A method and system for deep denoising of hyperspectral images based on a two-stage learning framework

Through a two-stage learning framework and a multi-scale cross-fusion network, combined with the global spatial spectral module and feature cross-fusion transformer, the problem of insufficient mining of global spatial and spectral related information in hyperspectral image denoising is solved, and non-independent homogeneous noise is realized The prediction and image depth reduction improve image quality.

CN114677300BInactive Publication Date: 2024-11-01XI AN JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210302853.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-25
Publication Date
2024-11-01
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing hyperspectral image denoising methods have shortcomings in mining global space and spectral related information, resulting in image distortion and affecting subsequent application effects.

Method used

Using a two-stage learning framework method, the objective function is constructed through noise estimation and image denoising, and the conditional estimation subnet and multi-scale cross-fusion noise denoising subnet are constructed. The objective function is used to train the network, infer the noise level and perform hyperspectral. Image denoising, combined with global spatial spectral module and feature cross-fusion transformer, enhances the network's information fusion capabilities.

Benefits of technology

The prediction of non-independent homogeneous noise and depth reduction of hyperspectral images are achieved, which solves the problem of image distortion and improves the quality of the image and the effect of subsequent applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114677300B_ABST
    Figure CN114677300B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for deep denoising of hyperspectral images based on a dual-stage learning framework, which belongs to the technical field of hyperspectral image denoising. A three-dimensional deep universal model 3D-DUSSD for hyperspectral image denoising includes the following steps: constructing an objective function based on noise estimation and image denoising; constructing a conditional estimation subnetwork (CENet) and a multi-scale cross-fusion denoising subnetwork based on the objective function; training the conditional estimation subnetwork (CENet) and the multi-scale cross-fusion denoising subnetwork using the objective function; inferring the noise level of the hyperspectral image using the trained conditional estimation subnetwork (CENet), and then passing the information into the multi-scale cross-fusion denoising subnetwork for denoising the hyperspectral image. The method can realize the prediction of non-iid noise distribution and realize deep denoising of hyperspectral images at the same time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperspectral image denoising, and relates to a method and system for deep denoising of hyperspectral images based on a dual-stage learning framework. Background Art

[0002] Hyperspectral data provides continuous or discontinuous 10nm bands in the 400–2500nm region of the electromagnetic spectrum. This three-dimensional (3D) data, composed of two-dimensional (2D) spatial information and one-dimensional (1D) spectral information, has been widely used in fields such as environmental monitoring, precision agriculture, mineral identification, and military surveillance. However, HSI acquisition and transmission are often affected by various factors and are inevitably contaminated by various types of noise, including Gaussian noise, deadline noise, streak noise, and impulse noise, which seriously affects subsequent HSI classification and target recognition operations. Therefore, HSI denoising is a very important image preprocessing task.

[0003] Numerous HSI denoising methods have been proposed. Initially, these methods were based on improvements to conventional image denoising methods. However, unlike conventional images, HSI takes the form of a three-dimensional cube, and the noise distribution in HSI is much more complex. As a result, these improved methods may overlook spectral features, lose spatial and spectral information, and often result in significant spectral distortion. Therefore, to fully utilize the spatial and spectral information in HSI, denoising methods specifically tailored to the HSI structure are needed. Existing traditional HSI denoising methods can be categorized into two main categories: those based on spatial-spectral transformations and those based on transform domains. Spatial-domain methods include a spectral-space-adaptive total variation denoising model proposed by Yuan et al., a total variation regularized low-rank matrix factorization model for HSI denoising proposed by He et al., and a non-negative Tucker decomposition model based on nonlocal similarity proposed by Bai et al. These methods have the advantage of effectively leveraging hyperspectral image priors, determining the regularization term and denoising method based on prior knowledge, and preserving the original HSI characteristics. Methods based on variation and summation include principal component analysis, wavelet transform, and Fourier transform. These methods have strong mathematical theoretical support and good interpretability. However, they often assume independent and identically distributed (iid) Gaussian noise, which deviates from the real HSI, which contains various complex noises. Furthermore, model hyperparameters are difficult to select, and implementation is slow.

[0004] In recent years, due to the increase in data volumes and computing power, convolutional neural network-based HSI denoising algorithms have been proposed to replace costly optimization and complex a priori algorithms. For example, Xie et al. proposed a deep convolutional neural network (DCNN) with trainable nonlinear functions for HSI denoising, but each spectral band image was still denoised separately. Chang et al. pioneered the application of deep convolutional neural networks for HSI denoising. First, they directly learned filters in each layer without destroying the spectral spatial structure and simultaneously accommodated various noise types in HSI. Dong et al. proposed a computationally efficient implementation of a three-dimensional U-net based on a separable filtering strategy. By decomposing 3D filtering into 2D spatial filtering and 1D spectral filtering, the network parameters were simplified and the model's generalization ability was enhanced. This not only addresses the problem that existing DCNNs cannot fully exploit the spatial-spectral correlation of three-dimensional hyperspectral images, but also can handle unknown noise levels and mixed noise types. Zhang et al. designed a spatial spectral gradient network (SSGN) that uses a spatial spectral gradient learning strategy and incorporates complementary information to compensate for the unique spatial structure, directionality, and spectral variations of sparse noise. This effectively extracts the intrinsic and deep features of HSI and is effective in removing mixed noise. Therefore, deep learning-based algorithms are more effective than traditional HSI denoising methods. Among deep learning-based algorithms, 3D convolution-based HSI denoising methods produce better results than those using 2D convolution. Unfortunately, these methods have significant shortcomings in mining global spatial and spectral information related to HSI, resulting in image distortion and affecting subsequent applications. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems in the prior art and provide a method and system for deep denoising of hyperspectral images based on a two-stage learning framework, aiming to solve the technical problems in the prior art of mining HSI global spatial and spectral related information, which leads to image distortion and affects the subsequent application effect of the image.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention proposes a method for deep denoising of hyperspectral images based on a two-stage learning framework, comprising the following steps:

[0008] Construct objective function based on noise estimation and image denoising;

[0009] Construct a conditional estimation sub-network and a multi-scale cross-fusion denoising sub-network based on the objective function;

[0010] The objective function is used to train the conditional estimation sub-network and the multi-scale cross-fusion denoising sub-network;

[0011] The noise level of the hyperspectral image is inferred using the trained conditional estimation subnetwork;

[0012] The noise level information of the hyperspectral image is passed into the trained multi-scale cross-fusion denoising sub-network to denoise the hyperspectral image, thereby achieving denoising of the hyperspectral image.

