A hyperspectral image denoising method based on content-aware sparse prompting

Abstract

This invention relates to the field of hyperspectral image processing, specifically to a hyperspectral image denoising method based on content-aware sparse cues. The method includes: constructing multiple typical noise degradation patterns and their linguistic descriptions; introducing a pre-trained text encoder to generate cue vectors; and adaptively evaluating the importance of the cue channel in conjunction with image content. Within the encoder-decoder framework of the denoising network, the cue is mapped to a feature map with a resolution consistent with the visual feature space, and fused with features from the corresponding level of the encoder at skip connections, thereby participating in the multi-scale feature modeling and reconstruction process. Through this approach, the denoising network can adaptively adjust the feature reconstruction path according to the image content when facing unknown noise types or mixed noise degradation, thus improving the network's adaptability. By sparsely modeling the high-dimensional cue features, the scale of redundant features involved in computation is reduced, improving computational efficiency while maintaining denoising performance.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral image processing, specifically to a hyperspectral image denoising method based on content-aware sparse cues. Background Technology

[0002] Hyperspectral images are three-dimensional data cubes that simultaneously contain two-dimensional spatial information and one-dimensional spectral information. They are typically composed of dozens to hundreds of continuous narrow bands and can provide detailed spectral feature descriptions of ground features. They have important application value in fields such as remote sensing monitoring, resource surveys, and environmental analysis.

[0003] However, in actual imaging processes, hyperspectral images are inevitably subject to various noise interferences due to limitations in the light flux of the imaging spectrometer, sensor characteristics, atmospheric transmission environment, and uncertainties in signal acquisition and transmission. Common noise types include Gaussian noise, stripe noise, dead-line noise, and impulse noise, which often exhibit complex degradation characteristics in real-world scenes, characterized by mixing, superposition, and uneven spatial distribution. These noises destroy the spatial structure information and spectral consistency of the image, severely affecting the accuracy of subsequent applications such as demixing, classification, and target detection. Therefore, hyperspectral image denoising, as a crucial preprocessing step, has significant research and application value.

[0004] To address noise degradation in hyperspectral images, existing techniques mainly fall into two categories: model-driven methods and data-driven methods. Model-driven methods typically utilize prior knowledge such as low-rank, sparse, total variational, and nonlocal similarity to construct mathematical models and achieve noise suppression by solving optimization problems. While offering some interpretability, they have limited adaptability to complex mixed noise scenarios and incur significant computational costs. Data-driven methods, on the other hand, rely on deep learning frameworks to learn the mapping relationship between noise distribution and image features from large-scale training data, achieving better denoising performance.

[0005] However, most existing deep learning models rely on supervised training based on known noise types or fixed noise intensity conditions. Their network structures and parameters are typically optimized around a single noise degradation pattern. In real-world applications, hyperspectral images are often affected by the random superposition of multiple noise types or unknown noise intensity perturbations, making it difficult to accurately model noise degradation patterns beforehand. In the aforementioned blind denoising scenarios, existing models lack the ability to adaptively perceive noise types and degradation levels, making it difficult to achieve stable denoising and reconstruction of complex degraded hyperspectral images.

[0006] To improve the adaptability of denoising models under various noise degradation modes, some existing techniques have introduced cue-based conditional modeling. By providing additional degradation information to the denoising network, these techniques guide the network to perform corresponding feature reconstruction processes for different noise modes. However, existing cueing typically relies on predefined noise category labels or fixed text descriptions, lacking the ability to perceive the features of the input image content. Furthermore, directly using high-dimensional dense vectors as guiding information can easily introduce redundant feature channels and increase computational complexity, thus affecting the stability and robustness of denoising reconstruction. Summary of the Invention

