A method for reconstructing spectral information of RGB images based on dictionary atom embedding

Abstract

A method for reconstructing spectral information of RGB images based on dictionary atom embedding belongs to the field of computer vision. The present invention obtains a hyperspectral image of the corresponding scene by reconstructing RGB image information. First, training samples are obtained from a high-resolution spectral image library. Then, based on the Bayesian non-parametric dictionary learning theory and considering the smoothness of spectral changes, a Gaussian process is used to learn a complete spectral dictionary; then, for each dictionary atom, low-dimensional pixels and their corresponding high-dimensional pixels are selected according to the feature similarity with the low-dimensional space, and their neighborhoods are used to form a mapping matrix from the low-dimensional space to the high-dimensional space to complete the training phase; finally, the mapping matrix from the low-dimensional space to the high-dimensional space is selected and reconstructed according to the relationship between the test image and the spectral dictionary, and the reconstruction of the hyperspectral image can be completed. The present invention can reconstruct high-resolution hyperspectral images through color images while ensuring the calculation speed, and can be applied in medical imaging, geological exploration, agricultural production and other fields.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and relates to a method for reconstructing a high spectral resolution image by a neighborhood embedding method, using Gaussian process learning taking into account the smoothness of spectral changes, obtaining a complete dictionary based on a Bayesian non-parametric analysis model, and restoring a camera spectral sensitivity function to optimize spectral information reconstruction. Background Art

[0002] Hyperspectral images are approximately continuous spectral curves of target objects obtained in a large number of electronic bands such as visible light, ultraviolet light, and near-infrared light, which simultaneously reflect the spectral information and spatial relationship information of the spectral reflectance characteristics of the target object. The application of hyperspectral images has now expanded to the fields of criminal investigation, medical imaging, agricultural detection, remote sensing, etc. Since the imaging instrument carrier is mobile but the imaging technology is limited and requires a long imaging time, it is difficult to obtain high-resolution hyperspectral images. At the same time, the expensive instruments cause high costs for obtaining hyperspectral images. The topic of reconstructing high-resolution spectral images of the same scene based on RGB images has become a hot topic, but this is an inverse pathological problem of reconstructing multiple spectra from a small amount of spectral information.

[0003] Currently, the two main methods of hyperspectral reconstruction are prior-driven methods and data-driven methods, in which the prior knowledge of hyperspectra usually includes sparsity, spatial structure similarity, spectral correlation, etc. Arad et al. proposed to use sparse coding to reconstruct hyperspectral images from RGB, but this method did not consider the local structural similarity, so the performance was limited. The dual dictionary algorithm, which was also based on sparse coding but used a non-negative one, could improve the reconstruction performance, but it was assumed that the camera spectral sensitivity function (CSS) was known, otherwise it could not be reconstructed. In addition to the dictionary learning algorithm, there is also a manifold-based reconstruction. Jia et al. reduced the dimension of the hyperspectral image to a three-dimensional manifold, that is, the reconstruction was transformed from three-dimensional to multi-dimensional into a three-dimensional to three-dimensional mapping, which simplified the reconstruction problem and shortened the algorithm running time. However, the proof of the hyperspectral image to the three-dimensional manifold needs to be further improved. Akhtar et al. used a Gaussian process to solve the dictionary, while considering the non-local similarity of the hyperspectral image, adopted a group learning dictionary method, used the same group selection dictionary to perform sparse coding on the RGB image, and directly used the code in the high-dimensional space, but this method requires the known CSS, and directly uses the same sparse coefficient when the training image and the test image are weakly connected, which lacks a strong theoretical basis. Data-driven methods are the application of the development of neural networks in the field of hyperspectral image reconstruction, such as convolutional neural networks, adversarial neural networks, and attention networks. However, these methods lack interpretability, their performance depends on parameter settings, and they have problems with generalization of data sets. Summary of the invention

[0004] In view of the problems existing in existing algorithms, such as lack of interpretability and insufficient consideration of hyperspectral image features, the present invention discloses a method for reconstructing spectral information of RGB images based on dictionary atom embedding to solve the following technical problems: on the basis of considering the smoothness of spectral changes, a Bayesian non-parametric analysis model is combined with a Gaussian process to learn a complete dictionary, a neighborhood is used to establish a mapping relationship from a low-dimensional space to a high-dimensional space, and then a mapping relationship is selected according to the feature similarity between the RGB image and the complete dictionary to achieve reconstruction of the hyperspectral image. In addition, CSS is restored through data and CSS characteristics, so as to accurately achieve dimensionality reduction in high-dimensional space and reduce reconstruction errors.

