CN111784572B - Image fusion and super-resolution joint implementation method based on discriminant dictionary learning - Google Patents
Image fusion and super-resolution joint implementation method based on discriminant dictionary learning Download PDFInfo
- Publication number
- CN111784572B CN111784572B CN202010425926.0A CN202010425926A CN111784572B CN 111784572 B CN111784572 B CN 111784572B CN 202010425926 A CN202010425926 A CN 202010425926A CN 111784572 B CN111784572 B CN 111784572B
- Authority
- CN
- China
- Prior art keywords
- image
- low
- dictionary
- resolution
- sparse
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 68
- 230000004927 fusion Effects 0.000 title claims abstract description 59
- 239000011159 matrix material Substances 0.000 claims abstract description 41
- 238000006243 chemical reaction Methods 0.000 claims abstract description 10
- 230000009466 transformation Effects 0.000 claims abstract description 9
- 238000000926 separation method Methods 0.000 claims abstract description 7
- 239000004576 sand Substances 0.000 claims description 39
- 230000014509 gene expression Effects 0.000 claims description 33
- 238000012549 training Methods 0.000 claims description 26
- 238000000354 decomposition reaction Methods 0.000 claims description 24
- 238000005457 optimization Methods 0.000 claims description 14
- 238000007500 overflow downdraw method Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 9
- 239000013598 vector Substances 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 8
- 230000000877 morphologic effect Effects 0.000 claims description 7
- 230000009977 dual effect Effects 0.000 claims description 6
- 230000007547 defect Effects 0.000 claims description 4
- OAICVXFJPJFONN-UHFFFAOYSA-N Phosphorus Chemical compound [P] OAICVXFJPJFONN-UHFFFAOYSA-N 0.000 claims description 3
- XOFYZVNMUHMLCC-ZPOLXVRWSA-N prednisone Chemical compound O=C1C=C[C@]2(C)[C@H]3C(=O)C[C@](C)([C@@](CC4)(O)C(=O)CO)[C@@H]4[C@@H]3CCC2=C1 XOFYZVNMUHMLCC-ZPOLXVRWSA-N 0.000 claims description 3
- 239000011541 reaction mixture Substances 0.000 claims description 3
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims 1
- 238000012545 processing Methods 0.000 abstract description 4
- 230000000007 visual effect Effects 0.000 abstract description 4
- 238000011156 evaluation Methods 0.000 description 10
- 238000002474 experimental method Methods 0.000 description 7
- 230000000295 complement effect Effects 0.000 description 4
- 238000011387 Li's method Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000011158 quantitative evaluation Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000593 degrading effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000004438 eyesight Effects 0.000 description 1
- 229920000314 poly p-methyl styrene Polymers 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000016776 visual perception Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention relates to a method for jointly realizing image fusion and super-resolution based on discriminative dictionary learning, belonging to the technical field of digital image processing. Specifically, two pairs of low-rank, sparse dictionaries and a high-resolution and low-resolution image coding coefficient transformation matrix are jointly trained first. One pair of dictionaries is used for representing low-rank and sparse components of an input image, the other pair is used for reconstructing a high-resolution fusion low-rank and sparse components, and a conversion matrix is used for establishing a potential relation between a high-resolution image and a low-resolution image. Then, a sparse and low-rank separation model is constructed, the input image is effectively decomposed into low-rank and sparse components, and therefore a high-resolution fusion image can be constructed through different dictionaries. The invention realizes the fusion of images and super-resolution reconstruction jointly. The experimental result shows that the invention has better fusion performance in both visual effect and objective index.
Description
Technical Field
The invention relates to a method for realizing image fusion and super-resolution combination based on discriminative dictionary learning, belonging to the technical field of digital image processing.
Background
Image fusion can integrate complementary information acquired by different sensors about the same scene into one image, and can provide more comprehensive and accurate description for the scene, thereby facilitating identification of events and objects. In recent years, this technology has been receiving more and more attention from researchers, and has made significant research progress.
Existing image fusion methods can be roughly classified into three categories, namely, a multiscale transform (MST) -based method, a Dictionary Learning (DL) -based method, and a deep learning-based method. Among the MST based methods, commonly used MSTs include wavelet transforms, dual tree complex transforms (DTCWT), shear transforms, curvelet transforms, contourlet transforms, and non-subsampled contourlet transforms (NSCT). Usually, the base of MST is fixed, sparseness is weak, and local information of an image cannot be represented adaptively sparsely. Different from the MST method, the Sparse Representation (SR) technology based on dictionary learning can effectively overcome the defects of the MST method and show good fusion performance. With the development of deep learning, image fusion based on deep learning is more and more concerned by people, and accordingly, some excellent fusion methods also appear. However, the above method has good performance only when the source image has a high resolution. If the input image is of low resolution, the resolution of the fusion result will also be low, which hinders the application of the fusion result.
In order to improve the resolution of the fused image, a common solution is to implement fusion and super-resolution reconstruction step by step. However, this method is likely to introduce the workpiece created in the first step into the next step, thereby degrading the visual quality of the final result.
Disclosure of Invention
The invention aims to provide a multi-source image fusion method based on discriminant dictionary learning and morphological component decomposition aiming at the defects and shortcomings in the prior art.
