A hyperspectral image denoising method based on unpaired unsupervised neural network

By combining unpaired unsupervised neural networks with U-net and 3D wavelet transform networks, the problem of relying on paired data in existing methods is solved, achieving efficient hyperspectral image denoising and improving image quality and local details.

CN114897737BActive Publication Date: 2024-11-12NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210572277.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-25
Publication Date
2024-11-12
Estimated Expiration
2042-05-25

AI Technical Summary

Technical Problem

Existing deep learning-based hyperspectral image denoising methods rely on paired data training, resulting in poor generalization performance. Furthermore, unsupervised methods cannot effectively learn mapping relationships, leading to low quality or data loss in the restored data.

Method used

An unpaired unsupervised neural network is employed, combining the U-net network and the three-dimensional wavelet transform network. The model is trained using the total loss function, and the outputs of multiple networks are fused by utilizing the subspace representation of hyperspectral data and principal component analysis to improve the denoising effect.

Benefits of technology

It achieves efficient denoising under unpaired data conditions, improves the quality of hyperspectral images, avoids the influence of noise, and maintains image clarity and local details.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114897737B_ABST
    Figure CN114897737B_ABST
Patent Text Reader

Abstract

The present invention discloses a hyperspectral image denoising method based on an unpaired unsupervised neural network, which specifically includes: generating a noisy data set through simulation of a hyperspectral image, and performing a matrix decomposition operation to form a low-dimensional projection matrix; calculating the unsupervised network loss function based on an unbiased risk estimate of minimizing the mean square error of the low-dimensional projection matrix; constructing a U‑net network for the low-dimensional projection matrix; taking another high-quality clean data set to construct a three-dimensional wavelet transform network; combining the above-constructed U‑net network with a three-dimensional wavelet transform network, obtaining a total loss function based on the loss function of the U‑net network and the reconstruction loss function of the three-dimensional wavelet transform network, and training an unpaired unsupervised neural network model. The present invention forms an unpaired input by inputting data of different contents, and uses additional information to guide the enhanced restored hyperspectral image, so as to restore a clean image with clearer edges, thereby solving the problem that it is difficult for unsupervised methods to learn good mapping.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a hyperspectral image denoising method based on an unpaired unsupervised neural network, and belongs to the field of hyperspectral image processing. Background Art

[0002] At present, thanks to the development of remote sensing technology and the enhancement of spectral resolution, the exploration and cognition of the spectral characteristics of ground objects have gradually deepened, and more and more hidden information in a narrow spectral range has been discovered, prompting hyperspectral remote sensing to become one of the important research directions in the field of remote sensing technology in the 21st century. In the application of hyperspectral remote sensing, after the imager is used to acquire data, the hyperspectral remote sensing data needs to be processed and analyzed. The main purpose of data preprocessing is to obtain a hyperspectral image with the best quality and the most realistic, so as to provide guarantee for subsequent data analysis and data application. Data preprocessing and data analysis are the premise and guarantee of data application, and are the key to expanding the breadth and depth of hyperspectral applications. Therefore, it is of great theoretical significance and practical application value to carry out research on key technologies in data preprocessing and data analysis. The present invention focuses on the research on denoising technology in hyperspectral remote sensing image data preprocessing.

[0003] In the past decade, many deep learning-based hyperspectral denoising methods have been proposed. Deep learning-based methods have also been used for hyperspectral denoising due to their high representation and strong learning ability. Deep learning-based methods use a learnable general network architecture to parameterize the model and introduce various loss functions to directly learn the nonlinear mapping from noisy hyperspectral data to clean hyperspectral data, such as HSID-CNN, HSI-SDeCNN and other methods, but these methods do not consider the spatial-spectral correlation of joint hyperspectral data and do not effectively utilize global correlation. In 2021, Wei et al. proposed a three-dimensional quasi-recurrent neural network QRNN3D with alternating direction for hyperspectral image denoising. The network can effectively embed domain knowledge with structural space, spectral correlation and global correlation along the spectral direction. In addition, the introduction of the alternating direction structure eliminates causal correlation without increasing the computational cost. The model can model spatial-spectral dependencies and is applicable to hyperspectral data with any number of channels.

