A Hyperspectral Image Deblurring Method, System, and Storage Medium

Abstract

A hyperspectral image deblurring method, system and storage medium. The method includes obtaining a first hyperspectral image and a second hyperspectral image of the same scene from an image library, where the clarity of the second hyperspectral image is higher than that of the first hyperspectral image; performing normalization processing on the first hyperspectral image and the second hyperspectral image; adding noise to the second hyperspectral image to obtain a third hyperspectral image, constructing a deep denoising neural network and training the deep denoising neural network using a paired image set to obtain a deep denoising prior network; constructing an objective function with a regularization term, decoupling the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem and a dual variable update sub-problem, and iteratively solving the above three sub-problems until a termination condition is reached. The hyperspectral image denoising sub-problem is solved through the deep denoising prior network to complete the deblurring of the first hyperspectral image. The present invention restores spectral and spatial information to the greatest extent, and the obtained hyperspectral image has high quality.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and particularly to a hyperspectral image deblurring method, system and storage medium. Background Art

[0002] Hyperspectral imaging technology can acquire image data containing hundreds of narrow and spectrally continuous bands. Due to carrying rich spatial and spectral information, hyperspectral images can be used to distinguish targets that cannot be recognized by traditional grayscale and color images, and thus have been successfully applied to many fields such as earth remote sensing, military reconnaissance, precision agriculture, and industrial sorting. Hyperspectral images are affected by various factors during acquisition, and there are various types of blurring and degradation in the spatial and spectral dimensions. It is very important to perform deblurring and restoration processing on them. However, the three-dimensional structure of hyperspectral images increases the scale of data, making the restoration processing very challenging.

[0003] Currently, the most advanced mainstream hyperspectral deblurring technologies can be roughly divided into two types: traditional optimization methods and deep learning methods. Traditional optimization methods can flexibly handle different problems according to the physical model of image degradation, but usually require artificial definition of appropriate image prior knowledge to ensure better results. However, constructing artificial priors usually increases the complexity of problem solving and does not fully utilize the rich information contained in the existing data. Deep learning methods use a data-driven approach and can better mine image characteristics, and can obtain excellent processing results when the characteristics of training and test data match. However, current deep neural networks have many parameters, low interpretability, and less consideration of the physical mechanism of data generation, resulting in limited generalization ability of such methods, so the application scope is also limited. Summary of the Invention

[0004] The purpose of the present invention is to provide a hyperspectral image deblurring method, system and storage medium for the above problems in the prior art, which can restore spectral and spatial information to the greatest extent, effectively avoid or reduce image distortion and aberration generated during the reconstruction process, so as to obtain high-quality clear hyperspectral images, and have high flexibility, fast operation speed and good applicability.

[0005] To achieve the above purpose, the present invention has the following technical solutions:

[0006] In a first aspect, a hyperspectral image deblurring method is provided, including:

[0007] Obtain a first hyperspectral image Y and a second hyperspectral image X of the same scene from an image library, where the clarity of the second hyperspectral image X is higher than that of the first hyperspectral image Y; perform normalization processing on the first hyperspectral image Y and the second hyperspectral image X;

[0008] Add noise to the second hyperspectral image X to obtain the third hyperspectral image Z. Construct a deep denoising neural network and use the paired image set (Z; X) to train the deep denoising neural network to obtain a deep denoising prior network.

[0009] Construct an objective function with a regularization term, decouple the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem, and iteratively solve the above three sub-problems until the termination condition is reached. The hyperspectral image denoising sub-problem is solved by the deep denoising prior network to complete the deblurring of the first hyperspectral image Y.

[0010] As a preferred solution of the hyperspectral image deblurring method of the present invention, the steps of constructing a deep denoising neural network and using the paired image set (Z; X) to train the deep denoising neural network include:

[0011] Construct a three-dimensional network module: from input to output are three-dimensional convolutional layer a 1 , batch normalization layer b, activation layer c 1 ;

[0012] Construct a deep denoising neural network: from input to output are three-dimensional convolutional layer a 2 , activation layer c 2 , B three-dimensional network modules and three-dimensional convolutional layer a 3 , and the network input is input into the stacked layer d 3 together with the output of the three-dimensional convolutional layer a 1 ; The three-dimensional convolutional layer a 3 contains 1 filter of 3×3×3 with a stride of 1;

