Multispectral image denoising method and device based on structured tensor sparse model

Abstract

Disclosed are a multispectral image denoising method and device based on a structured tensor sparse model. The structured tensor sparse model is based on tensor sparse representation and integrates non-local centralized tensor sparse constraints. The model can sparsely represent multidimensional signals in a group domain, simultaneously realizes local intrinsic tensor sparse regularization and non-local similarity of multidimensional signals in a unified framework, learns an orthogonal dictionary from a group of tensors to generate a basis of a space to which the group of tensors belongs, instead of decomposing each tensor, and can reduce the number of dictionary atoms. A compromise parameter is adaptively determined by using statistical characteristics of data, and an effective algorithm is designed to solve an optimization problem of the proposed model. Therefore, structural similarity between a group of full-band image blocks can be utilized, non-local centralized tensor sparse constraints can be integrated, multidimensional signals in a group domain can be sparsely represented, and local intrinsic tensor sparse regularization and non-local similarity of multidimensional signals can be simultaneously realized in a unified framework.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a multispectral image denoising method based on a structured tensor sparse model and a multispectral image denoising device based on the structured tensor sparse model. BACKGROUND

[0002] With the development of science and technology, as an information carrier, multispectral images play an increasingly important role in life. A multispectral image is a three-dimensional image that not only records the spatial information of an object but also records the spectral information of the object. Due to the influence of device precision and the surrounding environment, the image is often damaged by various noises in the process of acquisition, transmission and storage, and people are difficult to obtain high-quality images that restore the real scene. Because the camera sensor of a multispectral image has higher sensitivity, it is more susceptible to noise, which greatly reduces the application effect based on the multispectral image. Therefore, it is necessary to restore the multispectral image contaminated by noise to improve its quality.

[0003] Because a multispectral image is composed of a series of two-dimensional images of different frequency bands, it can be directly reconstructed by a two-dimensional image denoising method. At present, many one-dimensional two-dimensional image denoising methods have been proposed, such as K-SVD, NCSR based on sparsity, SAIST, WNNM based on low rank, NLM and BM3D based on filter, etc. However, these methods convert two-dimensional images into one-dimensional vectors for processing, which destroys the inherent two-dimensional structure of the image and leads to insufficient utilization of local correlation. In order to avoid the vectorization of two-dimensional images, two-dimensional methods such as 2DNSR are proposed. However, using one-dimensional or two-dimensional denoising methods to denoise three-dimensional multispectral images will also destroy the internal structure of the multispectral image.

[0004] At present, many denoising methods based on tensor representation have been proposed to protect the three-dimensional structure of the multispectral image. A multilinear algebraic method is proposed in LRTA, which reveals the denoising ability of low-rank tensor approximation. The PARAFAC method based on CANDECOMP / PARAFAC (CP) uses parallel factor analysis for denoising. TenSR uses a series of adaptive separable structure dictionaries to approximate the inherent structure of each dimension in a multidimensional signal. These methods make use of the global correlation of the spectral dimension of the multispectral image and obtain good denoising performance.

[0005] Recently, the low-rank tensor dictionary learning (LTDL) method was proposed to train a redundant dictionary for full-band patch (FBP) groups for denoising. The optimal low-rank tensor (OLRT) method discovered the low-rank property differences along the spatial, spectral, and nonlocal self-similarity dimensions in multispectral images and proposed an optimal low-rank tensor model for multispectral image restoration. Both LTDL and OLRT achieved advanced denoising performance. In particular, it is noted that the KBR method based on Tucker decomposition and CP decomposition not only utilizes the global correlation in the spectral dimension of multispectral images but also utilizes the nonlocal self-similarity in the spatial dimension and achieves good denoising performance. However, each full-band patch group in this method is decomposed, which increases the number of subspace bases.