[0004] Many supervised image denoising methods based on neural networks usually perform well and have better performance than traditional denoising algorithms. However, these denoising methods based on convolutional neural networks rely heavily on a large amount of paired noisy and clean hyperspectral image data for training. In addition, the models trained using synthetic paired noisy and clean hyperspectral image data have poor generalization performance in simulated data sets due to the large difference between synthetic noise and real noise. Unsupervised methods only use low-quality data and cannot learn the mapping relationship well, resulting in problems such as low quality and data loss in the recovered data. Summary of the invention

[0005] In view of the above problems, the present invention provides a hyperspectral image denoising method based on an unpaired unsupervised neural network. By using noisy hyperspectral data as input, the end-to-end training of the neural network can be quickly completed, solving the problem that the supervised method requires a large amount of paired noisy and clean data. At the same time, through the unpaired input form and the method of fusing multiple networks, a three-dimensional wavelet transform neural network branch is introduced, and the low-quality projection matrix output by the unsupervised network is enhanced through the network. The training is performed through the total loss function to guide the unsupervised network to learn better mapping.

[0006] The technical solution adopted by the present invention to solve the above technical problems is as follows:

[0007] A hyperspectral image denoising method based on an unpaired unsupervised neural network is characterized by comprising the following steps:

[0008] S1: The initial hyperspectral image generates a noisy data set through simulation, and the noisy data set is processed into image data in a general format of 128*128*31 size;

[0009] S2: performing a matrix decomposition operation on the image data after completing the above step S1 and representing it as a low-dimensional projection matrix;

[0010] S3: Calculating the unsupervised network loss function based on the unbiased risk estimation of minimizing the mean square error of the low-dimensional projection matrix in step S2;

[0011] S4: Construct U-net network for low-dimensional projection matrix;

[0012] S5: Take another high-quality clean data set, build a three-dimensional wavelet transform network to extract high-quality style features, and use principal component analysis to preprocess the additional information data input into the three-dimensional wavelet transform network;

[0013] S6: combining the U-net network described in step S4 with the three-dimensional wavelet transform network described in step S5, combining the loss function of the U-net network and the reconstruction loss function of the three-dimensional wavelet transform network as the total loss function, and training the unpaired unsupervised neural network model;

[0014] S7: Taking the synthesized noisy test set ICVL hyperspectral image as input, the denoised clear hyperspectral image is obtained through the output of the unpaired unsupervised neural network model described in step S6.

[0015] The specific implementation steps of the matrix decomposition operation described in step 2 are as follows:

[0016] A clean hyperspectral image It consists of B spectral vectors containing n pixels, which are added with Gaussian noise. The noisy hyperspectral observation image obtained by Decompose X into a spectral low-rank model X=DZ in the form of two matrix factor products, where the orthogonal matrix basis D=U is obtained by estimating the number of subspaces K based on the Hysime algorithm and the algorithm SVD K , U is the left singular value orthogonal matrix with non-increasing properties obtained by the SVD algorithm of Y, and Z is the coefficient matrix of X under the subspace basis D; considering the additive Gaussian noise under the condition of non-independent and identically distributed, it is necessary to process the non-independent and identically distributed Y into independent and identically distributed Right now Σ is the noise covariance matrix obtained by the HySime algorithm, and the clean hyperspectral data after denoising is restored by matrix multiplication and They are Corresponding The obtained orthogonal matrix basis and coefficient matrix;

[0017] The formula for calculating the unsupervised loss function in step S3 is as follows:

[0018] L SURE =L MSE +L div -σ 2

[0019] Among them, L SURE is the Stein unbiased risk estimation loss function based on subspace representation, L MSE is the mean square error loss, L div is the divergence loss, σ is the data noise level;

[0020] The L MSE The solution process is:

[0021]

[0022] Assumptions It is based on the subspace basis The projection matrix, the L div The solution process is:

[0023]

[0024] In the formula, is a Gaussian distribution matrix with mean 0 and variance 1, εb is the deviation of the projection matrix, and are the estimates of the two projection matrices output by the network. By changing the size of ε, the projection matrix has a very small deviation. ε is set in the range of 10 -9 to 10 -5 The noise level σ is obtained by The channel estimate is averaged, where Use the robust wavelet estimation function of Gaussian standard deviation to calculate and get the noise standard deviation list corresponding to each channel

[0025] The U-net network in step S4 is a basic architecture that is hierarchically constructed by a feature extraction layer, 5 pairs of encoder-decoder modules and a reconstruction layer. A jump connection of the same resolution is introduced between each pair of encoders and decoders. The reconstruction layer uses residual connection and three-dimensional deconvolution to restore the underlying clean hyperspectral projection image.