[0005] To achieve the above object, the present invention adopts the following technical solutions.

[0006] A method for reconstructing spectral information of RGB images based on dictionary atom embedding, by reconstructing RGB image information to obtain a hyperspectral image of the corresponding scene, firstly obtain training samples from a high-resolution spectral image library, then use Gaussian process based on Bayesian non-parametric dictionary learning theory and considering the smoothness of spectral changes to learn a complete spectral dictionary; then for each dictionary atom, select low-dimensional pixels and their corresponding high-dimensional pixels according to the feature similarity with the low-dimensional space, use their neighborhood to form a mapping matrix from low-dimensional space to high-dimensional space, and complete the training phase; finally, select and reconstruct the mapping matrix from low-dimensional space to high-dimensional space according to the relationship between the test image and the spectral dictionary, and complete the reconstruction of the hyperspectral image. The present invention can reconstruct high-resolution hyperspectral images through color images while ensuring the calculation speed, and can be applied in the fields of medical imaging, geological exploration, agricultural production, etc.

[0007] A method for reconstructing spectral information of RGB images based on dictionary atom embedding includes the following steps:

[0008] Step 1: Obtain a high-spectral resolution image training set and transform the structure into two-dimensional data Y train , the online dictionary learning method is used to set the dictionary size and sparsity to obtain the initial complete dictionary D, and the Lasso method is used to obtain the initial sparse coefficient A.

[0009] The objective function of learning a complete dictionary by the online dictionary learning method described in step 1 is:

[0010]

[0011] The objective function of learning sparse coefficients using the Lasso method is:

[0012]

[0013] In equations (1) and (2), λ1 is the sparsity constraint parameter, D is the initial complete dictionary, A is the initial sparse coefficient, and Yh Training images for high spectral resolution.

[0014] Step 2: Take the initial complete dictionary D and initial sparse coefficient A obtained in step 1 as initial values, use the Bayesian nonparametric analysis model, and consider the smoothness of spectral changes. Use Gibbs sampling to set the number of iterations and the initial value of the uninformative parameter to obtain the high-dimensional complete dictionary Φ h .

[0015] In the Bayesian nonparametric analysis model described in step 2, a high-dimensional complete dictionary is used to sparsely encode the image, and the image contains Gaussian noise:

[0016]

[0017] In formula (3) is a high spectral resolution pixel, Φ h ∈R L×K is a high-dimensional complete dictionary, ε i ∈R L is the noise, α i ∈R K is the sparse coefficient; where the sparse coefficient is represented by α i =s i ⊙z i , z i Indicates dictionary atom Is it expressed s i instruct The weight of the dictionary atom is The prior is set as a Gaussian process The Gaussian kernel uses a square exponential kernel with respect to the wavelength. in is the wavelength in nm, and the parameter η k The prior for z is set to a meaningless gamma distribution; i As the activation item of the i-th dictionary atom, the prior is set to Bernoulli(z ik |π k ), where the parameter π k The prior is set to Beta(π k |c / K,d(K-1) / K), where parameter K is the dictionary size; s i and noise ε i The prior of is set to Gaussian distribution; the prior of hyperparameters is set to gamma distribution without information; so far, the entire Bayesian nonparametric analysis model is derived according to the prior distribution and likelihood function. The posterior distribution of each parameter of the model is trained according to the Gibbs sampling method, and the dictionary mean μ is used as the high-dimensional complete dictionary Φ hOutput. Step 3: Obtain the low spectral resolution image Y corresponding to the scene of the high spectral resolution image in step 1 l , considering the smoothness and non-negativity of CSS, the objective function of solving C is established and solved, and the camera spectral sensitivity function C for reconstructing high spectral resolution scenes is obtained.

[0018] Step 3: Obtain the low spectral resolution image Y corresponding to the scene of the high spectral resolution image in step 1 l , considering the smoothness and non-negativity of the camera spectral sensitivity function C, the objective function for solving C is established and solved to obtain the camera spectral sensitivity function C for reconstructing high spectral resolution scenes;

[0019] The objective function considering the smoothness and non-negativity of the camera spectral sensitivity function described in step 3 is:

[0020]

[0021] Where C is the CSS matrix, and TV is used to constrain the smoothness of the CSS. It is used to constrain the non-negativity of CSS, and λ2 is the regularization parameter.