[0013] Select two-thirds of the image pairs in the image set (Z; X) as the training set, and the remaining one-third as the test set; randomly select M image patches of size 1 / h of the image itself in each image of the training set, and then randomly flip and mirror the M image patches in each image; use the image patches in the first hyperspectral image Y after the above processing as the network input, and use the image patches in the second hyperspectral image X after the above processing as the label image;

[0014] The loss function is defined as follows:

[0015] ||F(Z; Θ)-X|| 1

[0016] where F(*) represents the deep denoising network mapping, and Θ represents the deep denoising network model parameters;

[0017] Using the Adam optimization algorithm, with an initial learning rate of e, randomly selecting f samples during each forward propagation, and iterating the algorithm for g generations; after training is completed, the parameters of the deep denoising prior network model are obtained, and then the deep denoising prior network is obtained.

[0018] As a preferred embodiment of the hyperspectral image deblurring method of the present invention, the three-dimensional convolutional layer a 1 and the three-dimensional convolutional layer a 2 both contain 32 3×3×3 filters, and the stride is 1 for both; the activation layer c 1 and the activation layer c 2 are both ReLU functions; the number of three-dimensional network modules B = 8; randomly select M image patches of 1 / h of the image size itself from each image in the training set, where h = 4 and M = 100; the initial learning rate e = 0.0002, the number of samples f = 16, and the number of algorithm iterations g = 500.

[0019] As a preferred embodiment of the hyperspectral image deblurring method of the present invention, in the step of constructing the objective function with a regularization term, the constructed objective function is:

[0020]

[0021] where H represents the blur matrix, represents the data fidelity term, φ(X) represents the regularization term, and λ represents the regularization term parameter.

[0022] As a preferred embodiment of the hyperspectral image deblurring method of the present invention, decoupling the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem is completed using the plug-and-play algorithm based on the alternating direction method of multipliers; the termination condition is t = T, where t represents the number of iterations and T represents the threshold of the number of iterations.

[0023] As a preferred embodiment of the hyperspectral image deblurring method of the present invention, the iteration threshold T = 20.

[0024] As a preferred embodiment of the hyperspectral image deblurring method of the present invention, the step of decoupling the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem, and iteratively solving the above three sub-problems until the termination condition is reached includes:

[0025] Input the first hyperspectral image Y, the blur matrix H, the regularization term parameter λ, and the penalty parameter ρ; initialize the deep denoising network mapping F(*), the deep denoising network model parameters Θ; initialize the second hyperspectral image X 0 = Y, the auxiliary variable Z 0 = X 0 , the dual variable U 0= 0; The number of iterations t = 0;

[0026] Solve the least squares sub - problem according to the following formula and update X t+1 :

[0027] X t+1 = (H T H + ρI) -1 (H T Y + ρ(Z t - U t ))

[0028] Solve the hyperspectral image denoising sub - problem using F(*) and Θ according to the following formula and update Z t+1 :

[0029] Z t+1 = F(X t+1 + U t ; Θ)

[0030] Update the dual variable U according to the following formula t+1 :

[0031] U t+1 = U t + X t+1 - Z t+1

[0032] Update the number of iterations t ← t + 1; Judge whether the termination condition is satisfied. If satisfied, output the deblurred image X of the first hyperspectral image Y, otherwise return to continue solving the least squares sub - problem, the hyperspectral image denoising sub - problem and the dual variable update sub - problem. T , otherwise return to continue solving the least squares sub - problem, the hyperspectral image denoising sub - problem and the dual variable update sub - problem.

[0033] As a preferred solution of the hyperspectral image deblurring method of the present invention, the regularization term parameter λ = 9.6×10 -5 , and the penalty parameter ρ = 0.06.

[0034] In the second aspect, a hyperspectral image deblurring system is provided, including:

[0035] A paired image set construction module, configured to obtain the first hyperspectral image Y and the second hyperspectral image X of the same scene from the image library, where the clarity of the second hyperspectral image X is higher than that of the first hyperspectral image Y; perform normalization processing on the first hyperspectral image Y and the second hyperspectral image X;

[0036] A deep denoising neural network construction and training module, configured to add noise to the second hyperspectral image X to obtain a third hyperspectral image Z, construct a deep denoising neural network and train the deep denoising neural network using the paired image set (Z; X) to obtain a deep denoising prior network;

[0037] The deblurring module is used to construct an objective function with a regularization term, decouple the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem, and iteratively solve the above three sub-problems until the termination condition is reached. The hyperspectral image denoising sub-problem is solved by a deep denoising prior network to complete the deblurring of the first hyperspectral image Y.

