CN108985320A - Based on the multisource image anastomosing method for differentiating that dictionary learning and anatomic element decompose - Google Patents

Abstract

The invention proposes a kind of based on the multisource image anastomosing method for differentiating that dictionary learning and anatomic element decompose.For that can separate the cartoon of different shape structure in source images-texture ingredient, we are converted into the resolution problem of image the classification problem of image, and devise cartoon texture discrimination dictionary learning model.In view of picture breakdown is not only related with dictionary, the also fact related with the strategy of decomposition devises a kind of new picture breakdown model.In the model, texture ingredient regards the noise being superimposed upon on source images cartoon ingredient as, and introduces the consistency regular terms of non-local mean similitude, to constrain the solution space of sparse coding coefficient.Finally, according to the code coefficient l of tie element₁Norm value maximum chooses the code coefficient of blending image.The result shows that the present invention has preferably fusion performance no matter in visual effect or in objective indicator.

Description

Multi-source image fusion method based on discriminant dictionary learning and morphological component decomposition

Technical Field

The invention relates to a multi-source image fusion method based on discriminative dictionary learning and morphological component decomposition, and belongs to the technical field of image fusion data processing.

Background

Due to the unity of the image information obtained by different sensors, accurate description of the target object is difficult to achieve. To address this problem, image fusion techniques may be used to integrate image information from different sensors about the same scene to generate a description of the scene. Which cannot be obtained from single source image information. The technology can effectively integrate the complementarity of image information obtained by different sensors and provide a more accurate description for an observed object, so the technology is successfully applied to the fields of medical imaging, machine vision, remote sensing, safety monitoring and the like.

In recent years, image fusion techniques have received wide attention from researchers, and many effective fusion methods have been proposed. Among these methods, the fusion method based on multi-scale variation is the most representative. Among the multi-scale Transform methods, the Discrete Wavelet Transform (DWT) is the most commonly used. However, the DWT has no translation invariance and is easy to introduce some false information into the fusion result. In addition, DWT can only represent high frequency detail information of an image in three directions, and cannot achieve effective representation of information such as image edges, contours, curves, and the like. For this reason, some multi-scale geometric analysis tools with excellent performance have been proposed, including Ridgelet, Curvelet, Contourlet, non-sampled Contourlet Transform (NSCT), and the like. Among these methods, NSCT is widely used in image fusion and exhibits excellent fusion performance because it has not only multi-directionality, anisotropy, but also translational invariance. However, the image fusion method based on multi-scale transformation has weak robustness to image registration errors. In addition, the basis used by the multi-scale analysis method for expressing different structural features of the image is fixed and has no self-adaptability, and accurate representation of complex structures is difficult to realize.

Compared with the traditional fusion method based on multi-scale decomposition, image fusion based on sparse representation has received extensive attention of researchers because the image fusion can realize more effective representation of the content of the source image. In such methods, learning and construction of an overcomplete dictionary are one of the important factors affecting the quality of the final fused image. In general, the construction of an overcomplete dictionary may be accomplished in two ways. One is generated by adopting an analytic method, and the other is generated by adopting a learning mode. The dictionary configured by the analysis method is commonly referred to as a DCT dictionary, a Wavelet dictionary, a Curvelet dictionary, or the like. Such dictionaries often cannot adaptively express complex structural information in natural images. In contrast, a dictionary obtained by learning has a stronger expressive power according to a set of training samples. Wherein, the most representative dictionary learning method is a K-SVD algorithm. The method is widely applied to image fusion based on sparse representation.

