CN104200451A - 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 algorithm

Technical field

The invention belongs to Computer Image Processing field, relate to a kind of image interfusion method based on non-local sparse K-SVD algorithm.

Background technology

Image co-registration is to become piece image the image that same target or scene are obtained with different sensors or by the image co-registration of different modes acquisition, in this width image, can react the information in multiple original image, can must be than any single source images all more accurately, more comprehensively to scene description.Image co-registration is mainly by the processing of redundant data between multiple image being improved to the reliability of image, by the processing of complementary information between multiple image being improved to the sharpness of image.

The fusion process of image can occur in the different layers that information is described.Conventionally image co-registration is divided into Pixel-level fusion, feature level fusion and decision level fusion.Multi-Sensory Image Fusion at Pixel Level is most important, the most basic image interfusion method, and it merges the respective pixel in the changing image of each width source images or source images, thereby obtains the new image of a width.

The rarefaction representation theory of signal, because it has good signal representation ability, is convenient to storage and the processing of signal, obtains a wide range of applications in recent years at image processing method face, comprises 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, a kind of pixel-level image fusion algorithm is disclosed 13 (1): 10-19 ", be referred to as SOMP (Simultaneous Orthogonal Matching Pursuit, parallel orthogonal matching pursuit) algorithm, this algorithm is based on Its Sparse Decomposition, and realized the concurrency of signal decomposition, can effectively improve stability and the reconfigurability of signal decomposition, available dictionary comprises the study dictionaries such as the parsing dictionaries such as DCT dictionary and K-SVD dictionary, image syncretizing effect improves a lot than other algorithm.Yet, resolve structural that dictionary has had, but the adaptivity not had, and the adaptivity that study dictionary has had, what still do not had is structural, and this is a relevant very large defect can using dictionary in this algorithm.

Document " Double Sparsity:Learning Sparse Dictionaries for Sparse Signal Approximation, IEEE Trans.on Signal Processing, 2010, a kind of dictionary creation algorithm that is referred to as sparse K-SVD is disclosed 58 (3): 1553-1564 ", this algorithm is by being used general parsing dictionary to carry out Its Sparse Decomposition to dictionary, the adaptivity of resolving the structural of dictionary and study dictionary is organically combined, make dictionary there is better ability to express to signal, in original text with removing picture noise, obtained good effect.

Based on study dictionary, in the rarefaction representation process of signal, inevitably can be subject to the impact of the similarity between training sample.But traditional dictionary based on resolving and the dictionary based on study all do not make full use of the non local self-similarity of image.Due in image is processed, rarefaction representation and non local self-similarity have very large advantage, Mairal etc. have proposed a kind of image denoising model, be referred to as the sparse noise reduction model of non local self-similarity, this model has utilized the non local self-similarity of image fully, adopted joint sparse to represent, improved sparse stability of solution, denoising effect is pretty good.Therefore, make full use of the non local self-similarity of image, form new sample, on new samples, carry out the study of sparse dictionary, can effectively improve the performance of dictionary.

Summary of the invention

The technical matters solving

For fear of the deficiencies in the prior art part, the present invention proposes a kind of image interfusion method based on non-local sparse K-SVD algorithm, and the adaptivity of resolving the structural of dictionary and study dictionary in SOMP algorithm is combined, and improves the performance of dictionary simultaneously.

Technical scheme

An image interfusion method based on non-local sparse K-SVD algorithm, is characterized in that step is as follows:

Step 1: choose at random m the piece of size, to each selected block, divides the window of p * q size by pixel with the order of raster scanning, obtains individual piece, then based on Euclidean distance, calculates current selected block corresponding with it the distance of individual piece, according to distance from small to large, selects a most similar piece r piece; The new vector of end to end composition successively after each piece and similar stretching one-tenth column vector thereof, obtains the matrix of (r+1) n * m size;

Step 2: use sparse K-SVD algorithm to carry out dictionary learning, obtain sparse K-SVD dictionary;

Step 3: according to the piece of size by the upper left corner to the lower right corner, with the order of raster scanning by pixel to image I _kdivide, after division, piece is carried out stretchingly, obtain a matrix wherein, k represents to treat the label of fused images, and i represents the label of column vector, and n represents the dimension of column vector, and M and N represent respectively to treat line number and the columns of fused images;

Step 4: v in whole matrix _j, j=1tok, in searching and first matrix each column vector a most similar r column vector, calculates the Euclidean distance of a certain column vector and the column vector in whole matrix of first matrix, according to distance from small to large, select r column vector, and be divided into one group, the vector in group according to order from small to large of distance, new vector of end to end composition successively; For second matrix and matrix v afterwards thereof _j, j=2tok, gets the grouping corresponding with first matrix, like this, obtains K new matrix

Step 5: the vector for K different images at same position i place the dictionary that uses sparse K-SVD algorithm to generate according to SOMP algorithm carries out Its Sparse Decomposition, obtains its rarefaction representation separately

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

Step 6: try to achieve the Its Sparse Decomposition coefficient that merges i place, rear position, the vector after being merged according to the maximum principle of absolute value wherein, F represents the image after fusion, and D represents dictionary

Step 7: by the vector obtaining after merging carry out r+1 decile, each decile vector is rearranged into piece, place it in successively in fused images and the position for the treatment of that fused images is corresponding, lap is got average, Image Reconstruction obtains fused images I _f.

Parameter while using sparse K-SVD algorithm to produce dictionary is set to: the size of DCT base dictionary is (r+1) * 64 * 100, and 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, and the number of times of iteration is 10.

Beneficial effect

A kind of image interfusion method based on non-local sparse K-SVD algorithm that the present invention proposes, sparse K-SVD algorithm is that the script that puts forward of Ron Rubinstein is for the dictionary creation algorithm of image denoising.Training sample generative process based on the non local self-similarity of image can improve the performance of dictionary effectively.The present invention by the dictionary application producing based on non-local sparse K-SVD algorithm in the image interfusion method based on SOMP algorithm, thereby reach the object that produces better syncretizing effect.The invention has the beneficial effects as follows: the image interfusion method based on non-local sparse K-SVD algorithm, according to the thought of signal Its Sparse Decomposition, in Pixel-level, image is merged, the dictionary that uses sparse K-SVD algorithm to produce, effectively combine the adaptivity of resolving the structural of dictionary and study dictionary, the signal representation ability of dictionary is improved, and the sample based on non local method selects to have improved the performance of dictionary simultaneously, and image syncretizing effect is also improved.

The present invention uses the dictionary that occupy the generation of non-local sparse K-SVD algorithm of up-to-date proposition to carry out Its Sparse Decomposition to picture signal, improves the precision of SOMP algorithm.Experimental result shows, the present invention with traditional based on resolving the image interfusion method of dictionary and SOMP algorithm and comparing and more access better image syncretizing effect with the image interfusion method of SOMP algorithm based on study dictionary.

Accompanying drawing explanation

Fig. 1 is the processing flow chart that the present invention is based on the image interfusion method of non-local sparse K-SVD algorithm

Fig. 2 is used sparse K-SVD algorithm to produce the process flow diagram of dictionary

Embodiment

Now in conjunction with the embodiments, the invention will be further described for accompanying drawing:

The treatment scheme of the image interfusion method based on non-local sparse K-SVD algorithm (with reference to accompanying drawing 1,2):

1) choose at random m the piece of size, to each selected block, the window of p * q size of take is that restriction selects r piece the most similar to it, by the first new vector of composition that joins successively after each piece and similar stretching one-tenth column vector thereof, finally obtains the matrix of (r+1) n * m size;

2) use sparse K-SVD algorithm to carry out dictionary learning, obtain sparse K-SVD dictionary;

3) to an image I _kdivide: according to the piece of size by the upper left corner to the lower right corner order with raster scanning by pixel, image is divided, after division, piece is carried out stretchingly, obtain a matrix wherein, k represents to treat the label of fused images, and i represents the label of column vector, and n represents the dimension of column vector, and M and N represent respectively to treat line number and the columns of fused images;

4) for each column vector in first matrix, in whole matrix, find r column vector the most similar to it except self, these vectors are divided into one group, vector in group successively the first joining is formed to a new vector, for second matrix and matrix afterwards thereof, get the grouping corresponding with first matrix, like this, just can obtain K new matrix;

5) vector at same position i place for K different images the dictionary that uses sparse K-SVD algorithm to generate according to SOMP algorithm carries out Its Sparse Decomposition, obtains its rarefaction representation separately

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

6) according to the maximum principle of absolute value, try to achieve the Its Sparse Decomposition coefficient that merges i place, rear position:

Further, our vector after just can being merged wherein, F represents the image after fusion, and D represents dictionary.

7) Image Reconstruction, obtains fused images I _f: first by the vector obtaining after merging carry out r+1 decile, each decile vector is rearranged into piece, place it in successively in fused images and the position for the treatment of that fused images is corresponding, lap is got average.