[0006] The English full names and Chinese translations of the above English abbreviations are as follows:

[0007] K-SVD: K-Singular Value Decomposition

[0008] NCSR: Nonlocally Centralized Sparse Representation

[0009] SAIST: spatially adaptive iterative singular-value thresholding

[0010] WNNM: Weighted Nuclear Norm Minimization

[0011] BM3D: Block Matching 3D

[0012] NLM: Adaptive non-local means

[0013] 2DNSR: 2D Nonlocal Sparse Representation

[0014] LRTA: lower rank tensor approximation

[0015] CANDECOMP: canonical decomposition

[0016] PARAFAC: parallel factors

[0017] TenSR: Tensor Sparse Representation

[0018] LTDL: Low-Rank Tensor Dictionary Learning

[0019] OLRT: optimal low-rank tensor

[0020] FBP: full band patch

[0021] Tucker: Tucker

[0022] KBR: Kronecker-basis-representation based tensor sparsity measure SUMMARY

[0023] To overcome the defects of the prior art, the technical problem to be solved by the present application is to provide a multispectral image denoising method based on a structured tensor sparse model, which can utilize the structural similarity between full-band image block groups, integrate non-local centralized tensor sparse constraints, and sparsely represent multidimensional signals in the group domain, thereby simultaneously realizing local intrinsic tensor sparse regularization and non-local similarity of multidimensional signals in a unified framework.

[0024] The technical solution of the present application is that the multispectral image denoising method based on the structured tensor sparse model comprises the following steps:

[0025] (1) inputting a noisy multispectral image to the system, and for the multispectral image contaminated by noise where I H and I W denote the spatial dimension, I S denotes the spectral dimension of the image;

[0026] (2) obtaining a plurality of similar three-dimensional full-band image blocks by similar block matching, dividing the image into a plurality of image blocks I h ×I w ×I S and performing similar block matching thereon, so that M four-dimensional full-band image block groups I h ×I w ×I S ×C can be obtained, wherein C is the number of similar image blocks;

[0027] (3) All three-dimensional full-band image blocks are sorted according to spatial dimension I h with I w Expand it to convert it into a two-dimensional full-band image patch;

[0028] (4) Stitch similar two-dimensional full-band image blocks into a three-dimensional full-band image block group.

[0029] (5) Cluster the three-dimensional full-band image patch group into K classes using K-means clustering, and then train an orthogonal dictionary for each class using high-order singular value decomposition;

[0030] (6) Input each class into the structured tensor sparse model. The structured tensor sparse model is as follows:

[0031]

[0032] in For the desired reconstruction tensor, For tensors corrupted by noise, It is a sparse tensor that satisfies a multidimensional sparse model, where I is the identity matrix, when When it is an orthogonal dictionary, through Where T represents the matrix transpose operation, X (j) (j = 1, 2, 3) is a tensor mode-j expansion, β and α are compromise parameters;

[0033] (7) Output the reconstructed three-dimensional full-band image patch group;

[0034] (8) The reconstructed three-dimensional full-band image patch group is split and folded to reconstruct a reconstructed multispectral image.

[0035] (9) Output the reconstructed multispectral image.

[0036] The structured tensor sparse model of the application is based on tensor sparse representation, and integrates non-local centralized tensor sparse constraint. The model can sparsely represent multi-dimensional signals in a group domain, and simultaneously realizes local intrinsic tensor sparse regularization and non-local similarity of multi-dimensional signals in a unified framework. The application learns an orthogonal dictionary from a group of tensors to generate a basis of a space to which the group of tensors belongs, instead of decomposing each tensor, so that the number of dictionary atoms can be reduced. In addition, the application adaptively determines a trade-off parameter using statistical characteristics of data, and designs an effective algorithm to solve an optimization problem of the proposed model. Therefore, the structural similarity between group domains of full-band image blocks can be utilized, the non-local centralized tensor sparse constraint is integrated, the multi-dimensional signals in the group domain can be sparsely represented, and the local intrinsic tensor sparse regularization and the non-local similarity of the multi-dimensional signals can be simultaneously realized in the unified framework.

[0037] A multispectral image denoising device based on the structured tensor sparse model is also provided, and the device comprises:

[0038] an input module configured to input a noisy multispectral image to the system, for the multispectral image contaminated by noise wherein I H and I W denotes a spatial dimension, I S denotes a spectral dimension of the image;

[0039] a matching module configured to obtain a plurality of groups of similar three-dimensional full-band image blocks by similar block matching, divide the image into a plurality of image blocks I h ×I w ×I S and perform similar block matching on the image blocks, so that M groups of four-dimensional full-band image blocks I h ×I w ×I S ×C, wherein C is the number of similar image blocks;

[0040] a conversion module configured to convert all three-dimensional full-band image blocks into two-dimensional full-band image blocks by unfolding the three-dimensional full-band image blocks along the spatial dimension I h and I w ;