[0022] The method for solving the objective function in step 3 is the alternating operator multiplier method. After solving, the camera spectral sensitivity function C for reconstructing the high spectral resolution scene is obtained.

[0023] Step 4: Use the camera spectral sensitivity function obtained in step 3 to calculate the high-dimensional complete dictionary Φ h Dimensionality reduction, obtain low-dimensional complete dictionary Φ l According to the distance relationship between the low-dimensional complete dictionary atoms and the low spectral resolution image pixels, the low-dimensional neighborhood and the corresponding high-dimensional neighborhood are selected to construct the mapping matrix, and the low-dimensional complete dictionary Φ is obtained. l The corresponding mapping matrix.

[0024] Step 4 describes the high-dimensional complete dictionary Φ h The dimensionality reduction operation performed is:

[0025] Φ l =CΦ h (5)

[0026] Where C is the camera spectral sensitivity function obtained in step 2.

[0027] The distance relationship described in step 4 adopts the Euclidean distance. For each dictionary atom, l Find the low-dimensional pixel with the smallest Euclidean distance, and the mapping matrix corresponding to the dictionary atom is composed of the low-dimensional neighborhood and its corresponding high-dimensional neighborhood. The calculation equation of the mapping matrix is:

[0028]

[0029] Where P i is the mapping matrix corresponding to the i-th dictionary atom, N l is a low-dimensional neighborhood, N h is a high-dimensional neighborhood, I is the unit matrix, and λ3 is the constraint parameter; thus, we get the low-dimensional complete dictionary Φ l The corresponding mapping matrix.

[0030] Step 5, obtain an RGB test image, and arrange it into two-dimensional data with spatial dimension and spectral dimension as dimensions; for each pixel of the test image, find the nearest dictionary atom in the low-dimensional complete dictionary, and reconstruct the corresponding pixel in the high-dimensional space through the mapping matrix corresponding to the dictionary atom, and finally obtain the entire reconstructed high-spectral resolution image.

[0031] The distance in step 5 is the Euclidean distance, and the operation of reconstructing high-dimensional pixels based on the mapping matrix and low-dimensional pixels is:

[0032] y h =P i y l (7)

[0033] Where P i is the mapping matrix obtained in step 4, y h and l are high-dimensional and low-dimensional spatial pixels respectively, thus obtaining a reconstructed high spectral resolution image.

[0034] Beneficial effects of the present invention:

[0035] 1. The present invention discloses a method for reconstructing spectral information of RGB images based on dictionary atom embedding, which uses a Bayesian non-parametric analysis model and considers the smoothness of spectral changes in hyperspectral images. It learns spectral features under the action of a Gaussian process, improves the ability of the dictionary to represent spectral changes, and can be applied to large-scale data without pre-setting parameters such as dictionary size.

[0036] 2. The present invention discloses a method for reconstructing spectral information of RGB images based on dictionary atom embedding, which combines the results of the Bayesian non-parametric analysis model applied in dictionary learning with the neighborhood embedding theory, and reconstructs high-spectral resolution images from RGB based on the local similarity of hyperspectral images, thereby reducing the acquisition cost of hyperspectral images and the algorithm running time is very short.

[0037] 3. The present invention discloses a method for reconstructing spectral information of RGB images based on dictionary atom embedding, which takes into account the smoothness and non-negativity of CSS and uses the alternating operator multiplier method to restore CSS, effectively solving the problem from the perspective of reconstructing hyperspectral images and reducing the error of reconstructing hyperspectral images.

[0038] 4. The present invention discloses a method for reconstructing spectral information of RGB images based on dictionary atom embedding, which can reconstruct large-scene high-spectral resolution images while ensuring that the error is very small, and can be effectively applied in remote sensing, agriculture, medical treatment, reconnaissance and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is an overall flow chart of a method for reconstructing spectral information of RGB images based on dictionary atom embedding of the present invention, wherein the straight arrow represents the training phase and the dotted arrow represents the testing phase.

[0040] Figure 2 It is the reconstruction result implemented in the ICVL database of the present invention, showing pictures with an interval of 20 nm from 400 nm to 700 nm.