[0026] The high-quality clean data set in step S5 is the Harvard data set.

[0027] The reconstruction loss function L of the three-dimensional wavelet transform network in step S6 is WT is defined as follows:

[0028]

[0029] Where X PCA It is a low-dimensional feature projection matrix obtained through principal component analysis, i.e. PCA dimensionality reduction. represents a three-layer three-dimensional discrete wavelet transform neural network;

[0030] The total loss function L T =L SURE +λL WT , adding a weight term λ to the three-dimensional wavelet transform network branch to control the fusion enhancement of additional information;

[0031] Using the total loss function L T Train unpaired unsupervised network models.

[0032] A computer device comprises a memory, a processor and a computer program, wherein the processor is connected to the memory, the memory is used to store the computer program for implementing the hyperspectral image denoising method according to any one of claims 1 to 6, and when the computer device is running, the processor loads and executes the computer program stored in the memory, so that the computer device implements each step of the hyperspectral image denoising method according to any one of claims 1 to 6.

[0033] Compared with the prior art, the technical solution of the present invention has the following technical effects:

[0034] The present invention utilizes a subspace representation method to project hyperspectral data into a low-dimensional subspace, which speeds up network training while simplifying the Stein loss function's direct use of high-dimensional data for calculation. At the same time, different content data sets are introduced by constructing wavelet network branches and unsupervised network branches to form unpaired inputs, and the principal component analysis method is used to control the data dimension of the wavelet transform network output, so that the outputs of the two branches are fused in a low-dimensional subspace. The influence of the wavelet network branch is controlled by a weight factor, so that the branch is prevented from reducing the quality of the hyperspectral image finally reconstructed. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall framework diagram of the unpaired unsupervised neural network model described in the present invention;

[0036] Figure 2 Channel 15 is selected for the ICVL test set of the embodiment of the present invention to obtain a grayscale image of the experimental results obtained by the comparative test and the method proposed in the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them.

[0038] Example 1

[0039] In this embodiment, Figure 1 FIG. 1 is an overall framework diagram of the unpaired unsupervised neural network model of the present invention. The hyperspectral image denoising method based on an unpaired unsupervised neural network described in this embodiment specifically includes the following steps:

[0040] Step 1: Generate a noisy dataset from the original ICVL (Interdisciplinary Computational Vision Laboratory) dataset through simulation, and process the noisy dataset into image data in a general format of 128*128*31;

[0041] Step 2: Perform matrix decomposition operation on the image data completed in step 1 above;

[0042] In order to make full use of the global low rank and non-local self-similarity of hyperspectral data, the following matrix decomposition method is selected. Assuming a clean hyperspectral image X is corrupted by additive Gaussian noise N, the noisy hyperspectral observation image is obtained. The observation image X = X + N is decomposed into a spectral low rank model Y = DZ + N in the form of two matrix factor products. Among them, It is the K basis matrix of the generated subspace, K<<n, which is used to capture redundant spectral information data. D and K can be obtained by the minimum error-based hyperspectral signal subspace identification HySime algorithm. is the coefficient matrix of X under the subspace basis D. The denoising problem can be defined as a least squares problem under constraints:

[0043]

[0044] Considering the case of additive Gaussian noise under non-IID conditions, it is necessary to first convert the independent and identically distributed observation data into independent and identically distributed data and perform whitening processing is the noise covariance matrix obtained by the HySime algorithm through Y.