[0041] a splicing module configured to splice similar two-dimensional full-band image blocks into three-dimensional full-band image block groups

[0042] a clustering module configured to cluster the three-dimensional full-band image block groups into K classes using K-means clustering, and then train an orthogonal dictionary for each class using high-order singular value decomposition;

[0043] The reconstruction module is configured to input each class into a structured tensor sparse model, which is as follows:

[0044]

[0045] in For the desired reconstruction tensor, For tensors corrupted by noise, It is a sparse tensor that satisfies a multidimensional sparse model, where I is the identity matrix, when When it is an orthogonal dictionary, through Where T represents the matrix transpose operation, X (j) (j = 1, 2, 3) is a tensor mode-j expansion, β and α are compromise parameters;

[0046] The output module is configured to output a group of reconstructed 3D full-band image patches.

[0047] The stitching module is configured to stitch similar two-dimensional full-band image patches into a three-dimensional full-band image patch group, and then split and fold the reconstructed three-dimensional full-band image patch group to reconstruct a reconstructed multispectral image. Output the reconstructed multispectral image. Attached Figure Description

[0048] Figure 1 This is a flowchart of the multispectral image denoising method based on the structured tensor sparse model according to the present invention.

[0049] Figure 2 In the image, (a) is a clean image of the chart and staffed toy in two frequency bands (400 nm and 700 nm); (b) is an image corrupted by Gaussian noise with a variance of 0.2; and (c)-(n) are reconstructed images using 12 multispectral denoising methods.

[0050] Figure 3 In the image, (a) is a clean image of watercolor in two frequency bands (400nm and 700nm); (b) is an image corrupted by Gaussian noise with a variance of 0.3; and (c)-(n) are reconstructed images using 12 multispectral denoising methods.

[0051] Figure 4 This is the main view of six multispectral images from the CAVE dataset, including 'balloons', 'beads', 'cd', 'chart and staffed toy', 'cloth', and 'watercolor'.

[0052] Figure 5The average objective denoising results of 12 comparison methods on the CAVE dataset. Detailed Implementation

[0053] like Figure 1 As shown, this multispectral image denoising method based on a structured tensor sparse model includes the following steps:

[0054] (1) Input a noisy multispectral image into the system. For multispectral images contaminated by noise... Where I H with I W I represents spatial dimension. S Represents the spectral dimension of the image;

[0055] (2) Several sets of similar three-dimensional full-band image blocks are obtained through similar block matching, and the images are... Divided into several image blocks I h ×I w ×I S By performing similar block matching, M groups of 4D full-band image blocks I can be obtained. h ×I w ×I S ×C, where C is the number of similar image patches;

[0056] (3) All three-dimensional full-band image blocks are sorted according to spatial dimension I h with I w Expand it to convert it into a two-dimensional full-band image patch;

[0057] (4) Stitch similar two-dimensional full-band image blocks into a three-dimensional full-band image block group.

[0058] (5) Cluster the three-dimensional full-band image patch group into K classes using K-means clustering, and then train an orthogonal dictionary for each class using high-order singular value decomposition;

[0059] (6) Input each class into the structured tensor sparse model. The structured tensor sparse model is as follows:

[0060]

[0061] in For the desired reconstruction tensor, For tensors corrupted by noise, It is a sparse tensor that satisfies a multidimensional sparse model, where I is the identity matrix, when When it is an orthogonal dictionary, through Where T represents the matrix transpose operation, X (j) (j = 1, 2, 3) is a tensor mode-j expansion, β and α are trade-off parameters;

[0062] (7) output the reconstructed three-dimensional full-band image block group;

[0063] (8) split, fold and reconstruct the reconstructed three-dimensional full-band image block group into a reconstructed multi-spectral image

[0064] (9) output the reconstructed multi-spectral image.