In recent years, in order to learn a more compact dictionary while ensuring dictionary expression ability, researchers are concerned with learning, and a series of image fusion and recovery methods based on dictionary learning are proposed. However, in these methods, different components of the image are represented by a dictionary. However, since different components of an image have different morphological characteristics, it is difficult for a dictionary to achieve efficient representation of the different components of the image. To solve this problem, a dictionary with a larger size is required, which increases the calculation amount of the algorithm and reduces the efficiency of the algorithm. To overcome this problem, researchers have proposed an image fusion method based on morphological component analysis. In this method, the texture and cartoon components of an image are expressed using DCT and Curvelet dictionaries, respectively. However, since such a dictionary is constructed by parsing, it is not adaptive enough to effectively depict very complex image structure information. More importantly, the decomposition and representation of different components of the image are not only related to the performance of the dictionary, but also related to the decomposition model of the image. However, the important factor is not considered in the traditional image fusion method based on sparse representation and dictionary learning.

Disclosure of Invention

The invention aims to provide a multi-source image fusion technical scheme 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 a multi-source image fusion method based on discriminant dictionary learning and morphological component decomposition, which comprises the following steps:

step 1, firstly, collecting training sample data and decomposing the training sample data to obtain cartoon training data and texture training data, and constructing a group of diversified training samples including character images, medical images, food images and the like, so as to ensure that a learnt dictionary has better discriminative performance. Collecting more than one gray level image from the internet as a training sample, and then collecting data of the training sample in the form of sliding windows, wherein each window (n multiplied by n) collects data as a column vector (n multiplied by n)²X 1), n is the size of a sliding window, the collected data are decomposed through MCA algorithm to obtain cartoon training data and texture training data, and all the collected cartoon training data and texture training data are two n²A matrix of dimensions.

Step 2, learning cartoon training data and texture training data respectively through a K-SVD algorithm to obtain an initial cartoon dictionary and an initial texture dictionary; training an initial cartoon dictionary and an initial texture dictionary by adopting a recognition dictionary learning model to obtain a cartoon dictionary D_cAnd texture dictionary D_t；

The distinguishing dictionary learning model provided by the invention is as follows:

in which Y is ∈ R^m×nData collected for a sliding window is used as a matrix formed by column vectors, R is a spatial domain, m is the dimension of a vector space, n is the number of image blocks, D_t、D_cRespectively representing texture dictionaries and cartoon dictionaries, A_t＝[α_1，t，α_2，t，…，α_L，t]representing the corresponding texture sparse coding coefficient of the texture training data, wherein α_l，tFor the sparse coding coefficient corresponding to the texture component of the ith image block, A_c＝[α_1，c，α_2，c，…，α_L，c]representing the cartoon sparse coding coefficient corresponding to the cartoon training data, wherein α_l，cA matrix D for the sparse coding coefficient corresponding to the cartoon component of the ith image block_cA_cFor cartoon components separated from Y, matrix D_tA_tT is the transpose of the matrix, which is the texture component separated from Y,as gradient operator, λ₁，λ₂To balance the parameters, | ·| luminance_FIs F norm operator, | ·| non-woven phosphor₁Is 1₁Norm operator, | ·| tory²Is the norm square operator.

Step 2.1 optimization solution of dictionary learning model, variable D to be solved_t、D_c、A_t、A_c. When the other variables are fixed, the optimization problem (1) is convex. Therefore, we can use an alternating iterative approach to solve problem (1).

Step 2.1.1 first, we solve the optimal coding coefficient A_tAnd A_c. At this time, solve for A_tCan be written as equation (2)

Equation (2) is typical for₁The norm optimization problem can be solved by using an iterative shrinkage algorithm.

Solving for A_cThe objective function of (d) can be written as:

for A_cWe first introduce two auxiliary variablesAndso thatAndsubstituting the auxiliary variable into (3) yields:

thenAnd A_cThis can be achieved by solving the minimization problems (4), (5) and (6), respectively, it is clear that the optimization problems (4) and (6) can be solved by iterative shrinkage algorithms, while equation (5) can be solved directly by gradient descent methods.