[0038] In a third aspect, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the hyperspectral image deblurring method as described in the first aspect are implemented.

[0039] Compared with the prior art, the first aspect of the present invention has at least the following beneficial effects:

[0040] The deblurring method of the present invention first uses a hyperspectral database to train a deep denoising neural network to learn the prior information of clear hyperspectral images and obtain a deep denoising prior network; then decouples the hyperspectral image deblurring problem into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem; finally, reconstructs a clear hyperspectral image by iteratively solving the three sub-problems. The present invention restores spectral and spatial information to the greatest extent, thereby obtaining high-quality hyperspectral images. The present invention designs a deep denoising neural network capable of processing noisy hyperspectral images. Compared with common two-dimensional convolution kernels, three-dimensional convolution kernels are more suitable for the data characteristics of hyperspectral images. The present invention can simultaneously process different convolution kernels and noise conditions, with high flexibility, which is beneficial to the adaptability to different input data and enhances the applicability of the present invention.

[0041] Furthermore, the plug-and-play algorithm based on the alternating direction method of multipliers of the present invention decouples the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem. The present invention embeds the learned deep prior information into the physical degradation model through the plug-and-play algorithm to constrain the objective function, achieving a better deblurring effect.

[0042] It can be understood that the beneficial effects of the above second aspect to the third aspect can refer to the relevant descriptions in the above first aspect and will not be elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0044] Figure 1Schematic flow diagram of the hyperspectral image deblurring method according to an embodiment of the present invention;

[0045] Figure 2 Schematic diagram of the three-dimensional network module according to an embodiment of the present invention;

[0046] Figure 3 Schematic diagram of the deep denoising neural network according to an embodiment of the present invention. Detailed implementation manners

[0047] In the following description, for the purpose of illustration rather than limitation, specific details such as specific system architectures and technologies are set forth in order to provide a thorough understanding of the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present application. Additionally, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0048] As Figure 1 shown, the hyperspectral image deblurring method according to an embodiment of the present invention includes the following steps:

[0049] Step 1: Obtain a first hyperspectral image Y and a second hyperspectral image X of the same scene from an image library, where the clarity of the second hyperspectral image X is higher than that of the first hyperspectral image Y; perform normalization processing on the first hyperspectral image Y and the second hyperspectral image X;

[0050] Step 2: Add noise to the second hyperspectral image X to obtain a third hyperspectral image Z, construct a deep denoising neural network, and train the deep denoising neural network using the paired image set (Z; X) to obtain a deep denoising prior network;

[0051] Step 2-1: As Figure 2 shown, construct a three-dimensional network module: from input to output are a three-dimensional convolutional layer a 1 , a batch normalization layer b, and an activation layer c 1 ;

[0052] Step 2-2: As Figure 3 shown, construct a deep denoising neural network, from input to output are in sequence the three-dimensional convolutional layer a 2 , the activation layer c 2 , B three-dimensional network modules, and the three-dimensional convolutional layer a 3 , the network input is input into the stacked layer d 3 together with the output of the three-dimensional convolutional layer a 1 through a skip connection; the three-dimensional convolutional layer a 3It contains a 3×3×3 filter with a step size of 1;

[0053] Step 2-3: Select two-thirds of the image pairs in the image set (Z; X) as the training set, and the remaining one-third as the test set; randomly select M image patches of size 1 / h of the image itself in each image of the training set, and then randomly flip and mirror the M image patches in each image; use the image patches in the noisy hyperspectral image Z after the above processing as the input of the network, and use the image patches in the clear hyperspectral image X after the above processing as the label images;

[0054] The loss function is defined as follows:

[0055] ||F(Z; Θ)-X|| 1

[0056] where F(*) represents the mapping of the deep denoising network, and Θ represents the parameters of the deep denoising network model;

[0057] Step 2-4: Use the Adam optimization algorithm with an initial learning rate of e, and randomly select f samples each time for forward propagation. The algorithm iterates g generations; after training is completed, obtain the parameters Θ of the deep denoising prior network model;

[0058] Step 3: Construct an objective function with a regularization term:

[0059]

[0060] where H represents the blurring matrix, represents the data fidelity term, φ(X) represents the regularization term, and λ represents the regularization term parameter;