[0065] The structured tensor sparse model of the present application is based on tensor sparse representation, and integrates non-local centralized tensor sparse constraint. The model can sparsely represent multi-dimensional signals in a group domain, and simultaneously realizes local intrinsic tensor sparse regularization and non-local similarity of multi-dimensional signals in a unified framework. The present application learns an orthogonal dictionary from a group of tensors to generate a basis of a space to which the group of tensors belongs, instead of decomposing each tensor, so that the number of dictionary atoms can be reduced. In addition, the present application adaptively determines a trade-off parameter by using statistical characteristics of data, and designs an effective algorithm to solve the optimization problem of the proposed model. Therefore, the structural similarity between the full-band image block group can be utilized, the non-local centralized tensor sparse constraint is integrated, the multi-dimensional signals in the group domain can be sparsely represented, and the local intrinsic tensor sparse regularization and the non-local similarity of the multi-dimensional signals can be simultaneously realized in the unified framework.

[0066] Preferably, in the step (6), each class Ω k of orthogonal dictionary is obtained by solving the following problem

[0067]

[0068] where I is an identity matrix, is a sparse core tensor of the class Ω

[0069] Preferably, in the step (6), two auxiliary variables and are introduced into the model, and the L0 and low-rank terms are relaxed in the form of logarithmic sum, and the obtained model is as follows:

[0070]

[0071] wherein,

[0072]

[0073]

[0074] wherein is a tensor of the elements, ε is a small positive number, σ j (M) denotes the j-th singular value of the matrix M.

[0075] Preferably, the step (6), the augmented Lagrangian function of the problem is written as:

[0076]

[0077]

[0078] where, and are the Lagrange multipliers, μ and γ are positive numbers, the above problem is decomposed into three sub-problems by using the alternating direction method of multipliers solved.

[0079] Preferably, in the step (6), when other parameters are unchanged, updated by the following formula:

[0080]

[0081] where

[0082] D b.ε (·) is a threshold operation, defined as follows:

[0083]

[0084] where,

[0085] When other parameters are unchanged, updated by the following formula:

[0086]

[0087] where S d (·) = sign(·) max(|·|-d, 0).

[0088] When other parameters are unchanged, updated by the following formula:

[0089]

[0090] where is the SVD decomposition of the matrix

[0091] Preferably, the maximum a posteriori estimation is used to determine the parameter α of the centralized constraint term of the sparse coefficient, defined as ​For a given

[0092]

[0093] where the likelihood term follows a Gaussian distribution

[0094]

[0095] where θ and are assumed to be independent, and θ follows a Laplace distribution,

[0096]

[0097] where θ(i) is the i-th element of θ, σ i is the standard deviation of θ(i). Substituting equations (12) and (13) into equation (11), we have

[0098]

[0099] Substituting into equation (14),

[0100]

[0101] Comparing equation (15) with equation (3), we have

[0102] It is appreciated by those skilled in the art that all or part of the steps of the above-mentioned embodiment methods can be completed by programs instructing relevant hardware, and the programs can be stored in a computer readable storage medium, and the storage medium can be ROM / RAM, a magnetic disc, an optical disc, a memory card, etc. Therefore, the present application also comprises a multispectral image denoising device based on structured tensor sparse model, which is usually represented in the form of functional modules corresponding to the steps of the method. The device comprises:

[0103] An input module configured to input a noisy multispectral image to the system, and for the multispectral image contaminated by noise where I H and I W denote the spatial dimension, and I S denotes the spectral dimension of the image;

[0104] A matching module configured to obtain a plurality of groups of similar three-dimensional full-band image blocks by similar block matching, and divide the image into a plurality of image blocks I h × I w × I Sand similar block matching is performed on them, M groups of 4D full-band image blocks I h ×I w ×I S ×C, where C is the number of similar image blocks.

[0105] a conversion module configured to convert all 3D full-band image blocks into 2D full-band image blocks by unfolding them along the spatial dimension I h and I w ;

[0106] a stitching module configured to stitch similar 2D full-band image blocks into groups of 3D full-band image blocks

[0107] a clustering module configured to cluster the groups of 3D full-band image blocks into K classes using K-means clustering, and then train an orthogonal dictionary for each class using higher-order singular value decomposition;

[0108] a reconstruction module configured to input each class into a structured tensor sparse model, which is as follows:

[0109]

[0110] where is the reconstructed tensor, is the tensor corrupted by noise, is a sparse tensor satisfying the multi-dimensional sparse model, and I is an identity matrix when is an orthogonal dictionary, by where T denotes the matrix transpose operation, X (j) (j = 1, 2, 3) is the mode-j unfolding of the tensor x, β and α are trade-off parameters.

