A Hyperspectral Image Denoising Method Based on Residual Learning and Hybrid-Domain Attention

Abstract

The present invention discloses a hyperspectral image denoising method based on residual learning and hybrid-domain attention, mainly aiming at the problem of hyperspectral image noise pollution caused by factors such as electromagnetic interference, atmospheric disturbance, and imaging device limitations. First, perform overlapping band grouping operations on all bands of the hyperspectral image with complex mixed noise, and then use a sparse feature extraction module and a hybrid-domain attention-based module for each group to extract and process features, obtaining local spectral-spatial features. After connecting the local features, further use a module with the same network structure as the same branch to extract deep spectral-spatial features, and finally obtain the denoised hyperspectral image through a residual learning strategy.

Description

Technical Field

[0001] The present invention belongs to the field of hyperspectral image processing, and particularly relates to a hyperspectral image hybrid noise removal method based on residual learning and hybrid domain attention. Background Art

[0002] Hyperspectral imaging technology is a "spectrum-image integration" technology that combines spectral information reflecting the radiation of substances with image information reflecting the two-dimensional space of substances. The spectral bands of hyperspectral images cover the visible light and infrared spectra from 400 nm to 2500 nm, and contain dozens to hundreds of channels. Its rich spectral information has irreplaceable applications in many fields. For example, detecting the quality of food, exploring the distribution of geological minerals, monitoring environmental pollution, and military target recognition, etc. The insufficient optical imaging energy caused by the dense band channels of the imaging spectrometer and the influence of external environmental factors introduce noise, resulting in problems such as low quality and poor visual effects in the hyperspectral images we can observe. In addition, the existence of noise will also cause the spectral characteristics of the ground objects obtained to be "distorted" to a certain extent, thereby affecting the accuracy in subsequent hyperspectral image analysis and applications. Therefore, the denoising task of hyperspectral images is of great necessity and cannot be ignored.

[0003] At present, the common denoising methods for this task can be divided into three categories: filter-based methods, model optimization-based methods, and data-driven deep learning methods. 1) Filter-based methods aim to separate clean signals from noisy signals through filter operations, including Fourier transform, wavelet transform, non-local mean transform, and PCA principal component analysis transform, etc. This type of method is mainly for two-dimensional image denoising problems. When dealing with hyperspectral image data of high-dimensional cubes, it often leads to poor denoising effects due to ignoring the strong correlations between spectra. 2) Model optimization-based methods consider reasonable assumptions or priors of hyperspectral data, mapping noisy hyperspectral images to clean hyperspectral images to preserve spectral-spatial features, such as sparse low-rank algorithms, spatial adaptive total variation denoising algorithms, etc. The use of this type of method is currently a relatively common one, and it has strong assumptions about Gaussian noise. However, it cannot be ignored that this method ultimately boils down to the optimization problem of complex problems and requires multiple iterations to obtain the optimal solution, which has a high computational cost and time cost in the actual process. 3) Deep learning-based methods construct an automatic and hierarchical learning model from the data itself in an end-to-end manner to adapt to noise feature learning. Compared with the first two methods, learning-based methods can jointly consider spatial similarity and spectral correlation, better obtain the feature representation of hyperspectral images, and thus obtain high-quality denoising results. However, the current research on learning-based methods is still relatively scarce, and most of them focus on removing specific noise levels or types, and the research on mixed noise denoising is not sufficient enough. Summary of the Invention

[0004] The present invention provides a hyperspectral image denoising method based on residual learning and hybrid-domain attention, which extracts the unique correlation between the spectral and spatial information of hyperspectral images through a residual learning strategy and hybrid-domain attention, so as to achieve the purpose of recovering high-quality hyperspectral images from noise-contaminated hyperspectral images.

[0005] To solve the above technical problems, the present invention adopts the following technical solutions:

[0006] Step 1: Prepare hyperspectral image data as the original clean image data, and crop the downloaded hyperspectral image into two parts according to a ratio of 8:2: a training part and a test part, and perform dataset enhancement operations on each part respectively, and use this part as the noisy hyperspectral image X;

[0007] Step 2: Add simulated noise to the training part and the test part to obtain noisy image data, and use this part as the noisy hyperspectral image Y, and construct clean-noisy hyperspectral image pairs;

[0008] Step 3: Construct the hyperspectral image denoising network RMDAN based on residual learning and hybrid-domain attention: Group the high-dimensional hyperspectral data using a grouping strategy and perform specific feature processing operations on each group; adjust the obtained spectral-spatial features using hybrid-domain attention to obtain richer deep high-frequency features; finally, obtain the denoising result through residual connection;

[0009] Step 4: Train the hyperspectral image denoising network RMDAN based on residual learning and hybrid-domain attention using clean-noisy image pairs;

[0010] Step 5: Use the trained RMDAN network to perform denoising tasks on the test dataset.

[0011] Furthermore, the detailed operations for preparing and augmenting the dataset in Step 1 are as follows:

[0012] Step 1.1: Prepare hyperspectral image data as the original clean image data X, and crop the downloaded hyperspectral image into two parts according to the ratio of 8:2: the training part and the test part:

[0013] Step 1.2: To obtain a more diverse dataset, take the spatial pixel position (0, 0) as the origin for both the training part and the test data, and divide the entire hyperspectral image data into several sub-image blocks of N*N*C with a stride of stride, and further perform the following dataset augmentation operations:

[0014] data_augmentation(images, aug_num),

[0015] where aug_num includes flipping up and down, rotating clockwise / counterclockwise by a specified angle, etc.

[0016] Furthermore, the detailed operations for preparing clean-noisy hyperspectral image pairs in Step 2 are as follows:

[0017] Step 2.1: Model the noisy image data: In the present invention, various noise types will be converted into additive noise for approximate representation and processing. Therefore, the noise model considered in the present invention is:

[0018] Y = X + N,

[0019] where N represents a mixture of various different types of noise, Y represents the noisy hyperspectral image, and X represents the original clean hyperspectral image.

[0020] Step 2.2: Obtain the noisy data sample Y: Using the noise model established in 2.1, add simulated noise N to the original clean image to obtain the noisy hyperspectral image data Y.

[0021] Furthermore, the construction operation of the hyperspectral image denoising network RMDAN based on residual learning and hybrid-domain attention in step 3 is as follows:

[0022] Step 3.1: Group the noisy hyperspectral image according to the correlations and differences existing between its bands to obtain G groups. For the hyperspectral image Y, its spectral dimension can be divided into G mutually overlapping groups, i.e.:

[0023] {Y 1 , Y 2 ,......, Y G} = Group(Y),

[0024] where Group(·) represents the band grouping operation.

[0025] Step 3.2: Process each group of data Y g (g = 1, 2,......, G) respectively using a branch network composed of a sparse feature extraction module SFENet and a hybrid-domain attention module MDAB to obtain local spectral-spatial features. The detailed operation is as follows:

[0026] (1) First, use the sparse feature extraction module to extract sparse features Specifically, the sparse feature extraction module SFENet is composed of R sparse blocks SFEB connected in series. Each sparse block contains 8 convolutional blocks CB and atrous convolutional blocks ACB. Among them, the CB block represents a common convolutional block composed of Convolution + Batch Normalization + ReLU activation function, as Figure 2 shown in a); the ACB block represents an atrous convolutional block composed of Atrous Convolution + Batch Normalization + ReLU activation function, as Figure 2 shown in b). In addition, a cross-layer connection is added between the input and output of this module, and the input information is summed with the output of this module as the input of the next module to ensure better stability during the model training process. This process can be formally described as:

[0027] where f SFENet (·) represents the sparse feature extraction module SFENet.

[0028] (2) For the obtained sparse spectral-spatial features Use the hybrid-domain attention module MDAB to adjust the already learned features, capture cross-domain interaction features, and enable the network to pay attention to more high-frequency noise features to obtain deep local spectral-spatial features

[0029] Among them, f MDAB (·) represents the mixed-domain attention module MDAB. Specifically, MDAB contains three branches, which are used to model and extract the interaction features between three groups of two different dimensions.

[0030] Step 3.3: For the local spectral-spatial features obtained by all branch networks After further processing by convolution and connection in the spectral dimension, a global feature representation F with the same size as the input hyperspectral image is obtained G0 :

[0031] Among them, Cat(·) represents the feature concatenation operation, and Conv(·) represents the ordinary convolution operation.

[0032] Step 3.4: For the obtained global spectral-spatial feature F G0 , a global network with the same structure as the branch network is used for global feature extraction to further obtain the deep global feature F G1 , that is:

[0033] F G1 = f MDAB (f SFENet (F G0 ) + F G0 ),

[0034] Step 3.5: After the global feature F G1 is processed by convolution, the predicted residual noise estimate I res is obtained, and then the denoised hyperspectral image I fin is obtained by using the residual learning strategy:

[0035] I fin = I res + Y = Conv(F G1 ) + Y,

[0036] The hyperspectral image denoising method based on residual learning and hybrid-domain attention provided by the present invention learns the noise residual image through the sparse feature extraction, hybrid-domain attention module, etc., and then uses the global cross-layer connection to restore the final denoised result. The hyperspectral image denoising method based on residual learning and hybrid-domain attention provided by the present invention has the following advantages: 1) The present invention proposes a hyperspectral image denoising network RMDAN based on residual learning and hybrid-domain attention, which is mainly composed of a sparse feature extraction module SFENet, a hybrid-domain attention module MDAB, and residual connections. The sparse feature extraction module SFENet uses ordinary convolutional blocks CB and atrous convolutional blocks ACB to extract sparse spectral-spatial features, expanding the receptive field while not excessively increasing the network parameters. The hybrid-domain attention module MDAB is composed of three cross-domain branches in parallel, used to extract deep spectral-spatial features, aiming to make the network training pay more attention to the noise features, thereby obtaining more noise features and improving the restoration effect of hyperspectral images. 2) In view of the high-dimensional characteristics of hyperspectral image data, based on the characteristics that there are similarities between adjacent bands and differences between bands with a relatively large distance in hyperspectral images, the high-dimensional hyperspectral image is divided into multiple groups, and each group is separately subjected to feature processing to achieve the purpose of reducing the data dimension and the network parameters while making full use of the spectral correlation. 3) The present invention uses a variety of different types of noise for training to improve the denoising generalization ability of the network. Description of the Drawings

[0037] Figure 1 It is a flowchart of the operation steps of the present invention.

[0038] Figure 2 It is a schematic structural diagram of the ordinary convolutional block CB and the atrous convolutional block ACB of the present invention.

[0039] Figure 3 It is an architecture diagram of the hyperspectral image denoising network RMDAN based on residual learning and hybrid-domain attention of the present invention. Specific Implementation Method

[0040] Combined with Figure 1 、 3 As shown in, the specific steps of the present invention are as follows:

[0041] Step 1: Prepare the hyperspectral image data as the original clean image data, and cut the downloaded hyperspectral image into two parts according to the ratio of 8:2: the training part and the test part, and perform data set enhancement operations on them respectively. This part is used as the clean data set X.

[0042] Step 1.1: Download the open-source hyperspectral image dataset on the network, and divide it according to the 8:2 ratio in a specific direction of the image spatial dimension to obtain the training part and the test part of the data. For example, if the size of a hyperspectral image is W*H*C, it is divided into the training part and the test part according to the 8:2 ratio of W or H.

[0043] Step 1.2: In order to obtain a more diverse dataset, take the training part and the test part respectively with the spatial pixel position (0, 0) as the origin, and divide the entire hyperspectral image data into several sub-image blocks of N*N*C with a stride of stride, and perform the following dataset augmentation operations to expand the hyperspectral image dataset:

[0044] dataaugmentation(images, aug_num), where aug_num includes flipping up and down, rotating clockwise / counterclockwise by a specified angle, etc.

[0045] Step 2: Add simulated noise to the training part and the test part to obtain noisy image data, and construct a clean-noisy hyperspectral image pair. Take this part as the noisy hyperspectral image Y.

[0046] Step 2.1: Model the noise data:

[0047] The noise contained in the currently observable hyperspectral image can be divided into additive noise and multiplicative noise according to the relationship between the noise and the signal. The additive noise is not related to the signal intensity of the image itself and can be approximated as Gaussian white noise. Therefore, this type of noisy hyperspectral image Y can be regarded as the sum of the original clean hyperspectral image X and the noise A:

[0048] Y = X + A (1),

[0049] However, the multiplicative noise is related to the image signal and often changes with the change of the image signal. The sensor output signal is very sensitive to the photon signal, so the multiplicative noise part in the hyperspectral image cannot be ignored. In the present invention, it is converted into an approximate representation of additive noise:

[0050] Y = X + A*X (2),

[0051] To sum up, the noise data considered in the present invention can be formally expressed as:

[0052] Y = X + N (3),

[0053] Where Y represents the degraded hyperspectral image we can observe; X represents the clean hyperspectral image; N represents various types of noise such as Gaussian noise, impulse noise, sparse stripe noise, and dead lines. In this noise data model, X and N are independent of each other.

[0054] Step 2.2: Add simulated noise to the dataset X according to the noise data model described in 2.1 to obtain the input data Y of the model. Therefore, in the present invention, the modeling process of the hyperspectral image with mixed noise can be described as follows:

[0055] Y = X + N, where Y, X, N ∈ R N*N*C (4),

[0056] where N*N represents the spatial dimension size of the hyperspectral image block, i.e., the number of spatial pixels, and C represents the spectral dimension size of the hyperspectral image, i.e., the number of spectral channels.

[0057] Step 3: Combine ideas such as the grouping strategy, residual learning strategy, and hybrid-domain attention to construct a hyperspectral image denoising network RMDAN based on residual learning and hybrid-domain attention.

[0058] Step 3.1: In order to utilize the correlation between adjacent spectral bands of the hyperspectral image while reducing the data dimension and the number of model parameters, the noisy hyperspectral image Y is processed using the grouping strategy to obtain G groups Y 1 , Y 2 ,......, Y G , that is:

[0059] {Y 1 , Y 2 ,......, Y G} = Group(Y) (5),

[0060] where Group(·) represents the band grouping operation. The detailed grouping strategy is as follows:

[0061] (1) Calculate the correlation coefficient matrix between the bands of the hyperspectral image Y. The calculation method of the correlation coefficient between the i-th band and the j-th band is as follows:

[0062] where COV(i, j) represents the covariance between the i-th band and the j-th band, and D(i) represents the variance of the i-th band. The correlation coefficient matrix P Y of the hyperspectral image Y is:

[0063] P Y = (ρ i，j ) (7),

[0064] Obviously, P Y is a C*C symmetric triangular matrix with a diagonal of 1;

[0065] (2) Divide the hyperspectral image Y: According to the correlation coefficient matrix P Y, the hyperspectral image Y is divided into G groups. At the same time, the present invention takes into account that there is still a certain correlation between adjacent groups, so there are ovl overlapping bands set between adjacent groups. The specific grouping strategy can be described as follows:

[0066] 1) Input: correlation coefficient matrix (P Y ) C*C , overlap coefficient ovl;

[0067] 2) Initialization: i = 1, j = i + 1, start = [], end = [], G = 0;

[0068] 3) Group the spectral bands according to the process shown in the following pseudocode:

[0069]

[0070]

[0071] In this way, Y is divided into G groups, and the start band index number and end band index number of each group are respectively recorded in the lists start and end, that is:

[0072] Y = [Y 1 , Y 2 ,......, Y G (8),

[0073] where [...] represents the concatenation operation.

[0074] Step 3.2: Process each group of data respectively using the branch network composed of the sparse feature extraction module SFENet and the mixed-domain attention module MDAB to obtain local spectral-spatial features. The detailed steps are as follows:

[0075] (1) First, use the sparse feature extraction module to extract sparse features Specifically, the sparse feature extraction module SFENet is composed of R sparse blocks SFEB connected in series. Each sparse block contains 8 convolutional blocks CB and atrous convolutional blocks ACB. Among them, the CB block represents a common convolutional block composed of Convolution + Batch Normalization + ReLU activation function, as Figure 2 shown in a); the ACB block represents an atrous convolutional block composed of Atrous Convolution + Batch Normalization + ReLU activation function, as Figure 2 shown in b). In addition, cross-layer connections are added in this module, and the input information is summed with the output of this module as the input of the next module to ensure better stability during the model training process. The formal description of this process is:

[0076] Among them, f SFENet (·) represents the sparse feature extraction module SFENet. Specifically, the SFENet module is composed of 8 convolutional blocks in series. Therefore, the sparse spectral-spatial features extracted by SFENet are expressed as:

[0077]

[0078] Among them, 1, 2, 6, and 8 are CB ordinary convolutional blocks, represented by f CB (·); 3, 4, 5, and 7 are ACB dilated convolutional blocks, represented by f ACB (·).

[0079] In the sparse feature extraction module SFENet, dilated convolution is used and different dilation rates are set, and receptive fields of different sizes will be obtained. That is, without introducing additional parameters, multi-scale information is obtained. The use of dilated convolution needs to consider two issues: 1) When convolutional kernels with dilation rates greater than or equal to 2 are stacked multiple times, local information loss occurs in the obtained convolutional results, resulting in the grid effect; 2) Dilated convolution sparsely samples the input signal, making the information obtained from distant convolutions lack correlation, thus affecting the accuracy of the final task result.

[0080] To avoid the above problems, the dilation rates of the 8 convolutions in the SFENet module are set to 1, 1, 3, 5, 7, 1, 3, and 1 respectively to extract features of different scales. In addition, when the dilation rate is 1, the dilated convolution is equivalent to an ordinary convolution at this time.

[0081] (2) For the obtained sparse spectral-spatial features The hybrid domain attention module MDAB is used to adjust the learned features, capture cross-domain interaction features, make the network focus on more high-frequency noise features, and obtain a deep local spectral-spatial feature representation

[0082] Among them, f MDAB (·) represents the hybrid domain attention module MDAB. Specifically, MDAB contains three cross-domain branches, which are used to model and extract the interaction features between three groups of two different dimensions.

[0083] In hyperspectral images, different bands contribute differently to the final denoising result. Therefore, during the feature extraction process, it is necessary to adaptively and dynamically adjust the dependence of local band features and global features through an attention mechanism, enabling the network to pay more attention to noise features during training, extract richer noise information hidden in complex backgrounds, and thereby improve the efficiency of training the noise model and reduce complexity.

[0084] Specifically, the overall process of the three cross-domain branch attentions included in the mixed-domain attention module MDAB can be described as follows:

[0085]

[0086]

[0087]

[0088]

[0089]

[0090] Among them, Rot1 and Rot2 represent rotating 90° counterclockwise around the H and W dimensions of the space respectively, and represent their inverse processes. Z_Pool(·) represents connecting the average pooling feature and the maximum pooling feature on the first dimension of the input image. This operation can retain its rich features while reducing its depth and computational amount. For example, for the feature map X to perform the Z_Pool operation, it can be further described as:

[0091] T = Z_Pool (X) = [MaxPool(X); AvgPool(X)] (18),

[0092] Step 3.3: After the local spectral-spatial features obtained by all branch networks are processed by convolution and then connected in the spectral dimension, a global feature representation F of the same size as the input hyperspectral image is obtained G0 :

[0093] Among them, Cat(·) represents the feature connection operation, and Conv(·) represents the ordinary convolution operation. It should be noted that since there are ovl overlapping bands in adjacent groups, when performing the connection operation, the feature values in the overlapping bands need to be averaged to obtain the final global spectral-spatial features.

[0094] Step 3.4: For the obtained global spectral-spatial feature F G0, use the global network that is consistent with the branch network structure to extract global features, and obtain the deep global feature F G1 , that is:

[0095] F G1 = f MDAB (f SFENet (F G0 ) + F G0 ) (20),

[0096] Step 3.5: After processing the global feature F G1 using convolution, obtain the predicted residual noise estimate I res , and then use the residual learning strategy to obtain the denoised hyperspectral image I fin :

[0097] The residual learning strategy can learn a residual image that is more conducive to propagation in the network, that is, the difference between the input and the output. At the same time, it can largely avoid the problem of gradient disappearance in the network. Therefore, a global residual connection is added between the input and the output in the present invention. The noisy hyperspectral image of the input is added to the output end for co-located pixel-level summation operation to obtain the finally predicted denoised result I fin :

[0098] I fin = Y + I res = Y + Conv(F G1 ) (21),

[0099] Step 4: Use the clean-noisy image pair to train the hyperspectral image denoising network RMDAN based on residual learning and hybrid-domain attention.

[0100] Input the noisy training samples made in 2.2 into the denoising network for training, use the loss function described in Step 4.1 as the training objective, and Adam as the optimizer of the denoising network.

[0101] Step 4.1: Design the loss function: The hyperspectral image HSI contains rich spatial information and spectral information, and the noise is distributed in both the spatial domain and the spectral domain. Therefore, the loss function needs to measure both the spatial information and the spectral information. The present invention uses the mean square error L MSE and the spectral angle measurement error L Spec . In addition, a penalty term L punish for noise estimation is added to measure the prediction error of the residual noise. As follows:

[0102]

[0103]

[0104] Among them, N represents the amount of data in one training process, and X i ∈R H*W*C represents the clean original image, and Y i ∈R H*W*C represents the noisy image. Y i，j , X i，j respectively represent the spectral vectors of the noisy HSI and the clean HSI at the spatial position (i, j). ||·||2 represents the L2-norm, and <·, ·> represents the inner product operation of two vectors. Z represents the estimated residual noise, represents the truly added noise.

[0105] In summary, the complete loss function used in the present invention can be defined as:

[0106] L total = L MSE + L Spec + L punish (25)

[0107] Step 4.2: Input the training sample Y into the denoising network RMDAN based on residual learning and hybrid-domain attention for training. Use the loss function shown above as the optimization objective, and combine the optimizer Adam to optimize the network.

[0108] Step 5: Use the trained RMDAN network to perform the image denoising task on the test sample.

[0109] The above has made a detailed description of the hyperspectral image denoising method based on residual learning and hybrid attention mechanism provided by the present invention. However, obviously, the specific implementation form of the present invention is not limited thereto. For those of ordinary skill in the art, all obvious changes made without departing from the scope of the claims of the present invention are within the protection scope of the present invention.