[0045] Directly from Learning subspaces in The first k left singular vectors of approximately span the signal subspace yes The left singular value orthogonal matrix with non-increasing properties obtained by the SVD algorithm is Combined with HySime calculation to get the value of k, we can get the orthogonal matrix subspace basis Solve the coefficient matrix Z by the above expression 1:

[0046]

[0047] Define the neural network U-net that processes the projection matrix as Θ(·), let the projection matrix be the network input and pass the expression 2 to the coefficient matrix The solution can be expressed as:

[0048]

[0049] The estimated denoised hyperspectral data is obtained by the following inverse transformation:

[0050]

[0051] Step 3: Based on the unbiased risk estimation of minimizing the mean square error of the projected image sparse matrix, the divergence term is calculated using Monte Carlo to calculate the unsupervised network loss function;

[0052] By introducing Stein's unbiased risk estimation, an unsupervised loss function is constructed based on subspace representation. The calculation method of the SURE unsupervised convolutional neural network loss function based on subspace representation is as follows:

[0053] L SURE =L MSE +L div -σ 2 (5)

[0054] L SURE is the unbiased estimated loss function of the unsupervised network branch, L MSE is the mean square error loss, L div is the divergence loss under a minimum image difference εb;

[0055] In the present invention, the first item L MSE The solution process is:

[0056]

[0057] Assumptions It is based on the subspace basis The projection matrix, the second term L div The solution process is:

[0058]

[0059] In the formula, is a Gaussian distribution matrix with mean 0 and variance 1, εb is the deviation of the projection matrix, and are the estimates of the two projection matrices output by the network. By changing the size of ε, the projection matrix has a very small deviation. In this experiment, ε is set in the range of 10 -9 to 10 -5 , the noise level σ can be obtained by The channel estimate is averaged, where Use the robust wavelet estimation function of Gaussian standard deviation to calculate and get the noise standard deviation list corresponding to each channel

[0060] Step 4: Construct a U-net network for the projection image sparse matrix;

[0061] The spatial information processing module of the U-net network consists of a feature extraction layer, 5 pairs of encoder-decoder modules and a reconstruction layer. In order to better learn the information of each layer, a jump connection is introduced between the encoder and decoder of each layer. The U-net network architecture of the spatial information processing module is as follows: Figure 1 As shown; the purpose of the feature extraction layer is to extract shallow features from the input projection matrix through three-dimensional convolution. The three-dimensional convolution adaptively processes the projection matrix of any channel, while capturing the local spatial correlation and channel statistical correlation between the projection matrices. By cascading multiple downscale blocks, the encoder gradually reduces the spatial resolution by half and doubles the feature dimension to produce a multi-scale spatial representation; the decoder restores the resolution of the feature map through a series of upscaling and deconvolution blocks. Each upscale block doubles the spatial resolution of the feature map and reduces the number of channels by half. In addition, a jump connection of the same resolution is established between the encoder and the decoder to propagate shallow features to deeper layers, which is convenient for maintaining information during training and avoiding gradient disappearance. The reconstruction layer uses residual connections and three-dimensional deconvolution to restore the underlying clean projection matrix.

[0062] Step 5: Take the clean hyperspectral Harvard dataset and build a three-dimensional wavelet transform network to extract high-quality style features;

[0063] By moving and cropping each data, 50 128*128*31 hyperspectral data blocks are obtained as additional clean information data sets, and 100 of them are randomly selected as the 100 ICVL data corresponding to the introduced additional information for training;

[0064] The principal component analysis (PCA) method is used to project the additional information data of the three-dimensional wavelet transform network input into a low-dimensional matrix. The high-quality style features obtained by the wavelet network branch are fused with the output of the spatial information processing module. The high-quality style feature low-dimensional matrix obtained by this branch enhances the estimation of the projection matrix of the unsupervised network data. Finally, the inverse transformation is used to Get the estimated clean hyperspectral data.

[0065] The wavelet reconstruction loss function in the total loss function L is defined as follows:

[0066]

[0067] Where X PCA It is the low-dimensional feature projection matrix obtained by PCA dimensionality reduction. Represents a three-layer three-dimensional discrete wavelet transform neural network.

[0068] Step 6: Combine the U-net network described in step 4 with the three-dimensional wavelet transform network described in step 5, and use the loss function of the U-net network and the reconstruction loss function of the three-dimensional wavelet transform network as the total loss function. T =L SURE +λL WT Train unpaired unsupervised network models.

[0069] In order to verify the effectiveness of the present invention, we use different mainstream denoising methods to test on the same data set and compare them with the embodiments of the present invention.