The technical scheme adopted by the invention is as follows: a multi-source image fusion method based on discriminant dictionary learning and morphological component decomposition comprises the following steps:
1) and constructing a training sample of dictionary learning. Selecting 8 HR training samples for training a high-resolution image discrimination dictionary pair, then performing down-sampling, and performing up-sampling by a bicubic interpolation method to restore the images to be the same as the high-resolution images as corresponding LR training images;
2) an initial dictionary is randomly generated. A new dictionary learning method is then proposed for decomposing the input image into low rank and sparse components. For synchronously realizing image fusion and super-resolution reconstruction, a pair of HR dictionaries and a pair of LR dictionaries are jointly trained, and a coefficient conversion matrix H between coding coefficients of high-resolution image blocks and low-resolution image blocks is trained, wherein the pair of HR dictionaries is D h,lAnd Dh,sOne pair of LR dictionaries is Dl,lAnd Dl,s;
3) Obtaining a low-rank dictionary D from the step 2)l,lAnd sparse dictionary Dl,sThen, taking the LR image in pairs, decomposing the LR image block YlObtaining a low-rank sparse coding coefficient Al,lAnd Al,s;
4) LR coding coefficient A obtained by step 3)l,s、Al,lAnd the conversion matrix H obtained in the step 2) adopts a method of maximum absolute value to construct the coding coefficient of the HR fusion imageAndand finally obtaining a high-resolution fusion image block.
Specifically, the step 2) includes the following steps:
step2.1 utilizes the constructed training sample learning dictionary, and the proposed discriminant dictionary learning model is as follows:
andlow rank dictionary and sparse dictionary that are HR images。XhIs a high resolution training image, XlIs a corresponding low resolution image. In addition, D is used to low resolution low rank, sparse dictionaryl,lAnd Dl,sIs shown in which M × K represents the matrix dimension of M rows and K columns, ε1、ε2、ε3And ε4A constant controlling the amplitude of each atom in the different dictionaries. Zl,lAnd Zl,sIs XlIn dictionary Dl,lAnd Dl,sCoding coefficient ofh,lAnd Zh,sAre each XhIn dictionary Dh,lAnd Dh,sThe coding coefficients of (1). H is Zl,iAnd Zh,i(i ═ l, s); i | · | purple windFIs an F norm operator; Ψ (H, Z)l,l,Zh,l,Zl,s,Zh,s) And Φ (D)l,l,Dh,l,Zl,l,Zh,l,Zl,s,Zh,s) Is a discriminant regular term to ensure the learning dictionary D l,l、Dh,l、Dl,sAnd Dh,sThe discrimination ability of (1).
For convenience of processing, the LR image is set to the same size as the HR image by the bicubic interpolation process. Then, the relationship between the encoding coefficients of the LR image and the HR image can be described as:
in the formula (2) < lambda >1,λ2And λ5Is a regularization parameter.
In the formula (2), the reaction mixture is,for the purpose of avoiding over-fitting,is a relational transformation term used to represent the relationship between sparse component coding coefficients, where | · |. includesFIs the F-norm operator. In the LR image and its corresponding HR image,for establishing Zh,lAnd Zl,lThe relationship between them. It is clear that the low rank components have strong linear correlation between them and that each element in the same coding coefficient vector has similar values. Based on this fact, an all-1 matrix is introducedAnd using regularization termsTo characterize the low rank component coded coefficients.
To improve the discrimination of the learning dictionary, the following regularization terms are defined:
in formula (3) < lambda >3And λ4Is a regularization parameter. I Dh,lZh,l||*And | | | Dl,lZl,l||*D for ensuring separation from an input image pairh,lZh,lAnd Dl,lZl,lIs low-rank, wherein | · |. non-woven phosphor*Is the kernel norm operator. After obtaining the low rank component, it can also pass Xi-Di,lZi,lObtaining a sparse component, wherein XiRepresenting input source data. Therefore, the objective function that discriminates the dictionary learning can be expressed as:
And (4) optimizing and solving the Step2.2 dictionary learning model.
Variable D to be solvedl,l、Dl,s、Dh,l、Dh,s、Zh,s、Zh,l、Zl,s、Zl,lAnd H, which are non-convex and difficult to solve directly. If one of the variables is solved while the other variables are fixed, each sub-problem is a convex function. Therefore, each variable in (4) is solved in turn by an alternating iterative method.
Step2.2.1 for ease of optimization, four variable matrices X were introducedh,s、Xh,l、Xl,sAnd Xl,lThat is, training the sparse component and the low-rank component of the input high-resolution image and the sparse component and the low-rank component of the low-resolution image, the optimization problem in equation (4) is converted into:
step2.2.2 update Xh,s、Xh,lAnd a coding coefficient matrix Zh,sAnd Zh,l. First, fixing other variables, updating Xh,sThen, equation (5) is converted into the following expression:
the above formula has the following closed solution form:
similarly, fix other variables, Xh,lThe solution form of (c) is as follows:
equation (8) can be solved efficiently by the SVT algorithm.
After X is updatedh,sAnd Xh,lThen, the coding coefficient matrix Z is further updatedh,sAnd Zh,l:
In the formula (9)Formula (9) belongs to1Minimizing the problem by using a Two-Step Iterative Shrinkage/Thresholding algorithm (Two-Step Iterative Shrinkage/Thresholding algorithm, TwinT)[57]To solve. For equation (10), the solution is as follows:
step2.2.3 updating X l,s,Xl,lAnd a coding coefficient matrix Zl,sAnd Zl,l. First, fixing other variables, optimizing Xl,sAnd Xl,lSolving expressions are respectively as follows:
Further, by fixing other variables, Z is obtainedl,sUpdating expression sum Zl,lThe closed solution is in the form:
Step2.2.4 updates the transformation matrix H. The other variable matrix is fixed, and the updated expression of H obtained by the formula (5) is as follows:
the formula (16) is F norm, and the following expression about H is obtained by calculating partial derivative of H:
the above equation can effectively solve H by Sylvester function of MATLAB.