[0041] Figure 3 This is the original image of the present invention implemented in the ICVL database, showing images with an interval of 20 nm from 400 nm to 700 nm. DETAILED DESCRIPTION

[0042] In order to better illustrate the purpose and advantages of the present invention, further detailed description is given below with reference to the accompanying drawings and examples.

[0043] like Figure 1 As shown, a method for reconstructing spectral information of RGB images based on dictionary atom embedding includes the following steps:

[0044] Step 1: Obtain a high-spectral resolution image training set and transform the structure into two-dimensional data Y train , use the online dictionary learning method to set the dictionary size and sparsity to obtain the initial complete dictionary D, and use the Lasso method to obtain the initial sparse coefficient A;

[0045] The data dimension of the high-spectral resolution image training set obtained in step 1 is N×M×L×Q, where N and M are spatial dimensions, L is the spectral dimension, and Q is the number of high-spectral images in the data set. In this example, any t images in ICVL are selected as the data set for training the high-dimensional complete dictionary. The data dimension obtained by changing the data dimension is NM×L×t. Each image is sampled with a block length of patchsize and a step length of step (patchsize≤step) to obtain N y =floor(N / step)×floor(M / step) blocks, where floor(·) is a rounding operation. A pixel is randomly sampled from each block. The sampling method should ensure that the sampled pixels account for at least 1% of the entire image. At this time, the training sample Y train The dimension is tN y×L, this sampling method not only takes into account the local similarity of the hyperspectral image, but also considers minimizing the amount of data processed by the method to speed up the method.

[0046] The online dictionary learning algorithm and Lasso method described in step 1 are implemented in this example using the SPAMs library (see http: / / thoth.inrialpes.fr / people / mairal / spams / for details), thereby obtaining an initial complete dictionary D and an initial sparse coefficient A.

[0047] Step 2: Take the initial complete dictionary D and initial sparse coefficient A obtained in step 1 as initial values, use the Bayesian nonparametric analysis model, and consider the smoothness of spectral changes. Use Gibbs sampling to set the number of iterations and the initial value of the uninformative parameter to obtain the high-dimensional complete dictionary Φ h .

[0048] In step 2, the initial complete dictionary D and initial sparse coefficient A obtained in step 1 are used as initial values, wherein the initial complete dictionary D is used as the initial value of the dictionary and the Gaussian process mean, the initial sparse coefficient A is used as the initial value of S, and Z is initialized according to the value of the S matrix, that is, non-zero values ​​correspond to 1 and zero values ​​correspond to 0.

[0049] In the Bayesian nonparametric analysis model described in step 2, a high-dimensional complete dictionary is used to sparsely encode the high-spectral resolution image.

[0050] Include Gaussian noise:

[0051]

[0052] In formula (3) is a high spectral resolution pixel, Φ h ∈R L×K is a high-dimensional complete dictionary, ε i ∈R L is the noise, α i ∈R K is the sparse coefficient; where the sparse coefficient is represented by α i =s i ⊙z i , z i Indicates dictionary atom Is it expressed s i instruct The weight of the dictionary atom is The prior is set as a Gaussian process The Gaussian kernel uses a square exponential kernel with respect to the wavelength. in is the wavelength in nm, and the parameter η kThe prior for z is set to a meaningless gamma distribution; i As the activation item of the i-th dictionary atom, the prior is set to Bernoulli(z ik |π k ), where the parameter π k The prior is set to Beta(π k |c / K,d(K-1) / K), where K is the dictionary size; s i and noise ε i The prior is set to be a Gaussian distribution, that is and Where I is the identity matrix and the hyperparameter λ s and λ ε The priors of are all set to uninformative gamma distribution; this is the entire Bayesian nonparametric analysis model, and the joint probability distribution of the model is:

[0053]

[0054] According to the joint probability distribution, the posterior density function is deduced and the analytical formula of Gibbs sampling is obtained. express from Medium Sampling in From Gamma (η k |a,b) in the sample η k ,in From Bernoulli ik |π ko ξ / (1-π ko +π ko ξ)) in sampling z ik ,in from Medium sampling ik ,in From Beta(π k |c,d) in which the sample π k ,in From Gamma(λ s |e,f) in sampling λ ε ,in From Gamma(λ ε |g,h) in the sample λ ε ,in After setting the number of iterations and the initial value of the hyperparameter, the Y obtained in step 1 train Data training obtains the dictionary mean matrix μ k is a high-dimensional complete dictionary Φh .