Step 2.1.2 updating the sparse coding coefficient A_tAnd A_cThen, we can solve D by solving the optimization problem (7)_t：

If orderThe optimization problem (7) can be written as:

the problem is a standard least squares problem and has an analytical solution of the form:

step 2.1.3 similarly, we can fix D_t、A_tAnd A_cTo solve for D_c，D_cThe objective function of (a) is:

to facilitate the solution, we introduce the auxiliary variables Z and h such that Z ═ D_cA_c，At this point, the optimization problem (10) can be rewritten:

from this, the optimal X can be solved_cAnd the objective functions of the auxiliary variable g are:

and

optimization problem (1)2) Can be solved by a gradient descent method, and the optimization problem (13) is a standard l₁The norm optimization problem can be solved by adopting an iterative shrinkage algorithm.

Similarly, solve the optimal dictionary D_cThe objective function of (d) can be written as:

since the above problem is described by the F-norm optimization problem, we have the following closed form solution:

wherein,

all the solving processes need to be iteratively updated to obtain an optimal solution, wherein the two dictionaries input in the first iteration are an initial cartoon dictionary and an initial texture dictionary which are obtained through K-SVD algorithm learning, and a cartoon dictionary D is obtained through a formula (9)_tAnd then substituting the variables into equations (10) - (15) to solve other variables, setting the introduced auxiliary variables to be 0, performing iteration for the second time, wherein all the variables are data obtained after the first iteration is updated, and repeating the iteration and updating in the same way.

Step 3, preprocessing the image to be fused, firstly adding white Gaussian noise, then collecting the data of the image to be fused in a sliding window mode, and collecting the data in each window (n multiplied by n) as a column vector (n²X 1), n is the size of the sliding window, and the size is decomposed by MCA algorithm to obtain a cartoon part and a texture part which are two n²The matrix (16) is used for solving the source image into a cartoon part and a texture part, and is calculated by the following formula:

wherein X ∈ R^m×nData collected for sliding windows as a matrix of column vectors, R being the spatial domain, D_t、D_cRespectively representing the texture dictionary and the cartoon dictionary learned in step 1, A_t＝[α_1，t，α_2，t，…，α_L，t]represents the texture sparse coding coefficient corresponding to the texture sample data, where α_l，tFor the sparse coding coefficient corresponding to the texture component of the ith image block, A_c＝[α_1，c，α_2，c，…，α_L，c]represents the cartoon sparse coding coefficient corresponding to the cartoon sample data, wherein α_l，csparse coding coefficients, eta corresponding to cartoon components of the ith image block₁，η₂To balance the parameters, | ·| luminance_FIs F norm operator, | ·| non-woven phosphor₁Is 1₁Norm operator, | ·| tory²In the form of the square operator of the norm,for sparsely encoding coefficients alpha_cIs estimated value ofThe calculation formula of the formed matrix is as follows:

wherein omega_ito code the coefficient alpha_i，jCorresponding to the local area where the image block is located. According to the idea of the non-local mean method, the calculation formula is as follows:

wherein alpha is_c，iCoding coefficients for texture components of the ith image block,is the local region omega where the texture components of the i image blocks are located_iH is a normalization factor, and H is a preset scale size.

And 3.1, solving the decomposition model, namely solving the image decomposition model (16) by adopting an alternative iteration method. Thus A is_t，A_cCan be obtained by solving the following optimization problem respectively:

obviously, the minimization problem (19) can be solved by iterative shrinkage algorithms, while the minimization problem (20), which is transformed and solved with alternative algorithms, is solved.

All the above solving processes need to be iteratively updated to obtain an optimal solution.