Step2.2.5 updating dictionary pair Dh,s、Dh,lAnd Dl,s、Dl,l. Fixing other variables to be unchanged, solving dictionary variables by using a Lagrange dual method to obtain Dh,sAnd Dh,lThe update expression of (a) is as follows:
in the same way, Dl,sAnd Dl,lThe closed solution expression is as follows:
in the above four expressions, Λ1、Λ2、Λ3And Λ4Is a diagonal matrix of the corresponding optimal dual variables.
Specifically, the steps of step 3) are as follows:
from step 2), in order to ensure the successful separation of different components, a new image decomposition model is provided, and a source image is decomposed into a low-rank part and a sparse part.
Obtaining a low-rank dictionary D from Step3.1 l,lSparse dictionary Dl,sThen, taking paired LR images, collecting N image block data by using a sliding window with the size of (N multiplied by N), wherein the number of each image block is used as a column vector, and then forming a matrix by the column vectors. Decomposition of LR image blocks YlObtaining low rank, sparse coding coefficient Al,lAnd Al,s. The low rank, sparse decomposition model is as follows:
in the formula Al,l(i, j) is Al,lThe (i, j) th value of (a). If A is to bel,lAs a group, the value in each row of (a) is minimizedl,l||2,1Will be such that Al,lThe values in each row of (a) are the same. From the inequalities R (AB) ≦ min { R (A), R (B) } (R denotes the rank of the matrix), it is known that the best is obtainedMiniaturization | | | Al,l||2,1Can further ensure Dl,lAl,lLow rank and avoid D due to minimization | |l,lAl,l||*Causing some drawbacks.
The solution of the Step3.2 decomposition model can also adopt an alternative iteration method to solve the image decomposition model (22) and update Al,sAnd Al,l。
Step3.2.1 for ease of optimization, variable Y was introducedl,lAnd Yl,sI.e., the low rank component and the sparse component of the LR image block, equation (22) can be transformed into:
step3.2.2 updating Yl,sAnd Yl,l: first fix Yl,l,Al,lAnd Al,s,Yl,sThe update expression is as follows:
similarly, using the TwinT algorithm to solve for Yl,sL of1And (4) optimizing the norm.
Further, Y is fixedl,s,Al,lAnd Al,s,Yl,lThe update expression is as follows:
Step3.2.3 optimized sparse coding coefficient Al,sAnd Al,l. The update expression is as follows:
formula (26) is1And (3) solving the norm optimization problem by adopting a TwinT algorithm in the section. Formula (27) is2,1Norm optimization problem.
Specifically, the steps of step 4) are as follows:
assuming N LR images to be fused, the sparse dictionary D learned through step 2l,lLow rank dictionary Dl,sAnd a transformation matrix H; the decomposition model in the step3 is used for decomposing the image to be fused to obtain a coding coefficient A of low-rank sparse componentsl,lAnd Al,s. Order toAndrespectively representing low-rank, sparse coding coefficients of the LR input image. Then low rank, sparse fusion coding coefficients of HRAndis constructed by the following steps:
m in the formula (28) represents the number of input images, such thatA high resolution fused image block can be obtained
The invention has the beneficial effects that:
(1) the invention provides an effective discriminant dictionary learning model, which can learn a conversion dictionary between high and low resolutions and two pairs of discriminant dictionaries and jointly represent low-rank and sparse components of a low-resolution and high-resolution image pair.
(2) The conversion dictionary in the invention reveals the relationship between the coding coefficient of the low-resolution image and the corresponding high-resolution version image, and can improve the resolution of the low-resolution image after fusion, so that the image details are clearer.
(3) The invention realizes the fusion and the high-resolution reconstruction of the low-resolution images at the same time, and can obtain the fusion images with better vision and objective evaluation performance, thereby providing more comprehensive and accurate image description for the identification of events and objects.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is an HR source image in which the first row is a pair of infrared and visible light images (240X 320) of "street", respectively; the second row is the MR-T1/MR-T2 medical image pair (256 × 256), respectively; the third row is a "bookshelf" and "clock" multi-focal image pair (240 × 320);
FIG. 3 is an LR source image, in which (a) is an infrared and visible light image pair (120 × 160); (b) for a medical image pair (128 x 128); (c) for a pair of multifocal images (120 × 160);
FIG. 4 shows the fusion and high resolution results (240X 320) of infrared and visible images of "street", in which (a) - (d) represent the fusion and super resolution reconstruction result images of Ours, Li's, Zhu's (bicubic), Zhu's (SRSR), respectively;
FIG. 5 shows the fusion and high resolution results (256X 256) of MR-T1/MR-T2 medical images, and (a) to (d) respectively show the fusion and super resolution reconstruction result images of Ours, Li's, Zhu's (bicubic), Zhu's (SRSR);
Fig. 6 shows the fusion and high resolution results (240 × 320) of the multi-focus image of "bookshelf", and in the drawings, (a) to (d) respectively show the fusion and super-resolution reconstruction result images of Ours, Li's, Zhu's (bicubic), Zhu's (srsr).
Detailed Description
The invention is described in further detail below with reference to the figures and the specific embodiments.
Example 1: as shown in fig. 1 to 6, a multi-source image fusion method based on discriminative dictionary learning and morphological component decomposition includes the following steps:
1) and constructing a training sample for dictionary learning. Selecting 8 HR training samples for training a high-resolution image discrimination dictionary pair, then performing down-sampling, and performing up-sampling by a bicubic interpolation method to restore the images to be the same as the high-resolution images as corresponding LR training images;
2) an initial dictionary is randomly generated. A new dictionary learning method is then proposed for decomposing the input image into low rank and sparse components. For synchronously realizing image fusion and super-resolution reconstruction, a pair of HR dictionaries and a pair of LR dictionaries are jointly trained, and a coefficient conversion matrix H between coding coefficients of high-resolution image blocks and low-resolution image blocks is trained, wherein the pair of HR dictionaries is Dh,lAnd Dh,sOne pair of LR dictionaries is D l,lAnd Dl,s;
3) Obtaining a low-rank dictionary D from the step 2)l,lAnd sparse dictionary Dl,sThen, taking the LR image in pairs, decomposing the LR image block YlObtaining low rank, sparse coding coefficient Al,lAnd Al,s;
4) LR coding coefficient A obtained by step 3)l,s、Al,lAnd the conversion matrix H obtained in the step 2) adopts a method of maximum absolute value to construct the coding coefficient of the HR fusion imageAndand finally obtaining a high-resolution fusion image block.