[0055] Step 3: Obtain the low spectral resolution image Y corresponding to the scene of the high spectral resolution image in step 1 l , considering the smoothness and non-negativity of the camera spectral sensitivity function C, the objective function of restoring C is established and solved to obtain the camera spectral sensitivity function C for reconstructing the high spectral resolution scene;

[0056] The objective function in step 3 is solved using the alternating operator multiplier method, and equation (4) is written as:

[0057]

[0058] Where R is The variable is Φ = {C, V1, V2, V3}, and the augmented Lagrangian function can be obtained from equation (10):

[0059]

[0060] Solving the updated analytical expressions of each variable is:

[0061]

[0062] The soft(·) operation is y→sign(y)max{|y|-τ,0}, and the number of iterations and error constraints are set. At this point, the CSS is restored based on the training data.

[0063] Step 4: Use the camera spectral sensitivity function obtained in step 3 to reduce the dimension of the high-dimensional complete dictionary to obtain a low-dimensional complete dictionary Φ l , based on the distance relationship between the low-dimensional complete dictionary atoms and the low spectral resolution image pixels, a low-dimensional neighborhood and a corresponding high-dimensional neighborhood are selected to construct a mapping matrix, and a mapping matrix corresponding to the low-dimensional complete dictionary is obtained;

[0064] The dimensionality reduction operation described in step 4 is:

[0065] Φ l =CΦ h (13)

[0066] Where C is the CSS obtained in step 2.

[0067] The distance relationship described in step 4 is based on the Euclidean distance. The mapping matrix P corresponding to the i-th dictionary atom i The calculation method of is deduced from the neighborhood embedding theory as follows:

[0068]

[0069] Where N h is the neighborhood of the high-dimensional pixel, y l ,Nl are the corresponding low-dimensional space pixels and neighborhoods respectively, and λ3 represents the constraint coefficient; thus, the relationship between the low-dimensional complete dictionary and the mapping matrix is ​​established. At the same time, the mapping matrix takes into account the local similarity of the image, which is composed of high-dimensional space and low-dimensional space.

[0070] Step 5, obtain an RGB test image, and arrange it into two-dimensional data with spatial dimension and spectral dimension as dimensions; for each pixel of the test image, find the nearest dictionary atom in the low-dimensional complete dictionary, and reconstruct the corresponding pixel in the high-dimensional space through the mapping matrix corresponding to the dictionary atom, and finally obtain the entire reconstructed high-spectral resolution image.

[0071] The distance described in step 5 is based on the Euclidean distance in this example, and the reconstruction operation is:

[0072] y h =P i y l (15)

[0073] Where P i is the mapping matrix obtained in step 4, i is determined based on the distance between the low-dimensional complete dictionary atom and the low-dimensional space pixel, y h and l are high-dimensional and low-dimensional spatial pixels respectively, thus obtaining a reconstructed high spectral resolution image.

Claims

1. A method for reconstructing spectral information of RGB images based on dictionary atom embedding, characterized in that: The following steps are involved: Step 1: Obtain a high-spectral resolution image training set and transform the structure into two-dimensional data Y train , use the online dictionary learning method to set the dictionary size and sparsity to obtain the initial complete dictionary D, and use the Lasso method to obtain the initial sparse coefficient A; Step 2: Take the initial complete dictionary D and initial sparse coefficient A obtained in step 1 as initial values, use the Bayesian nonparametric analysis model, and consider the smoothness of spectral changes. Use Gibbs sampling to set the number of iterations and the initial value of the uninformative parameter to obtain the high-dimensional complete dictionary Φ h ; Step 3: Obtain the low spectral resolution image Y corresponding to the scene of the high spectral resolution image in step 1 l , considering the smoothness and non-negativity of the camera spectral sensitivity function C, the objective function for solving C is established and solved to obtain the camera spectral sensitivity function C for reconstructing high spectral resolution scenes; Step 4: Use the camera spectral sensitivity function obtained in step 3 to calculate the high-dimensional complete dictionary Φ h Dimensionality reduction to obtain a low-dimensional complete dictionary Φ l , with a low-dimensional complete dictionary Φ l The distance relationship between atoms and low spectral resolution image pixels is used to select low-dimensional neighborhoods and corresponding high-dimensional neighborhoods to construct a mapping matrix, and a low-dimensional complete dictionary Φ is obtained. l The corresponding mapping matrix; Step 5: obtain a single-frame RGB test image and arrange it into two-dimensional data with spatial dimension and spectral dimension as dimensions; for each pixel of the test image, l The nearest dictionary atom is found, and the corresponding pixel in the high-dimensional space is reconstructed through the mapping matrix corresponding to the dictionary atom, and finally the entire reconstructed high-spectral resolution image is obtained.