[0061] Step 4: Use the Plug-and-Play algorithm based on the alternating direction method of multipliers to decouple the objective function in Step 3 into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem, and iteratively solve the three sub-problems until the termination condition is reached:

[0062] t = T (2)

[0063] where t represents the number of iterations, and T represents the threshold of the number of iterations;

[0064] Step 4-1: Input the blurred hyperspectral image Y, the blurring matrix H, the regularization term parameter λ, and the penalty parameter ρ; initialize the mapping F(*) of the deep denoising network, the parameters Θ of the deep denoising network; initialize the clear hyperspectral image X 0 = Y, the auxiliary variable Z 0 = X 0 , the dual variable U 0 = 0; the number of iterations t = 0;

[0065] Step 4-2: Solve the least squares sub-problem and update X t+1 :

[0066] X t+1 =(H T H + ρI) -1 (H T Y + ρ(Z t -U t )) (3)

[0067] Step 4-3: Use F(*) and Θ to solve the hyperspectral image denoising sub-problem and update Z t+1 :

[0068] Z t+1 = F(X t+1 + U t ; Θ) (4)

[0069] Step 4-4: Update the dual variable U t+1 :

[0070] U t+1 = U t + X t+1 - Z t+1 (5)

[0071] Step 4-5: Update the iteration number t ← t + 1; judge whether the termination condition expression (2) is satisfied. If it is satisfied, output X T , otherwise continue to execute Step 4-2, Step 4-3, and Step 4-4;

[0072] Step 5: Obtain the estimated hyperspectral image X = X T , and complete the hyperspectral image deblurring.

[0073] In the embodiment, the three-dimensional convolutional layer a 1 、a 2 both contain 32 3×3×3 filters, and the stride is 1 for both.

[0074] In the embodiment, the activation layers c 1 、c 2 are both ReLU functions.

[0075] In the embodiment, B = 8.

[0076] In the embodiment, h = 4 and M = 100.

[0077] In the embodiment, the initial learning rate e = 0.0002, the number of samples f = 16, and the number of algorithm iterations g = 500.

[0078] In the embodiment, the iteration number threshold T = 20.

[0079] In an embodiment, the regularization parameter λ = 9.6×10 -5 , and the penalty parameter ρ = 0.06.

[0080] The effectiveness of the method proposed by the present invention is verified through actual cases below.

[0081] This embodiment conducts experiments on the CAVE database. This database contains 32 hyperspectral images taken under controlled indoor light sources. The spatial resolution of each hyperspectral image is 512×512 (pixels), and there are 31 spectral channels in total. The present invention divides the CAVE dataset into a training set (the first 20 images) and a test set (the last 12 images), takes the original hyperspectral images as the real clear hyperspectral images, and uses four types of blur kernels (a), (b), (c), and (d) to blur the clear hyperspectral images, and then adds Gaussian white noise with a standard deviation of 0.01 to obtain blurred hyperspectral images. Then, the method of the present invention is used for deblurring.

[0082] Among them:

[0083] (a) A Gaussian blur kernel with a spatial resolution of 15×15 (pixels) and a standard deviation of 1.6;

[0084] (b) A Gaussian blur kernel with a spatial resolution of 15×15 (pixels) and a standard deviation of 2.4;

[0085] (c) A circular blur kernel with a diameter of 7 pixels;

[0086] (d) A square blur kernel with a side length of 5 pixels.

[0087] To prove the effectiveness and novelty of the method, 2 comparison methods are selected for comparison: HLP and SSP.

[0088] The present invention uses the root mean square error (RMSE: Root Mean Square Error), peak signal-to-noise ratio (PSNR: Peak Signal to Noise Ratio), spectral angle mapper (SAM: Spectral Angle Mapper), and structural similarity (SSIM: Structural Similarity Index) to comprehensively evaluate the deblurring performance of hyperspectral images.

[0089] Table 1 Average values of RMSE, PSNR, SAM, and SSIM of the present invention and the prior art for the last 12 images in the CAVE database

[0090]

[0091] As can be seen from Table 1, the average values of each performance index of the present invention are the best under various fuzzy kernel conditions. This is because the present invention makes full use of the prior information of the deep denoising prior network to learn the clear hyperspectral image and combines the physical degradation model well. Therefore, this method is more effective and robust than other methods, further verifying the effectiveness of the present invention.