Specifically, the step 2) includes the following steps:
step2.1 utilizes the constructed training sample learning dictionary, and the proposed discriminant dictionary learning model is as follows:
andare the low rank dictionary and sparse dictionary of the HR image. XhIs a high resolution training image, XlIs a corresponding low resolution image. In addition, D is used to low resolution low rank, sparse dictionaryl,lAnd Dl,sIs shown in which M × K represents the matrix dimension of M rows and K columns, ε1、ε2、ε3And ε4A constant controlling the amplitude of each atom in the different dictionaries. Zl,lAnd Zl,sIs XlIn dictionary Dl,lAnd Dl,sCoding coefficient ofh,lAnd Zh,sAre each XhIn dictionary Dh,lAnd Dh,sThe coding coefficients of (1). H is Zl,iAnd Zh,i(i ═ l, s); i | · | purple windFIs an F norm operator; Ψ (H, Z)l,l,Zh,l,Zl,s,Zh,s) And Φ (D)l,l,Dh,l,Zl,l,Zh,l,Zl,s,Zh,s) Is a discriminant regular term to ensure the learning dictionary Dl,l、Dh,l、Dl,sAnd Dh,sThe discrimination ability of (1).
For convenience of processing, the LR image is set to the same size as the HR image by the bicubic interpolation process. Then, the relationship between the encoding coefficients of the LR image and the HR image can be described as:
in the formula (2) < lambda >1,λ2And λ5Is a regularization parameter.
In the formula (2), the reaction mixture is,for the purpose of avoiding over-fitting,is a relational transformation term used to represent the relationship between sparse component coding coefficients, where | · |. includesFIs the F-norm operator. In the LR image and its corresponding HR image,for establishing Zh,lAnd Zl,lThe relationship between them. It is clear that the low rank components have strong linear correlation between them and that each element in the same coding coefficient vector has similar values. Based on this fact, an all-1 matrix is introducedAnd using regularization termsTo characterize the low rank component coded coefficients.
To improve the discrimination of the learning dictionary, the following regularization terms are defined:
in formula (3) < lambda >3And λ4Is a regularization parameter. I Dh,lZh,l||*And | | | Dl,lZl,l||*D for ensuring separation from an input image pairh,lZh,lAnd Dl,lZl,lIs low-rank, wherein | · |. non-woven phosphor*Is the kernel norm operator. After obtaining the low rank component, it can also pass Xi-Di,lZi,lObtaining a sparse component, wherein XiRepresenting input source data. Therefore, the objective function that discriminates the dictionary learning can be expressed as:
And (4) optimizing and solving the Step2.2 dictionary learning model.
Variable D to be solvedl,l、Dl,s、Dh,l、Dh,s、Zh,s、Zh,l、Zl,s、Zl,lAnd H, which are non-convex and difficult to solve directly. If one of the variables is solved while the other variables are fixed, each sub-problem is a convex function. Therefore, each variable in (4) is solved in turn by an alternating iterative method.
Step2.2.1 for ease of optimization, four variable matrices X were introducedh,s、Xh,l、Xl,sAnd Xl,lThat is, training the sparse component and the low-rank component of the input high-resolution image and the sparse component and the low-rank component of the low-resolution image, the optimization problem in equation (4) is converted into:
step2.2.2 update Xh,s,Xh,lAnd a coding coefficient matrix Zh,sAnd Zh,l. First, fixing other variables, updating Xh,sThen, equation (5) is converted into the following expression:
the above formula has the following closed solution form:
similarly, fix other variables, Xh,lThe solution form of (c) is as follows:
equation (8) can be solved efficiently by the SVT algorithm.
After X is updatedh,sAnd Xh,lThen, the coding coefficient matrix Z is further updatedh,sAnd Zh,l:
In the formula (9)Formula (9) belongs to1Minimizing the problem by using a Two-Step Iterative Shrinkage/Thresholding algorithm (Two-Step Iterative Shrinkage/Thresholding algorithm, TwinT)[57]To solve. For equation (10), the solution is as follows:
step2.2.3 updating X l,s、Xl,lAnd a coding coefficient matrix Zl,sAnd Zl,l. First, fixing other variables, optimizing Xl,sAnd Xl,lSolving expressions are respectively as follows:
Further, by fixing other variables, Z is obtainedl,sUpdating expression sum Zl,lThe closed solution is in the form:
Step2.2.4 updates the transformation matrix H. The other variable matrix is fixed, and the updated expression of H obtained by the formula (5) is as follows:
the formula (16) is F norm, and the following expression about H is obtained by calculating partial derivative of H:
the above equation can effectively solve H by Sylvester function of MATLAB.
Step2.2.5 updating dictionary pair Dh,s、Dh,lAnd Dl,s、Dl,l. Fixing other variables to be unchanged, solving dictionary variables by using a Lagrange dual method to obtain Dh,sAnd Dh,lThe update expression of (a) is as follows:
in the same way, Dl,sAnd Dl,lThe closed solution expression is as follows:
in the above four expressions, Λ1、Λ2、Λ3And Λ4Is a diagonal matrix of the corresponding optimal dual variables.
