blending image is obtained,；

(1) image Fusion and I are obtained by formula₁And I₂The difference figure DiffF of mean value,

(2) the difference figure DiffF of acquisition is subjected to morphological dilations, is as a result used as parameter adaptive Pulse Coupled Neural Network (PA-PCNN) input obtains image DiffF_PCNN, then to DiffF_PCNNNormalization；

(3) difference figure DiffF is handled according to formula, is combined with Fusion, obtains final blending image Finish,

Finish (a, b)=Fusion (a, b)+DiffF_dilate (a, b) × DiffF_PCNN(a,b)。

DiffF_dilate is that difference figure DiffF passes through the image that morphological dilations obtain.

In order to which quantitative assessment difference fusion method is used for Medical image fusion performance, the present invention uses entropy (EN), difference phase The sum of closing property (SCD), average gradient (AVG), edge strength (EI), spatial frequency (SF), graphic definition (FD), mutual information (MI) Etc. evaluation parameters, for comparison fusion method use convolution rarefaction representation image co-registration (CSR), be based on parameter adaptive arteries and veins The non-sub-sampling shear transformation domain Medical image fusion (NSST-PAPCNN) of coupled neural network is rushed, direction discrete cosine is based on The image co-registration (DDCT-PCA) of transformation and principle component analysis, the Laplacian-pyramid image fusion based on discrete cosine transform Method (DCT-LP), as shown in table 1, parameter is bigger, and expression effect is better, and fused image quality is better for record.

1 distinct methods image co-registration result parameter contrast table of table

	CSR	NSST_PAPCNN	DDCT-PCA	DCT-LP	the proposed
						EN	0.999537	0.974531156	0.998245	0.779888	0.999999933
MI	1.999074	1.949062312	1.99649	1.559775	1.999999866
						SCD	1.077217	1.539213138	1.034647	0.911093	1.392087796
edge	87.55595	83.39720191	52.4083	99.62058	90.89056079
						G	10.66156	10.01598777	6.480024	14.66764	11.00615204
AVG	8.637632	8.223030571	5.25503	10.71748	8.962417008
						SF	36.64251	32.75371082	19.36494	38.21401	40.12292627

As shown in data in table 1, first is classified as evaluation parameter, and second is classified as the image co-registration (CSR) of convolution rarefaction representation, Third is classified as the non-sub-sampling shear transformation domain Medical image fusion (NSST- based on parameter adaptive Pulse Coupled Neural Network PAPCNN), the 4th it is classified as the image co-registration (DDCT-PCA) based on direction discrete cosine transform and principle component analysis, the 5th is classified as Laplacian-pyramid image fusion method (DCT-LP) based on discrete cosine transform, the 6th is classified as side proposed by the present invention Method, it can be seen that this method is a kind of unique method that front three is all come in all indexs, shows that this method can enhance The details expressive ability of image extracts more representational characteristics of image from source images and is saved, so in all fields It all does well, obtains preferably syncretizing effect.

In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show The description of example " or " some examples " etc. means specific features, structure, material or spy described in conjunction with this embodiment or example Point is included at least one embodiment or example of the invention.In the present specification, schematic expression of the above terms are not Centainly refer to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be any One or more embodiment or examples in can be combined in any suitable manner.

Although an embodiment of the present invention has been shown and described, it will be understood by those skilled in the art that: not A variety of change, modification, replacement and modification can be carried out to these embodiments in the case where being detached from the principle of the present invention and objective, this The range of invention is by claim and its equivalent limits.

Claims (6)

1. a kind of Method of Medical Image Fusion based on dictionary learning and low-rank representation, which is characterized in that this method specifically includes that

Step 1, to two registered width source images [I₁,I₂] using sliding window technique it is divided into image block, after calculating segmentation Image block histograms of oriented gradients (HOG), and classified according to HOG feature to image block, it is assumed that I₁And I₂Size be M×N；

Step 2, sorted image is subjected to rarefaction representation, it is more using successive ignition singular value decomposition method (K-SVD) training A dictionary；

Step 3, multiple dictionaries of acquisition are synthesized into a dictionary, indicates source images using low-rank representation (LRR) method, obtains low The coefficient that order indicates；

Step 4, the selection of low-rank coefficient is carried out by 1 norm maximum principle, and is drawn according to activity magnitude based on the improvement of neighborhood The weighted sum WSEML of general Laplacian operater calculates weight and is merged, and obtains preliminary blending image；

Step 5, image compensation is carried out to the preliminary blending image of acquisition, obtains final blending image.

2. a kind of Method of Medical Image Fusion based on dictionary learning and low-rank representation according to claim 1, feature It is, the step 1 specifically includes:

(1) to two registered width source images [I₁,I₂] use sliding window technique progress image segmentation, it is assumed that I₁And I₂Ruler Very little is M × N, and window size is n × n, step-length s, so will be divided intoA fritter P_i(j=1, 2,···,Q)；

(2) in extracting HOG characteristic procedure, it is assumed that have L chest { θ₁,θ₂,···,θ_L, G_i(θ_j) (j=1,2, L it) represents in j-th of chest, the gradient value of i-th of fritter, defines J_iAs fritter P_iGrade, the classification method of grade is such as Under,

G_imax=max { G_i(θ_i), J=arg max_J{G_i(θ_i) represent P_iDominant gradient, T be a threshold value be used to determination be No fritter P_iThere are dominant gradient, J_i=0 represents P_iThere is no dominant gradient, that is to say, that P_iIt is random.