[0111] an output module configured to output the reconstructed groups of 3D full-band image blocks;

[0112] a stitching module configured to stitch similar 2D full-band image blocks into groups of 3D full-band image blocks, and to split, fold, and reconstruct the reconstructed groups of 3D full-band image blocks into a reconstructed multi-spectral image output the reconstructed multi-spectral image.

[0113] To verify the effectiveness of the proposed structured tensor sparse model, the multispectral image denoising method proposed in the present application was tested on the CAVE dataset. Four widely used multispectral image quality evaluation indexes were used to compare the denoising performance, including the peak signal-to-noise ratio (PSNR), structural similarity (SSIM), feature similarity (FSIM), and the relative dimensionless global error (ERGAS). The higher the PSNR, SSIM, and FSIM or the lower the ERGAS of the reconstructed multispectral image, the better the quality. Figure 5 The average denoising performance of 12 competing methods on the 32 multispectral images in the CAVE dataset was demonstrated. It can be seen that the denoising performance of the tensor-based methods is superior to the vector-based methods because they preserve the three-dimensional structure of the multispectral image. Moreover, the denoising performance of the low-rank-based method OLRT is superior to other methods proposed in the past under the four different evaluation indexes. In addition, the denoising method proposed in the present application (STSP) has better performance than the OLRT method under the PSNR and FSIM evaluation indexes. And the performance under the SSIM and ERGAS evaluation indexes is similar to the OLRT method under three different noise variances.

[0114] To visualize the subjective denoising performance, the reconstructed results of the multispectral images chart and staffed toy and watercolor at 400 nm (top) and 700 nm (bottom) spectra are shown in Figure 2 and Figure 3 respectively. To facilitate the observation of details, two regions in these images were selected and magnified by 4 times. It can be seen that the denoising method provides the best subjective performance with fewer artifacts and smoother edges. The following is a specific description.

[0115] 1 Dataset

[0116] Experiments were conducted on the Columbia University CAVE dataset to verify the effectiveness of the proposed multispectral image denoising method. The CAVE dataset includes 32 real-world images. Their resolution size in the spatial dimension is 512×512, and the size in the spectral dimension is 31, which contains full-spectral resolution reflectance data from 400 nm to 700 nm with a step of 10 nm. The main view of some multispectral images in the dataset is shown in Figure 4 .

[0117] 2 Comparison algorithms

[0118] The proposed method (STSP) is compared with tensor-based methods OLRT, TDL, LRTA, tSVD, KBR and vector-based methods K-SVD, BM3D, ANLM3D, PARAFAC, BM4D and MSt SVD on the CAVE dataset. The parameters of these comparison methods are set optimally according to the descriptions in their cited papers.

[0119] 3 Experimental configuration

[0120] In the experiment, the multispectral images in the CAVE dataset are gray-scaled to [0, 1] as groundtruth. Then Gaussian noise with variance of v = 0.1, 0.2 and 0.3 is added to these clean multispectral images as the noise corrupted image input to the proposed model. The cluster number K of the full-band image block group is set to 1000, 700 and 700 corresponding to three different noise variances; the parameter γ in the algorithm is set to 7, 4 and 4. The parameters ρ, μ in the algorithm are set to 1.05 and 250 respectively. The other two parameters are directly set to λ = 10 and In addition, the parameters are adaptively estimated according to the statistical characteristics of the data

[0121] The above description is only the preferred embodiment of the present application, and does not limit the present application in any form. Any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiment are still within the protection scope of the technical solution of the present application.

Claims

1. A multispectral image denoising method based on a structured tensor sparse model, characterized in that: It includes the following steps: (1) Input a noisy multispectral image into the system. For multispectral images contaminated by noise... Among them I H with I W I represents spatial dimension. S Represents the spectral dimension of the image; (2) Several sets of similar three-dimensional full-band image blocks are obtained through similar block matching, and the images are... Divided into several image blocks I h ×I w ×I S By performing similar block matching, M groups of 4D full-band image blocks I can be obtained. h ×I w ×I S ×C, where C is the number of similar image patches; (3) All three-dimensional full-band image blocks are sorted according to spatial dimension I h with I w Expand it to convert it into a two-dimensional full-band image patch; (4) Stitch similar two-dimensional full-band image blocks into a three-dimensional full-band image block group. (5) Cluster the three-dimensional full-band image patch group into K classes using K-means clustering, and then train an orthogonal dictionary for each class using high-order singular value decomposition; (6) Input each class into the structured tensor sparse model. The structured tensor sparse model is as follows: in For the desired reconstruction tensor, For tensors corrupted by noise, It is a sparse tensor that satisfies a multidimensional sparse model, where I is the identity matrix, when When it is an orthogonal dictionary, through Where T represents the matrix transpose operation, X (j) (j = 1, 2, 3) is a tensor mode-j expansion, β and α are compromise parameters; (7) Output the reconstructed three-dimensional full-band image patch group; (8) The reconstructed three-dimensional full-band image patch group is split and folded to reconstruct a reconstructed multispectral image. (9) Output the reconstructed multispectral image.

