CN104200451B - Image fusion method based on non-local sparse K-SVD algorithm - Google Patents

Abstract

The invention relates to an image fusion method based on a non-local sparse K-SVD algorithm. The sparse K-SVD algorithm is the dictionary generation algorithm which is presented by Ron Rubinstein and used for image de-noising. A training sample generation process based on image non-local self-similarity is capable of effectively improving the dictionary performance. The image fusion method based on the non-local sparse K-SVD algorithm uses the dictionary which is generated based on the non-local sparse K-SVD algorithm for the image fusion method based on SOMP algorithm so as to generate a better fusion effect. The image fusion method based on the non-local sparse K-SVD algorithm has beneficial effects that the image is fused on a pixel level according to the signal sparse decomposition idea, the dictionary generated based on the sparse K-SVD algorithm effectively combines the analytic dictionary structure and learning dictionary adaptability to improve the signal presentation skill of the dictionary, and meanwhile, the sample selection based on the non-local method improves the dictionary performance, and the image fusion effect is improved.

Description

A kind of image interfusion method based on non-local sparse K-SVD algorithms

Technical field

The invention belongs to Computer Image Processing field, is related to a kind of image based on non-local sparse K-SVD algorithms and melts Conjunction method.

Background technology

Image co-registration is the image obtained to same target or scene different sensors or differently obtains Image co-registration into piece image, the information in multiple original image can be reacted in diagram picture, scene description be able to must be compared Any single source images are all more accurately, more comprehensively.Image co-registration mainly by the process redundant data between multiple image come The reliability of image is improved, the definition of image is improved by the process complementary information between multiple image.

The fusion process of image can occur the different layers in information description.Generally image co-registration is divided into into Pixel-level to melt Conjunction, feature-based fusion and decision level fusion.Multi-Sensory Image Fusion at Pixel Level is most important, most basic image interfusion method, it Respective pixel in the changing image of each width source images or source images is merged, the image new so as to obtain a width.

The sparse representation theory of signal has good signal representation ability due to it, is easy to the storage of signal and processes, Obtain a wide range of applications in terms of image procossing in recent years, including signal reconstruction, image denoising and image co-registration etc..Document “B.Yang,S.Li,Pixel-level image fusion with simultaneous orthogonal matching pursuit,Information Fusion,2012,13(1):A kind of pixel-level image fusion algorithm is disclosed in 10-19 ", is claimed Be SOMP (Simultaneous Orthogonal Matching Pursuit, parallel orthogonal match tracing) algorithm, the calculation Method realizes the concurrency of signal decomposition based on Its Sparse Decomposition, can effectively improve signal decomposition stability and Reconfigurability, available dictionary includes the parsing study dictionary such as dictionary and K-SVD dictionaries such as DCT dictionaries, image syncretizing effect compared with Others algorithm improve a lot.However, parsing dictionary had it is structural, but the adaptivity not had, And learn the adaptivity that had of dictionary, but do not had it is structural, this is about using dictionary in the algorithm One very big defect.

Document " Double Sparsity:Learning Sparse Dictionaries for Sparse Signal Approximation,IEEE Trans.on Signal Processing,2010,58(3):One is disclosed in 1553-1564 " The dictionary creation algorithm for being referred to as sparse K-SVD is planted, the algorithm using general parsing dictionary to dictionary by carrying out sparse point Solution, structural and study dictionary the adaptivity of parsing dictionary is organically combined so that dictionary has more preferable table to signal Danone power, is used for removing picture noise in original text, achieves preferable effect.

During based on rarefaction representation of the study dictionary to signal, inevitably by similar between training sample The impact of property.But, traditional does not all make full use of the non local of image based on the dictionary for parsing and based on the dictionary of study Self-similarity.Because in image procossing, there is very big advantage, Mairal etc. to propose for rarefaction representation and non local self-similarity A kind of image denoising model, the sparse noise reduction model of referred to as non local self-similarity, this model is sufficiently used image Non local self-similarity, employ joint sparse and represent, improve sparse stability of solution, denoising effect is pretty good.Therefore, fill Divide using the non local self-similarity of image, constitute new sample, the study of sparse dictionary is carried out on new samples, can be effective Improve the performance of dictionary in ground.

The content of the invention

The technical problem to be solved

In order to avoid the deficiencies in the prior art part, the present invention proposes a kind of figure based on non-local sparse K-SVD algorithms As fusion method, the adaptivity that the structural of dictionary and study dictionary are parsed in SOMP algorithms is combined, while improving The performance of dictionary.