Specifically, the steps of step 3) are as follows:
from step 2), in order to ensure the successful separation of different components, a new image decomposition model is provided, and a source image is decomposed into a low-rank part and a sparse part.
Obtaining a low-rank dictionary D from Step3.1 l,lSparse dictionary Dl,sThen, a pair of LR images is taken, which is (n × n) largeThe small sliding window collects N image block data, each image block number is used as a column vector, and then the column vectors form a matrix. Decomposition of LR image blocks YlObtaining low rank, sparse coding coefficient Al,lAnd Al,s. The low rank, sparse decomposition model is as follows:
in the formula Al,l(i, j) is Al,lThe (i, j) th value of (a). If A is to bel,lAs a group, the value in each row of (a) is minimizedl,l||2,1Will be such that Al,lThe values in each row of (a) are the same. From the inequalities R (AB) ≦ min { R (A), R (B) } (R represents the rank of the matrix), it may be known to minimize | | | Al,l||2,1Can further ensure Dl,lAl,lLow rank and avoid D due to minimization | |l,lAl,lSome defects are caused.
The solution of the Step3.2 decomposition model can also adopt an alternative iteration method to solve the image decomposition model (21) and update Al,sAnd Al,l。
Step3.2.1 for ease of optimization, variable Y was introducedl,lAnd Yl,sI.e., the low rank component and the sparse component of the LR image block, equation (22) can be transformed into:
step3.2.2 updating Yl,sAnd Yl,l: first fix Yl,l,Al,lAnd Al,s,Yl,sThe update expression is as follows:
similarly, using the TwinT algorithm to solve for Yl,sL of1And (4) optimizing the norm.
Further, Y is fixedl,s、Al,lAnd Al,s、Yl,lThe update expression is as follows:
Step3.2.3 optimized sparse coding coefficient Al,sAnd Al,l. The update expression is as follows:
formula (26) is1And (3) solving the norm optimization problem by adopting a TwinT algorithm in the section. Formula (27) is2,1And (5) carrying out norm optimization.
Specifically, the steps of step 4) are as follows:
assuming N LR images to be fused, the sparse dictionary D learned through step 2l,lLow rank dictionary Dl,sAnd a transformation matrix H; the decomposition model in the step3 is used for decomposing the image to be fused to obtain a coding coefficient A of low-rank sparse componentsl,lAnd Al,s. Order toAndrespectively representing low-rank, sparse coding coefficients of the LR input image. Then low rank, sparse fusion coding coefficients of HRAndis constructed by the following steps:
m in the formula (28) represents the number of input images, such thatA high resolution fused image block can be obtained
The present invention will be described in detail with reference to specific examples.
In the dictionary learning part, 8 training samples are selected as HR training images for training images, then down sampling is carried out, and then up sampling is carried out by a bicubic interpolation method to restore the images with the same size as the high-resolution images as corresponding LR training images. And as a training sample, obtaining a required low-rank dictionary, a sparse dictionary and a conversion matrix by iterative updating according to a proposed dictionary learning algorithm. For the test image, a pair of IR and visible images of HR (as shown in the first row of FIG. 2), a pair of medical images of HR (as shown in the second row of FIG. 2) and a pair of multi-focus images of HR (as shown in the last row of FIG. 2) are selected, with the low resolution versions of these source images being shown in FIG. 3. In dictionary learning and image decomposition, lambda is involved in total 1,λ2,λ3,λ4And λ5And 5 regularization parameters and the maximum iteration number K, wherein six parameters are required to be set. According to experimental experience, setting lambdaiThe numerical values of (i ═ 1,2, …,5) are: 1, 0.01, 1.5, 0.01 and 0.00001; k is 10. In order to verify the effectiveness of the method, experiments are respectively carried out on medical images, infrared and visible light images and multi-focus image images.
In order to verify the superiority of the invention, the proposed method is compared with the latest image fusion and super-resolution methods at present. The Li method can realize image fusion and super-resolution at the same time, so that the method is one of comparison methods. However, due to the limitations of such methods, fusion results produced by excellent fusion algorithms such as the method of Zhu are super resolved. For super resolution, high resolution results are constructed using bicubic (bicubic) and sparse representation based super resolution (SRSR). Therefore, in the present invention, the super resolution results of the method of Zhu were compared with the obtained results.
In order to objectively and fairly evaluate the quality of the fusion result generated by different methods, the invention adopts four objective evaluation indexes to measure the quality of the fusion result besides comparing the fusion result in visual effect. These metrics include a Spatial Frequency (SF) based metric Q SFQuality-aware clustering (Q) methodQACCommon entropy of image information QENTAnd image mean gradient (gradient) QGD。QSFThe ratio of the SF errors is used to measure the quality of the fused image. Specifically, if QSFIf the value of (a) is lower than zero, it means that the active image information is lost in the fusion result; and if QSFA value greater than 0, and in the absence of any artifacts and noise, indicates that the details of the source image are enhanced. The higher the objective evaluation value of the other three indexes, the better the fusion quality.
Experiment 1: and (3) fusion and super-resolution reconstruction of the infrared and visible light images.
The first set of experiments was performed on the low resolution infrared and visible images of fig. 3(a) as test images. The fusion and super-resolution reconstruction results of the different methods are shown in fig. 4.
"Ours" in fig. 4 indicates the fusion and hyper-resolution reconstruction effect of the proposed method. As can be seen from fig. 4 and table 1 objective evaluation index data, the fusion results of Ours are superior to other methods in terms of visual perception and objective evaluation. The digital detail in the upper left corner of fig. 4 shows that the image after fusion reconstruction by the method of the present invention is clearer and brighter. The visual result of the method of the invention at the details proves to be better, i.e. the proposed method is very efficient and rational. In addition, three objective evaluation results of different methods are shown in table 1. From these data, it can be seen that objective evaluation leads to a conclusion that is consistent with subjective evaluation, which further verifies the superiority of the method of the invention.