Step 4, reconstructing the image to be fused

Assuming that N multi-source images with fusion are provided, the cartoon dictionary D learned through the step 2_cAnd texture dictionary D_tAnd step 3, decomposing the image to be fused to obtain a sparse coding coefficient A of the cartoon-texture component by using the decomposition model_cAnd A_t. Is provided withFor the coding coefficient of the l image block fusing the texture and cartoon components of the image, l is adopted₁The norm maximum principle is selectedSelecting coding coefficients for fusing different components of the image, wherein the fusion scheme is as follows:

wherein,a l-th column, l ═ 1, 2, a.Coding coefficient A representing texture component of the r-th image to be fused_tThe ith column of (1) is a column vector, r is {1, 2., N }, N is the number of images to be fused, and K is the number of image blocks;

wherein,a column l, l ═ 1, 2, K, representing the coding coefficients of the cartoon components of the fused image,coding coefficient A of cartoon component representing s-th image to be fused_cIs a column vector, s ═ 1, 2,.., N, thenN is the number of images to be fused, and K is the number of image blocks;

is obtained byAndthen, assuming that there are K image blocks in total,encoding coefficients representing texture components of the fused image,representing the coding coefficients of the cartoon components of the fused image, the fused cartoon-texture components can be represented asAndthe matrix formed by the block vectors of the fused image can thus be represented asRearranging the result into image blocks to obtain the final fusion image.

The principle of the invention is as follows: in the method, the decomposition problem of the image is regarded as the classification problem of the image, and an effective dictionary learning model is designed. In the model, not only the relation between different component dictionaries but also the morphological characteristics of cartoon components are considered. In addition, in order to effectively separate different components of the image, an effective image decomposition model is designed based on the learned dictionary. In order to improve the decomposition performance of a model, texture components of an image are regarded as oscillation noise superposed on cartoon components of the image, and similarity based on non-local image blocks is introduced to be used as a regular term of sparse coding of the cartoon components. Finally, using l₁And selecting coding coefficients of different components of the fused image for fusion processing according to the principle of maximum norm to form a sparse coding coefficient of the fused image.

The invention has the beneficial effects that:

1. the method not only considers the irrelevance between dictionaries of different components, but also considers the weak expression characteristics of the dictionaries of different components, and introduces the gradient minimization constraint as the regular term of the cartoon components, so that the cartoon dictionary has strong expression capability.

2. The invention not only considers the functions of different component dictionaries in image decomposition, but also considers the influence of the design of an image decomposition model on the decomposition result, thereby obtaining more satisfactory fusion effect.

3. Compared with other methods, the image fusion method provided by the invention has the advantage that the fusion performance is obviously improved.

Drawings

FIG. 1 is a flow chart of the present invention;

FIGS. 2(a) and 2(b) are multi-modality medical images to be fused in example 1;

FIGS. 3(a) - (f) are the medical image fusion effects obtained using different methods in example 1;

FIGS. 4(a) and 4(b) are multi-focus images to be fused in example 2;

FIGS. 5(a) - (f) are multi-focus image fusion effects obtained using different methods in example 2;

FIGS. 6(a) and 6(b) are the infrared and visible images to be fused in example 3;

fig. 7(a) - (f) are the infrared and visible light image fusion effects obtained using different methods in example 3.

Detailed Description

The invention is described in further detail below with reference to the figures and the specific embodiments.

Several groups of images are fused according to the specific scheme in the specification and the flow shown in FIG. 1, in whichIn the process of training the dictionary, 8 multi-source images are collected and used as training samples, and the required cartoon dictionary and texture dictionary are obtained through iterative updating according to the provided dictionary learning algorithm. And obtaining the cartoon part and the texture part of the source image through the proposed decomposition algorithm of the image. In dictionary learning and image decomposition, lambda and lambda are involved₁、λ₂and η, η₁、η₂six parameters need to be set, and according to experimental experience, the lambda and η are set to be 0.01, and the lambda is set to be₁、λ₂and η₁、η₂Set to 0.005. To verify the validity of the method, we performed experiments on multi-focus images, medical images, infrared and visible light images, respectively. In order to objectively and fairly evaluate the quality of the fusion result produced by different methods, a plurality of objective evaluation indexes are used for measuring the quality of the fusion result in addition to the comparison in visual effect. Among these objective evaluation indexes, we use normalized mutual information Q_MIEvaluation index Q based on linear correlation information entropy_NCIEAnd evaluation index Q based on phase consistency_P. Wherein Q is_MICan be used to measure how much information in the source image is transferred to the fused image; q_NCIEThe evaluation of the quality of the fusion image is realized by measuring the correlation between the fusion image and a source image; q_PWhich is used to measure the degree to which salient features in the source image remain in the fused image. The larger the value of these evaluation indexes, the better the quality of the fusion result.