[0013] Preferably, the specific process of constructing the objective function based on noise estimation and image denoising is as follows:

[0014] The noisy hyperspectral image HSI is linearly modeled as shown in formula (1):

[0015] y=x+n (1)

[0016] Where y, x, n∈R H×W×B , y, x, n represent the noisy hyperspectral image HSI, the noise-free hyperspectral image HSI and the random added noise that causes the degradation of the hyperspectral image HSI, respectively. H, W and B represent the spatial height, spatial width and the number of spectral bands, respectively. The hyperspectral image HSI restoration task is a traditional statistical inference problem. A maximum a posteriori MAP inference model is established as shown in formula (2):

[0017]

[0018] A new random variable is introduced to calculate the logarithmic posterior probability k, which contains the prior information of the denoising task. The reformulated form of p(x|y) is shown in formula (2):

[0019] logp(x|y)=log∫ k p(x|y,k)p(k|y) (3)

[0020]

[0021]

[0022] Use the point estimate of the logarithmic posterior probability k, that is, When p k When (k|y) has a unimodal distribution with a sharp peak, the estimated value accurately estimates the value of k; the maximum a posteriori MAP inference model is transformed into two sub-stages, first obtaining a point estimate of the logarithmic posterior probability k Reuse point estimate calculate As shown in formula (6) and formula (7):

[0023]

[0024]

[0025] According to formula (6) and formula (7), two CNNs are designed to process the two sub-stages of the maximum a posteriori MAP inference model respectively, which can be expressed as:

[0026]

[0027]

[0028] Among them, h δ (.) represents the conditional estimation subnetwork with δ parameters, f θ (.) represents the multi-scale cross-fusion denoising sub-network with θ parameters.

[0029] Preferably, the conditional estimation subnetwork is used to infer the point estimate Learn prior knowledge that is helpful for denoising tasks; the multi-scale cross-fusion denoising sub-network is used to integrate prior knowledge, simplify the learning difficulty of noise distribution, and thus obtain a clear image.

[0030] Preferably, the conditional estimation subnetwork consists of 3×3×3 convolutional layers, each of which has 64 feature channels, and a LeakyReLU nonlinearity is added after each convolution layer except the last convolution layer; at the same time, a global spatial spectrum module GSSM is added after the third and fifth convolutional layers.

[0031] Preferably, the feature map input to the global spatial spectrum module GSSM is F in , the dimension is expressed as C represents the number of feature channels, H represents the height, W represents the width, and B represents the number of spectral bands. in , on each feature channel, pooling encoding in different directions is performed to obtain long-distance fusion information. The c-th channel along the spectral band is expressed as:

[0032]

[0033] The c-th channel encoded along the spatial dimension W is expressed as:

[0034]

[0035] The c-th channel encoded along the spatial dimension H is expressed as:

[0036]

[0037] Then F b 、F h 、F wThe features of the three dimensions are spliced ​​and convolved using the 1×1×1 convolution kernel ω. Then the sigmoid activation function is used to learn the internal relationship of the nonlinear channel to obtain the intermediate feature map F mid As shown in formula (13):

[0038] F mid =σ(ω([F b , F h , F w ])) (13)

[0039] in, σ represents the sigmoid activation function, r represents the decay rate; then the obtained F mid Split into Then By integrating global information through point multiplication, a new 1×1×1 convolution kernel is used Convolution, followed by superposition of the original input feature map F in , and obtain the output F of the global spatial spectrum module GSSM in the conditional estimation subnetwork GSSM As shown in formula (14):

[0040]

[0041] Preferably, the multi-scale cross-fusion denoising sub-network can accept the variable parameter k generated in the first stage and perform denoising on the hyperspectral image in combination with prior knowledge; the multi-scale cross-fusion denoising sub-network is composed of a backbone network improved from 3D-UNet and a conditional encoder.

[0042] Preferably, the backbone network includes four encoders and three decoders; the encoder is used to extract features, and the decoder is used for image reconstruction; each encoder consists of two consecutive 3×3×3 convolution layers and a linear rectifier unit (ReLU) [Conv3d 3*3*3+ReLU], and except for the last encoder, each encoder is followed by an average pooling [Average Pooling 1*2*2] for downsampling; each decoder contains a 3D transposed convolution and two consecutive 3×3×3 3D convolution kernels and a ReLU activation layer [Transpose Conv3d 1*2*2+Conv3d+ReLU], and the affine transformation is applied to the representation after the first encoder as shown in formula (15):

[0043]

[0044] Among them, F a and F b Represents the feature map before and after affine transformation, Indicates element-wise multiplication, each decoder in U-net fuses the underlying features through a skip connection structure;

[0045] The feature cross-fusion transformer (FCFT) designed based on the characteristics of HSI includes the following steps:

[0046] Channel statistics F obtained by spatial compression s , channel statistics Channel statistics F s The information calculation of the cth channel in is shown in formula (16):

[0047]

[0048] Denote the encoder feature map and decoder feature map as F e and F d ,in Output fusion vector M e As shown in formula (17):

[0049] M e =σ(μ×(γ1·ω(F e ))+v×(γ2·ω(F d )) (17)

[0050] Among them, δ represents the sigmoid activation function, and is the weight of the linear layer, μ and ν are the proportions of fusion; the fused vector M e Used to recalibrate the low-level features F e , the calibrated fusion feature is expressed as The final calibrated fusion features Then compare it with the feature map F input by the decoder d The output F of FCFT is obtained by concatenation s .