3. a kind of Method of Medical Image Fusion based on dictionary learning and low-rank representation according to claim 1, feature It is, the step 2 specifically includes:

(1) after classifying, all sorted image blocks is reconstructed into column vector, constitute corresponding matrix V_j(j=0,1 ..., L)；

(2) matrix V_jDictionary D_jIt can be obtained by KSVD, subordinate dictionary D_jIt, will after (j=0,1 ..., L) is respectively obtained It is combined into Global Dictionary D according to figure, and Global Dictionary D will be used to carry out the process of image co-registration, and the dictionary as LRR is defeated Enter.

4. a kind of Method of Medical Image Fusion based on dictionary learning and low-rank representation according to claim 1, feature It is, the step 3 specifically includes:

(1) during image co-registration, each source images are divided into Q image block first, are then sorted by dictionary by institute Some image blocks are converted into vector, by these Vector Groups composograph matrix Vs I_A, to source images I₂Identical operation is carried out, is obtained Image array VI_B；

(2) LRR parameter matrix Z_AAnd Z_BIt can be by calculating VI_A, VI_BAnd following formula,

s,t.,VI_C=DZ_C+E_C,

VI_C, C ∈ { A, B } representative representative is by I₁Or I₂The image array of acquisition, D represent Global Dictionary, Z_C, C ∈ { A, B } represents VI_C Corresponding LRR coefficient matrix, | | | |_*Indicating nuclear norm, it is the summation of singular values of a matrix,Referred to as l_2,1- norm, y, x respectively represent E_CY row and xth column, λ > 0 be balance system Number, E_C, C ∈ { A, B } represents VI_CCorresponding error matrix.

5. a kind of Method of Medical Image Fusion based on dictionary learning and low-rank representation according to claim 1, feature It is, the step 4 specifically includes:

(1)Z_CI, C ∈ { A, B } represent Z_CThe i-th column vector i={ 1,2..., Q }, by 1 norm and MAXIMUM SELECTION strategy come really Fixed correlation LRR coefficient vector, fusion LRR related coefficient vector can be obtained by following formula,

Z_fRepresent the fusion LRR correlation matrix obtained by formula, Z_fiRepresent Z_fThe i-th column vector i={ 1,2..., Q }, | |·||₁Indicate 1 norm；

(2) blending image block matrix V_fIt can be obtained by formula, D represents Global Dictionary, Z in this formula_fIt represents by public affairs The fusion LRR correlation matrix that formula obtains,

V_f=DZ_f

Vector V_fI represents matrix V_fThe i-th column vector, i={ 1,2..., Q }, by vector V_fI is reconstructed into the image block of n × n, It is defined as block_i, i={ 1,2, Q }；

(3) assume that two image blocks for having lap are respectively g_aAnd g_b, improvement La Pula of the calculating activity magnitude based on neighborhood The weighted sum WSEML of this operator_s,

S ∈ { A, B }, q and p are respectively represented in g_aOr g_bIn q row pth column, q={ 1,2, n }, p=1, 2, n }, W is one (2r+1) × (2r+1) weight matrix, and r is radius, and for each element in W, value is 2^2r-d, d is distance of four neighborhoods to center, and it is exactly g that S represents the image to be calculated herein_aOr g_b,

(4) work as W_A<W_B,

Fusion (a, b)=g_A(q,p)×W₁+g_B(q,p)×W₂, a={ 1,2,3..., M }, b={ 1,2,3..., N }

W in formula_A=WSEMLA (q, p), W_B=WSEMLB (q, p), W₁, W₂The weight calculated is represented, if W_A>W_B, exchange public affairs W in formula_AAnd W_BPosition,

Fusion (a, b)=g_B(q,p)×W₁+g_A(q,p)×W₂, a={ 1,2,3..., M }, b={ 1,2,3..., N }

Fusion is defined as preliminary blending image, and Fusion (a, b) represents the pixel in preliminary blending image at a row b column Point, q={ 1,2, M }, p={ 1,2, N }, g_a(q, p) and g_b(q, p) is two image blocks for having lap, Q and p are respectively represented in g_aOr g_bIn q row pth column, q={ 1,2, n }, p={ 1,2, n }.

6. a kind of Method of Medical Image Fusion based on dictionary learning and low-rank representation according to claim 1, feature It is, the step 5 specifically includes:

(1) image Fusion and I are obtained by formula₁And I₂The difference figure DiffF of mean value,

(2) the difference figure DiffF of acquisition is subjected to morphological dilations, is as a result used as parameter adaptive Pulse Coupled Neural Network (PA- PCNN input) obtains image DiffF_PCNN, then to DiffF_PCNNNormalization；

(3) difference figure DiffF is handled according to formula, is combined with Fusion, obtains final blending image Finish,

Finish (a, b)=Fusion (a, b)+DiffF_dilate (a, b) × DiffF_PCNN(a,b)

DiffF_dilate is that difference figure DiffF passes through the image that morphological dilations obtain.