Technical scheme

A kind of image interfusion method based on non-local sparse K-SVD algorithms, it is characterised in that step is as follows：

Step 1：Randomly select mThe block of size, to each selected block, with the order of raster scanning by picture Element is divided to the window of p × q sizes, is obtainedIndividual block, is then based on Euclidean distance, calculates Currently selected piece correspondingThe distance of individual block, according to distance from small to large, selects most phase As r block；Each block and its similar block stretch into after column vector the new vector of end to end composition successively, obtain (a r+ 1) matrix of n × m sizes；

Step 2：Dictionary learning is carried out using sparse K-SVD algorithms, sparse K-SVD dictionaries are obtained；

Step 3：According toThe block of size by the upper left corner to the lower right corner, with the order of raster scanning pixel-by-pixel to figure As I_kDivided, stretched block after division, obtained K matrixK=1 ..., K.Each square Battle array beWherein, k represents the label of image to be fused, has K image, i The label of column vector is represented, n represents the dimension of column vector, and M and N represents respectively the line number and columns of image to be fused；

Step 4：In first matrixFind each column vectorWith in addition to itself Column vectorR j ≠ i most like column vector, that is, calculate first matrix A certain column vector and whole matrix in column vector (in addition to itself) Euclidean distance, according to distance from small to large, select r Column vector, and be divided into one group, the vector in group according toDistance order from small to large, end to end composition one successively Individual new vector, obtains a new matrixFor Second matrix and its matrix afterwardsJ=2 ..., K, to each column vectorPacket corresponding with first matrix is taken, so, K new square is obtained Battle arrayK=1 ..., K；

Step 5：For vector of the K different images at same position iSparse K- is used according to SOMP algorithms The dictionary that svd algorithm is generated carries out Its Sparse Decomposition, obtains its respective rarefaction representation

Step 6：The sparse decomposition coefficients after fusion at the i of position are tried to achieve according to maximum absolute value principle, after being merged VectorWherein, F represents the image after fusion, and D represents dictionary

Step 7：The vector that will be obtained after fusionR+1 deciles are carried out, each decile vector is rearranged into's Block, places it in successively in fused images at position corresponding with image to be fused, and lap takes average, and Image Reconstruction is obtained To fused images I_F。

Parameter when producing dictionary using sparse K-SVD algorithms is set to：The size of DCT base dictionaries be (r+1) * 64 × 100, the number of target atoms is 200, and the degree of rarefication of echo signal is 20, and the degree of rarefication of target atoms is 10, the number of times of iteration For 10.

Beneficial effect

A kind of image interfusion method based on non-local sparse K-SVD algorithms proposed by the present invention, sparse K-SVD algorithms are What Ron Rubinstein put forward is originally used for the dictionary creation algorithm of image denoising.Based on the non local self-similarity of image Training sample generating process can effectively improve the performance of dictionary.The present invention will be produced based on non-local sparse K-SVD algorithms Dictionary application to based on SOMP algorithms image interfusion method in, so as to reach the purpose for producing more preferable syncretizing effect.This Bright beneficial effect is：Based on the image interfusion method of non-local sparse K-SVD algorithms, according to the thought of signal Its Sparse Decomposition, Image is merged on pixel level, the dictionary produced using sparse K-SVD algorithms effectively combines the structure of parsing dictionary Property and study dictionary adaptivity so that the signal representation ability of dictionary is improved, while the sample based on non local method is selected The performance that improve dictionary is selected, image syncretizing effect is also improved.

The present invention is carried out dilute using the dictionary for occuping the generation of non-local sparse K-SVD algorithms of newest proposition to picture signal Dredge and decompose, improve the precision of SOMP algorithms.Test result indicate that, it is of the invention with traditional based on parsing dictionary and SOMP algorithms Image interfusion method and based on study dictionary compare with the image interfusion method of SOMP algorithms more access better image fusion Effect.

Description of the drawings

Fig. 1 is process chart of the present invention based on the image interfusion method of non-local sparse K-SVD algorithms

Fig. 2 is the flow chart for using sparse K-SVD algorithms to produce dictionary

Specific embodiment

In conjunction with embodiment, accompanying drawing, the invention will be further described：

Handling process (referring to the drawings one, two) based on the image interfusion method of non-local sparse K-SVD algorithms：

1) m is randomly selectedThe block of size, to each selected block, with the window of p × q sizes as restriction choosing Most like with it blocks of r are selected, each block and its similar block are stretched into into after column vector first place are successively connected and constitute new vector, Finally give the matrix of (r+1) n × m size；

2) dictionary learning is carried out using sparse K-SVD algorithms, obtains sparse K-SVD dictionaries；

To an image I_kDivided：According toThe block of size by the upper left corner to the lower right corner, with raster scanning Order is pixel-by-pixel to image I_kDivided, stretched block after division, obtained K matrixK= 1,...,K.Each matrix isWherein, k represents the mark of image to be fused Number, K image is had, i represents the label of column vector, and n represents the dimension of column vector, and M and N represents respectively the row of image to be fused Number and columns；

3) for each column vector in first matrix, most like with it r in addition to itself is found in the matrix Individual column vector, these vectors one group is divided into, and the vector in group successively first place is connected and constitutes a new vector, for second Individual matrix and its matrix afterwards, take packet corresponding with first matrix, in this manner it is possible to obtain K new matrix；

4) vector for K different images at same position iCalculated using sparse K-SVD according to SOMP algorithms The dictionary that method is generated carries out Its Sparse Decomposition, obtains its respective rarefaction representation

5) sparse decomposition coefficients after fusion at the i of position are tried to achieve according to maximum absolute value principle：

Further, we can be obtained by the vector after mergingWherein, F represents the image after fusion, D

Represent dictionary.

6) Image Reconstruction, obtains fused images I_F：The vector that will be obtained after fusion firstR+1 deciles are carried out, by each Decile vector is rearranged intoBlock, place it in successively in fused images at position corresponding with image to be fused, overlap Part takes average.

The flow process (referring to the drawings two) of dictionary is produced using sparse K-SVD algorithms：

Input：Signal X ∈ R^N×R, base dictionary Φ, initialization sparse dictionary represent A₀, target atoms degree of rarefication p, mesh

Mark signal degree of rarefication t, iterations k.

Output：Sparse dictionary represents A and sparse signal representation Γ, meets X ≈ Φ A Γ.

Initialization：A:=A₀

The following process of repetition meets until stop condition：

1) the sparse coding stage：For each sample x_i, according to following object function

Obtain its corresponding rarefaction representation Γ_i；

2) sparse dictionary represents the more new stage：Each row A in for A_j, j=1,2 ..., L update as follows：

A_j:=0

I:={ atom a used in signal set X_jSignal call number

g:=g/ | | g | |₂

z:=X_Ig-ΦaΓ_Ig

a:=a/ | | Φ a | |₂

A_j:=a

Wherein, parameter is set to：M is that 2000, n is 64, p and q is 3 for 10, r.Dictionary is produced using sparse K-SVD algorithms When parameter be set to：The size of DCT base dictionaries is 256 × 100, and the number of target atoms is 200, the degree of rarefication of echo signal For 20, the degree of rarefication of target atoms is 10, and the number of times of iteration is 10.When carrying out Its Sparse Decomposition to signal using SOMP algorithms, eventually Only condition can according to actual needs be set as the threshold value of iterations or Setting signal residual error.

Claims (2)

1. a kind of image interfusion method based on non-local sparse K-SVD algorithms, it is characterised in that step is as follows：

Step 1：Randomly select mThe block of size, to each selected block, with the order of raster scanning pixel-by-pixel to p The window of × q sizes is divided, and is obtainedIndividual block, is then based on Euclidean distance, calculates current Selected block is correspondingThe distance of individual block, according to distance from small to large, selects most like r Individual block；Each block and its similar block stretch into after column vector the new vector of end to end composition successively, obtain (r+1) in × The matrix of m sizes；

Step 2：Dictionary learning is carried out using sparse K-SVD algorithms, sparse K-SVD dictionaries are obtained；

Step 3：According toThe block of size by the upper left corner to the lower right corner, with the order of raster scanning pixel-by-pixel to image I_kEnter Row is divided, and is stretched block after division, obtains K matrixEach matrix isWherein, k represents the label of image to be fused, has K image, and i is represented The label of column vector, n represents the dimension of column vector, and M and N represents respectively the line number and columns of image to be fused；

Step 4：In first matrixFind each column vectorWith the column vector in addition to itselfR most like column vector, that is, calculate a certain row of first matrix It is vectorial with whole matrix in column vector in addition to itself Euclidean distance, according to distance from small to large, select r column vector, and Be divided into one group, the vector in group according toDistance order from small to large, successively end to end composition one it is new to Amount, obtains a new matrixFor second matrix and its afterwards MatrixTo each column vector Packet corresponding with first matrix is taken, so, K new matrix is obtained

Step 5：For vector of the K different images at same position iCalculated using sparse K-SVD according to SOMP algorithms The dictionary that method is generated carries out Its Sparse Decomposition, obtains its respective rarefaction representation

${\mathrm{\α}}_F^i\left(t\right)={\mathrm{\α}}_{\hat{k}}^i\left(t\right),\hat{k}=\arg \underset{k=1,2,\mathrm{...},K}{max}\left(\right|{\mathrm{\α}}_k^i\left(t\right)\left|\right)$

$;$

Step 6：The sparse decomposition coefficients after fusion at the i of position, the vector after being merged are tried to achieve according to maximum absolute value principleWherein, F represents the image after fusion, and D represents dictionary；

Step 7：The vector that will be obtained after fusionR+1 deciles are carried out, each decile vector is rearranged intoBlock, according to Secondary to place it in fused images at position corresponding with image to be fused, lap takes average, and Image Reconstruction is merged Image I_F。

2. the image interfusion method of non-local sparse K-SVD algorithms is based on according to claim 1, it is characterised in that：Use Parameter when sparse K-SVD algorithms produce dictionary is set to：The size of DCT base dictionaries is (r+1) * 64 × 100, target atoms Number is 200, and the degree of rarefication of echo signal is 20, and the degree of rarefication of target atoms is 10, and the number of times of iteration is 10.