[0070] This example uses the Pytorch library in the Python environment on a GeForce GTX 2080 Ti processor. In the experiment, the Adam optimizer with a learning rate of 0.001 is used. In the training phase, the number of iterations is set to 400 and the learning rate decay is set to 0.35. The learning rate decay step is set to 80. After the training is completed and the network weights are saved, the results on the validation set are as follows: Figure 2 shown.

[0071] Comparative Example 1: According to the LRTA algorithm in (Renard N, Bournane S, Blanc-Talon J.Denoising and Dimensionality Reduction Using Multilinear Tools for Hyperspectral Images[J].IEEE Geoscience&Remote Sensing Letters, 2008, 5(2):138-142.);

[0072] Comparative Example 2: According to the PARAFAC algorithm in (Liu X, Bournane S, Fossati C. Denoising of Hyperspectral Images Using the PARAFAC Model and Statistical Performance Analysis [J]. IEEE Transactions on Geoscience & Remote Sensing, 2012, 50 (10): 3717-3724.);

[0073] Comparative Example 3: According to the FORPDN algorithm in (Rasti B, Sveinsson JR, Ulfarsson MO, et al. Hyperspectral image denoising using first order spectral roughness penalty in wavelet domain [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2013, 7 (6): 2458-2467.);

[0074] Comparative Example 4: According to the LTDL algorithm in (Gong X, Chen W, Chen JA Low-Rank Tensor Dictionary Learning Method for Hyperspectral Image Denoising[J]. IEEE Transactions on Signal Processing, 2020, 68: 1168-1180);

[0075] Comparative Example 5: According to the NGMEET algorithm in (He W, Yao Q, Li C, et al. Non-local Meets Global: An Iterative Paradigm for Hyperspectral Image Restoration [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44 (4): 2089-2107.);

[0076] Comparative Example 6: According to the SURE-CNN algorithm in (Nguyen HV, Ulfarsson MO, Sveinsson J R. Hyperspectral image denoising using SURE-based unsupervised convolutional neural networks [J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 59 (4): 3369-3382.);

[0077] Comparative Example 7: According to the QRNN3D algorithm in (K. Wei, Y. Fu and H. Huang. 3-D Quasi-Recurrent Neural Network for Hyperspectral Image Denoising. IEEE Transactions on Neural Networks and Learning Systems, 2021, 32(1): 363-375).

[0078] Table 1

[0079]

[0080] Table 1 is a comparison of the indicators on hyperspectral data obtained by testing the present invention and other methods on 50 ICVL data in the test set.

[0081] In order to quantitatively analyze the effectiveness of the proposed method, a result graph in the ICVL dataset is selected. In order to visualize the denoising effect, the result graph is synthesized by the pseudo-color image of the three channels (25, 18, 5), which can intuitively evaluate the clarity of the image and the local detail information. It can be observed that the results of LRTA and PARAFAC have obvious stripe traces. Although the results of the LTDL method have no obvious stripe traces, they are blurred overall and locally. The overall visual effects of the results obtained by the deep learning methods SURE-CNN, QRNN3D and the proposed method are not much different, but through local details, it can be seen that SURE-CNN contains more unremoved noise points. Although the indicators of the results obtained by QRNN3D and the method proposed in this embodiment are not much different, through observation, the results obtained by the method proposed in this embodiment are clearer locally.

[0082] This embodiment also provides a computer device, including a memory, a processor and a computer program, wherein the processor is connected to the memory, and the memory is used to store the computer program for implementing the above-mentioned hyperspectral image denoising method. When the computer device is running, the processor loads and executes the computer program stored in the memory, so that the computer device implements each step of the above-mentioned hyperspectral image denoising method.

[0083] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.