[0051] The present invention proposes a system for deep denoising of hyperspectral images based on a two-stage learning framework, comprising:

[0052] An objective function construction module, wherein the objective function construction module is used to construct an objective function based on noise estimation and image denoising;

[0053] A network acquisition module, configured to construct a conditional estimation subnetwork and a multi-scale cross-fusion denoising subnetwork according to an objective function;

[0054] A network training module, wherein the network training module is used to train the conditional estimation subnetwork and the multi-scale cross-fusion denoising subnetwork using an objective function;

[0055] A noise level acquisition module, wherein the noise level acquisition module is used to infer the noise level of the hyperspectral image using the trained conditional estimation subnetwork;

[0056] The denoising module of the hyperspectral image is used to pass the noise level information of the hyperspectral image into the trained multi-scale cross-fusion denoising sub-network to denoise the hyperspectral image, thereby achieving denoising of the hyperspectral image.

[0057] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a method for deep denoising of hyperspectral images based on a dual-stage learning framework are implemented.

[0058] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for deep denoising of hyperspectral images based on a dual-stage learning framework.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] This paper proposes a method for deep denoising of hyperspectral images based on a two-stage learning framework. This method can blindly denoise HSI photos with random noise. First, an objective function is constructed based on noise estimation and image denoising. Then, a conditional estimation subnetwork and a multi-scale cross-fusion denoising subnetwork are constructed based on the objective function. The conditional estimation subnetwork and the multi-scale cross-fusion denoising subnetwork are trained using the objective function. The noise level of the hyperspectral image is inferred using the trained conditional estimation subnetwork, and this information is then passed to the multi-scale cross-fusion denoising subnetwork for denoising the hyperspectral image. This method can predict non-iid noise distributions and achieve deep denoising of hyperspectral images, solving the problem of image distortion.

[0061] Furthermore, the global spatial spectral module (GSSM) is used to capture the correlation of long-range information in the spectral dimension and space, making up for the lack of information between the spatial and spectral dimensions.

[0062] Furthermore, to allow the network to effectively incorporate conditional information and avoid the network's inference of each pixel position relying on a small local area, we apply an affine transformation after the first encoder.

[0063] Furthermore, each decoder in U-net fuses the underlying features through a skip connection structure to allow the network to retain richer underlying feature information and avoid gradient vanishing. However, the rough fusion of underlying features and high-level features will damage network performance. Therefore, a feature cross-fusion transformer (FCFT) is designed based on the characteristics of HSI.

[0064] The present invention proposes a system for deep denoising of hyperspectral images based on a two-stage learning framework. By dividing the system into an objective function construction module, a network acquisition module, a network training module, a noise level acquisition module and a hyperspectral image denoising module, a modular concept is adopted to make each module independent of each other, facilitating unified management of each module. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0066] Figure 1 This is a flowchart of the hyperspectral image deep denoising method based on the dual-stage learning framework of the present invention.

[0067] Figure 2 This is a structural diagram of the network in the present invention.

[0068] Figure 3 This is a diagram of the GSSM module architecture in the conditional estimation subnetwork of the present invention.

[0069] Figure 4 This is a diagram of the transformer (FCFT) architecture for feature cross-fusion in the multi-scale cross-fusion network of the present invention;

[0070] Figure 5 This is the denoising effect diagram of the present invention on the ICVL dataset with a noise level of 50 ((a) real noise-free image; (b) noisy image; (c) BM4D; (d) LRMR; (e) FastHyDe; (f) DnCNN; (g) HSID-CNN; (h) HSI-SDeCNN; (i) QRNN3D; (j) Partial-DNet; (k) 3D-DUSSD).

[0071] Figure 6This is the denoising effect of the CAVE doll dataset with a noise level of 70 in the present invention. The color-inaccurate photo is obtained by merging 15(R)23(G)30(B) basebands ((a) real noise-free image; (b) noisy image; (c) BM4D; (d) LRMR; (e) FastHyDe; (f) DnCNN; (g) HSID-CNN; (h) HSI-SDeCNN; (i) QRNN3D; (j) Partial-DNet; (k) 3D-DUSSD).

[0072] Figure 7 This is a comparison of the denoising effects of different models on the CAVE doll in 30 frequency bands under case 3 of the present invention ((a) real noise-free image; (b) noisy image; (c) BM4D; (d) LRMR; (e) FastHyDe; (f) DnCNN; (g) HSID-CNN; (h) HSI-SDeCNN; (i) QRNN3D; (j) Partial-DNet; (k) 3D-DUSSD).

[0073] Figure 8 This is a comparison of the denoising effects of different models on the ICVL dataset on 24 frequency bands under case 4 of the present invention ((a) real noise-free image; (b) noisy image; (c) BM4D; (d) LRMR; (e) FastHyDe; (f) DnCNN; (g) HSID-CNN; (h) HSI-SDeCNN; (i) QRNN3D; (j) Partial-DNet; (k) 3D-DUSSD).

[0074] Figure 9 The denoising results on frequency band 2 of the Indian pines dataset of the present invention are ((a) original image; (b) BM4D; (c) LRMR; (d) FastHyDe; (e) DnCNN; (f) HSID-CNN; (g) HSI-SDeCNN; (h) QRRN3D; (i) Partial-DNet; (j) 3D-DUSSD).

[0075] Figure 10 Classification results on frequency band 2 of the Indian pines dataset of the present invention ((a) correctly labeled classification; (b) original image; (c) BM4D; (d) LRMR; (e) FastHyDe; (f) DnCNN; (g) HSID-CNN; (h) HSI-SDeCNN; (i) QRRN3D; (j) Partial-DNet; (k) 3D-DUSSD).

[0076] Figure 11The denoising results of the Roche University of Pavia dataset in frequency band 1 of the present invention ((a) original image; (b) BM4D; (c) LRMR; (d) FastHyDe; (e) DnCNN; (f) HSID-CNN; (g) HSI-SDeCNN; (h) QRRN3D; (i) Partial-DNet; (j) 3D-DUSSD).