[0092] Another embodiment of the present invention further provides a hyperspectral image deblurring system, including:

[0093] A paired image set construction module, configured to obtain a first hyperspectral image Y and a second hyperspectral image X of the same scene from an image library, where the clarity of the second hyperspectral image X is higher than that of the first hyperspectral image Y; perform normalization processing on the first hyperspectral image Y and the second hyperspectral image X to construct a paired image set (Y; X);

[0094] A deep denoising neural network construction and training module, configured to construct a deep denoising neural network and train the deep denoising neural network using the paired image set (Y; X) to obtain a deep denoising prior network;

[0095] A deblurring module, configured to construct an objective function with a regularization term, decouple the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem, and iteratively solve the above three sub-problems until a termination condition is reached. The hyperspectral image denoising sub-problem is solved by the deep denoising prior network to complete the deblurring of the first hyperspectral image Y.

[0096] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps in the hyperspectral image deblurring method as described above are implemented.

[0097] Exemplarily, the computer program can be divided into one or more modules / units, and the one or more modules / units are stored in the computer-readable storage medium and executed by the processor to complete the steps in the hyperspectral image deblurring method of the present application. The one or more modules / units can be a series of computer-readable instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the server.

[0098] The server can be a computing device such as a smart phone, a notebook, a palm computer, and a cloud server. The server may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the server may further include more or fewer components, or combine certain components, or different components. For example, the server may further include input / output devices, network access devices, a bus, etc.

[0099] The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0100] The memory may be an internal storage unit of the server, such as the hard disk or memory of the server. The memory may also be an external storage device of the server, such as a plug-in hard disk equipped on the server, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Further, the memory may also include both the internal storage unit and the external storage device of the server. The memory is used to store the computer-readable instructions and other programs and data required by the server. The memory may also be used to temporarily store data that has been output or is to be output.

[0101] It should be noted that for the information interaction, execution process, etc. between the above-mentioned device / units, due to the same concept as the method embodiment, for the specific functions and the technical effects brought, reference may be specifically made to the method embodiment part, and details will not be repeated here.

[0102] Those skilled in the art can clearly understand that for the sake of convenience and brevity of description, only the above-mentioned division of each functional unit and module is used as an example. In actual applications, the above-mentioned functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction and do not limit the protection scope of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the foregoing method embodiment, and details will not be repeated here.

[0103] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, to implement all or part of the processes in the above-mentioned embodiment methods of this application, a computer program can be used to instruct the relevant hardware to complete. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the photographing device / terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk, or an optical disc, etc.

[0104] In the above embodiments, the descriptions of the various embodiments have their own emphases. For the parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0105] The above-described embodiments are only used to illustrate the technical solutions of this application, rather than to limit them; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for deblurring hyperspectral images, characterized in that, it includes: Obtain a first hyperspectral image Y and a second hyperspectral image X of the same scene from an image library, where the clarity of the second hyperspectral image X is higher than that of the first hyperspectral image Y; Perform normalization processing on the first hyperspectral image Y and the second hyperspectral image X; Add noise to the second hyperspectral image X to obtain a third hyperspectral image Z, construct a deep denoising neural network, and train the deep denoising neural network using the paired image set (Z; X) to obtain a deep denoising prior network; Construct an objective function with a regularization term, decouple the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem, and iteratively solve the above three sub-problems until the termination condition is reached. The hyperspectral image denoising sub-problem is solved by the deep denoising prior network to complete the deblurring of the first hyperspectral image Y.

2. The hyperspectral image deblurring method according to claim 1, characterized in that, The steps of constructing the deep denoising neural network and training the deep denoising neural network using the paired image set (Z; X) include: Construct a three-dimensional network module: From input to output, there are a three-dimensional convolutional layer 1 , a batch normalization layer b, and an activation layer c 1 ; Construct a deep denoising neural network: from input to output, it is a three-dimensional convolutional layer a 2 , activation layer c 2 , B 3D network modules and 3D convolutional layer a 3 , the network input is connected to the 3D convolutional layer a through a skip connection 3 The output of the stack is connected together with the input 1 ; 3D convolutional layer a 3 Contains a 3×3×3 filter with a step size of 1; Select two-thirds of the image pairs in the image set (Z; X) as the training set, and the remaining one-third as the test set; randomly select M image patches of size 1 / h of the image itself in each image of the training set, and then randomly flip and mirror the M image patches in each image; use the image patches in the first hyperspectral image Y after the above processing as the input of the network, and use the image patches in the second hyperspectral image X after the above processing as the label image; The loss function is defined as follows: ||F(Z; Θ) - X|| 1 where F(*) represents the mapping of the deep denoising network, and Θ represents the model parameters of the deep denoising network; Use the Adam optimization algorithm, with an initial learning rate of e, randomly select f samples each time for forward propagation, and the algorithm iterates g generations; after training is completed, obtain the model parameters of the deep denoising prior network, and thus obtain the deep denoising prior network.

