Hyperspectral image restoration method and device based on cyclic decoupling model

Abstract

This application relates to a method and apparatus for hyperspectral image restoration based on a cyclic decoupling model. The method comprises: constructing a mathematical model for hyperspectral image hybrid degradation based on pre-introduced descriptive factors; deriving multiple solution objectives using the mathematical model; constructing a hyperspectral image restoration network with a cyclic decoupling structure; designing multiple sub-loss functions based on the multiple solution objectives; training the hyperspectral image restoration network based on the multiple sub-loss functions to obtain a trained hyperspectral image restoration network; and restoring the image to be restored based on the trained hyperspectral image restoration network. This method can simultaneously achieve hyperspectral image denoising and defect information completion, significantly improving image quality.

Description

Technical Field

[0001] The present application relates to the technical field of spectral image restoration, and in particular to a hyperspectral image restoration method and device based on a cyclic decoupling model. Background Art

[0002] Hyperspectral images are susceptible to various types of noise contamination, including Gaussian noise, banding noise, dead pixels, and dead lines, due to a complex imaging chain. These degradations can lead to image quality degradation and affect subsequent hyperspectral image analysis and interpretation. Furthermore, due to the complex high-dimensional structure of hyperspectral data, different parts of the spatial domain or different bands in the spectral dimension may experience varying degrees of degradation, further exacerbating the challenges of hyperspectral image restoration.

[0003] Currently, the proposed hyperspectral image restoration techniques mainly treat denoising and missing information restoration as separate tasks. Denoising usually separates noise and radiation signals based on their texture appearance, while missing information restoration mainly uses consistent geometric structures, textures with repeated patterns, and their combination to complete spatial information. Various techniques such as frequency domain representation, total variation, sparse representation, low-rank matrix approximation, and deep networks have been used as basic models to restore hyperspectral denoising under different noise types and intensities. On the other hand, techniques such as tensor completion and total variation are often used for missing information restoration. Existing methods can handle some specific cases of hyperspectral degradation, but they ignore the following key issues: (1) The universal additive noise model cannot adapt to hyperspectral image degradation in more complex real-world scenarios. For example, occlusion caused by bad pixels or dead lines may completely block the entire area, resulting in information loss; (2) The spatial and spectral information of hyperspectral images are inherently related and complementary, but existing methods usually treat them as separate entities, and the potential benefits of jointly utilizing these two types of information have not been fully explored. Therefore, how to fully exploit the effective features of the spatial spectrum, develop robust hyperspectral image restoration technology, and achieve missing information completion while achieving high-fidelity denoising is of great research value. Summary of the Invention

[0004] Based on this, it is necessary to provide a hyperspectral image restoration method and device based on a cyclic decoupling model that can simultaneously achieve hyperspectral image denoising and defect information completion, thereby significantly improving image quality, in order to address the above technical problems.

[0005] A hyperspectral image restoration method based on a cyclic decoupling model, the method comprising:

[0006] A mathematical model of mixed degradation of hyperspectral images is constructed based on pre-introduced describable factors; multiple solution objectives are derived using the mathematical model of mixed degradation of hyperspectral images;

[0007] A hyperspectral image restoration network with a cyclic decoupling structure is constructed. Multiple sub-loss functions are designed based on multiple solution objectives. The hyperspectral image restoration network is trained according to the multiple sub-loss functions to obtain a trained hyperspectral image restoration network.

[0008] The image to be restored is restored according to the trained hyperspectral image restoration network.

[0009] A hyperspectral image restoration device based on a cyclic decoupling model, the device comprising:

[0010] A solution target deduction module is used to construct a hyperspectral image mixed degradation mathematical model based on pre-introduced describable factors; and deduce multiple solution targets using the hyperspectral image mixed degradation mathematical model;

[0011] An image restoration network training module is used to construct a hyperspectral image restoration network with a cyclic decoupling structure, design multiple sub-loss functions based on the multiple solution objectives, and train the hyperspectral image restoration network according to the multiple sub-loss functions to obtain a trained hyperspectral image restoration network;

[0012] The image restoration module is used to restore the image to be restored according to the trained hyperspectral image restoration network.

[0013] The above-mentioned hyperspectral image restoration method and device based on the cyclic decoupling model, the present application constructs a mathematical model of hyperspectral image mixed degradation according to a pre-introduced descriptive factor, the model is suitable for describing various hyperspectral image degradations including noise occlusion, so as to simultaneously solve the problems of hyperspectral image mixed noise and information loss degradation; it explores the characteristics of spatial consistency and spectral independence of hyperspectral images, and constructs a high-performance cyclic decoupling restoration network; it decouples the hyperspectral image restoration task into degradation type classification, texture sharing based on spectral coherence, and detail recovery and reconstruction based on feature fusion, and obtains more refined image restoration results. The present application can accurately predict the location and type of degradation suffered by the hyperspectral image, and the restored hyperspectral image has clear texture details and retains the intrinsic characteristics of the hyperspectral image. It has better quality than traditional methods in terms of qualitative and quantitative results. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 1 is a flow chart of a hyperspectral image restoration method based on a cyclic decoupling model in one embodiment;

[0015] Figure 2A schematic diagram of a hyperspectral image restoration network with a cyclic decoupling structure in one embodiment;

[0016] Figure 3 is a structural diagram of a binary degradation classification module in one embodiment;

[0017] Figure 4 This is a structural diagram of a feature fusion module in another embodiment;

[0018] Figure 5 In one embodiment, a hyperspectral image restoration device based on a cyclic decoupling model is provided;

[0019] Figure 6 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0021] In one embodiment, Figure 1 As shown, a hyperspectral image restoration method based on a cyclic decoupling model is provided, comprising the following steps:

[0022] Step 102: construct a hyperspectral image mixed degradation mathematical model based on the pre-introduced describable factors; and deduce multiple solution objectives using the hyperspectral image mixed degradation mathematical model.

[0023] The observed noisy hyperspectral image is denoted as Y, and the universal additive noise model is usually described as:

[0024] ;

[0025] in, represents the band index, and , is the total number of bands, represents additive noise, and Represents a clean image, which is the goal of hyperspectral image restoration;

[0026] Since the degradation of real hyperspectral images is more complex, the noise model cannot describe the problem of key pixel loss caused by noise occlusion. Therefore, this application designs a factor that can describe whether a pixel is defective. , and defines a new noise-sensitive hyperspectral image mixed degradation noise model as follows:

[0027] ;

[0028] Among them, D represents the distribution map of the noise occlusion, and is the corresponding occluded foreground, defined as follows:

[0029] ;

[0030] in, Indicates the area where the original signal is lost, that is, where the noise is blocked.

[0031] The above degradation model is converted into a pixel-level description as follows:

[0032] ;

[0033] in, Indicates the spatial position of the corresponding pixel point, Indices representing spectral dimensions.

[0034] Further, the deduction and solution target includes performing the following sub-steps: solving the above formula , this application defines the function Used for calculation , the function can be described as follows:

[0035] ;

[0036] in, for Neighborhood pixels of

[0037] The solution can be divided into the following two cases: , then the noise The estimated use function Described as:

[0038] ;

[0039] The target noise-free image The solution can be obtained by the following formula:

[0040] .

[0041] like , then the observed The signal at , the missing information needs to be inferred based on adjacent pixels in the spatial dimension or adjacent bands in the spectral dimension. The estimation of can be done by the function The implementation formula is as follows:

[0042] ;

[0043] The target noise-free image The solution formula is as follows:

[0044] ;

[0045] in, Represents adjacent pixels in the spatial dimension or the same pixel in adjacent bands in the spectral dimension without information defects. .

[0046] By pre-introducing describable factors, a mathematical model for the mixed degradation of hyperspectral images that is sensitive to noise occlusion is established, uniformly describing various degradation types such as Gaussian noise, stripe noise, and bad pixel or dead line occlusion. Compared with the traditional universal additive noise model, this model can adapt to more complex real-world scenarios, effectively solving the problems of mixed noise and information loss degradation in hyperspectral images, and providing a more practical theoretical basis for subsequent image restoration. Based on the mathematical model of mixed degradation, multiple solution objectives are derived. These objectives clarify the specific direction of image restoration, integrating the two major tasks of denoising and defect information completion into a unified framework, making the restoration process more targeted and providing target guidance for high-quality image restoration.

[0047] Step 104: construct a hyperspectral image restoration network with a cyclic decoupling structure, design multiple sub-loss functions based on multiple solution objectives, and train the hyperspectral image restoration network according to the multiple sub-loss functions to obtain a trained hyperspectral image restoration network.

[0048] like Figure 2 As shown in Figure 2, a hyperspectral image restoration network with a cyclic decoupling structure is constructed. Input the observed noisy hyperspectral image ,in, represents the noisy hyperspectral image at time K+1, represents the noisy hyperspectral image at time K-1, Indicates the next band, i.e. At this moment, the calculated fusion features are first used to build a convolutional neural network block FE based on residual connections. FE is composed of multiple convolutional layers combined with dense residual connections to extract shallow features of each band. :

[0049] ;

[0050] Establish a binary degradation classification module B, such as Figure 3 As shown, the B is composed of S convolution blocks and softmax activation function, where the S convolution block consists of two Convolutional layer, with a ReLU activation function layer in the middle. This module B inputs shallow features , and the fused features calculated in the previous band After the two S convolution blocks are cascaded, the output of the softmax layer is the representation of the occluded foreground image. , further, an S convolution block is used to estimate the intermediate layer features Continue passing it down:

[0051] ;

[0052] Establish a feature fusion module FU, the structure is as follows Figure 4 As shown, and Respectively represent the latent space features after the update gate and reset gate in the feature fusion module. Combined with the shallow features calculated on the current band image , intermediate layer features And the fused features calculated in the previous band Three parts of input are used to calculate the fusion features of the current band :

[0053] ;

[0054] Establish a noise estimation module E, which is consistent with the binary degradation classification module B. The noise estimation module E can be used to calculate :

[0055] ;

[0056] The texture feature preservation module T is established, and its structure is consistent with the binary degradation classification module B. Combining the characteristics of static imaging of hyperspectral images, the fused features calculated based on the previous band are , the texture features shared between bands are calculated through the texture feature preservation module T , used to fill in pixel defect areas in other bands:

[0057] ;

[0058] Establish an aggregation reconstruction module R, which is consistent with the binary degradation classification module B. The reconstructed noise-free band image is estimated by the aggregation reconstruction module R :

[0059] ;

[0060] Aggregate the reconstructed images from all bands , and the restored noise-free hyperspectral image is obtained .

[0061] This application constructs a high-performance recurrent decoupled restoration network that exploits the spatial consistency and spectral independence of hyperspectral images. This decouples the hyperspectral image restoration task into three submodules: degradation type classification, texture sharing based on spectral coherence, and detail recovery and reconstruction based on feature fusion. This architecture fully utilizes the spatial and spectral information of hyperspectral images, avoiding treating them as separate entities. This allows for cross-domain fusion of spatial and spectral information, resulting in more refined image restoration results.

[0062] According to each solution goal, multiple sub-loss functions are designed specifically, including the following:

[0063] For the reconstructed clean image , loss function The calculation formula is as follows:

[0064] ;

[0065] in, A clean image for reference;

[0066] For the estimated noise , loss function The calculation formula is as follows:

[0067] ;

[0068] in, is the difference between the reference noisy image and the reference clean image;

[0069] For the calculated texture sharing features , loss function The calculation formula is as follows:

[0070] ;

[0071] in, , is a high-frequency characteristic filter;

[0072] Foreground map of the estimated occluded area , loss function The calculation formula is as follows:

[0073] ;

[0074] in, Can be obtained based on reference noise;

[0075] Combining the above loss functions, we get the hybrid optimization objective :

[0076] ;

[0077] in, is the weight of each loss function.

[0078] Based on multiple objectives, multiple sub-loss functions are designed to supervise the training of the hyperspectral image restoration network from multiple dimensions, including pixels, structure, and spectrum. This multi-sub-loss function avoids the overfitting problem caused by a single objective. Its recurrent structure allows the network to perform multiple rounds of corrections on difficult-to-recover areas, gradually approaching the true value, significantly improving the network's restoration accuracy and generalization capabilities.

[0079] Step 106: Restoring the image to be restored according to the trained hyperspectral image restoration network.

[0080] Image restoration using a trained network: Using a trained hyperspectral image restoration network to restore the image to be restored accurately predicts the location and type of degradation experienced by the hyperspectral image. The restored hyperspectral image exhibits clear texture details, effectively removes various types of noise, accurately fills in defect information, and preserves the intrinsic characteristics of the hyperspectral image. Both qualitative and quantitative results demonstrate superior quality compared to traditional methods.

[0081] The above-mentioned hyperspectral image restoration method based on the cyclic decoupling model, this application constructs a mathematical model of hyperspectral image mixed degradation according to a pre-introduced descriptive factor. The model is suitable for describing various hyperspectral image degradations including noise occlusion, so as to simultaneously solve the problems of hyperspectral image mixed noise and information loss degradation; it explores the characteristics of spatial consistency and spectral independence of hyperspectral images and constructs a high-performance cyclic decoupling restoration network; it decouples the hyperspectral image restoration task into degradation type classification, texture sharing based on spectral coherence, and detail recovery and reconstruction based on feature fusion, and obtains more refined image restoration results. This application can accurately predict the location and type of degradation suffered by the hyperspectral image. The restored hyperspectral image has clear texture details and retains the intrinsic characteristics of the hyperspectral image. It has better quality than traditional methods in terms of qualitative and quantitative results.

[0082] In one embodiment, a mathematical model of hyperspectral image mixing degradation is constructed based on pre-introduced describable factors, including:

[0083] Design a factor that describes whether a pixel is defective The definition of the mixed degradation noise model of hyperspectral image is as follows:

[0084] ;

[0085] ;

[0086] Where D represents the distribution map of the noise occlusion, Indicates the area where the original signal is lost, that is, where the noise is blocked. represents additive noise, Represents a clean image, which is the goal of hyperspectral image restoration. Indicates the spatial position of the corresponding pixel.

[0087] In one embodiment, a mathematical model of hyperspectral image hybrid degradation is used to deduce multiple solution objectives, including:

[0088] The mathematical model of hyperspectral image mixing degradation is converted into a pixel-level description as follows:

[0089] ;

[0090] in, Indicates the spatial position of the corresponding pixel point, represents the index of the spectral dimension, Indicates observed pixels;

[0091] Defining a function Used for calculation ,function The expression is:

[0092] ;

[0093] in, for Neighborhood pixels of Indicates observed pixels;

[0094] Calculate separately The noise and The target noise-free image is obtained according to the distribution map of the noise or noise occlusion, that is, the target is solved.

[0095] In one embodiment, if , using the function The noise The estimate of is described as:

[0096] .

[0097] In one embodiment, if , using the function The distribution map of the noise occlusion The estimate of is described as:

[0098] .

[0099] In one embodiment, the target noise-free image is obtained according to the noise calculation:

[0100] .

[0101] In one embodiment, the target noise-free image is calculated based on the distribution map of the noise occlusion area as follows:

[0102] ;

[0103] in, Represents adjacent pixels in the spatial dimension or the same pixel in adjacent bands in the spectral dimension without information defects. express The solution function of .

[0104] In one embodiment, the hyperspectral image restoration network with a cyclic decoupling structure includes a convolutional neural network block based on residual connection, a binary degradation classification module, a feature fusion module, a noise estimation module, a texture feature preservation module, an aggregation reconstruction module and a reconstruction fusion module; the convolutional neural network block based on residual connection is used to extract the shallow features of each band in the input noisy hyperspectral image; the binary degradation classification module is used to cascade the shallow features and the fused features calculated in the previous band through two S convolution blocks to output the representation of the occluded foreground image; the S convolution block is used to represent the occluded foreground image The extraction is performed to obtain the intermediate layer features; the feature fusion module is used to calculate the fused features of the current band from the shallow features, the intermediate layer features and the fused features calculated from the previous band; the noise estimation module is used to perform noise estimation on the shallow features; the texture feature preservation module is used to calculate the texture features shared between bands based on the fused features calculated from the previous band; the aggregation reconstruction module is used to reconstruct the fused features of the current band to obtain a reconstructed noise-free band image; the reconstruction fusion module is used to aggregate the reconstructed noise-free band images on all bands to obtain a restored noise-free hyperspectral image.

[0105] In one embodiment, multiple sub-loss functions are designed based on multiple solution objectives, including:

[0106] For the reconstructed noise-free band image , loss function The calculation formula is as follows:

[0107] ;

[0108] in, A clean image for reference;

[0109] For the estimated noise , loss function The calculation formula is as follows:

[0110] ;

[0111] in, is the difference between the reference noisy image and the reference clean image;

[0112] For the calculated texture sharing features , loss function The calculation formula is as follows:

[0113] ;

[0114] in, , is a high-frequency characteristic filter;

[0115] Foreground map of the estimated occluded area , loss function The calculation formula is as follows:

[0116] ;

[0117] in, is the foreground image of the occluded area.

[0118] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0119] In one embodiment, a hyperspectral image restoration device based on a cyclic decoupling model is provided, the device comprising:

[0120] The solution target deduction module 502 is used to construct a hyperspectral image mixed degradation mathematical model based on the pre-introduced describable factors; and deduce multiple solution targets using the hyperspectral image mixed degradation mathematical model;

[0121] An image restoration network training module 504 is configured to construct a hyperspectral image restoration network having a cyclic decoupling structure, design a plurality of sub-loss functions based on the plurality of solution objectives, and train the hyperspectral image restoration network according to the plurality of sub-loss functions to obtain a trained hyperspectral image restoration network.

[0122] The image restoration module 506 is configured to restore the image to be restored according to the trained hyperspectral image restoration network.

[0123] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, memory, network interface, display screen and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a hyperspectral image restoration method based on a cyclic decoupling model is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0124] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0125] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0126] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and such modifications and improvements are intended to fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A hyperspectral image restoration method based on a cyclic decoupling model, characterized in that: The method comprises: Constructing a hyperspectral image mixed degradation mathematical model based on pre-introduced describable factors; and deducing multiple solution objectives using the hyperspectral image mixed degradation mathematical model; Constructing a hyperspectral image restoration network with a cyclic decoupling structure, designing multiple sub-loss functions based on the multiple solution objectives, and training the hyperspectral image restoration network according to the multiple sub-loss functions to obtain a trained hyperspectral image restoration network; Restoring the image to be restored according to the trained hyperspectral image restoration network; A mathematical model of hyperspectral image mixed degradation is constructed based on the pre-introduced descriptive factors, including: Design a factor that describes whether a pixel is defective The definition of the mixed degradation noise model of hyperspectral image is as follows: ； ； Where D represents the distribution map of the noise occlusion, Indicates the area where the original signal is lost, that is, where the noise is blocked. represents additive noise, Represents a clean image, which is the goal of hyperspectral image restoration. Indicates the spatial position of the corresponding pixel; The hyperspectral image restoration network with a cyclic decoupling structure includes a convolutional neural network block based on residual connection, a binary degradation classification module, a feature fusion module, a noise estimation module, a texture feature preservation module, an aggregation reconstruction module and a reconstruction fusion module; the convolutional neural network block based on residual connection is used to extract the shallow features of each band in the input noisy hyperspectral image; the binary degradation classification module is used to cascade the shallow features and the fused features calculated in the previous band through two S convolution blocks to output the representation of the occlusion foreground map; the S convolution block is used to extract the representation of the occlusion foreground map, and obtain to the intermediate layer features; the feature fusion module is used to calculate the fused features of the current band from the shallow features, the intermediate layer features and the fused features calculated from the previous band; the noise estimation module is used to perform noise estimation on the shallow features; the texture feature preservation module is used to calculate the texture features shared between bands based on the fused features calculated from the previous band; the aggregation reconstruction module is used to reconstruct the fused features of the current band to obtain a reconstructed noise-free band image; the reconstruction fusion module is used to aggregate the reconstructed noise-free band images on all bands to obtain a restored noise-free hyperspectral image.

2. The method according to claim 1, characterized in that The hyperspectral image mixed degradation mathematical model is used to deduce multiple solution objectives, including: The mathematical model of hyperspectral image mixing degradation is converted into a pixel-level description as follows: ； in, Indicates the spatial position of the corresponding pixel point, represents the index of the spectral dimension, Indicates observed pixels; Defining a function Used for calculation , the function The expression is: ； in, for Neighborhood pixels of Indicates observed pixels; Calculate separately The noise and The target noise-free image is obtained by calculating the distribution map of the noise or the noise occlusion, that is, solving the target.

3. The method according to claim 2, characterized in that The method further comprises: like , using the function The noise The estimate of is described as: 。 4. The method according to claim 2, characterized in that The method further comprises: like , using the function The distribution map of the noise occlusion The estimate of is described as: 。 5. The method according to claim 3, characterized in that The method further comprises: The target noise-free image obtained by the noise calculation is: 。 6. The method according to claim 4, characterized in that The method further comprises: The target noise-free image is calculated based on the distribution map of the noise occlusion area: ； in, Represents adjacent pixels in the spatial dimension or the same pixel in adjacent bands in the spectral dimension without information defects. express The solution function of .

7. The method according to claim 1, characterized in that Designing multiple sub-loss functions based on the multiple solution objectives includes: For the reconstructed noise-free band image , loss function The calculation formula is as follows: ； in, A clean image for reference; For the estimated noise , loss function The calculation formula is as follows: ； in, is the difference between the reference noisy image and the reference clean image; For the calculated texture sharing features , loss function The calculation formula is as follows: ； in, , is a high-frequency characteristic filter; Foreground map of the estimated occluded area , loss function The calculation formula is as follows: ； in, is the foreground image of the occluded area.

8. A hyperspectral image restoration device based on a cyclic decoupling model, characterized in that: The device comprises: The target deduction module is used to construct a mathematical model of hyperspectral image mixed degradation based on pre-introduced descriptive factors, including: Design a factor that describes whether a pixel is defective The definition of the mixed degradation noise model of hyperspectral image is as follows: ； ； Where D represents the distribution map of the noise occlusion, Indicates the area where the original signal is lost, that is, where the noise is blocked. represents additive noise, Represents a clean image, which is the goal of hyperspectral image restoration. Indicates the spatial position of the corresponding pixel point; derives multiple solution targets using the hyperspectral image mixed degradation mathematical model; An image restoration network training module is used to construct a hyperspectral image restoration network with a cyclic decoupling structure, design multiple sub-loss functions based on the multiple solution objectives, and train the hyperspectral image restoration network according to the multiple sub-loss functions to obtain a trained hyperspectral image restoration network; the hyperspectral image restoration network with a cyclic decoupling structure includes a convolutional neural network block based on residual connection, a binary degradation classification module, a feature fusion module, a noise estimation module, a texture feature preservation module, an aggregation reconstruction module and a reconstruction fusion module; the convolutional neural network block based on residual connection is used to extract the shallow features of each band in the input noisy hyperspectral image; the binary degradation classification module is used to convert the shallow features and the features calculated in the previous band into the features. The fused features are cascaded through two S convolution blocks to output a representation of the occluded foreground image; the S convolution block is used to extract the representation of the occluded foreground image to obtain the intermediate layer features; the feature fusion module is used to calculate the fused features of the current band from the shallow features, the intermediate layer features and the fused features calculated from the previous band; the noise estimation module is used to perform noise estimation on the shallow features; the texture feature preservation module is used to calculate the texture features shared between bands based on the fused features calculated from the previous band; the aggregation reconstruction module is used to reconstruct the fused features of the current band to obtain a reconstructed noise-free band image; the reconstruction fusion module is used to aggregate the reconstructed noise-free band images on all bands to obtain a restored noise-free hyperspectral image; The image restoration module is used to restore the image to be restored according to the trained hyperspectral image restoration network.

Images