[0007] This invention aims to overcome the shortcomings of existing hyperspectral image denoising methods, such as the lack of adaptive perception of noise degradation type and image content in blind denoising scenarios, and the problem that directly introducing high-dimensional cues leads to feature redundancy and high computational cost. It creatively proposes a hyperspectral image denoising method based on content-aware sparse cues. By introducing an adaptive content-aware sparse cue mechanism and combining it with a deep learning network, this invention effectively improves the denoising accuracy, stability, and generalization ability of complex degraded hyperspectral images while reducing computational efficiency.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] A hyperspectral image denoising method based on content-aware sparse cues includes the following steps:

[0010] Step 1: Acquire the original hyperspectral image and preprocess it to obtain hyperspectral image sample blocks for training; during the training process, generate noisy hyperspectral image samples based on the hyperspectral image sample blocks using a random degradation strategy, and record the corresponding noise degradation category labels.

[0011] Step 2: Encode the pre-built text describing various noise types using a pre-trained text encoder to obtain the corresponding cue vector set; construct the corresponding cue weight vector based on the noise degradation category label corresponding to the hyperspectral image, and perform a linear weighted summation of the cue vector set using this weight vector to obtain the high-dimensional cue vector corresponding to the current degradation mode;

[0012] Step 3: Perform global average pooling on the input hyperspectral image to obtain global statistical features for each spectral channel. Input these global statistical features into a content-aware visual gating network to obtain importance score vectors corresponding to each channel of the cue vector. Sort the importance score vectors by channel dimension and select the highest-scoring vectors according to a preset sparsity ratio. The first channel is reserved as a channel, and the second channel is used as the reserved channel. The large score is used as a dynamic threshold to construct a channel selection mask. The mask is then used to filter the high-dimensional cue vector channel by channel to obtain a sparse cue vector.

[0013] Step 4: Embed the sparse cue vector as guiding information into the hyperspectral image denoising network to guide the feature reconstruction of the hyperspectral image and output the denoised hyperspectral image.

[0014] Preferably, step 1 specifically includes:

[0015] Step 101: Obtain the hyperspectral image dataset as the original noise-free image data required for training. Crop the original hyperspectral image using a sliding window method and divide it into local noise-free image blocks with a spatial size of 64×64.

[0016] Step 102: During the network training phase, a noisy hyperspectral image sample is generated using a random degradation strategy, and the corresponding noise degradation category label is recorded.

[0017] Specifically, the data loading module reads noise-free image patches of a preset batch size from the training data, and independently and randomly selects a predefined degradation task mode for each image patch for real-time degradation processing. This allows the same training batch to contain samples of multiple noise degradation types, enabling the denoising network to learn the noise features and removal rules of multiple degradation modes simultaneously during parameter updates. At the same time, the system synchronously records the degradation task label corresponding to each image patch. The degradation modes include Gaussian noise degradation at different intensity levels, non-independent and identically distributed Gaussian noise degradation, and various mixed noise degradation forms formed by superimposing strip noise, dead-line noise, or impulse noise on non-independent and identically distributed Gaussian noise.

[0018] Preferably, step 2 specifically includes:

[0019] Step 201: For different types of noise degradation patterns in hyperspectral images, construct a set of linguistic description texts containing Gaussian noise, strip noise, dead line noise, and mixed noise;

[0020] Step 202: Load the pre-trained text encoder, encode the above language description text set, normalize the encoded feature vectors, arrange all normalized feature vectors in order of noise category to form the cue basis matrix, and store it as a fixed parameter.

[0021] Step 203: Based on the noise degradation category labels corresponding to the input hyperspectral image, construct a task weight vector with the same dimension as the number of degradation categories. The task weight vector is represented using one-hot encoding; then the task weight vector and the cue basis matrix are used. Perform matrix multiplication to obtain the initial high-dimensional hint vector: ;

[0022] in, Represents the task weight vector. This represents the cue basis matrix composed of the cue vectors corresponding to each noise degradation category. This represents a matrix multiplication operation between a row vector and a matrix. Specifically, it involves weighting and summing the cue vectors corresponding to each noise degradation category in the cue basis matrix according to the task weights to obtain an initial high-dimensional cue vector.

[0023] Preferably, step 3 specifically includes:

[0024] Step 301: Channel Importance Scoring Modeling:

[0025] Construct a content-aware visual gating network that includes global average pooling layers, layer normalization layers, and multilayer perceptrons; input hyperspectral images The input is processed by a visual gating network, which compresses the spatial dimension and maps it to a feature space consistent with the channel dimension of the cue vector, thus obtaining the cue channel importance score vector. :

[0026] ;

[0027] in, This represents the features of the input hyperspectral image. This indicates a global average pooling operation. These represent the weight matrices of the first and second layers in a multilayer perceptron, respectively. Presentation layer normalization operation, This represents the activation function. This represents the channel importance score vector;

[0028] Step 302: Adaptive sparsity threshold determination:

[0029] Based on preset sparse ratio parameters And hint vector channel dimension Calculate the number of reserved channels. :

[0030] ;

[0031] in, This represents the floor function; it then applies to the channel importance score vector. Sort by channel dimension in descending order, and select the first... Large values ​​are used as sparsity thresholds ;

[0032] Step 303: Generation of continuous differentiable sparse masks:

[0033] Using sparsity threshold and channel importance score vector Through a scaling factor Function generates soft binary selection mask :

[0034] ;

[0035] in, express Activation function This represents a scaling factor used to control the steepness of the activation function. Indicates the value in Channel selection mask vector for the interval;

[0036] Step 304: Generate sparse cue vectors:

[0037] Obtaining high-dimensional cue vectors Combine it with the channel selection mask vector Element-wise multiplication yields the content-aware sparse cue vector. :

[0038] ;

[0039] Here, ⊙ represents channel-by-channel multiplication.

[0040] Preferably, step 4 specifically includes:

[0041] Obtaining content-aware sparse cue vectors As guidance information, it is embedded in the encoder-decoder hyperspectral image denoising network structure with multi-cue guidance, and a cross-modal interaction mechanism is introduced to enable the sparse cue vector to participate in the image feature modeling process.

[0042] Specifically, the content-aware sparse cue vector is input to the spatial cue generation module, which is used to establish the interaction relationship between the content-aware sparse cue vector and the learnable visual cue features inside the network. By spatially modulating the sparse cue vector, it is mapped to a spatial cue feature map with the same spatial resolution as the current feature layer.

[0043] Secondly, in the feature interaction path between the encoder and decoder of the denoising network, for each scale level, the spatial cue feature map and the encoder output features of the corresponding level are concatenated in the channel dimension, and the concatenated features are jointly modeled by the Transformer-based feature modeling unit to obtain the encoding-side enhanced features that fuse cue guidance information.

[0044] Secondly, the enhanced features are concatenated with the decoded features of the current level through skip connections, and after channel compression and convolution transformation, they are used as input to the corresponding decoder module to participate in the subsequent feature reconstruction process. Finally, the denoised hyperspectral image is output.

[0045] In this way, the content-aware sparse cue vector is embedded into the multi-scale decoding and reconstruction process of the denoising network in the form of spatially consistent cue features. Through cross-modal attention interaction and feature fusion mechanism, the denoising network is guided to adaptively adjust the feature reconstruction process according to the current image content, thereby realizing the denoising reconstruction of hyperspectral images and outputting denoised hyperspectral images.

[0046] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for hyperspectral image denoising based on content-aware sparse cues.

[0047] A computer device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the above-described method for hyperspectral image denoising based on content-aware sparse cues.

[0048] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0049] This invention not only enables the denoising network to adaptively select effective information based on the input image, but also achieves stable and reliable hyperspectral image denoising and reconstruction in scenarios with unknown noise types and mixed noise degradation. Simultaneously, by sparsely modeling high-dimensional cue features, it effectively reduces the scale of redundant features involved in computation, improving computational efficiency. Furthermore, by introducing a continuously differentiable soft sparse selection mechanism, this invention allows the feature selection process to be jointly trained end-to-end with the denoising network, ensuring sparse modeling effectiveness while improving model training stability and overall convergence efficiency. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart of a hyperspectral image denoising method based on content-aware sparse cues according to the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the generation of content-aware sparse prompts according to the present invention;

[0053] Figure 3 This is a schematic diagram of the hyperspectral image denoising network embedded in this invention. Detailed Implementation

[0054] 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.

[0055] Please see Figures 1-3 The present invention provides the following technical solution:

[0056] Example 1: As Figure 1 As shown, this invention provides a hyperspectral image denoising method based on content-aware sparse cues. Before implementing this method, a data environment for training the denoising network is first constructed, and a corresponding training sample generation strategy is designed to provide a data foundation for subsequent conditional modeling based on content-aware sparse cues.

[0057] Data Preparation: In this embodiment, the publicly available ICVL hyperspectral image dataset is selected as the original noise-free image data required for training. To accommodate the input scale of the deep learning network and facilitate subsequent model training, the original hyperspectral images are cropped using a sliding window method, dividing them into local noise-free image patches with a spatial size of 64×64. The resulting image patches are then constructed into an LMDB format database to improve data reading efficiency during the training phase.

[0058] Training Strategy: During the network training phase, this invention employs a random degradation strategy to generate noisy training samples. Specifically, in each training iteration, the data loading module reads clean image patch samples of a preset batch size from the training data and independently and randomly selects a predefined degradation task mode for each image patch for real-time degradation processing. This ensures that samples with multiple noise degradation types are included within the same training batch, enabling the denoising network to learn noise characteristics and removal rules under multiple degradation modes simultaneously during parameter updates. In this embodiment, multiple typical noise degradation modes are preset to simulate common complex noise interference forms in hyperspectral imaging. These degradation modes include Gaussian noise degradation at different intensity levels, non-independent and identically distributed Gaussian noise degradation, and various mixed noise degradation forms formed by superimposing strip noise, dead-line noise, or impulse noise on non-independent and identically distributed Gaussian noise, thereby covering diverse degradation scenarios from simple random noise to complex structural noise.

[0059] While generating noisy training samples, the system simultaneously records the degradation task label corresponding to each image block and uses the degradation task label as the basis for generating subsequent prompt weights, thereby realizing the correspondence between degradation mode, image sample and degradation description, and providing a data foundation for prompt modeling based on pre-trained text encoder.

[0060] A hyperspectral image denoising method based on content-aware sparse cues specifically includes the following steps:

[0061] Step 1: Acquire noisy hyperspectral images; perform normalization preprocessing on the acquired raw hyperspectral images and spatial cropping using a sliding window method to obtain hyperspectral image sample blocks for training; construct a set of degradation models containing multiple noise degradation modes, and independently and randomly select a degradation mode and its corresponding parameters for each noise-free sample in the batch during training to generate noisy hyperspectral image samples, while recording the corresponding noise degradation category labels.

[0062] Step 2: Generate a high-dimensional cue vector based on the noise degradation category.

[0063] Specifically, firstly, for different types of noise degradation patterns in hyperspectral images, a set of descriptive texts including Gaussian noise, stripe noise, dead-line noise, and mixed noise is constructed. Then, a pre-trained text encoder is loaded to encode the aforementioned descriptive text set, and the encoded feature vectors are normalized. All normalized feature vectors are arranged in order of noise category to form a cue basis matrix, which is stored as a fixed parameter. Based on the noise degradation category labels corresponding to the input hyperspectral image, a task weight vector with the same dimension as the number of degradation categories is constructed. This task weight vector is represented using one-hot encoding. Then, matrix multiplication is performed between the task weight vector and the cue basis matrix to obtain an initial high-dimensional cue vector. This high-dimensional cue vector contains the global prior information required for the current task.

[0064] Step 3: Dynamically generate sparse cue vectors based on image content.

[0065] Step 301, Channel Importance Scoring Modeling:

[0066] A content-aware visual gating network is constructed, comprising a global average pooling layer, a layer normalization layer, and a multilayer perceptron. The image features extracted in step 1 are input into the visual gating network to compress the spatial dimension and map them to a feature space consistent with the channel dimension of the cue vector, thereby obtaining the cue channel importance score vector. :

[0067] :

[0068] in, This indicates a global average pooling operation. These are the weight matrices for the first and second layers of a multilayer perceptron, respectively. Presentation layer normalization operation, This represents the activation function. This is the channel importance scoring vector.

[0069] Step 302: Determining the adaptive sparsity threshold:

[0070] Based on the preset sparsity ratio parameters And hint vector channel dimension Calculate the number of reserved channels. :

[0071] :

[0072] in, This indicates the floor function.

[0073] Channel importance score vector Sort by channel dimension in descending order, and select the first... Large values ​​are used as sparsity thresholds ;

[0074] Step 303: Generation of continuous differentiable sparse masks:

[0075] Using sparsity threshold and channel importance score vector , through with scaling factor Function generates soft binary selection mask :

[0076] ;

[0077] in, This represents the Sigmoid activation function. This is a scaling factor used to control the steepness of the activation function. To take values ​​in Channel selection mask vector for the interval.

[0078] Step 304: Generate sparse cue vectors:

[0079] The high-dimensional cue vector obtained in step 2 With channel selection mask vector Element-wise multiplication yields the content-aware sparse cue vector. :

[0080] ;

[0081] Here, ⊙ represents channel-by-channel multiplication.

[0082] Step 4: Use sparse cueing to guide the denoising network for feature reconstruction. This step is used to embed sparse cueing into the hyperspectral image denoising network to realize the cue-guided feature reconstruction process.

[0083] The content-aware sparse cue vectors obtained in step 3 are embedded as guiding information into a multi-cue-guided encoder-decoder hyperspectral image denoising network structure, introducing a cross-modal interaction mechanism so that the sparse cue vectors participate in the image feature modeling process. Specifically, the content-aware sparse cue vectors are input into a spatial cue generation module, which is used to establish the interaction relationship between the content-aware sparse cue vectors and the learnable visual cue features inside the network. By spatially modulating the sparse cue vectors, they are mapped into a spatial cue feature map with the same spatial resolution as the current feature layer.

[0084] Subsequently, in the feature interaction path between the encoder and decoder of the denoising network, for each scale level, the spatial cue feature map and the encoder output features of the corresponding level are concatenated in the channel dimension. The concatenated features are then jointly modeled by a Transformer-based feature modeling unit to obtain the encoder-side enhanced features that fuse cue guidance information. The enhanced features are then concatenated with the decoder features of the current level through skip connections, and after channel compression and convolution transformation, they are used as the input of the corresponding decoder module to participate in the subsequent feature reconstruction process.

[0085] In this way, the content-aware sparse cue vector is embedded into the multi-scale decoding and reconstruction process of the denoising network in the form of spatially consistent cue features. Through cross-modal attention interaction and feature fusion mechanism, the denoising network is guided to adaptively adjust the feature reconstruction process according to the current image content, thereby realizing the denoising reconstruction of hyperspectral images and outputting denoised hyperspectral images.

[0086] Figure 2 This is a schematic diagram of the content-aware sparse suggestion generation module used in this application example, where the adaptive filtering and sparse modeling module based on visual gating is the core innovative part of this invention. Specifically, the input hyperspectral image is input into the visual gating network, and the spatial dimension is compressed using a global average pooling layer to obtain a global feature vector containing only spectral channel information. Subsequently, this feature vector is sequentially passed through a fully connected layer, a layer normalization layer, and a nonlinear activation function for feature mapping, and a channel importance score vector is output through a second fully connected layer. Rating vector Each element in the table represents the importance of the corresponding cue channel in the current image denoising task.

[0087] Obtain the rating vector Then, based on the preset sparsity parameters... Calculate the number of channels that need to be retained. and the scoring vector Sort and select the first one. Large values ​​as adaptive thresholds To ensure stable training of the sparse selection process, a method based on... The soft binary mask generation method maps the difference between the scoring vector and the threshold to channel weights with values ​​between 0 and 1, thereby constructing an approximately binary channel selection mask vector. .

[0088] Finally, the sparse mask is... The high-dimensional cue vector generated in step 2 is then weighted channel-by-channel to filter out redundant information, resulting in a content-aware sparse cue vector that matches the current image content. This invention achieves a dynamic transformation of cue information from task-level prior to image-level adaptive prior, enabling cue information to adaptively participate based on the input image content.

[0089] like Figure 3 As shown, the method of this invention is constructed based on an existing hyperspectral image denoising network framework, including an encoder for extracting image features, a decoder for restoring image structure, and a cue embedding denoising network path set between the encoder and the decoder. The cue embedding path is implemented by a spatial cue generation module and a cue feature fusion module, and is used for the feature modeling process of embedding the content-aware sparse cue generated by this invention into the denoising network.

[0090] The encoder expands its receptive field through progressive downsampling to extract global and local features, while the decoder progressively restores spatial resolution and outputs a denoised hyperspectral image. The spatial cue generation module and cue feature fusion module are positioned on the skip connection path between the encoder and decoder, serving as interface components for cross-modal feature interaction. These modules primarily address the spatial dimension mismatch between sparse cue vectors and visual features, transforming prior information into guiding signals for image reconstruction through feature alignment and fusion.

[0091] Furthermore, the spatial cue generation module establishes the interaction between content-aware sparse cue vectors and learnable visual cue features within the network. By spatially modulating the sparse cue vectors, it maps them into a spatial cue feature map with the same spatial resolution as the current feature layer. Subsequently, the generated cue feature map is concatenated with the image features output by the encoder along the channel dimension and input into the subsequent feature modeling unit for joint processing. This allows prior information to participate in the image feature reconstruction process, thereby effectively guiding the feature modeling path of the denoising network.

[0092] Among them, the content-aware sparse cue embedding process is carried out in parallel at different levels of the network, so that prior information plays a guiding role in the global structure restoration and local texture reconstruction processes simultaneously, thereby improving the feature modeling ability and reconstruction stability of the denoising network in complex noise degradation scenarios.

[0093] To verify the denoising performance of the method of this invention, a multi-cue guided encoder-decoder hyperspectral image denoising network embedding the method of this invention was compared with other existing methods SST, SERT, HSDT, HyMatt, and DU-TRPCA on the ICVL hyperspectral dataset. Peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were calculated as evaluation indicators, where higher PSNR and SSIM values ​​indicate less image distortion and higher reconstruction quality. Tables 1 and 2 show the denoising results of the method of this invention and various existing methods on the ICVL dataset.

[0094] Table 1. Comparison of Gaussian noise denoising performance between the method of this invention and the comparative method on the ICVL dataset.

[0095]

[0096] Table 2. Comparison of the denoising performance of the method of the present invention and the comparative method on the ICVL dataset for complex noise reduction.

[0097]

[0098] As can be seen from Tables 1 and 2, for the ICVL dataset, the method of this invention achieves the best results in every metric.

[0099] Especially in blind denoising scenarios and complex degradation scenarios containing non-iid noise, stripe noise, dead-line noise, and impulse noise, traditional methods struggle to adaptively model different levels of degradation due to the random variation in noise intensity. This invention, however, introduces a content-aware sparse cueing mechanism, enabling the network to dynamically select effective prior information based on input image features, thus maintaining stable denoising performance under varying noise intensities. Therefore, this invention provides a feasible and efficient hyperspectral image denoising method that significantly improves denoising accuracy.

[0100] The above embodiments specifically describe the overall technical solution and network implementation process of the hyperspectral image denoising method based on content-aware sparse cues proposed in this invention. Through a detailed explanation of the training data construction method, cue generation mechanism, content-aware sparse filtering strategy, and cue-guided denoising and reconstruction process, the effectiveness and practicality of the method in complex noise degradation scenarios are further verified.

[0101] Example 2: The computer-readable storage medium of this example stores a computer program that, when executed by a processor, implements the steps of the hyperspectral image denoising method based on content-aware sparse cues in Example 1.

[0102] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0103] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0104] Example 3: The computer device of this example includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the hyperspectral image denoising method based on content-aware sparse cues in Example 1.

[0105] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0106] Those skilled in the art will understand that the content disclosed in the embodiments can be provided as a method, system, or computer program product. Therefore, this solution can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this solution can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage) containing computer-usable program code.

[0107] This solution is described with reference to flowchart illustrations and / or schematic diagrams of methods and computer program products according to embodiments of this solution. It should be understood that each block of the flowchart illustrations and / or schematic diagrams, and combinations of blocks of the flowchart illustrations and / or schematic diagrams, can be implemented by computer program instructions; these computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the flowchart illustrations and / or block combinations. Figure 1 One or more processes and / or methods are illustrated. Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or methods are illustrated. Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or methods are illustrated. Figure 1 The steps of the function specified in one or more boxes.

[0110] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0111] 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 content-aware sparse cues, characterized in that, Includes the following steps: Step 1: Acquire raw hyperspectral images and preprocess them to obtain hyperspectral image sample blocks for training; During training, noisy hyperspectral image samples are generated based on hyperspectral image sample blocks using a random degradation strategy, and the corresponding noise degradation category labels are recorded. Step 2: Encode the pre-built text describing various noise types using a pre-trained text encoder to obtain the corresponding cue vector set; construct the corresponding cue weight vector based on the noise degradation category label corresponding to the hyperspectral image, and perform a linear weighted summation of the cue vector set using this weight vector to obtain the high-dimensional cue vector corresponding to the current degradation mode; Step 3: Perform global average pooling on the input hyperspectral image to obtain global statistical features for each spectral channel. Input these global statistical features into a content-aware visual gating network to obtain importance score vectors corresponding to each channel of the cue vector. Sort the importance score vectors by channel dimension and select the highest-scoring vectors according to a preset sparsity ratio. The first channel is reserved as a channel, and the second channel is used as the reserved channel. The large score is used as a dynamic threshold to construct a channel selection mask. The mask is then used to filter the high-dimensional cue vector channel by channel to obtain a sparse cue vector. Step 4: Embed the sparse cue vector as guiding information into the hyperspectral image denoising network to guide the feature reconstruction of the hyperspectral image and output the denoised hyperspectral image.

2. The hyperspectral image denoising method based on content-aware sparse cues as described in claim 1, characterized in that, Step 1 specifically includes: Step 101: Obtain the hyperspectral image dataset as the original noise-free image data required for training. Crop the original hyperspectral image using a sliding window method and divide it into local noise-free image blocks with a spatial size of 64×64. Step 102: During the network training phase, a noisy hyperspectral image sample is generated using a random degradation strategy, and the corresponding noise degradation category label is recorded. Specifically, the data loading module reads noise-free image patches of a preset batch size from the training data, and independently and randomly selects a predefined degradation task mode for each image patch for real-time degradation processing. This allows the same training batch to contain samples with multiple noise degradation types, while the system synchronously records the degradation task label corresponding to each image patch. The degradation modes include Gaussian noise degradation at different intensity levels, non-independent and identically distributed Gaussian noise degradation, and various mixed noise degradation forms formed by superimposing strip noise, dead-line noise, or impulse noise on non-independent and identically distributed Gaussian noise.