Example 1: medical image fusion

In the first set of experiments, we first performed a fusion experiment on a set of multi-modal medical images as shown in fig. 2(a) and (b). In which FIG. 2(a) is an MR-T1 image and FIG. 2(b) is an MR-T2 image. As can be seen from fig. 2(a) and (b), due to the difference of the weights, the MR-T1 and MR-T2 images contain a great deal of complementary information, and if the information can be integrated to produce a fused image, the fused image is very beneficial for the diagnosis and treatment of the disease condition by the doctor and the subsequent image processing, such as image classification, segmentation and target identification.

FIGS. 3(a) - (f) show the fusion results of NSCT, NSCT-SR, Kim's, Zhu-KSVD, ASR, and the methods presented herein, in that order. It can thus be seen that different fusion methods have different properties in preserving image edge detail information. The NSCT-based fusion method can effectively retain complementary information of the source image, but has slightly poorer capability of retaining the detailed information of the source image than that of the Zhu-KSVD method. The method proposed by Kim fuses the detailed information of the result more vaguely. In contrast, the Zhu-KSVD method and the ASR method can effectively protect edge detail information of a source image, but cannot well protect the contrast of the image. This is very disadvantageous for medical images with high quality requirements, and also for subsequent image processing and recognition tasks. On the contrary, through comparison, it can be found that the fusion method provided by the present disclosure not only can effectively protect edge detail information of the source image, but also can maintain the contrast of the source image, which is mainly benefited from the fact that the method of loop iteration is adopted to solve sparse coding coefficients of different components. Meanwhile, artificial false information is not introduced in the fusion process, so that the visual effect of the result generated by the method is better. In addition, three objective evaluation results of different methods are shown in table 1. From these data we can see that objective evaluation leads to a conclusion that is consistent with subjective evaluation. This further demonstrates the superiority of the methods herein and over conventional methods.

TABLE 1 medical image fusion Performance comparison of different fusion methods

Example 2: multi-focus image fusion

In a second set of experiments, we performed fusion experiments on a set of multi-focused images as shown in fig. 4(a) and (b). As can be seen from fig. 4(a) and (b), when the lens of the camera is focused on a certain object, the object can be clearly imaged, whereas the object imaging away from the focal plane is blurred. However, in reality, for some computer vision tasks or image processing tasks such as object segmentation, image classification, object recognition, etc., it is necessary to obtain an image with clear objects. This problem can be solved by multi-focus image fusion. The method proposed herein can be used not only to solve the fusion problem of medical images, but also to solve the fusion of multi-focus images.

FIGS. 5(a) - (f) show visual effect comparisons of NSCT, NSCT-SR, Kim's, Zhu-KSVD, ASR and fusion results of the methods herein. Therefore, all the methods can effectively extract the information of the clear target object in the source image and reserve the information into the fusion image. The fusion solution generated by different methods is locally amplified, and the information contained in the amplified region shows that the NSCT, NSCT-SR, Kim and Zhu-KSVD-based method can effectively retain the edge detail information of the image to obtain the fusion result similar to the method. The ASR method blurs part of the edge information in the focus area while preserving the edge detail information of the image. Although it is difficult to visually determine the evaluation results for NSCT, NSCT-SR, Kim, Zhu-KSVD, the objective evaluation results shown in Table 2 reflect the effectiveness and superiority of the algorithm in comparison with the conventional methods.