TABLE 1 quantitative evaluation of different fusion and super-resolution reconstruction methods for infrared and visible light images of "streets
Experiment 2: fusion and super-resolution reconstruction of medical images
In a second experiment, a low resolution medical image pair such as that of fig. 3(b) was fused and super-resolved reconstructed. From these source images, it can be seen that the image information of the same scene obtained by different devices is complementary. The purpose of medical image fusion and super-resolution is to extract complementary information in a low-resolution source image, inject the complementary information into a fused image and improve the resolution of a fusion result. Fig. 5 shows the fusion and super-resolution reconstruction results of different methods. As can be seen from fig. 5(a) to 5(d), the proposed Ours are visually clearly superior to the Li method, and the step-wise operation method: zhu's (bicubic) and Zhu's (SRSR). In addition, it can be seen from the table that the proposed method excludes the index QENTSlightly lower than the other three results, and better than the results of the Li method and the zhu method after the ultra-separation treatment in other evaluation indexes.
TABLE 2 quantitative evaluation of the Performance of different fusion and super-resolution reconstruction methods for MR-T1/MR-T2 medical images
Experiment 3: fusion and super-resolution reconstruction of multi-focus images
In a third experiment, the image pair of fig. 3(c) was tested. In the focused images, each focused image is captured by the same sensor modality, but the sensor modes are different, and the focus areas are different, and the multi-focus image fusion is to obtain images of all focused objects and fuse the images into one image. The fusion and super-resolution results for the different fusion methods are given in fig. 6. Comparing fig. 6(a) with fig. 6(c) - (d), it can be found that the resolution of the alarm clock at the near position and the resolution of the bookshelf at the far position are both higher in the result image of the method provided by the present invention, which proves that the method sufficiently fuses the images with different focuses. And the proposed Ours method is superior to other methods in terms of image contrast. As can be seen in Table 3, in the proposed method, QSFAlso less than 0, but still loses less information than the other three methods. The results of the objective evaluation of the data in table 3 further demonstrate the superiority of the proposed method over other methods.
TABLE 3 quantitative evaluation of different fusion and super-resolution reconstruction method performances of "Bookshelf" multi-focus image
While the present invention has been described in detail with reference to the embodiments shown in the drawings, the present invention is not limited to the embodiments, and various changes can be made without departing from the spirit and scope of the present invention.
Claims (4)
1. A multi-source image fusion method based on discriminant dictionary learning and morphological component decomposition is characterized in that: the method comprises the following specific steps:
1) constructing a training sample for dictionary learning: selecting 8 HR training samples to train a high-resolution image discrimination dictionary pair, then performing down-sampling, and performing up-sampling by a bicubic interpolation method to restore the images to be the same as the high-resolution images as corresponding LR training images;
2) randomly generating an initial dictionary: a new dictionary learning method is provided for decomposing an input image into low-rank and sparse components, jointly training a pair of HR dictionaries, a pair of LR dictionaries and a coefficient conversion matrix H between coding coefficients of high-resolution and low-resolution image blocks for synchronously realizing image fusion and super-resolution reconstruction, wherein the pair of HR dictionaries is Dh,lAnd Dh,sOne pair of LR dictionaries is Dl,lAnd Dl,s;
3) Obtaining a low-rank dictionary D from step 2)l,lAnd sparse dictionary Dl,sThen, taking the LR image pair, decomposing the LR image block YlObtaining low rank, sparse coding coefficient Al,lAnd Al,s;
2. The multi-source image fusion method based on discriminative dictionary learning and morphological component decomposition according to claim 1, wherein the step 2) comprises the following steps:
step2.1 utilizes the constructed training sample to learn the dictionary, and the proposed discriminant dictionary learning model is as follows:
andlow rank dictionary and sparse dictionary, X, being HR imageshIs a high resolution training image, XlIs a corresponding low resolution image, and further, is a low resolution low rank, sparse dictionary for Dl,lAnd Dl,sIs shown in which M × K represents the matrix dimension of M rows and K columns, ε1、ε2、ε3And ε4Constants controlling the amplitude of each atom in different dictionaries, Zl,lAnd Zl,sIs XlIn dictionary Dl,lAnd Dl,sCoding coefficient ofh,lAnd Zh,sAre each XhIn dictionary Dh,lAnd Dh,sH is Zl,iAnd Zh,i(i ═ l, s); i | · | purple windFIs an F norm operator; Ψ (H, Z)l,l,Zh,l,Zl,s,Zh,s) And Φ (D)l,l,Dh,l,Zl,l,Zh,l,Zl,s,Zh,s) Is a discriminant regular term to ensure the learning dictionary Dl,l、Dh,l、Dl,sAnd Dh,sThe discrimination ability of (1);
the LR image is set to the same size as the HR image by the bicubic interpolation process, and then the relationship between the encoding coefficients of the LR image and the HR image is described as:
in the formula (2) < lambda >1,λ2And λ5Is a regularization parameter;
in the formula (2), the reaction mixture is,for the purpose of avoiding over-fitting, Is a relational transform term used to represent the relationship between sparse component coding coefficients, where | · | | | calvert |FFor the F norm operator, in the LR image and its corresponding HR image,for establishing Zh,lAnd Zl,lThe relationship between the low rank components is obviously strong in linear correlation, and each element in the same coding coefficient vector has similar value, based on the fact that all 1 matrixes are introducedAnd using regularization termsTo characterize the low rank component coding coefficients;
the following regularization term is defined:
in formula (3) < lambda >3And λ4Is the regularization parameter, | Dh,lZh,l||*And | | | Dl,lZl,l||*D for ensuring separation from an input image pairh,lZh,lAnd Dl,lZl,lIs low-rank, wherein | · |. non-woven phosphor*For the kernel norm operator, after the low rank component is obtained, pass Xi-Di,lZi,lObtaining a sparse component, wherein XiRepresenting input source data, and therefore, discriminating the objective function representing dictionary learning is expressed as:
optimized solution of Step2.2 dictionary learning model
Variable D to be solvedl,l、Dl,s、Dh,l、Dh,s、Zh,s、Zh,l、Zl,s、Zl,lAnd H, which are non-convex and difficult to solve directly, if one of the variables is solved while the other variables are fixed, each sub-problem is a convex function, and therefore each variable in (4) is solved in turn by an alternating iterative method;
step2.2.1 introduces four variable matrices X h,s、Xh,l、Xl,sAnd Xl,lThat is, training the sparse component and the low-rank component of the input high-resolution image and the sparse component and the low-rank component of the low-resolution image, the optimization problem in equation (4) is converted into:
step2.2.2 update Xh,s、Xh,lAnd a coding coefficient matrix Zh,sAnd Zh,l: first, fixing other variables, updating Xh,sThen, equation (5) is converted into the following expression:
the above formula has the following closed solution form:
similarly, fix other variables, Xh,lThe solution form of (c) is as follows:
the formula (8) can be solved effectively through SVT algorithm;
after X is updatedh,sAnd Xh,lThen, the coding coefficient matrix Z is further updatedh,sAnd Zh,l:
In the formula (9)Formula (9) belongs to1Minimum sizeSolving the problem by using a two-step iterative shrinkage/threshold algorithm, wherein for the formula (10), the solving result is as follows:
step2.2.3 updating Xl,s,Xl,lAnd a coding coefficient matrix Zl,sAnd Zl,lFirst, fixing other variables, optimizing Xl,sAnd Xl,lSolving the expressions respectively as follows:
further, by fixing other variables, Z is obtainedl,sUpdating expression sum Zl,lThe closed solution is in the form:
step2.2.4 updates the transformation matrix H: the other variable matrix is fixed, and the updated expression of H obtained by the formula (5) is as follows:
The formula (16) is F norm, and the following expression about H is obtained by calculating partial derivative of H:
according to the formula, H can be effectively solved through a Sylvester function of MATLAB;
step2.2.5 updating dictionary pair Dh,s、Dh,lAnd Dl,s、Dl,lFixing other variables to be unchanged, solving dictionary variables by using a Lagrange dual method to obtain Dh,sAnd Dh,lThe update expression of (a) is as follows:
in the same way, Dl,sAnd Dl,lThe closed solution expression is as follows:
in the above four expressions, Λ1、Λ2、Λ3And Λ4Is the most correspondingDiagonal matrix of the dual-superior variable.
3. The multi-source image fusion method based on the discriminant dictionary learning and morphological component decomposition as claimed in claim 2, wherein the step 3) comprises the steps of:
obtaining a low rank dictionary D from Step3.1l,lSparse dictionary Dl,sThen, taking paired LR images, collecting N image block data by a sliding window with the size of (N multiplied by N), taking the number of each image block as a column vector, forming the column vectors into a matrix, decomposing an LR image block YlObtaining low rank, sparse coding coefficient Al,lAnd Al,sThe low rank, sparse decomposition model is as follows:
in the formula Al,l(i, j) is Al,lIf A is equal to the (i, j) th value of (a)l,lAs a group, the value in each row of (a) is minimizedl,l||2,1Will be such that Al,lIs different from the equation R (AB) ≦ min { R (A), R (B) }, R represents the rank of the matrix, it can be known to minimize | | A l,l||2,1Can further ensure Dl,lAl,lLow rank of (c), and avoid D due to minimization | | |l, lAl,l||*Some defects are caused;
the solution of Step3.2 decomposition model is carried out by adopting an alternative iteration method for the image decomposition model (22) and updating Al,sAnd Al,l;
Step3.2.1 introduction of variable Yl,lAnd Yl,sI.e., the low rank component and the sparse component of the LR image block, equation (22) can be transformed into:
Step3.2.2 update Yl,sAnd Yl,l: first fix Yl,l,Al,lAnd Al,s,Yl,sThe update expression is as follows:
similarly, using the TwinT algorithm to solve for Yl,sL of1A norm optimization problem;
further, Y is fixedl,s,Al,lAnd Al,s,Yl,lThe update expression is as follows:
Step3.2.3 optimized sparse coding coefficient Al,sAnd Al,lThe update expression is as follows:
formula (26) is1The norm optimization problem is solved by adopting a TwinT algorithm, and the equation (27) is l2,1Norm optimization problem.