Use sparse K-SVD algorithm to produce the flow process (with reference to accompanying drawing 2) of dictionary:

Input: signal X ∈ R ^{n * R}, base dictionary Φ, the sparse dictionary of initialization represents A ₀, target atoms degree of rarefication p, order

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:=A ₀

Repeat following process until stop condition meets:

1) the sparse coding stage: for each sample x _i, according to following objective function

${\mathrm{\Γ}}_i:=\arg \underset{\mathrm{\γ}}{\min }{\left|\right|x_i-\mathrm{\Φa\γ}\left|\right|}_2^{2}s.t.{\left|\right|\mathrm{\γ}\left|\right|}_0^0\≤t$

Obtain its corresponding rarefaction representation Γ _i;

2) sparse dictionary represents the renewal stage: for each the row A in A _j, j=1,2 ..., L, upgrades as follows:

A _j:＝0

In I:={ signal set X, use atom a _jthe call number of signal

$g:={\mathrm{\Γ}}_{j,I}^T$

g:＝g/||g|| ₂

z:＝X _Ig-ΦaΓ _Ig

$a:=\arg \underset{a}{\min }{\left|\right|z-\mathrm{\Φa}\left|\right|}_2^{2}s.t.{\left|\right|\mathrm{a\γ}\left|\right|}_0^0\≤p$

a:＝a/||Φa|| ₂

A _j:＝a

${\mathrm{\Γ}}_{j,I}:={(X_I^T\mathrm{\Φa}-{\left({\mathrm{\Φa\Γ}}_I\right)}^T\mathrm{\Φa})}^T$

Wherein, parameter is set to: m is that 2000, n is that 64, p and q are that 10, r is 3.Parameter while using sparse K-SVD algorithm to produce dictionary is set to: the size of DCT base dictionary is 256 * 100, and 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, and the number of times of iteration is 10.While using SOMP algorithm to carry out Its Sparse Decomposition to signal, end condition can be set as the threshold value of iterations or given signal residual error according to actual needs.

Claims (2)

1. the image interfusion method based on non-local sparse K-SVD algorithm, is characterized in that step is as follows:

Step 1: choose at random m the piece of size, to each selected block, divides the window of p * q size by pixel with the order of raster scanning, obtains individual piece, then based on Euclidean distance, calculates current selected block corresponding with it the distance of individual piece, according to distance from small to large, selects a most similar piece r piece; The new vector of end to end composition successively after each piece and similar stretching one-tenth column vector thereof, obtains the matrix of (r+1) n * m size;

Step 2: use sparse K-SVD algorithm to carry out dictionary learning, obtain sparse K-SVD dictionary;

Step 3: according to the piece of size by the upper left corner to the lower right corner, with the order of raster scanning by pixel to image I _kdivide, after division, piece is carried out stretchingly, obtain a matrix wherein, k represents to treat the label of fused images, and i represents the label of column vector, and n represents the dimension of column vector, and M and N represent respectively to treat line number and the columns of fused images;

Step 4: v in whole matrix _j, j=1tok, in searching and first matrix each column vector a most similar r column vector, calculates the Euclidean distance of a certain column vector and the column vector in whole matrix of first matrix, according to distance from small to large, select r column vector, and be divided into one group, the vector in group according to order from small to large of distance, new vector of end to end composition successively; For second matrix and matrix v afterwards thereof _j, j=2tok, gets the grouping corresponding with first matrix, like this, obtains K new matrix

Step 5: the vector for K different images at same position i place the dictionary that uses sparse K-SVD algorithm to generate according to SOMP algorithm carries out Its Sparse Decomposition, obtains its rarefaction representation separately

Step 6: try to achieve the Its Sparse Decomposition coefficient that merges i place, rear position, the vector after being merged according to the maximum principle of absolute value wherein, F represents the image after fusion, and D represents dictionary

Step 7: by the vector obtaining after merging carry out r+1 decile, each decile vector is rearranged into piece, place it in successively in fused images and the position for the treatment of that fused images is corresponding, lap is got average, Image Reconstruction obtains fused images I _f.

2. the image interfusion method based on non-local sparse K-SVD algorithm according to claim 1, it is characterized in that: the parameter while using sparse K-SVD algorithm to produce dictionary is set to: the size of DCT base dictionary is (r+1) * 64 * 100, the number of target atoms is 200, the degree of rarefication of echo signal is 20, the degree of rarefication of target atoms is 10, and the number of times of iteration is 10.