TABLE 2 comparison of fusion performance of multiple focus images for different fusion methods

Example 3: infrared and visible image fusion

In the third set of experiments we performed fusion experiments using different methods on the infrared and visible images shown in fig. 6(a) and (b). Fig. 6(a) shows an infrared image, and fig. 6(b) shows a visible light image. As seen from the source images, the visible light image can clearly reflect the background detail information of the scene, but cannot clearly image the thermal target () such as a pedestrian and a vehicle); in contrast, infrared images can clearly reflect hot objects (e.g., pedestrians, vehicles), but cannot clearly image backgrounds and the like without higher temperatures. In order to obtain an image with clear background and hot targets, the method plays an important role in target tracking, identification, segmentation and detection.

Fused images obtained by different methods are shown in fig. 7(a) - (f). Wherein FIGS. 7(a) and (b) are fusion results of NSCT and NSCT-SR methods, respectively; FIGS. 7(c) - (f) are the results of the Kim's method, the Zhu-KSVD method, the ASR method, and the algorithm herein, respectively. From these images, it can be seen that all the comparison methods can effectively retain the thermal target and background information of the source image. However, as can be seen from the locally enlarged region, the different methods exhibit different fusion properties. Simple NSCT-based fusion and NSCT-SR-based fusion achieve a relatively similar fusion effect, while Kim's fusion, Zhu-KSVD, and ASR fusion methods are effective in retaining background information in visible images (FIG. 6(b), but are less effective in retaining brightness and thermal target information in infrared images (FIG. 6 (a)), as shown in FIG. 7 (f)). Overall, the method proposed herein can effectively retain not only background information in visible light images, but also target objects in infrared images, and at the same time, can maintain the contrast of source images, thus having better visual effect. Table 3 shows the objective evaluation results of different fusion methods when fusing the infrared and visible light images as shown in fig. 6(a) and (b). From the data presented in table 3, we can see that the objective evaluation index used herein can be basically consistent with the visual effect evaluation. The fusion method has better visual effect, and the objective evaluation result is better on the whole. This also reflects that the objective assessment index chosen here is reasonably feasible. The data in table 3 show that the method proposed herein has better fusion performance.

TABLE 3 comparison of Infrared and visible image fusion Performance for different fusion methods

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 of the present invention within the knowledge of those skilled in the art.

Claims (6)

1. The multi-source image fusion method based on the discriminant dictionary learning and morphological component decomposition is characterized by comprising the following steps of:

(1) firstly, collecting training sample data and decomposing to obtain cartoon training data and texture training data;

(2) learning cartoon training data and texture training data respectively through a K-SVD algorithm to obtain an initial cartoon dictionary and an initial texture dictionary;

(3) training an initial cartoon dictionary and an initial texture dictionary by adopting a recognition dictionary learning model to obtain a cartoon dictionary D_cHarmonic waveDictionary D_t；

(4) Based on cartoon dictionary D_cAnd texture dictionary D_tAnd forming an image decomposition model, decomposing the image to be fused by using the image decomposition model to obtain a corresponding cartoon part and a corresponding texture part, and fusing.

2. The multi-source image fusion method based on discriminative dictionary learning and morphological component decomposition of claim 1, wherein: the specific process of the step (1) is as follows: collecting a plurality of gray level images from the Internet as training samples, then collecting data of the training samples in a sliding window mode, and collecting the data as a column vector n through a window n multiplied by n²And (4) multiplying by 1, wherein n is the size of a sliding window, and the acquired data is decomposed through an MCA algorithm to obtain cartoon training data and texture training data.

3. The multi-source image fusion method based on discriminative dictionary learning and morphological component decomposition of claim 1, wherein: the objective function for distinguishing the dictionary learning model in the step (3) is as follows:

wherein Y ∈ R^m×nData collected for a sliding window is used as a matrix formed by column vectors, R is a spatial domain, m is the dimension of a vector space, n is the number of image blocks, D_t、D_cRespectively representing texture dictionaries and cartoon dictionaries, A_t＝[α_1,t,α_2,t,…,α_L,t]representing the corresponding texture sparse coding coefficient of the texture training data, wherein α_l,tFor the sparse coding coefficient corresponding to the texture component of the ith image block, A_c＝[α_1,c,α_2,c,…,α_L,c]representing the cartoon sparse coding coefficient corresponding to the cartoon training data, wherein α_l,cis a sparse coding coefficient corresponding to the cartoon component of the ith image block, T is the transpose of the matrix, and v is a gradient calculationSub, λ₁，λ₂To balance the parameters, | ·| luminance_FIs F norm operator, | ·| non-woven phosphor₁Is 1₁Norm operator, | ·| tory²Is the norm square operator.