3. The hyperspectral image deblurring method according to claim 2, characterized in that: The three-dimensional convolutional layer a 1 and the three-dimensional convolutional layer a 2 both contain 32 3×3×3 filters with a stride of 1; the activation layer c 1 and the activation layer c 2 are both ReLU functions; the number of three-dimensional network modules B = 8; in each image of the training set, M image patches with a size of 1 / h of the image size itself are randomly selected, where h = 4 and M = 100; the initial learning rate e = 0.0002, the number of samples f = 16, and the number of algorithm iterations g = 500.

4. The hyperspectral image deblurring method according to claim 1, characterized in that, In the step of constructing the objective function with a regularization term, the constructed objective function is: where \(H\) represents the fuzzy matrix, represents the data fidelity term, \(\varphi(X)\) represents the regularization term, and \(\lambda\) represents the regularization parameter.

5. The hyperspectral image deblurring method according to claim 1, characterized in that, The decoupling of the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem is completed using the Plug-and-Play algorithm based on the alternating direction method of multipliers; the termination condition is t = T, where t represents the number of iterations and T represents the threshold of the number of iterations.

6. The hyperspectral image deblurring method according to claim 5, characterized in that, The iteration number threshold T = 20.

7. The hyperspectral image deblurring method according to claim 5, characterized in that, The steps of decoupling the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem, and iteratively solving the above three sub-problems until the termination condition is reached include: Input the first hyperspectral image Y, the blur matrix H, the regularization parameter λ, and the penalty parameter ρ; initialize the deep denoising network mapping F(*), the deep denoising network model parameters Θ; initialize the second hyperspectral image X 0 = Y, the auxiliary variable Z 0 = X 0 , the dual variable U 0 = 0; the iteration number t = 0; Solve the least-squares subproblem according to the following formula to update X t+1 : X t+1 = (H T H + ρI) -1 (H T Y + ρ(Z t - U t )) Solve the hyperspectral image denoising sub-problem using F(*) and Θ according to the following formula, and update Z t+1 : Z t+1 = F(X t+1 + U t ; Θ) Update the dual variable U according to the following formula t+1 :[[]]END]] U t+1 = U t + X t+1 - Z t+1 Update the iteration count \(t\leftarrow t + 1\); Determine whether the termination condition is met. If it is met, output the deblurred image \(X\) of the first hyperspectral image \(Y\). T Otherwise, return to continue solving the least - squares sub - problem, the hyperspectral image denoising sub - problem, and the dual variable update sub - problem.

8. The hyperspectral image deblurring method according to claim 7, characterized in that, The regularization term parameter λ = 9.6×10 -5 , and the penalty parameter ρ = 0.

06.

9. A hyperspectral image deblurring system, characterized in that, it includes: A paired image set construction module, configured to obtain a first hyperspectral image Y and a second hyperspectral image X of the same scene from an image library, where the clarity of the second hyperspectral image X is higher than that of the first hyperspectral image Y; Perform normalization processing on the first hyperspectral image Y and the second hyperspectral image X; A deep denoising neural network construction and training module, configured to add noise to the second hyperspectral image X to obtain a third hyperspectral image Z, construct a deep denoising neural network and train the deep denoising neural network using the paired image set (Z; X) to obtain a deep denoising prior network; A deblurring module, configured to construct an objective function with a regularization term, decouple the objective function into a least squares sub-problem, a hyperspectral image denoising sub-problem, and a dual variable update sub-problem, and iteratively solve the above three sub-problems until a termination condition is reached. The hyperspectral image denoising sub-problem is solved by the deep denoising prior network to complete the deblurring of the first hyperspectral image Y.

10. A computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps in the hyperspectral image deblurring method described in any one of claims 1 to 8.

Images