4. The multi-source image fusion method based on the discriminant dictionary learning and morphological component decomposition as claimed in claim 3, wherein the step 4) comprises the steps of:
let us assume that N treats the fused LR image, learn through step 2To sparse dictionary Dl,lLow rank dictionary Dl,sAnd a transformation matrix H; the decomposition model in the step3 is used for decomposing the image to be fused to obtain a coding coefficient A of low-rank sparse components l,lAnd Al,sLet us orderAndrespectively representing low-rank and sparse coding coefficients of LR input image, and low-rank and sparse fusion coding coefficients of HRAndis constructed by the following steps:
m in the formula (28) represents the number of input images, such that A high resolution fused image block can be obtained
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010425926.0A CN111784572B (en) | 2020-05-19 | 2020-05-19 | Image fusion and super-resolution joint implementation method based on discriminant dictionary learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010425926.0A CN111784572B (en) | 2020-05-19 | 2020-05-19 | Image fusion and super-resolution joint implementation method based on discriminant dictionary learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111784572A CN111784572A (en) | 2020-10-16 |
CN111784572B true CN111784572B (en) | 2022-06-28 |
Family
ID=72754253
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010425926.0A Active CN111784572B (en) | 2020-05-19 | 2020-05-19 | Image fusion and super-resolution joint implementation method based on discriminant dictionary learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111784572B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112561842B (en) * | 2020-12-07 | 2022-12-09 | 昆明理工大学 | Multi-source damaged image fusion and recovery combined implementation method based on dictionary learning |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107563968A (en) * | 2017-07-26 | 2018-01-09 | 昆明理工大学 | A kind of method based on the group medicine image co-registration denoising for differentiating dictionary learning |
CN107977949A (en) * | 2017-07-26 | 2018-05-01 | 昆明理工大学 | A kind of method improved based on projection dictionary to the Medical image fusion quality of study |
CN108985320A (en) * | 2018-05-31 | 2018-12-11 | 昆明理工大学 | Based on the multisource image anastomosing method for differentiating that dictionary learning and anatomic element decompose |
CN109410157A (en) * | 2018-06-19 | 2019-03-01 | 昆明理工大学 | The image interfusion method with PCNN is decomposed based on low-rank sparse |
CN110706156A (en) * | 2019-09-16 | 2020-01-17 | 昆明理工大学 | Image fusion and super-resolution reconstruction combined implementation method based on multi-component analysis and residual compensation |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10409888B2 (en) * | 2017-06-02 | 2019-09-10 | Mitsubishi Electric Research Laboratories, Inc. | Online convolutional dictionary learning |
-
2020
- 2020-05-19 CN CN202010425926.0A patent/CN111784572B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107563968A (en) * | 2017-07-26 | 2018-01-09 | 昆明理工大学 | A kind of method based on the group medicine image co-registration denoising for differentiating dictionary learning |
CN107977949A (en) * | 2017-07-26 | 2018-05-01 | 昆明理工大学 | A kind of method improved based on projection dictionary to the Medical image fusion quality of study |
CN108985320A (en) * | 2018-05-31 | 2018-12-11 | 昆明理工大学 | Based on the multisource image anastomosing method for differentiating that dictionary learning and anatomic element decompose |
CN109410157A (en) * | 2018-06-19 | 2019-03-01 | 昆明理工大学 | The image interfusion method with PCNN is decomposed based on low-rank sparse |
CN110706156A (en) * | 2019-09-16 | 2020-01-17 | 昆明理工大学 | Image fusion and super-resolution reconstruction combined implementation method based on multi-component analysis and residual compensation |
Non-Patent Citations (4)
Title |
---|
"Analysis-synthesis dictionary pair learning and patch saliency measure for image fusion";Yafei Zhang等;《Signal Processing》;20200229;第167卷;1-13 * |
"Super-resolution images fusion via compressed sensing and low-rank matrix decomposition";Kan Ren等;《Infrared Physics & Technology》;20150131;第68卷;61-68 * |
"基于判别低秩稀疏字典学习的医学图像融合质量改善算法研究";和晓歌;《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》;20190115;I138-3746 * |
"基于判别字典学习与形态成分分解的多源图像融合";王一棠 等;《光学技术》;20190131;第45卷(第1期);63-69 * |
Also Published As
Publication number | Publication date |
---|---|
CN111784572A (en) | 2020-10-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Yin et al. | Simultaneous image fusion and super-resolution using sparse representation | |
CN103854262B (en) | Medical image denoising method based on documents structured Cluster with sparse dictionary study | |
CN106485656B (en) | A kind of method of image super-resolution reconstruct | |
CN107301630B (en) | CS-MRI image reconstruction method based on ordering structure group non-convex constraint | |
Chen et al. | A novel medical image fusion method based on rolling guidance filtering | |
CN106920214B (en) | Super-resolution reconstruction method for space target image | |
CN109522971B (en) | CS-MRI image reconstruction method based on classified image block sparse representation | |
CN105046672A (en) | Method for image super-resolution reconstruction | |
CN110047058A (en) | A kind of image interfusion method based on residual pyramid | |
CN106097253B (en) | A kind of single image super resolution ratio reconstruction method based on block rotation and clarity | |
CN104574456B (en) | A kind of super lack sampling K data imaging method of magnetic resonance based on figure regularization sparse coding | |
CN106251320A (en) | Remote sensing image fusion method based on joint sparse Yu structure dictionary | |
CN111487573B (en) | Enhanced residual error cascade network model for magnetic resonance undersampling imaging | |
CN105654425A (en) | Single-image super-resolution reconstruction method applied to medical X-ray image | |
CN105590296B (en) | A kind of single-frame images Super-Resolution method based on doubledictionary study | |
CN110570387A (en) | image fusion method based on feature level Copula model similarity | |
CN111784572B (en) | Image fusion and super-resolution joint implementation method based on discriminant dictionary learning | |
CN105931181B (en) | Super resolution image reconstruction method and system based on non-coupled mapping relations | |
CN104200439B (en) | Image super-resolution method based on adaptive filtering and regularization constraint | |
CN108596866B (en) | Medical image fusion method based on combination of sparse low-rank decomposition and visual saliency | |
CN111242873A (en) | Image denoising method based on sparse representation | |
CN109886869A (en) | A kind of unreal structure method of face of the non-linear expansion based on contextual information | |
CN110706156B (en) | Image fusion and super-resolution reconstruction combined implementation method based on multi-component analysis and residual compensation | |
CN117333365A (en) | Image super-resolution method based on hybrid transducer super-resolution network | |
CN107993205A (en) | A kind of MRI image reconstructing method based on study dictionary with the constraint of non-convex norm minimum |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |