A hyperspectral image denoising method based on spatial spectrum feature fusion

Abstract

The application discloses a hyperspectral image denoising method based on spatial spectrum feature fusion, and belongs to the technical field of hyperspectral remote sensing image processing. The method comprises the following steps: preparing and preprocessing a hyperspectral image dataset; constructing a network model composed of spatial spectrum feature extraction, feature fusion and reconstruction stages; selecting a suitable loss function and evaluation index; training and saving the network model. The application extracts spatial structure and texture features of an image by using a spatial path, and models spectral correlation among multiple bands by using a spectral path. Through a cross-guiding dual-frequency feature fusion module, bidirectional cross-guiding and frequency domain fusion of spatial and spectral features are realized. A band adaptive attention mechanism module is introduced at a bottleneck layer to suppress noise and enhance feature response. A spatial-spectral collaborative attention module in a Fourier domain is combined to further improve the collaborative representation ability and relevance of spatial and spectral dimensions, so that high-quality and high-fidelity hyperspectral image denoising is realized.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral remote sensing image processing technology, and more specifically relates to a hyperspectral image denoising method based on spatial spectral feature fusion. Background Technology

[0002] Hyperspectral image denoising aims to effectively separate noise signals from useful information by analyzing the spatial structure information and spectral features in images, thereby achieving the restoration and reconstruction of degraded images. It has significant application value in various fields such as remote sensing monitoring, military reconnaissance, medical imaging, and industrial inspection. Methods that have achieved good results in hyperspectral image denoising mainly include traditional filtering methods, low-rank matrix recovery, tensor decomposition, and deep learning. Deep learning-based methods, with their powerful feature learning capabilities, can automatically extract joint spatial and spectral features, making them widely popular in the field of hyperspectral image denoising. However, deep learning methods still face many challenges in practical applications. On the one hand, due to limitations in network structure, they often struggle to fully exploit the joint representation of space and spectrum, neglecting global frequency domain information. On the other hand, deep learning models require large amounts of training data, resulting in high network computational complexity and issues such as over-smoothing and loss of detail. Therefore, reducing computational complexity and improving feature representation capabilities while ensuring denoising performance is of great research significance in hyperspectral image denoising.

[0003] The Chinese patent authorization number is "CN114972075A", entitled "A Hyperspectral Image Denoising Method Based on Residual Learning and Hybrid Domain Attention". This method first performs band grouping processing on the hyperspectral image. Then, a sparse feature extraction module is used to obtain local spectral-spatial features. A hybrid domain attention module adjusts the features and captures cross-domain interaction information. The grouped local features are then concatenated and fused. Finally, a residual learning strategy is used to obtain the denoising result. While this method can effectively handle hyperspectral image noise, its network structure is complex, the grouping strategy has a high computational cost, the hybrid domain attention module has a large number of parameters, and its training efficiency is low. Summary of the Invention

[0004] To address the shortcomings of existing hyperspectral image denoising methods, such as insufficient band correlation modeling, inadequate spatial-spectral feature fusion, and high computational complexity, this invention proposes a hyperspectral image denoising method based on spatial spectral feature fusion. Through a four-stage parallel spatial spectral fusion architecture and a lightweight attention module design, this method significantly reduces computational complexity and the number of parameters while fully exploring band correlation and enhancing spatial-spectral correlation, thereby improving the efficiency of spatial-spectral feature fusion and denoising performance.

[0005] To achieve the above objectives, the present invention specifically adopts the following technical solution: A hyperspectral image denoising method based on spatial spectral feature fusion includes the following steps: Step 1, Prepare the dataset: Prepare five public datasets for hyperspectral image denoising, and divide them into training set, test set and validation set according to the proportion; Step 2, Preprocess the dataset: First, standardize the hyperspectral dataset, then crop the images into fixed-size image patches; Step 3, Constructing the Network Model: The entire hyperspectral image denoising network based on spatial spectral feature fusion consists of a feature extraction stage, a feature fusion stage, and a reconstruction stage. The feature extraction stage includes spatial feature extraction and spectral feature extraction. Spatial feature extraction includes a multi-scale convolution module and a feature enhancement module, while spectral feature extraction includes a spectral preprocessing module and a spectral enhancement module. The feature fusion stage consists of a cross-domain feature fusion module, a band-adaptive attention mechanism module, and a Fourier domain spatial-spectral collaborative attention module. The cross-domain feature fusion module includes a cross-guided mask generator, a dual-frequency gating mechanism, a frequency domain fusion module, and a feature refinement unit. The band-adaptive attention mechanism module includes a noise estimation module and a brightness statistics module. The Fourier domain spatial-spectral collaborative attention module fuses spatial channel attention with frequency domain modulation. Step 4: Select appropriate loss function and evaluation metrics: The loss function is a weighted combination of L1 loss function and structural similarity loss as the optimization objective for network training. At the same time, peak signal-to-noise ratio and structural similarity are selected to evaluate spatial quality, and spectral fidelity is quantified using a spectral angle mapper. Step 5, train the network model: train the hyperspectral image denoising network model. The publicly available hyperspectral image dataset obtained in Step 1 is processed through the preprocessing process set in Step 2. Then, the processed data is input into the denoising network model constructed in Step 3 for end-to-end training to complete the learning and optimization of network parameters. Step 6, Save the model: Based on the loss changes and performance metrics evaluation during training, save the parameters of the best-performing model to determine the final hyperspectral image denoising model.

[0006] Furthermore, in step 1, the dataset used is the ICVL hyperspectral dataset, the University of Pavia dataset, the Washington, D.C. shopping mall dataset, the EO-1 dataset, and the Gaofen-5 satellite dataset, and the dataset is divided into training set, validation set, and test set according to a preset ratio.

[0007] Furthermore, in step 2, the hyperspectral image dataset is first numerically normalized, and then the images are cropped into N×N pixel uniform-sized image blocks to adapt to the fixed-dimensional requirements of the network input and expand the number of training samples.

[0008] Furthermore, in step 3, the spatial feature extraction module consists of a multi-scale convolution module and a feature enhancement module, used to extract multi-scale spatial features. The multi-scale convolution module extracts multi-scale spatial features in parallel using convolution kernels of different sizes. The feature enhancement module models the target region structure through residual convolution and spatial attention module, while enhancing important spatial locations and suppressing redundant information.

[0009] Furthermore, in step 3, the spectral feature extraction module consists of a spectral preprocessing module and a spectral feature extraction enhancement module, used to extract and enhance spectral features; the spectral preprocessing module eliminates dimensional differences between different bands through band normalization, suppresses random noise through spectral domain smoothing, stabilizes training and prevents anomalous spectra through amplitude standardization, enhances the continuity of adjacent bands and prevents anomalous bands from disrupting training; the spectral feature extraction enhancement module learns the band correlation multiple times through multi-layer spectral domain convolution and pixel-by-pixel spectral modeling module.

[0010] Furthermore, in step 3, the dynamic feature fusion module is composed of a cross-domain feature fusion module, which can adaptively fuse spatial features and spectral features. This module controls the information interaction between spatial flow and spectral flow through a gating mechanism, effectively complementing and enhancing the two modal features, thereby obtaining a more discriminative fused feature representation of spatial and spectral features.

[0011] Furthermore, in step 3, the bottleneck processing module consists of a feature compression layer, a band adaptive attention mechanism module, and a Fourier domain spatial-spectral collaborative attention module. After the spatial and spectral features output by the encoder are fused and downsampled through convolution operations, the band adaptive attention mechanism module captures the correlation of different spectral bands and the dependency of each pixel position, while the Fourier domain spatial-spectral collaborative attention module collaboratively models the global correlation of spatial details and spectral features.

[0012] Furthermore, in step 3, the decoder module consists of a deconvolution upsampling layer, a feature concatenation layer, and a reconstructed convolutional layer; the deconvolution upsampling layer gradually restores the spatial resolution of the feature map, the feature concatenation layer makes skip connections between the features of the corresponding level of the encoder and the upsampled features, and the reconstructed convolutional layer fuses multi-level features and finally generates a denoised hyperspectral image.

[0013] Furthermore, in step 4, the loss function is selected as a weighted combination of the L1 loss function and the structural similarity loss as the optimization objective for network training, which measures the difference between the denoised image and the noisy image. At the same time, the peak signal-to-noise ratio and structural similarity are selected to evaluate the spatial quality, and the spectral fidelity is quantified using a spectral angle mapper.

[0014] Compared with existing technologies, this invention proposes a hyperspectral image denoising method based on spatial spectral feature fusion, which has the following beneficial effects: This invention provides a spatial-spectral feature fusion denoising method. First, it designs parallel spatial feature extraction paths and spectral feature extraction paths, independently optimizing and collaboratively enhancing the two modalities. The spatial path employs a hierarchical, multi-scale convolutional design, preserving texture features in shallow layers and enhancing spatial feature information through attention weights in deeper layers. The spectral path enhances 3D spectral modeling through spectral preprocessing and inter-spectral mapping residuals. The designed U-shaped architecture effectively overcomes the defect of mutual interference between spatial and spectral features in single-encoder networks, fully utilizing spatial-spectral structural information.

[0015] The cross-domain feature fusion module designed in this invention introduces a cross-guided mask, enhancing the correlation between spatial and spectral features through information interaction. A dynamic weight allocation mechanism adaptively modulates the weights through gating units. This module not only simply concatenates dual-path features but also recalibrates and selectively enhances spatial and spectral features through attention weights, ensuring that the fused features retain both the integrity of the spatial structure and the accuracy of the spectral curves. In complex noisy environments, this module effectively improves feature fusion performance and denoising generalization ability.

[0016] This invention designs a band-adaptive attention mechanism module at the bottleneck layer. By introducing noise estimation and adaptive channel weight adjustment mechanisms, it achieves synergistic optimization of feature enhancement and noise suppression. This module estimates the noise distribution of the input features and dynamically generates channel weights based on global statistical information, thereby suppressing high-noise channels and enhancing effective signal channels. Feature reconstruction is completed through residual methods, significantly reducing random noise and imaging noise interference while improving effective detail features. This module adaptively sets noise parameters, exhibits good stability, and is suitable for different noise levels, effectively improving denoising performance and robustness.

[0017] After suppressing inter-spectral differences, this invention introduces a spatial-spectral collaborative attention module based on the Fourier domain. This module adaptively modulates and fuses features in the frequency domain to achieve synergistic optimization of noise suppression and detail enhancement. Then, through inverse transformation, it restores the image to the spatial domain, achieving joint denoising of spatial and frequency domain information. This module can enhance edge and texture information, improve high-frequency details and edge information, and enhance detail preservation and visual quality without destroying the image structure. Attached Figure Description

[0018] Figure 1 Here is a flowchart of the hyperspectral image denoising method provided by the present invention; Figure 2 This is a general framework diagram of the hyperspectral image denoising method provided by the present invention; Figure 3 A schematic diagram illustrating the specific process of spatial feature extraction provided by this invention; Figure 4 A schematic diagram illustrating the specific process of spectral feature extraction provided by this invention; Figure 5 This is a schematic diagram of the fusion downsampling structure provided by the present invention; Figure 6 A schematic diagram showing the specific composition of the cross-domain feature fusion module provided by the present invention; Figure 7 A schematic diagram showing the specific components of the band-adaptive attention mechanism module provided by the present invention; Figure 8 A schematic diagram illustrating the specific composition of the Fourier domain spatial-spectral coordinated attention module provided by this invention; Figure 9 This is a schematic diagram of the decoder structure provided by the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 like Figure 1 The flowchart shown is a hyperspectral image denoising method based on spatial spectral feature fusion proposed in one embodiment of the present invention. The method specifically includes the following steps: Step 1, Prepare the dataset: Prepare five publicly available hyperspectral datasets, namely the ICVL hyperspectral dataset, the University of Pavia dataset, the Washington, D.C. shopping mall dataset, the EO-1 dataset, and the Gaofen-5 satellite dataset, and divide the datasets into training set, validation set, and test set according to a preset ratio.

[0021] Step 2, Preprocessing the dataset: The original hyperspectral images in each dataset are numerically normalized so that their pixel values ​​are linearly mapped to the [0, 1] interval, eliminating the numerical scale differences caused by different sensor imaging conditions. The normalized hyperspectral images are then cropped into image blocks of a fixed size of 64×64 to adapt to the fixed dimension requirements of the network input and expand the number of training samples.

[0022] Step 3, Construct the network model: The entire denoising network includes a spatial feature extraction module, a spectral feature extraction module, a cross-domain feature fusion module, a band adaptive attention mechanism module, a Fourier domain spatial-spectral collaborative attention module, and a decoder.

[0023] The spatial feature extraction module consists of a multi-scale convolution module and a feature enhancement module. The multi-scale convolution module is set in the shallow stage of spatial feature extraction. It uses three different sizes of convolution kernels (3×3, 5×5, and 7×7) to perform convolution operations on the input features in parallel, obtaining local spatial texture information and edge structure information under different receptive fields. After concatenation, the channel dimension is fused and compressed through 1×1 convolution. In the intermediate stage, the feature enhancement module uses 3×3 convolution, batch normalization, and non-linear activation functions to extract intermediate spatial features to ensure the effective transmission of feature information. Then, it uses a spatial attention mechanism to adaptively weight key regions in the spatial dimension, concatenates the input features with the convolution output, and then compresses them through 1×1 convolution to highlight important spatial structural features and suppress redundant noise information.

[0024] The spectral feature extraction module consists of a spectral preprocessing module and a spectral enhancement module. The spectral preprocessing module includes band normalization, spectral smoothing, and amplitude normalization. The spectral smoothing first expands the dimension to (B,1,D,H,W), then smooths the spectral noise using a 3D convolutional kernel, and finally compresses the dimension to (B,D,H,W). Amplitude normalization uses dimension transformation, normalization layers, and dimension recovery to restore a stable amplitude distribution. The spectral enhancement module extracts spectral features and captures inter-band dependencies through a 7×1×1 3D convolutional kernel, batch normalization, R-type activation function, and multilayer perceptron.

[0025] The cross-domain feature fusion module consists of a cross-guided mask generator, a dual-frequency gating mechanism, a frequency domain fusion module, and a feature refinement unit. The cross-guided mask generator is divided into a spatial-spectral guide and a spectral-spatial guide. After normalization, they are concatenated, and then concatenated again after global average pooling and global max pooling, respectively. 3×3 convolution and sigmoid activation function are used to generate spatial-to-spectral and spectral-to-spatial guide masks, respectively. The dual-frequency gating mechanism consists of a low-frequency gating submodule and a high-frequency gating submodule. Global average pooling, 1×1 convolution, and sigmoid activation function are used to extract channel attention weights from the real and imaginary parts of the frequency domain, respectively. The frequency domain fusion module consists of a fast Fourier transform layer, a complex reconstruction layer, and an inverse transform layer. The features are decomposed into real and imaginary parts through orthogonal Fourier transform. After applying dual-frequency gating weights, complex features are reconstructed and inversely transformed back to the spatial domain. Finally, the features are efficiently fused through a refined convolutional layer and residual connection.

[0026] The band-adaptive attention mechanism module consists of a noise estimation module and a brightness statistics module. The noise estimation module is composed of 3×3 convolutional layers, 5×5 convolutional layers, and 1×1 convolutional layers. It extracts noise features and generates a noise distribution map through multi-scale convolution. The brightness statistics module is composed of a global average pooling layer and two fully connected layers. It generates basic feature importance weights by statistically analyzing the brightness information in the spatial dimension and learning the nonlinear relationship between bands through a multilayer perceptron. Then, it enhances important band features and suppresses noise interference through learnable noise suppression coefficients and sigmoid activation functions.

[0027] The Fourier domain spatial-spectral collaborative attention module consists of an input preprocessing unit, a spatial channel attention mechanism, a dual-branch feature extraction module, and a frequency domain modulation fusion unit. The input preprocessing unit performs nonlinear transformation on the input features using a 1×1 convolutional layer and a G-type activation function. The spatial channel attention mechanism consists of a global average pooling layer and a 1×1 convolutional layer, generating channel attention weights to enhance important features. The dual-branch feature extraction module extracts spatial and frequency domain features respectively using two parallel 1×1 convolutional layers. The frequency domain modulation fusion unit transforms the frequency domain branch to the frequency domain using a fast Fourier transform, performs frequency domain modulation with the complex tensor constructed from the spatial branch, reconstructs the features using an inverse Fourier transform, and finally adaptively fuses the reconstructed features with the original features using two learnable parameters.

[0028] The decoder consists of a deconvolutional upsampling layer, a feature concatenation layer, and a reconstruction convolutional layer. The deconvolutional upsampling layer gradually restores the feature map resolution through four levels of transposed convolution. The feature concatenation layer concatenates the fused features of the encoder at the corresponding scale through skip connections, and compresses the redundant channels after concatenation through 3×3 convolution. At the same time, it refines the spatial-spectral joint features, suppresses artifacts caused by transposed convolution, and integrates cross-level information. The reconstruction convolutional layer restores the spatial resolution through 3×3 convolution and R-type activation function, and finally restores the spectral resolution through 1×1 convolution. It is then connected with the residual of the original image to output the reconstruction result.

[0029] Step 4: Select appropriate loss function and evaluation metrics: The loss function is selected as a weighted combination of L1 loss function and structural similarity loss as the optimization objective for network training. The L1 loss function preserves edge details and texture information by calculating the pixel-level absolute difference between the denoised image and the real image. The structural similarity loss evaluates the similarity of brightness, contrast and structural information by calculating local statistical features through Gaussian filtering. At the same time, the peak signal-to-noise ratio is selected to comprehensively and quantitatively evaluate the denoising effect from three dimensions: pixel-level accuracy, structural similarity from perceptual quality, and spectral angle mapping from spectral fidelity.

[0030] Step 5, Train the network model: Train the hyperspectral image denoising network model. The publicly available hyperspectral image dataset obtained in Step 1 is processed through the preprocessing procedure set in Step 2. The processed data is then input into the denoising network model constructed in Step 3 for end-to-end training to complete the learning and optimization of network parameters.

[0031] Step 6, Save the model: Based on the loss changes and performance metrics evaluation during training, save the parameters of the best-performing model to determine the final hyperspectral image denoising model.

[0032] Example 2 like Figure 1 The diagram shows a flowchart of a hyperspectral image denoising method based on spatial spectral feature fusion. The method specifically includes the following steps: Step 1, Prepare Datasets: Prepare five hyperspectral datasets: ICVL, University of Pavia, Washington D.C. Shopping Mall, EO-1, and Gaofen-5 satellite. The ICVL hyperspectral dataset has a wavelength range of 400-1000 nm, contains 519 bands, and has a spatial resolution of 0.1 m. During training, 201 hyperspectral images from the ICVL dataset were downloaded. 100 images were divided into a training set, 11 into a complex noise validation set, 10 into a Gaussian noise validation set, 40 into a complex noise test set, and 40 into a Gaussian noise test set for model training and evaluation. The University of Pavia dataset has a wavelength range of 0.43-0.86 μm, contains 115 bands, and has a spatial resolution of 1.3 m. This dataset contains noise-free images and was used to test the model. The Washington D.C. Shopping Mall dataset has wavelengths... The first dataset has a wavelength range of 0.4–2.5 μm, a spatial resolution of 3.7 m, and contains 191 bands. This dataset contains noise-free images and is used to validate the model. The second dataset has a wavelength range of 0.4–2.5 μm, contains 242 bands, and a spatial resolution of 30 m. This dataset contains real noise and is used to evaluate the denoising effect of the model. The third dataset has a wavelength range of 0.39–1.04 μm, contains 330 bands, and a spatial resolution of 30 m. This dataset contains sensor noise and is used to test the applicability of the model.

[0033] Step 2, Data Preprocessing: The hyperspectral images are standardized and noise simulated to enhance the model's generalization ability and robustness. Standardization uses a min-max normalization method to linearly transform the original data to the [0, 1] interval, as shown in the formula: , in, For raw hyperspectral data, and These are the global minimum and maximum values ​​of the data, respectively. This is to prevent division by zero errors. During the noise simulation phase, complex noise is added to the clean images in the training set to construct training samples, including Gaussian noise, salt-and-pepper noise, and stripe noise, simulating noise distribution in real-world scenes. Simultaneously, data augmentation operations such as random cropping, rotation, and flipping are applied to the training data to increase the diversity of the training samples.

[0034] Step 3, construct the network model: such as Figure 2 The diagram illustrates the overall framework of a hyperspectral image denoising method based on spatial spectral feature fusion. Specifically, it includes a spatial feature extraction module, a spectral feature extraction module, a cross-domain feature fusion module, a band-adaptive attention mechanism module, a Fourier domain spatial-spectral collaborative attention module, and a decoder. The spatial and spectral paths are synchronously downsampled at multiple scale levels and fused step-by-step through the cross-domain feature fusion module. At the bottleneck layer, the effective features are further enhanced through the band-adaptive attention mechanism module and the Fourier domain spatial-spectral collaborative attention module. Finally, the decoder reconstructs the denoised hyperspectral image step-by-step.

[0035] A schematic diagram illustrating the specific process of spatial feature extraction is shown below. Figure 3 As shown, it consists of three modules: a simple convolutional module, a multi-scale convolutional module 3(a), and a feature enhancement module 3(b). It employs a progressively differentiated design strategy to enhance spatial feature extraction information at different network stages. In the first stage of the encoder, a lightweight 3×3 convolutional layer ensures the receptive field covers the local spatial region, a batch normalization layer stabilizes the feature distribution, and an R-type activation function introduces a non-linear expression to form a basic convolutional block. This preprocesses the input feature X for spatial feature extraction, extracting preliminary spatial features, expressed as: , in, Represents the R-type activation function. For batch normalization.

[0036] Then, a multi-scale convolution module consisting of three parallel two-dimensional convolutional branches is used. The three branches adopt two-dimensional convolutional layers with kernel sizes of 3×3, 5×5, and 7×7, respectively. The parallel convolution at three scales extracts shallow spatial features through different receptive fields, enhancing the multi-scale representation capability of local textures and edges. Finally, the features are stitched together using channel dimensions. The number of channels is 3C. After being input into a batch normalization layer and an R-type activation function, the channels are restored to C output through a 1×1 convolution. The expression is: , Subsequently, in the intermediate stage, the input is fed into a feature enhancement convolutional module consisting of a 3×3 convolutional layer for feature refinement, a batch normalization layer, an R-type activation function, spatial attention for weight extraction, and a 1×1 convolution. After lightweight spatial attention, the input is multiplied element-wise to obtain the enhanced spatial discriminative ability and local context modeling ability of the features. Finally, the feature enhancement module is passed through again to obtain the final deep feature output of global spatial semantics.

[0037] A schematic diagram illustrating the specific process of spectral feature extraction is shown below. Figure 4 As shown, it consists of a spectral preprocessing module and a spectral enhancement module. Firstly, the spectral preprocessing module, such as... Figure 4 As shown in (a), the input is first band normalized, and mean-standard deviation normalization is performed on each band. Then, the spectral curve is learned through layer normalization. The dimension is then expanded to (B,1,D,H,W), and a 3×1×1 3D convolution is performed on the spectral dimension D to smooth the spectral noise. Then, the dimension is compressed to (B,D,H,W), and the dimension is rearranged to (B,H,W,D). Finally, after passing through the normalization layer, the dimension is restored to (B,D,H,W) to stabilize the amplitude distribution. After fusing with spatial features and downsampling, the signal is then branched through the first spectral block, as shown in (a). Figure 4 As shown in (b) and (c), the dimension (B,1,D,H,W) is first expanded using a 3D convolution kernel of size 7×1×1 to perform a one-dimensional convolution operation in the spectral dimension, thereby extracting local spectral features and capturing inter-band dependencies. Then, batch normalization and an R-type activation function are used to enhance the feature representation capability and compress the dimension (B,D,H,W), thus extracting local spectral correlation features without changing the spectral dimension length D. Subsequently, the shape (B,H×W,D) is transformed by performing nonlinear recombination of the spectral vector of each pixel through a pixel-wise spectral domain multilayer perceptron, further enhancing the spectral domain representation capability. The perceptron projection layer maps high-dimensional spectral features to low dimensions through a multilayer perceptron consisting of two fully connected layers and an R-type activation function. After mapping, two-dimensional batch normalization is used to further stabilize the feature distribution and restore the shape (B,D,H,W). The original spectral features are preserved through residual connections. Finally, the output of spatial block one is fused and downsampled, and then input into the next branch. The second and third spectral blocks follow the same steps.

[0038] The structure diagram of the fusion downsampling is as follows Figure 5 As shown, the spectral input is first convolved with 1×1 to adapt the spectral channels, and then concatenated with the spatial input features. Then, all channels after concatenation are linearly combined using 1×1 convolution, batch normalization, and R-type activation function. The correlation of spatial spectrum is captured by 3×3 convolution, batch normalization, and R-type activation function, and then downsampling is performed using average pooling.

[0039] A schematic diagram of the specific components of the cross-domain feature fusion module is shown below. Figure 6As shown, the system consists of a cross-guided mask generator, a dual-frequency gating mechanism, a frequency domain fusion module, and a feature refinement unit. First, the cross-guided mask generator normalizes the input spatial and spectral features through a normalization layer to eliminate differences in numerical distribution between different modal features. Then, the normalized spatial and spectral features are concatenated along the channel dimension. The concatenated features undergo global average pooling and global max pooling operations, and the results are concatenated again before being input into a 3×3 convolutional block with a stride of 1 and padding of 1 for feature compression and nonlinear mapping. A cross-guided mask with weights of [0, 1] is then generated using a sigmoid activation function. This mask is multiplied element-wise by both the spatial and spectral features and then added to obtain the guidance of spatial information on spectral features and the inverse constraint of spectral information on spatial features. The mask calculations for guiding spatial features from spectral features and guiding spectral features from spatial features are as follows: , , in, Represents the sigmoid activation function. Indicates splicing.

[0040] Then, the original features are weighted element-wise using a mask and summed to obtain the preliminary fused features, the expression of which is: , Subsequently, a two-dimensional fast Fourier transform is used to map the fused features to the frequency domain and decompose them into low-frequency structural components. With high frequency structural components Then, gating modules consisting of global average pooling, 1×1 convolution, R-type activation function, 1×1 convolution, and sigmoid activation function are constructed for the low-frequency and high-frequency structural components respectively to generate low-frequency and high-frequency weights, respectively. , , in and These are low-frequency weights and high-frequency weights, respectively.

[0041] Then, adaptive modulation is performed on the corresponding frequency domain components. Finally, the weighted low-frequency and high-frequency components are recombine into complex frequency domain features. Since the original signal is a real number signal, the real part of the inverse transform result is taken as the output, while the imaginary part is theoretically close to zero and can be ignored. A two-dimensional inverse Fourier transform is used to convert the frequency domain features back to the spatial domain. Subsequently, the features are refined through convolution, 1×1 convolutional layers, R-type activation functions, and 1×1 convolutional layers to retain more original information and enhance the feature representation. The expressions for the two-dimensional fast Fourier transform and the two-dimensional fast Fourier transform are: , , , in, and These represent the height and width of the feature map, respectively. Represents frequency domain characteristics, ,1,..., -1, =0, 1, ..., -1 indicates the frequency coordinate. The coefficients represent the spatial features obtained from frequency domain reconstruction. As the normalization factor, This represents the low-frequency structural information component. This represents the high-frequency texture and detail information components.

[0042] A schematic diagram of the specific components of the band-adaptive attention mechanism module is shown below. Figure 7 As shown, after completing the multi-level cross-domain fusion of spatial and spectral features, a band-adaptive attention mechanism module is further introduced to extract deep features. This module consists of a luminance statistics network, a noise estimation network, a multilayer perceptron, and an adaptive weighting unit. First, the luminance statistics network extracts the luminance statistics vector for each channel in the input feature map through global average pooling, expressed as: , in This represents the brightness statistics of the channel. and These represent the height and width of the feature map, respectively. Indicates position The characteristic value of the channel.

[0043] Meanwhile, the noise estimation network extracts noise distribution features and generates a noise distribution map through a multi-scale convolutional network consisting of a 3×3 kernel, a 1-stride 2D convolutional layer, an R-type activation function, a 5×5 kernel, a 1-stride 2D convolutional layer, an R-type activation function, a 1×1 dimensional convolutional layer, and a S-type activation function. The expression for this map is: , Subsequently, a global average pooling operation is performed on the noise distribution map to obtain the noise level vector. Next, the luminance statistical vector is nonlinearly mapped through a multilayer perceptron consisting of two fully connected layers to generate basic importance weights: , in , and , These represent the weights and biases of the two fully connected layers, respectively.

[0044] Finally, the final channel attention weights are calculated using an adaptive weighting unit, which introduces a learnable noise suppression coefficient. The adjusted channel weights are generated by subtracting the noise penalty term from the basic importance weights and then applying a sigmoid activation function. , in This is a learnable parameter, initially set to 0.8. The closer the value is to 1, the less noise is suppressed; the closer it is to 0, the more noise is suppressed, but gradient explosion may occur. An appropriate value can be set based on the training results. value.

[0045] The final output is achieved through two weighted operations: first, the input features are multiplied by the channel weights and then added to the original features to obtain the enhanced features; then, the enhanced features are adaptively fused with the noise distribution map. , , in This indicates an element-wise multiplication operation.

[0046] This design enables the module to adaptively adjust attention weights based on the noise level of each band, enhancing important features while suppressing noise interference and improving the denoising performance of hyperspectral images.

[0047] A schematic diagram illustrating the specific components of the spatial-spectral co-attention module in the Fourier domain is shown below. Figure 8 As shown, it consists of an input preprocessing unit, a spatial channel attention unit, a dual-branch feature extraction unit, a frequency domain modulation unit, and an adaptive fusion unit. First, the input preprocessing unit performs nonlinear transformation preprocessing on the input features using a 1×1 dimensional convolutional layer and a G-type activation function to obtain enhanced features. Then, the spatial channel attention unit uses a global average pooling layer and a 1×1 dimensional convolutional layer to generate channel attention weights by aggregating global spatial information. The generated channel attention weights are multiplied element-wise with the enhanced features to obtain the spatial channel enhancement features. The dual-branch feature extraction unit extracts spatial branch features and frequency domain branch features from the spatial channel enhancement features through two independent 1×1 dimensional convolutional layers. First, a two-dimensional fast Fourier transform is performed on the frequency domain branch features, while simultaneously constructing the spatial branch features into complex form. Then, modulation is performed in the frequency domain, multiplying the complex representation of the spatial branch element-wise with the Fourier transform result of the frequency domain branch. Finally, a two-dimensional inverse fast Fourier transform is performed on the modulated frequency domain features, and the reconstructed features are obtained by taking the real part. This is achieved through two learnable parameters. and , Enhanced control Control the original to achieve a weighted fusion of reconstructed features and original spatial channel enhancement features.

[0048] A schematic diagram of the hyperspectral image reconstruction structure is shown below. Figure 9 As shown, it consists of four levels of feature splicing and fusion blocks and a final reconstruction block. First, the first-level decoder is as follows... Figure 9 As shown in (a), the deconvolution upsampling block doubles the size of the input feature map and reduces the number of channels by using a transposed convolutional layer with a stride of 2 and a kernel size of 2×2. Then, the feature concatenation and fusion block concatenates the upsampled features with the fused features output from the fourth stage of the encoder along the channel dimension. If the feature sizes do not match before concatenation, bilinear interpolation is used for adjustment. The second and third deconvolution upsampling blocks follow the same operation process, upsampling the feature map to 1 / 4 and 1 / 2 of its original size, respectively. The fourth stage decoder is as follows: Figure 9 As shown in (b), the deconvolution upsampling block upsamples the feature map to the original spatial resolution, and the final reconstruction block maps the fused features to the output channel dimension through a 1×1 convolutional layer to generate a denoised hyperspectral map.

[0049] Step 4, Selecting an appropriate loss function and evaluation metric: For hyperspectral image denoising, during the training process, a suitable loss function and evaluation metric are typically used. and The norm is used as the loss function. However, in this design, in order to achieve a balance between denoising and detail preservation, a [missing information] is adopted. The weighted combination of the loss and the structural similarity loss serves as the model's loss function, defined as: , Among them, the weighting coefficient Set to 0.8 to balance pixel-level precision with structural similarity. The loss function calculates the pixel-level absolute difference between the predicted image and the target image: , in, This represents the total number of pixels in the image. For the first The actual value of each pixel. For the first The predicted value for each pixel. Structural similarity loss preserves perceptual image quality by calculating the similarity of brightness, contrast, and structural information: , in Represents the local mean. Indicates local standard deviation. Describing covariance, and It is the stability constant.

[0050] The evaluation metrics selected are Peak Signal-to-Noise Ratio (PSNR), Structural Similarity, and Spectral Angle Mapping. PSNR reflects the model's reconstruction quality at the pixel level; Structural Similarity measures the model's performance in perceptual quality, assessing the preservation of brightness, contrast, and structural information; and Spectral Angle Mapping quantifies the model's fidelity to spectral features, evaluating the consistency of the spectral curve shape before and after denoising. The expressions for PSNR, Structural Similarity, and Spectral Angle Mapping are as follows: , , , , in This represents the maximum possible value of the image pixels. , The local mean of the image. , For local variance, For covariance, , It is the stability constant. and These represent the original image and the denoised image at different locations. Spectral vector at that location, Represents the dot product of vectors. Representing vectors Norm. The higher the PSNR value, the better the denoising effect and the less distortion. The closer the SSIM value is to 1, the better the target denoising is and the closer it is to the original clean image. The smaller the SAM value, the more similar the spectra of the two images are, indicating better denoising effect in the spectral dimension.

[0051] Step 5, Training the Network Model: A phased, progressive training strategy is adopted, dividing the training process into three progressive stages. The first stage (rounds 1-10) uses fixed-intensity Gaussian noise, with a noise level of... The training process involves three stages: 1) Basic denoising capability training using a loss function; 2) Blind noise (randomly sampled within the range [10, 70]) introduced in the first stage (rounds 11-20) to enhance the model's adaptability to different noise intensities; 3) Non-independent, identically distributed mixed noise, including impulse noise, stripe noise, and dead-line noise, is used in the second stage (rounds 21-50) to improve the model's robustness to complex real-world noise through a composite loss function. The Adam optimizer is used during training, with a breakpoint-based training mechanism. The initial learning rate is set to 0.001, and a cosine annealing strategy is employed to dynamically adjust the learning rate. After each training round, model performance is evaluated on multiple validation sets, and the optimal weights are saved for final testing. The formula for the cosine annealing strategy is as follows: , in, Let be the learning rate at step t. The initial maximum learning rate, To minimize the learning rate, The number of steps in the current period. This represents the total number of steps in a single cycle.

[0052] Step 6, Save the model: Based on the loss changes and performance metrics evaluation during training, save the parameters of the best-performing model to determine the final hyperspectral image denoising model.

[0053] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hyperspectral image denoising method based on spatial spectral feature fusion, characterized in that, Includes the following steps: Step 1: Prepare publicly available hyperspectral datasets, namely the ICVL hyperspectral dataset, the University of Pavia dataset, the Washington, D.C. shopping mall dataset, the EO-1 dataset, and the Gaofen-5 satellite dataset, and divide the datasets into training set, validation set, and test set according to a preset ratio; Step 2: First, standardize the hyperspectral dataset, then crop the images into fixed-size image patches; Step 3: Construct a network model consisting of feature extraction, feature fusion and reconstruction stages. The feature extraction stage includes spatial feature extraction and spectral feature extraction, and the feature fusion stage includes a cross-domain feature fusion module, a band adaptive attention mechanism module and a Fourier domain spatial-spectral collaborative attention module. Step 4, select appropriate loss function and evaluation metric: adopt The weighted combination of the loss function and the structural similarity loss function is used to quantitatively evaluate the denoising effect through peak signal-to-noise ratio, structural similarity, and spectral angle mapping evaluation metrics; Step 5, Train the network model: Train the hyperspectral image denoising network model, using a composite loss function as the optimization objective, and using the Adam optimizer to update the network parameters. The gradient is calculated and the network weights are iteratively optimized through the backpropagation algorithm. During training, a learning rate decay strategy is used to adjust the training progress, the model performance is monitored using the validation set, and an early stopping mechanism is used to prevent overfitting. The denoising effect is quantitatively evaluated through evaluation metrics, and the network parameters are optimized until the model converges. Step 6, Save the model: Based on the loss changes and performance metrics evaluation during training, save the parameters of the best-performing model to determine the final hyperspectral image denoising model; The cross-domain feature fusion module consists of a cross-guided mask generator, a dual-frequency gating mechanism, and a feature refinement unit, used to fuse spatial and spectral features. The cross-guided mask generator learns the spatial-spectral interdependence by mutually guiding spatial and spectral features, thereby enhancing their correlation. The dual-frequency gating mechanism decomposes the fused features into real low-frequency components and imaginary high-frequency components using fast Fourier transform, and adaptively adjusts the weights of different frequency components using low-frequency gating and high-frequency gating to optimize the frequency domain feature components. The feature refinement unit achieves deep fusion and enhancement by optimizing feature and residual connections. The band-adaptive attention mechanism module integrates noise estimation with brightness statistics; it identifies feature channels rich in brightness perception, assesses noise levels by combining noise estimation, and adaptively generates attention weights using learnable noise suppression coefficients to enhance important features and suppress noise interference. The Fourier domain spatial-spectral collaborative attention module integrates spatial channel attention with frequency domain modulation; it identifies important channel features through spatial channel attention, extracts frequency domain features by combining frequency domain modulation, and uses learnable fusion coefficients to adaptively balance the reconstructed features with the original features, thereby enhancing and optimizing the collaboration between spatial and frequency domain information.

2. The hyperspectral image denoising method based on spatial spectral feature fusion according to claim 1, characterized in that: The spatial feature extraction module consists of a multi-scale convolution module and a feature enhancement module, which are used to fully extract spatial features at different scales. The multi-scale convolution module obtains rich local spatial features through different receptive fields, and the feature enhancement module enhances spatial features by using convolution blocks and spatial attention.

3. The hyperspectral image denoising method based on spatial spectral feature fusion according to claim 1, characterized in that: The spectral feature extraction module consists of a spectral preprocessing module and a spectral enhancement module, which are used to extract spectral features from shallow to deep layers. The spectral preprocessing module suppresses random spectral noise and maintains spectral continuity and stable amplitude distribution by band normalization, spectral domain smoothing filtering and amplitude normalization. The spectral enhancement module uses multi-layer spectral domain convolution and pixel-by-pixel spectral modeling module to learn the band correlation multiple times.

4. The hyperspectral image denoising method based on spatial spectral feature fusion according to claim 1, characterized in that: The network adopts a U-shaped architecture, extracting features layer by layer from shallow to deep in the spatial and spectral branches, and fusing downsampling. At the same time, after the spatial and spectral features are extracted at each stage, a cross-domain fusion module is introduced, and a collaborative attention mechanism is constructed in the bottleneck layer to realize deep interaction and extraction of spatial and spectral feature information.

5. The hyperspectral image denoising method based on spatial spectral feature fusion according to claim 1, characterized in that: In step 3, the cross-domain feature fusion module consists of a cross-guided mask generator, a dual-frequency gating mechanism, and a feature refinement unit. The cross-guided mechanism enables mutual enhancement of spatial spectral features, and the frequency domain decomposition technique is used to control low-frequency structural information and high-frequency detail components respectively. Finally, feature refinement and deep fusion are completed through residual connection.

6. The hyperspectral image denoising method based on spatial spectral feature fusion according to claim 1, characterized in that: In step 3, the band adaptive attention mechanism module integrates noise estimation and brightness statistics. Through the collaborative analysis of brightness perception and noise estimation, a feature quality assessment system is established. Learnable parameters are used to adaptively balance feature enhancement and noise suppression, and important features are selectively strengthened.

7. The hyperspectral image denoising method based on spatial spectral feature fusion according to claim 1, characterized in that: In step 3, the spatial-spectral collaborative attention module in the Fourier domain integrates spatial channel attention with frequency domain modulation. It processes spatial and frequency domain features separately through a dual-branch architecture, completes feature modulation and reconstruction in the frequency domain, and uses learnable fusion coefficients to achieve adaptive weighted integration of spatial and frequency domain information.

Images