3. The hyperspectral image denoising method based on content-aware sparse cues as described in claim 1, characterized in that, Step 2 specifically includes: Step 201: For different types of noise degradation patterns in hyperspectral images, construct a set of linguistic description texts containing Gaussian noise, strip noise, dead line noise, and mixed noise; Step 202: Load the pre-trained text encoder, encode the above language description text set, normalize the encoded feature vectors, arrange all normalized feature vectors in order of noise category to form the cue basis matrix, and store it as a fixed parameter. Step 203: Based on the noise degradation category labels corresponding to the input hyperspectral image, construct a task weight vector with the same dimension as the number of degradation categories. The task weight vector is represented using one-hot encoding; then the task weight vector and the cue basis matrix are used. Perform matrix multiplication to obtain the initial high-dimensional hint vector: ; in, Represents the task weight vector. This represents the cue basis matrix composed of the cue vectors corresponding to each noise degradation category. This represents a matrix multiplication operation between a row vector and a matrix. Specifically, it involves weighting and summing the cue vectors corresponding to each noise degradation category in the cue basis matrix according to the task weights to obtain an initial high-dimensional cue vector.

4. The hyperspectral image denoising method based on content-aware sparse cues as described in claim 1, characterized in that, Step 3 specifically includes: Step 301: Channel Importance Scoring Modeling: Construct a content-aware visual gating network that includes global average pooling layers, layer normalization layers, and multilayer perceptrons; input hyperspectral images The input is processed by a visual gating network, which compresses the spatial dimension and maps it to a feature space consistent with the channel dimension of the cue vector, thus obtaining the cue channel importance score vector. : ； in, This represents the features of the input hyperspectral image. This indicates a global average pooling operation. These represent the weight matrices of the first and second layers in a multilayer perceptron, respectively. Presentation layer normalization operation, This represents the activation function. This represents the channel importance score vector; Step 302: Adaptive sparsity threshold determination: Based on preset sparse ratio parameters And hint vector channel dimension Calculate the number of reserved channels. : ； in, This represents the floor function; it then applies to the channel importance score vector. Sort by channel dimension in descending order, and select the first... Large values ​​are used as sparsity thresholds ; Step 303: Generation of continuous differentiable sparse masks: Using sparsity threshold and channel importance score vector Through a scaling factor Function generates soft binary selection mask : ； in, express Activation function This represents a scaling factor used to control the steepness of the activation function. Indicates the value in Channel selection mask vector for the interval; Step 304: Generate sparse cue vectors: Obtaining high-dimensional cue vectors Combine it with the channel selection mask vector Element-wise multiplication yields the content-aware sparse cue vector. : ； Here, ⊙ represents channel-by-channel multiplication.

5. The hyperspectral image denoising method based on content-aware sparse cues as described in claim 1, characterized in that, Step 4 specifically includes: Obtaining content-aware sparse cue vectors As guidance information, it is embedded in the encoder-decoder hyperspectral image denoising network structure with multi-cue guidance, and a cross-modal interaction mechanism is introduced to enable the sparse cue vector to participate in the image feature modeling process. Specifically, the content-aware sparse cue vector is input to the spatial cue generation module, which is used to establish the interaction relationship between the content-aware sparse cue vector and the learnable visual cue features inside the network. By spatially modulating the sparse cue vector, it is mapped to a spatial cue feature map with the same spatial resolution as the current feature layer. Secondly, in the feature interaction path between the encoder and decoder of the denoising network, for each scale level, the spatial cue feature map and the encoder output features of the corresponding level are concatenated in the channel dimension, and the concatenated features are jointly modeled by the Transformer-based feature modeling unit to obtain the encoding-side enhanced features that fuse cue guidance information. Secondly, the enhanced features are concatenated with the decoded features of the current level through skip connections, and after channel compression and convolution transformation, they are used as input to the corresponding decoder module to participate in the subsequent feature reconstruction process. Finally, the denoised hyperspectral image is output.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps in the hyperspectral image denoising method based on content-aware sparse cues as described in any one of claims 1-5.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps in the hyperspectral image denoising method based on content-aware sparse cues as described in any one of claims 1-5.

Images