Claims

1. A hyperspectral image denoising method based on unpaired unsupervised neural network, characterized in that: The specific steps include: S1: The initial hyperspectral image generates a noisy data set through simulation, and the noisy data set is processed into image data in a general format of 128*128*31 size; S2: performing a matrix decomposition operation on the image data after completing the above step S1 and representing it as a low-dimensional projection matrix; S3: Calculating the unsupervised network loss function based on the unbiased risk estimation of minimizing the mean square error of the low-dimensional projection matrix in step S2; S4: Construct U-net network for low-dimensional projection matrix; S5: taking another high-quality clean data set, constructing a three-dimensional wavelet transform network, and preprocessing the additional information data input into the three-dimensional wavelet transform network by principal component analysis; S6: combining the U-net network described in step S4 with the three-dimensional wavelet transform network described in step S5, combining the loss function of the U-net network and the reconstruction loss function of the three-dimensional wavelet transform network as the total loss function, and training the unpaired unsupervised neural network model; S7: Taking the synthesized noisy test set ICVL hyperspectral image as input, the denoised clear hyperspectral image is obtained through the output of the unpaired unsupervised neural network model described in step S6.

2. The hyperspectral image denoising method based on unpaired unsupervised neural network according to claim 1 is characterized in that: The specific implementation steps of the matrix decomposition operation described in step 2 are as follows: A clean hyperspectral image It consists of B spectral vectors containing n pixels, which are added with Gaussian noise. The noisy hyperspectral observation image obtained by The X is decomposed into a spectrum low-rank model X=DZ in the form of two matrix factor products, wherein the subspace number K is estimated by the minimum error-based hyperspectral signal recognition algorithm (Hysime algorithm) and the orthogonal matrix basis D=U is obtained by the algorithm SVD K , U is the left singular value orthogonal matrix with non-increasing properties obtained by the SVD algorithm of Y, and Z is the coefficient matrix of X under the subspace basis D; considering the additive Gaussian noise under the condition of non-independent and identically distributed, it is necessary to process the non-independent and identically distributed Y into independent and identically distributed Right now Σ is the noise covariance matrix obtained by the HySime algorithm, and the clean hyperspectral data after denoising is restored by matrix multiplication and They are Corresponding The obtained orthogonal matrix basis and coefficient matrix.

3. The hyperspectral image denoising method based on unpaired unsupervised neural network according to claim 1 is characterized in that: The formula for calculating the unsupervised loss function in step S3 is as follows: L SURE =L MSE +L div -σ 2 Among them, L SURE is the Stein unbiased risk estimation loss function based on subspace representation, L MSE is the mean square error loss, L div is the divergence loss, σ is the data noise level; The L MSE The solution process is: Assumptions It is based on the subspace basis The projection matrix, the L div The solution process is: In the formula, is a Gaussian distribution matrix with mean 0 and variance 1, εb is the deviation of the projection matrix, and are the estimates of the two projection matrices output by the network. By changing the size of ε, the projection matrix has a very small deviation. ε is set in the range of 10 -9 to 10 -5 , the noise level σ is obtained by taking The channel estimate is averaged, where Use the robust wavelet estimation function of Gaussian standard deviation to calculate and get the noise standard deviation list corresponding to each channel 4. The hyperspectral image denoising method based on unpaired unsupervised neural network according to claim 1, characterized in that: The U-net network in step S4 is a basic architecture that is hierarchically constructed by a feature extraction layer, 5 pairs of encoder-decoder modules and a reconstruction layer. A jump connection of the same resolution is introduced between each pair of encoders and decoders. The reconstruction layer uses residual connection and three-dimensional deconvolution to restore the underlying clean hyperspectral projection image.

5. The hyperspectral image denoising method based on unpaired unsupervised neural network according to claim 4 is characterized in that: The high-quality clean data set in step S5 is the Harvard data set.

6. A hyperspectral image denoising method based on an unpaired unsupervised neural network according to any one of claims 3 to 5, characterized in that: The reconstruction loss function L of the three-dimensional wavelet transform network in step S6 is WT is defined as follows: Where X PCA It is a low-dimensional feature projection matrix obtained by principal component analysis PCA dimensionality reduction. represents a three-layer three-dimensional discrete wavelet transform neural network; The total loss function L T =L SURE +λL WT , adding a weight term λ to the three-dimensional wavelet transform network branch to control the fusion enhancement of additional information; Using the total loss function L T Train unpaired unsupervised network models.

7. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program, wherein the processor is connected to the memory, the memory is used to store a computer program for implementing the hyperspectral image denoising method according to any one of claims 1 to 6, and when the computer device is running, the processor loads and executes the computer program stored in the memory, so that the computer device implements each step of the hyperspectral image denoising method according to any one of claims 1 to 6.