[0077] Figure 12 Classification results of the University of Pavia dataset of the present invention ((a) correctly labeled classification; (b) original image; (c) BM4D; (d) LRMR; (e) FastHyDe; (f) DnCNN; (g) HSID-CNN; (h) HSI-SDeCNN; (i) QRRN3D; (j) Partial-DNet; (k) 3D-DUSSD).

[0078] Figure 13 The denoising results of the urban dataset of the present invention on 109 baseband ((a) original image; (b) BM4D; (c) LRMR; (d) FastHyDe; (e) DnCNN; (f) HSID-CNN; (g) HSI-SDeCNN; (h) QRRN3D; (i) Partial-DNet; (j) 3D-DUSSD).

[0079] Figure 14 The denoising results of the urban dataset of the present invention on 206 baseband ((a) original image; (b) BM4D; (c) LRMR; (d) FastHyDe; (e) DnCNN; (f) HSID-CNN; (g) HSI-SDeCNN; (h) QRRN3D; (i) Partial-DNet; (j) 3D-DUSSD).

[0080] Figure 15 This is a diagram of the hyperspectral image deep denoising system based on the dual-stage learning framework of the present invention. DETAILED DESCRIPTION

[0081] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0082] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0083] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0084] In the description of the embodiments of the present invention, it should be noted that if the terms "upper," "lower," "horizontal," "inner," etc. appear, the orientation or positional relationship indicated is based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the inventive product is typically placed when in use. These terms are merely for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first," "second," etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0085] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0086] In the description of the embodiments of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0087] The present invention is described in further detail below with reference to the accompanying drawings:

[0088] The present invention proposes a method for deep denoising of hyperspectral images based on a two-stage learning framework. Figure 1 As shown, the following steps are included:

[0089] Construct objective function based on noise estimation and image denoising;

[0090] Construct a conditional estimation subnetwork (CENet) and a multi-scale cross-fusion denoising subnetwork based on the objective function;

[0091] The objective function is used to train the conditional estimation sub-network (CENet) and the multi-scale cross-fusion denoising sub-network;

[0092] The noise level of the hyperspectral image is inferred using the trained conditional estimation subnetwork (CENet);

[0093] The noise level information of the hyperspectral image is passed into the trained multi-scale cross-fusion denoising sub-network to denoise the hyperspectral image, thereby achieving denoising of the hyperspectral image.

[0094] The present invention proposes a method for deep denoising of hyperspectral images based on a two-stage learning framework. Figure 2 The figure shows the structure of the network in the present invention.

[0095] The specific process of constructing the objective function based on noise estimation and image denoising is:

[0096] The noisy hyperspectral image HSI is linearly modeled as:

[0097] y=x+n (1)

[0098] Where y, x, n∈R H×W×B , y, x, n represent the noisy hyperspectral image HSI, the noise-free hyperspectral image HSI, and the random noise that causes the degradation of the hyperspectral image HSI, respectively. H, W, and B represent the spatial height, spatial width, and number of spectral bands, respectively. In this way, the hyperspectral image HSI restoration task can be regarded as a traditional statistical inference problem. A maximum a posteriori (MAP) inference model is established as follows:

[0099]

[0100] In order to reduce the difficulty of learning the noise distribution, a new random variable k is introduced to calculate the logarithmic posterior probability, which contains the prior information of the denoising task. p(x|y) can be reformulated as:

[0101] logp(x|y)=log∫ k p(x|y,k)p(k|y) (3)

[0102]

[0103]

[0104] As shown in Formula 3, by introducing the random variable k, a new marginal expression of the posterior probability is obtained. In order to approximate the integral, the point estimate of k is used, that is, When p kWhen (k|y) is a unimodal distribution with a peak, this estimate can accurately estimate the value of k. Therefore, the maximum a posteriori MAP inference model can be transformed into two sub-stages. First, obtain the point estimate of k Reuse calculate Specifically expressed as:

[0105]

[0106]

[0107] According to Formula 6 and Formula 7, two CNNs are designed to process the two sub-stages of the maximum a posteriori MAP inference model respectively, which can be expressed as:

[0108]

[0109]

[0110] Among them, h δ (.) represents the conditional estimation subnetwork with δ parameters, f θ (.) represents the multi-scale cross-fusion denoising sub-network with θ parameters.

[0111] Two-stage reasoning network, the first stage network is the conditional estimation subnetwork, which is used to infer the point estimate Learning prior knowledge that is helpful for denoising tasks; the second stage network is a multi-scale cross-fusion denoising sub-network, which is used to integrate prior knowledge and simplify the learning difficulty of noise distribution, thereby obtaining a clear image. The specific network architecture is as follows Figure 2 shown.

[0112] The Conditional Estimation Network (CENet) primarily consists of 3×3×3 convolutional layers, each with 64 feature channels. Except for the last convolutional layer, a LeakyReLU nonlinearity is added after each convolutional layer. Furthermore, a Global Spatial Spectral Module (GSSM) is added after the third and fifth convolutional layers.

[0113] The role of the Global Spatial Spectral Module (GSSM) is to capture the correlation of long-range information in the spectral dimension and space, and make up for the lack of information between the spatial and spectral dimensions. The architecture of the Global Spatial Spectral Module (GSSM) is as follows: Figure 3 shown.

[0114] Assume that the feature map input to the global spatial spectrum module (GSSM) is F in , its dimension is expressed as Where C represents the number of feature channels, H represents the height, W represents the width, and B represents the number of spectral bands. in, on each feature channel, pooling encoding in different directions is performed to obtain long-distance fusion information. It is expressed as follows:

[0115] The cth channel along the spectral band is expressed as:

[0116]

[0117] The c-th channel encoded along the spatial dimension W is expressed as:

[0118]

[0119] The c-th channel encoded along the spatial dimension H is expressed as:

[0120]

[0121] Then F b 、F h 、F w The features of the three dimensions are spliced ​​and convolved using the 1×1×1 convolution kernel ω. Then the sigmoid activation function is used to learn the internal relationship of the nonlinear channel to obtain the intermediate feature map F mid It can be expressed as:

[0122] F mid =σ(ω([F b ,F h ,F w ])) (13)

[0123] in, σ represents the sigmoid activation function, and r represents the decay rate. Then the obtained F mid Split into three parts, expressed as Then By integrating global information through point multiplication, a new 1×1×1 convolution kernel is used Convolution, followed by superposition of the original input feature map F in , and finally get the output F of the global spatial spectral module (GSSM) in the conditional estimation subnetwork GSSM , specifically expressed as follows:

[0124]

[0125] The multi-scale cross-fusion denoising sub-network in the second stage accepts the variable parameter k generated in the first stage and combines it with prior knowledge to denoise the hyperspectral image. It consists of two parts: a backbone network improved from 3D-UNet and a conditional encoder.