2. The method for reconstructing spectral information of RGB images based on dictionary atom embedding according to claim 1, characterized in that: The objective function of learning a complete dictionary by the online dictionary learning method described in step 1 is: The objective function of learning sparse coefficients using the Lasso method is: In equations (1) and (2), λ1 is the sparsity constraint parameter, D is the initial complete dictionary, A is the initial sparse coefficient, and Y h Training images for high spectral resolution.

3. The method for reconstructing spectral information of an RGB image based on dictionary atom embedding as claimed in claim 1, characterized in that: In the Bayesian nonparametric analysis model described in step 2, a high-dimensional complete dictionary Φ is used. h Sparsely encode a hyperspectral image while containing Gaussian noise: In formula (3) is a high spectral resolution pixel, Φ h ∈R L×K is a high-dimensional complete dictionary, ε i ∈R L is the noise, α i ∈R K is the sparse coefficient; in Sparse coefficient Represented as α i =s i ⊙z i , z i Indicates dictionary atom Is it expressed s i instruct The weight of the dictionary atom is The prior is set as a Gaussian process The Gaussian kernel uses a square exponential kernel with respect to the wavelength. in is the wavelength in nm, and the parameter η k The prior for z is set to a meaningless gamma distribution; i As the activation item of the i-th dictionary atom, the prior is set to Bernoulli(z ik |π k ), where the parameter π k The prior is set to Beta(π k |c / K,d(K-1) / K), where parameter K is the dictionary size; s i and noise ε i The prior of is set to Gaussian distribution; the prior of hyperparameters is set to gamma distribution without information; so far, the entire Bayesian nonparametric analysis model is derived according to the prior distribution and likelihood function. The posterior distribution of each parameter of the model is trained according to the Gibbs sampling method, and the dictionary mean μ is used as the high-dimensional complete dictionary Φ h Output.

4. The method for reconstructing spectral information of an RGB image based on dictionary atom embedding according to claim 1, characterized in that: The objective function in step 3 considering the smoothness and non-negativity of the camera spectral sensitivity function C is: where Y h is the high spectral resolution training image, Y l is the low spectral resolution training image of the corresponding scene, C is the camera spectral sensitivity function, and TV is used to constrain the smoothness term of the camera spectral sensitivity function. It is used to constrain the non-negativity of the camera spectral sensitivity function, and λ2 is the regularization parameter; The method for solving the objective function in step 3 is the alternating operator multiplier method. After solving, the camera spectral sensitivity function C for reconstructing the high spectral resolution scene is obtained.

5. The method for reconstructing spectral information of RGB images based on dictionary atom embedding according to claim 1, characterized in that: The high-dimensional complete dictionary Φ described in step 4 h The dimension reduction operation is: Φ l =CΦ h (5) Where C is the camera spectral sensitivity function obtained in step 2; The distance relationship described in step 4 adopts the Euclidean distance. For each dictionary atom, l Find the low-dimensional pixel with the smallest Euclidean distance, and the mapping matrix corresponding to the dictionary atom is composed of the low-dimensional neighborhood and its corresponding high-dimensional neighborhood. The calculation equation of the mapping matrix is: Where P i is the mapping matrix corresponding to the i-th dictionary atom, N l is a low-dimensional neighborhood, N h is a high-dimensional neighborhood, I is the unit matrix, and λ3 is the constraint parameter; thus, we get the low-dimensional complete dictionary Φ l The corresponding mapping matrix.

6. The method for reconstructing spectral information of RGB images based on dictionary atom embedding according to claim 5, characterized in that: The distance in step 5 is the Euclidean distance. The operation of reconstructing high-dimensional pixels based on the mapping matrix and low-dimensional pixels is: and h =P i and l (7) Where P i is the mapping matrix obtained in step 4, y h and l are high-dimensional and low-dimensional spatial pixels respectively, thus obtaining a reconstructed high spectral resolution image.

Images