2. The multispectral image denoising method based on a structured tensor sparse model according to claim 1, characterized in that: In step (6), each class Ω k orthogonal dictionary Obtained by solving the following problems Where I is the identity matrix. Is class sparse kernel tensor, 3. The multispectral image denoising method based on a structured tensor sparse model according to claim 2, characterized in that: In step (6), two auxiliary variables were introduced into the model. and The L0 and low-rank terms were relaxed using a logarithmic sum, resulting in the following model: in, in It is a tensor The elements, ε is a very small positive number, σ j (M) represents the j-th singular value of matrix M.

4. The multispectral image denoising method based on a structured tensor sparse model according to claim 3, characterized in that: In step (6), the augmented Lagrangian function of this problem is written as: in, and These are Lagrange multipliers, where μ and γ are positive numbers. Using the alternating direction multiplier method, the above problem can be decomposed into three subproblems. Please provide a solution.

5. The multispectral image denoising method based on a structured tensor sparse model according to claim 4, characterized in that: In step (6), when other parameters remain unchanged, Updated by the following formula: in D b，ε The · operator is a threshold operation, defined as follows: in, c1(x) = |x| - ε When other parameters remain unchanged Updated by the following formula: in S d (·)=sign(·)max(|·|-d, 0). When other parameters remain unchanged Updated by the following formula: in It is a matrix SVD decomposition.

6. The multispectral image denoising method based on a structured tensor sparse model according to claim 5, characterized in that: The parameter α of the centering constraint term for the sparse coefficients is determined using maximum a posteriori estimation, and is defined as follows: For a given The likelihood term follows a Gaussian distribution. Where θ and They are considered to be independent, and θ follows a Laplace distribution. Where θ(i) is the i-th element of θ, σ i It is the standard deviation of θ(i). Substituting equations (12) and (13) into equation (11), we get Will Substitute into equation (14), Comparing equation (15) and equation (3), we have 7. A multispectral image denoising device based on a structured tensor sparse model, characterized in that: It includes: The input module is configured to input noisy multispectral images into the system, for multispectral images contaminated by noise. Among them I H with I W I represents spatial dimension. S Represents the spectral dimension of the image; The matching module is configured to obtain several sets of similar 3D full-band image patches through similar block matching, and then... Divided into several image blocks I h ×I w ×I S By performing similar block matching, M groups of 4D full-band image blocks I can be obtained. h ×I w ×I S ×C, where C is the number of similar image patches; The conversion module is configured to convert all three-dimensional full-band image patches according to spatial dimension I. h with I w Expand it to convert it into a two-dimensional full-band image patch; The stitching module is configured to stitch similar two-dimensional full-band image patches into a three-dimensional full-band image patch group. The clustering module is configured to cluster groups of 3D full-band image patches into K classes using K-means clustering, and then train a set of orthogonal dictionaries for each class using higher-order singular value decomposition. The reconstruction module is configured to input each class into a structured tensor sparse model, which is as follows: in For the desired reconstruction tensor, For tensors corrupted by noise, It is a sparse tensor that satisfies a multidimensional sparse model, where I is the identity matrix, when When it is an orthogonal dictionary, through Where T represents the matrix transpose operation, X (j) (j = 1, 2, 3) is a tensor mode-j expansion, β and α are compromise parameters; The output module is configured to output a group of reconstructed 3D full-band image patches. The stitching module is configured to stitch similar two-dimensional full-band image patches into a three-dimensional full-band image patch group, and then split and fold the reconstructed three-dimensional full-band image patch group to reconstruct a multispectral image. Output the reconstructed multispectral image.

Images