Claims

1. A hyperspectral image denoising method based on residual learning and hybrid-domain attention, characterized in that The specific operation steps are as follows: Step 1: Prepare the hyperspectral image data as the original clean image data, and crop the downloaded hyperspectral image into two parts according to the ratio of 8:2: the training part and the test part, and perform hyperspectral dataset augmentation operations on each part respectively, and use this part as the original clean image X; Step 2: Add simulated noise to the training part and the test part to obtain the noisy image data, and use it as the noisy hyperspectral image Y; Step 3: Construct a hyperspectral image denoising network RMDAN based on residual learning and hybrid-domain attention: Combine the grouping strategy to group the high-dimensional hyperspectral data, and perform feature processing operations on each group respectively; Combine the hybrid-domain attention to adjust the obtained spectral-spatial features to obtain richer deep high-frequency features; Finally, obtain the denoising result through residual connection; Step 4: Train the hyperspectral image denoising network RMDAN based on residual learning and hybrid-domain attention; Step 5: Use the trained RMDAN network to perform denoising tasks on the test dataset; In step 3, the input high-dimensional hyperspectral image data is grouped in the spectral dimension according to the characteristics of correlation and difference between its bands to obtain G groups, that is: {Y 1 , Y 2 ,......, Y G} = Group(Y), Among them, Group(·) represents the band grouping operation, and the detailed steps include: (1) Calculate the correlation coefficient matrix between each band of the hyperspectral image Y. The calculation method of the correlation coefficient between the i-th band and the j-th band is as follows: Among them, COV(i, j) represents the covariance between the i-th band and the j-th band, and D(i) represents the variance of the i-th band; the correlation coefficient matrix P of the hyperspectral image Y v is as follows: P Y = (ρ i,j ) P Y is a C*C symmetric triangular matrix with a diagonal of 1; (2) Divide the hyperspectral image Y: According to the correlation coefficient matrix P Y , divide the hyperspectral image Y into G groups. At the same time, considering that there is still correlation between adjacent groups, so there are ovl overlapping bands set between adjacent groups. The specific grouping strategy is described as follows: 1) Input: Correlation coefficient matrix (P Y ) C*C , overlap coefficient ovl; 2) Initialization: i = 1, j = i + 1, start = [], end = [], G = 0; 3) Group the spectral bands according to the process shown in the following pseudocode: In this way, Y is divided into G groups, and the start band index number and the end band index number of each group are respectively recorded in the lists start and end; In step 3, the deep spectral-spatial feature extraction module based on hybrid-domain attention has the following specific processing process: (1) First, use the sparse feature extraction module to extract sparse features The sparse feature extraction module SFENet is composed of R sparse blocks SFEB connected in series. Each sparse block contains 8 convolutional blocks CB and atrous convolutional blocks ACB. The CB block represents a common convolutional block composed of Convolution + BatchNormalization + ReLU activation function, and the ACB block represents an atrous convolutional block composed of AtrousConvolution + BatchNormalization + ReLU activation function. In addition, cross-layer connections are added in this module, and the input information is summed with the output of this module as the input of the next module to ensure better stability during the model training process. The formal description of this process is as follows: Among them, f SFENet (·) represents the sparse feature extraction module SFENet, f CB (·) represents the CB ordinary convolution block, f ACB (·) represents the ACB atrous convolution block; (2) For the obtained sparse spectral-spatial features The learned features are adjusted using the Mixed Domain Attention Block (MDAB), while capturing cross-domain interaction features, enabling the network to focus on more high-frequency noise features and obtaining deep spectral-spatial features. Among them, f MDAB (·) represents the Mixed-Domain Attention Block (MDAB). The MDAB contains three branches and is used to model the interaction features between three groups of two different dimensions. Specifically, the overall process of the MDAB is described as follows: s1 = σ(Conv1(Z_Pool(T1))), s2 = σ(Conv2(Z_Pool(T2))), Among them, Rot1 and Rot2 represent counterclockwise rotations of 90° around the H and W dimensions of space respectively, and represent the reverse process thereof, and Z_Pool(·) represents concatenating the average pooling feature and the max pooling feature on the first dimension of the input image. This operation can retain its rich features while reducing its depth and computational amount; when the feature map X is subjected to the Z_Pool operation, it is further described as: T = Z_Pool(X) = [MaxPool(X); AvgPool(X)].

2. A hyperspectral image denoising method based on residual learning and hybrid-domain attention according to claim 1, characterized in that In step 1, the preprocessing operation of the hyperspectral image dataset and various hyperspectral image dataset augmentation means include the following detailed steps: Step 1.1: Download the open-source hyperspectral image dataset on the network, and crop it into two parts according to the ratio of 8:2 in a specific direction of the image spatial dimension. If the size of a hyperspectral image is W*H*C, then crop it into the training part and the test part according to the ratio of 8:2 of W or H; Step 1.2: In order to obtain a more diverse dataset, the training part and the test data are respectively centered at the spatial pixel position (0, 0), and the entire hyperspectral image data is divided into several N*N*C sub-image blocks with a stride of stride, and the following dataset augmentation operations are performed to expand the hyperspectral image dataset: data_augmentation(images, aug_num), Among them, the operations specified by aug_num include flipping up and down, rotating clockwise / counterclockwise by a specified angle.

Images