[0126] The backbone network consists of four encoders and three decoders. The encoder is used to extract features and the decoder is used for image reconstruction. The cleverness of this design is that it allows the network to obtain multi-scale features. Each encoder consists of two consecutive 3×3×3 convolution layers and a linear rectification unit (ReLU) [Conv3d 3*3*3+ReLU]. Except for the last encoder, each encoder is followed by an average pooling [Average Pooling 1*2*2] for downsampling. Each decoder contains a 3D transposed convolution and two consecutive 3×3×3 3D convolution kernels and a ReLU activation layer [TransposeConv3d 1*2*2+Conv3d+ReLU]. At the same time, in order to allow the network to effectively merge conditional information and avoid the network's reliance on local small areas for inference of each position pixel, an affine transformation is applied after the first encoder:

[0127]

[0128] Among them, F a and F b Represents the feature map before and after affine transformation, In U-net, each decoder fuses the underlying features through a skip connection structure to allow the network to retain richer underlying feature information and avoid gradient vanishing. However, the rough fusion of underlying features and high-level features will damage network performance. Therefore, a feature cross-fusion transformer (FCFT) is designed based on the characteristics of HSI. Its structure is as follows: Figure 4 shown.

[0129] First, the channel statistics F obtained by spatial compression s , F s The information calculation formula for the cth channel in is as follows:

[0130]

[0131] Denote the encoder feature map and decoder feature map as F e and F d ,in Output fusion vector M e As shown in formula (17):

[0132] M e =σ(μ×(γ1·ω(F e ))+v×(γ2·ω(F d )) (17)

[0133] Among them, δ represents the sigmoid activation function, and is the weight of the linear layer, μ and v are the fusion ratios, and the best performance is achieved when μ = 0.6 and ν = 0.4. e Used to recalibrate the low-level features F e , the calibrated fusion feature is expressed as The final calibrated fusion features Then compare it with the feature map F input by the decoder d The output F of FCFT is obtained by concatenation s .

[0134] In this multi-stage denoising network, the networks at different stages are trained separately. In order to prevent the generation of overly smoothed HSI after denoising, the L1 loss function is used as the loss function of the network. The L1 loss used by the first stage network is expressed as:

[0135]

[0136] The L1 loss used by the second stage network is expressed as:

[0137]

[0138] Among them, Y i and They represent hyperspectral images with random noise and correctly labeled noise-free medical hyperspectral images, respectively, N is the training batch, and σ is the noise type mapping.

[0139] The specific training process of the network is:

[0140] The same hyperparameter settings were used in both training phases, and the network parameters were updated using the Adam algorithm. The network hyperparameters were as follows: batch_size was 24, epochs was 100, and the initial learning rate was 2e-4. This learning rate was then reduced by half every 10 epochs until it dropped to 5e-6, at which point it remained at 5e-6 for subsequent epochs. All other Adam algorithm hyperparameters retained their default values.

[0141] Simulation experiment:

[0142] To train the proposed HSI denoising and noise estimation models, the proposed network was trained using the CAVE dataset. In this experiment, the results were compared with those of representative traditional HSI denoising methods and deep learning approaches. Traditional methods include the filtering-based method BM4D, low-rank methods, and FastHyDe, which uses low-rank sparse representations. Deep learning methods include DnCNN, HSID-CNN, HSI-SDeCNN, QRRN3D, and Partial-DNet. Parameters for all these methods were optimized according to their respective papers.

[0143] To validate the performance of the proposed model for HSI denoising, the well-known CAVE and ICVL datasets were used for training and testing. The CAVE dataset contains 32 hyperspectral images from everyday scenes, with an image size of 512×512×31, an image spacing of 10 nm, and a wavelength range of 400–700 nm. For the training set, 15 images from the CAVE dataset were randomly selected as training samples. These training data were then slidingly cropped with a step size of 25, followed by a data augmentation scheme, resulting in 19,440 images of 64×64×31 size. Real-world HSI is often contaminated by various types of noise, and the noise distribution is often extremely complex. To verify the practicality of the proposed method, experiments were conducted with various noise types, including Gaussian noise, streak noise, cutoff noise, and mixed noise. The denoising performance of different methods was evaluated from both a quantitative and visual quality perspective. Five noise cases were selected and described in detail below.

[0144] a) Case 1 (iid Gaussian noise): Add zero-mean additive white Gaussian noise to all frequency bands. The noise variance of each band is the same and is set to 30, 50, and 70, respectively, or Gaussian blind noise in the range [30, 70].

[0145] b) Case 2 (non-iid Gaussian noise): Add zero-mean additive white Gaussian noise to all frequency bands, and the noise level of each frequency band is randomly selected from 30, 50, and 70.

[0146] c) Case 3 (non-iid Gaussian + stripe noise): Non-iid Gaussian noise as in Case 2 is added to all frequency bands, and stripe noise is added to 1 / 3 of the frequency bands randomly selected.

[0147] d) Case 4 (non-iid Gaussian + deadline noise): The stripe noise in Case 3 is replaced by deadline noise, and the non-Gaussian noise remains unchanged.

[0148] e) Case 5 (Mixed Noise): Based on Case 2, stripe noise is randomly added to half of the spectral basebands, and deadline noise is added to the remaining spectral basebands. The stripes and deadlines in each spectral baseband account for 5% to 15%.

[0149] The same hyperparameter settings were used in both training phases, and the network parameters were updated using the Adam algorithm. The network hyperparameters were as follows: batch_size was 24, epochs was 100, and the initial learning rate was 2e-4. This learning rate was then reduced by half every 10 epochs until it dropped to 5e-6, at which point it remained at 5e-6 for subsequent epochs. All other Adam algorithm hyperparameters retained their default values.

[0150] Evaluation criteria:

[0151] The experiment selected mean peak signal-to-noise ratio (MPSNR), mean structural similarity (MSSIM), and spectral corneal map (SAM) as evaluation criteria. Larger MPSNR and MSSIM values ​​indicate less image distortion, while smaller SAM values ​​indicate better performance. MPSNR and MSSIM evaluate the similarity between the denoised HSI and the original HSI based on mean squared error (MSE) and structural consistency, respectively. SAM, on the other hand, is the average angle between the spectral vector of the denoised HSI and the reference space at all spatial locations, reflecting the spectral fidelity of the denoised HSI.

[0152] For real noise experiments, this method uses a classification task on the denoised hyperspectral image to evaluate the performance of the denoising method. Overall accuracy (OA) and kappa coefficient are used as evaluation indicators of the classification results.

[0153] Example 1

[0154] 1. HSI denoising experiment based on zero-mean noise

[0155] First, zero-mean Gaussian noise was added to the CAVE, ICVL, and Washington DC Mall datasets at noise levels of 30, 50, and 70, as well as random independent and identically distributed Gaussian noise. Quantitative evaluation results of various denoising methods on the CAVE, ICVE, and Washington DC Mall datasets using zero-mean Gaussian noise are shown in Table 1. It is clear that the proposed 3D-DUSSD model outperforms other methods on most evaluation metrics. Furthermore, FastHyDe, HSI-SDeCNN, and QRNN3D all exhibit excellent denoising performance, but FastHyDe does not perform well with blind noise. HSI-SDeCNN achieves excellent results when the noise level is known in advance. QRNN3D uses samples of varying noise intensities to train the network, but its model requires fine-tuning for different datasets. The proposed method uses only non-IID Gaussian noise to simulate training samples, training a single model and achieving superior performance on different IID Gaussian noise denoising tasks, further demonstrating its versatility and high fidelity.

[0156] Table 1 Quantitative analysis on CAVE, ICVL and Washington datasets with iid Gaussian noise

[0157]

[0158]

[0159] The noise reduction simulation results with noise variance of 50 and 70 are as follows Figure 5 and Figure 6 As shown in the figure, it is not difficult to find that the 3D-DUSSD of the present invention can correctly restore the image while well preserving the image structure and edge texture. In contrast, BM4D, LRMR, and some DNets cannot suppress all noise, while DnCNN shows a smooth transition phenomenon. FastHyDe, HSID-CNN, HSI SDeCNN, and QRNN3D show excellent performance in suppressing Gaussian noise, but some fine-grained details are lost and a small amount of artifacts are present.

[0160] 2. HSI denoising experiment based on complex noise conditions

[0161] In the denoising experiment under complex noise conditions (case 2-5), the experimental results of each model under different complex noises and different data sets are shown in Table 2.

[0162] Table 2 Quantitative analysis on CAVE, ICVL and Washington datasets with different complex noises

[0163]

[0164]

[0165] Under complex noise, the denoising visual effects of different methods are as follows Figure 7 and Figure 8 As shown. It can be clearly observed that although BM4D, DnCNN and FastHyDe can suppress Gaussian noise well, they perform poorly in denoising of mixed noise, and more noise information is still retained in the denoising results. LRMR, HSID-CNN, HSI SDeCNN and part of Dnet successfully removed most of the stripe / cutoff noise, but there is still obvious residual noise. QRNN3D and the 3D-DUSSD of the present invention show better noise reduction effects, while removing noise, they protect the fine-grained structural information of the original HSI. However, QRNN3D is trained separately according to different noises, while the present invention only needs non-iid Gaussian noise for training, which can effectively suppress noise with different characteristics, confirming the versatility of the present invention.

[0166] 3. Based on real HSI denoising experiments

[0167] To validate the practicality of this method, we conducted experiments in real-world scenarios. Because ground-truth noise-free images are unavailable in real-world data, the denoising effect was evaluated by assessing the classification accuracy after denoising. Three hyperspectral datasets were selected: an urban dataset, the AVIIS Indian pine dataset, and the RoSIS University of Pavia dataset. A support vector machine (SVM) classifier was chosen, with 10% of the samples used for training and the remainder for testing. The kappa coefficient and OA were selected as classification evaluation metrics to reflect the denoising effect.

[0168] In the denoising experiment using the AVIRIS Indian pine dataset, in order to facilitate visual evaluation, grayscale images with heavy noise pollution in two frequency bands were selected for visualization. Figure 9 As shown, some DNets perform poorly on frequency band 2, which is severely corrupted by various noises. LRMR and QRRN3D show good performance, but some dense noise is still present. BM4D and HSID-CNN show good denoising results, but a small amount of streak noise is present in the image, resulting in some information loss. FastHyDe, DnCNN, and HSI SDeCNN can remove dense and streak noise, but perform poorly in the trade-off between retaining detail information and removing noise. Compared to the comparison methods, 3D-DUSSD cleverly incorporates additional information, achieving an optimal balance between removing noise and preserving image details.

[0169] The OA and kappa coefficients obtained by classifying the original image and the denoised images generated by all the contrast methods are shown in Table 3. The classification results are shown in Figure 10 As shown in the figure, the classification results of the original image are discontinuous and contain some misclassifications, with OA and kappa reaching 0.7955 and 0.7656, respectively. The classification results of the image after denoising using FastHyDe are even weaker than those without denoising, with many misclassifications. The classification results of the remaining methods are improved compared to the original image, indicating that these algorithms effectively suppress noise in the image. Among them, the OA and kappa of the 3D-DUSSD proposed in this invention are 0.9262 and 0.9147, respectively, which is the best performance among all methods.

[0170] Table 3 SVM classification performance of Indian pine

[0171]

[0172] In the Roche University of Pavia dataset experiment, the noise-concentrated frequency band 1 is selected for visualization. The denoising result is as follows Figure 11 As shown in the figure, it can be noticed that DnCNN is over-smoothed, while BM4D and LRMR can only partially suppress noise, and noise still exists in the results. FastHyDe, HSI SDeCNN, HSID-CNN, Partial DNet, and QRRN3D all show good denoising performance. They remove almost all noise but lose image details, while the method performs well in suppressing noise and preserving image texture information.

[0173] The classification accuracy results of the University of Pavia dataset are listed in Table 4. Figure 12 The classification results are presented. Due to the low noise content of the University of Pavia dataset, the OA and kappa of the original images reached 0.9658 and 0.9584, respectively. The FastHyDe method performed poorly on real-world images, resulting in lower classification accuracy than the original images. The classification accuracy of the denoised data using DnCNN, HSID-CNN, HSI-SDeCNN, QRRN3D, Partial-DNet, and 3D-DUSSD was significantly improved, with 3D-DUSSD achieving the best classification results, with an OA and kappa coefficient of 0.9854 and 0.9823, respectively.

[0174] Table 4 SVM classification performance of the University of Pavia dataset

[0175]

[0176] In the urban dataset experiment, the dataset is contaminated by various types of noise including Gaussian noise and stripe noise. The original image and denoising results at band 109 and band 206 are shown in Figure 2. Figure 13 and Figure 14 As shown in the figure. It can be seen from the figure that FastHyDe, HSI-SDeCNN, QRRN3D, and 3D-DUSSD all achieve good noise reduction performance in both frequency bands; BM4D, LRMR, and HSID-CNN can effectively remove Gaussian noise, but the noise reduction effect on stripe noise is poor. The 3D-DUSSD of the present invention not only effectively suppresses various noises, but also retains the original image details. For example, the area marked in the figure effectively removes stripes and Gaussian noise and retains the image details, while other methods are unable to effectively remove stripe noise and other mixed noises, or lose the original image details.

[0177] 4. Ablation Experiment

[0178] To verify the effectiveness of the proposed modules, ablation experiments were conducted. The conditional estimation network is denoted by C, the GSSM by G, and the FCFT by F. To ensure experimental accuracy, a training strategy consistent with that used in this method was employed, and experiments were conducted on the CAVE dataset, which was mixed with Gaussian noise at a level of 70. The experimental results are shown in Table 5. As can be seen, the baseline model achieved MPSNR, MSSIM, and SAM after denoising, respectively, of 35.15, 0.9015, and 0.1456. This demonstrates that the baseline model effectively extracts spatial spectral structure and achieves good denoising results. After adding the conditional estimation network C, MPSNR increased by 1.22, MSSIM increased by 0.0112, and SAM decreased by 0.0308. This demonstrates that the conditional estimation subnetwork can effectively and accurately learn the noise distribution. When CG is also added, MPSNR and MSSIM increase, and SAM decreases, compared to adding C alone. This demonstrates that the GSSM effectively integrates global spatial spectral information, thereby improving denoising performance. Adding CF also increases MPSNR by 0.90, MSSIM by 0.0126, and SAM by 0.0146. This demonstrates that the CCFT module in this invention effectively integrates low-level and high-level features, effectively improving denoising performance. Finally, ablation experiments confirm that the proposed 3D-DUSSD method achieves the best noise reduction performance.

[0179] Table 5 Ablation experiment results

[0180]

[0181] The present invention proposes a system for deep denoising of hyperspectral images based on a two-stage learning framework. Figure 15 As shown, including:

[0182] An objective function construction module, wherein the objective function construction module is used to construct an objective function based on noise estimation and image denoising;

[0183] A network acquisition module, configured to construct a conditional estimation subnetwork and a multi-scale cross-fusion denoising subnetwork according to an objective function;

[0184] A network training module, wherein the network training module is used to train the conditional estimation subnetwork and the multi-scale cross-fusion denoising subnetwork using an objective function;

[0185] A noise level acquisition module, wherein the noise level acquisition module is used to infer the noise level of the hyperspectral image using the trained conditional estimation subnetwork;

[0186] The denoising module of the hyperspectral image is used to pass the noise level information of the hyperspectral image into the trained multi-scale cross-fusion denoising sub-network to denoise the hyperspectral image, thereby achieving denoising of the hyperspectral image.

[0187] An embodiment of the present invention provides a terminal device, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of each of the aforementioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of each module / unit in each of the aforementioned apparatus embodiments are implemented.

[0188] The computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to accomplish the present invention.

[0189] The terminal device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0190] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0191] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0192] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0193] The present invention discloses a method for deep denoising of hyperspectral images based on a two-stage learning framework. The three-dimensional deep universal model 3D-DUSSD for hyperspectral image denoising based on the two-stage learning framework constructs an objective function according to noise estimation and image denoising during specific operation; then the objective function constructs a conditional estimation subnetwork (CENet) and a multi-scale cross-fusion denoising subnetwork; the conditional estimation subnetwork infers a noise estimate from HSI containing random noise, and then passes the estimate into the multi-scale cross-fusion denoising subnetwork. After multiple encoding and decoding, the noise in the HSI is removed.

[0194] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for deep denoising of hyperspectral images based on a two-stage learning framework, It is characterized in that The following steps are involved: Construct objective function based on noise estimation and image denoising; Construct a conditional estimation sub-network and a multi-scale cross-fusion denoising sub-network according to the objective function; The objective function is used to train the conditional estimation sub-network and the multi-scale cross-fusion denoising sub-network; The noise level of the hyperspectral image is inferred using the trained conditional estimation subnetwork; The noise level information of the hyperspectral image is passed into the trained multi-scale cross-fusion denoising sub-network to denoise the hyperspectral image, thereby achieving denoising of the hyperspectral image; The specific process of constructing the objective function based on noise estimation and image denoising is as follows: The noisy hyperspectral image HSI is linearly modeled as shown in formula (1): (1) in, , They represent the noisy hyperspectral image HSI, the noise-free hyperspectral image HSI and the random added noise that causes the degradation of the hyperspectral image HSI. H, W and B represent the spatial height, spatial width and the number of spectral bands respectively. The hyperspectral image HSI restoration task is a traditional statistical inference problem. A maximum a posteriori MAP inference model is established as shown in formula (2): (2) A new random variable is introduced to calculate the logarithmic posterior probability k, which contains the prior information of the denoising task. The restated form is shown in formula (3) to formula (5): (3) (4) (5) Use the point estimate of the logarithmic posterior probability k, that is, ;when When the distribution is unimodal with a spike, the estimated value accurately estimates the value of k; the maximum a posteriori MAP inference model is transformed into two sub-stages, first obtaining a point estimate of the logarithmic posterior probability k , and then use the point estimate calculate , as shown in formula (6) and formula (7): (6) (7) According to formula (6) and formula (7), two CNNs are designed to process the two sub-stages of the maximum a posteriori MAP inference model respectively, which can be expressed as: (8) (9) in, Indicates that The conditional estimation subnetwork of parameters, Indicates that Parameterized multi-scale cross-fusion denoising sub-network.

2. The method for deep denoising of hyperspectral images based on a dual-stage learning framework according to claim 1, It is characterized in that The conditional estimation subnetwork is used to infer point estimates , learn prior knowledge that is helpful for denoising tasks; the multi-scale cross-fusion denoising sub-network is used to fuse prior knowledge and simplify the learning difficulty of noise distribution, thereby obtaining a clear image.

3. The method for deep denoising of hyperspectral images based on a dual-stage learning framework according to claim 1, It is characterized in that The conditional estimation subnetwork consists of 3×3×3 convolutional layers, each of which has 64 feature channels. Except for the last convolutional layer, a LeakyReLU nonlinearity is added after each convolutional layer. At the same time, a global spatial spectral module GSSM is added after the third and fifth convolutional layers.

4. The method for deep denoising of hyperspectral images based on a dual-stage learning framework according to claim 3, It is characterized in that Suppose the feature map input to the global spatial spectrum module GSSM is , the dimension is expressed as , C represents the number of feature channels, H represents the height, W represents the width, and B represents the number of spectral bands; for the feature map , on each feature channel, pooling encoding in different directions is performed to obtain long-distance fusion information. The cth channel along the spectral band is expressed as: (10) The cth channel encoded along the spatial dimension W is expressed as: (11) The cth channel encoded along the spatial dimension H is expressed as: (12) Then will , , The features of the three dimensions are concatenated and a 1×1×1 convolution kernel is used. Convolution, and then use the sigmoid activation function to learn the internal relationship of the nonlinear channel to obtain the intermediate feature map As shown in formula (13): (13) in, , represents the sigmoid activation function, r represents the decay rate; then the obtained Split into , , ; then , , The global information is integrated by dot product and a new 1×1×1 convolution kernel is used. Convolution, followed by superposition of the feature maps of the original input , and obtain the output of the global spatial spectral module GSSM in the conditional estimation subnetwork As shown in formula (14): (14)。 5. The method for deep denoising of hyperspectral images based on a dual-stage learning framework according to claim 1, It is characterized in that The multi-scale cross-fusion denoising sub-network can accept the variable parameter k generated in the first stage and combine prior knowledge to perform denoising on the hyperspectral image; the multi-scale cross-fusion denoising sub-network is composed of a backbone network and a conditional encoder improved from 3D-UNet.

6. The method for deep denoising of hyperspectral images based on a dual-stage learning framework according to claim 5, It is characterized in that The backbone network consists of four encoders and three decoders; the encoders are used to extract features and the decoders are used to reconstruct images; each encoder consists of two consecutive 3×3×3 convolutional layers and a linear rectifier unit ReLU [ ], except for the last encoder, each encoder is followed by an average pooling [ ] for downsampling; each decoder contains a 3D transposed convolution and two consecutive 3×3×3 3D convolution kernels and a ReLU activation layer[ ], the affine transformation is applied to the representation after the first encoder as shown in formula (15): (15) in, Represents the feature map before and after affine transformation, It represents element-wise multiplication. Each decoder in U-net fuses the underlying features through a skip connection structure. The feature cross-fusion transformer designed based on the characteristics of HSI includes the following steps: Channel statistics obtained through spatial compression , channel statistics , channel statistics The information calculation of the cth channel in is shown in formula (16): (16) The encoder feature map and decoder feature map are represented as and ,in{ , } ; Output fused vector As shown in formula (17): (17) in, represents the sigmoid activation function, and are the weights of the linear layer, is the ratio of fusion; the fused vector Used to recalibrate low-level features , the calibrated fusion feature is expressed as , the final calibrated fusion feature Then compare it with the feature map input by the decoder The output of FCFT is obtained by concatenation .

7. A system for deep denoising of hyperspectral images based on a two-stage learning framework, It is characterized in that The method for deep denoising of hyperspectral images based on a dual-stage learning framework according to any one of claims 1 to 6 comprises: An objective function construction module, wherein the objective function construction module is used to construct an objective function based on noise estimation and image denoising; A network acquisition module, wherein the network acquisition module is used to construct a conditional estimation subnetwork and a multi-scale cross-fusion denoising subnetwork according to an objective function; A network training module, wherein the network training module is used to train the conditional estimation subnetwork and the multi-scale cross-fusion denoising subnetwork using an objective function; A noise level acquisition module, wherein the noise level acquisition module is used to infer the noise level of the hyperspectral image using the trained conditional estimation subnetwork; A denoising module for a hyperspectral image is used to transfer the noise level information of the hyperspectral image into a trained multi-scale cross-fusion denoising sub-network to denoise the hyperspectral image, thereby achieving denoising of the hyperspectral image.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program. It is characterized in that When the processor executes the computer program, the steps of the method for deep denoising of hyperspectral images based on a dual-stage learning framework described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium storing a computer program. It is characterized in that When the computer program is executed by a processor, the steps of the method for deep denoising of hyperspectral images based on a dual-stage learning framework described in any one of claims 1 to 6 are implemented.