4. The multi-source image fusion method based on discriminative dictionary learning and morphological component decomposition of claim 1, wherein: the specific process of the step (4) is as follows: taking N pictures to be fused, preprocessing the pictures to be fused, firstly adding Gaussian white noise, then collecting data of the pictures to be fused in a sliding window mode, and collecting the data as a column vector N through a window N multiplied by N²X 1 and n is the size of a sliding window to obtain an image matrix to be fused, and decomposing the image to be fused to obtain a sparse coding coefficient A of the cartoon component_cAnd sparse coding coefficient A of texture component_tBy using a₁And selecting coding coefficients of different components of the fused image according to the norm maximum principle, and fusing the image.

5. The multi-source image fusion method based on discriminative dictionary learning and morphological component decomposition of claim 4, wherein: the decomposition of the image to be fused is realized by an image decomposition model, and the objective function is as follows:

wherein X ∈ R^m×nData collected for sliding windows as a matrix of column vectors, R being the spatial domain, D_t、D_cRespectively representing the texture dictionary and the cartoon dictionary learned in step 3, A_t＝[α_1,t,α_2,t,…,α_L,t]represents the texture sparse coding coefficient corresponding to the texture sample data, where α_l,tFor the sparse coding coefficient corresponding to the texture component of the ith image block, A_c＝[α_1,c,α_2,c,…,α_L,c]To indicate cartoon sample datacorresponding cartoon sparse coding coefficient, wherein alpha_l,csparse coding coefficients, eta corresponding to cartoon components of the ith image block₁，η₂To balance the parameters, | ·| luminance_FIs F norm operator, | ·| non-woven phosphor₁Is 1₁Norm operator, | ·| tory²In the form of the square operator of the norm,for sparsely encoding coefficients alpha_cIs estimated value ofThe calculation formula of the formed matrix is as follows:

wherein omega_ito code the coefficient alpha_i,jAccording to the idea of the non-local mean method, the calculation formula of the local area corresponding to the image block is as follows:

wherein alpha is_c,iCoding coefficients for texture components of the ith image block,is the local region omega where the texture components of the i image blocks are located_iH is a normalization factor, and H is a preset scale size.

6. The multi-source image fusion method based on discriminative dictionary learning and morphological component decomposition of claim 4, wherein: by means of₁The specific process of fusing images by selecting coding coefficients for fusing different components of the images according to the norm maximum principle is as follows:

is provided withIn order to fuse the texture of the image and the coding coefficient of the ith image block of the cartoon component, the fusion scheme is as follows:

wherein,column l of coding coefficients representing texture components of the fused image, {1, 2, …, K },coding coefficient A representing texture component of the r-th image to be fused_tColumn 1, where r is {1, 2, …, N }, N is the number of images to be fused, and K is the number of image blocks;

wherein,column l, which represents the coding coefficients of the cartoon component of the fused image, {1, 2, …, K },coding coefficient A of cartoon component representing s-th image to be fused_cIs a column vector, s ═ 1, 2, …, N, thenN is the number of images to be fused, and K is the number of image blocks;

is obtained byAndthen, assuming that there are K image blocks in total,encoding coefficients representing texture components of the fused image,representing the coding coefficients of the cartoon components of the fused image, the fused cartoon-texture components can be represented asAndthe matrix formed by the block vectors of the fused image can thus be represented asRearranging the result into image blocks to obtain the final fusion image.