Method, device and equipment for removing strip noise of remote sensing image and medium

Abstract

The application relates to the technical field of remote sensing image processing, in particular to a strip noise removal method and device for remote sensing images, equipment and a medium, wherein the method comprises the following steps: decomposing a remote sensing image into a low-frequency first frequency subband and a high-frequency second frequency subband, and extracting first image features and second image features respectively; inputting the first image features into a first processing network to extract multi-scale features and enhance a strip region through an attention mechanism, so as to obtain a first feature map; inputting the second image features into a second processing network, processing strip noise through sparse convolution and a strip window attention, and reconstructing through inverse wavelet transform, so as to obtain a second feature map; and finally fusing the two, so as to obtain a remote sensing image from which strip noise is removed. Thus, the problems that related technologies depend on manual parameter adjustment or modeling for specific noise, cannot retain details and cannot effectively remove noise suppression and the like are solved.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a method, apparatus, device and medium for stripe noise removal from remote sensing images. Background Technology

[0002] With the development of remote sensing technology, the demand for high-quality imagery is constantly increasing. However, during high-resolution satellite imaging and data transmission, stripe noise often appears in the images due to factors such as environmental conditions and sensor performance, affecting their subsequent applications.

[0003] To address the aforementioned issues, while existing technologies can suppress noise to some extent, they typically rely on manual parameter adjustments, require customized modeling for specific noise types, and exhibit strip noise in the spatial domain as a coupling of high-frequency directional brightness abrupt changes and low-frequency brightness fluctuations, making it difficult to effectively separate high and low frequency noise, thus increasing the difficulty of denoising high-resolution images. Summary of the Invention

[0004] This application provides a method, apparatus, device, and medium for strip noise removal from remote sensing images, to solve the problems of related technologies that rely on manual parameter adjustment or modeling for specific noise, which cannot preserve details and effectively remove noise.

[0005] The first aspect of this application provides a method for stripe noise removal from remote sensing images, comprising the following steps: decomposing the remote sensing image into a first frequency sub-band and a second frequency sub-band; extracting a first image feature from the first frequency sub-band and a second image feature from the second frequency sub-band, wherein the first frequency sub-band is a low-frequency sub-band with a frequency lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with a frequency higher than a second preset frequency; inputting the first image feature into a pre-constructed first processing network, the first processing network extracting multi-scale features from the first image feature and enhancing the stripe regions of the multi-scale features based on an attention mechanism to obtain a first feature map; inputting the second image feature into a pre-constructed second processing network, the second processing network processing the second image feature through sparse convolution and a striped window attention mechanism to process the stripe noise, and reconstructing the denoised image through inverse wavelet transform to obtain a second feature map; and fusing the first feature map and the second feature map to obtain a remote sensing image with stripe noise removed.

[0006] Optionally, in one embodiment of this application, the second frequency sub-band includes at least one of a horizontal high-frequency sub-band, a vertical high-frequency sub-band, and a diagonal high-frequency sub-band, and the expressions for the first frequency sub-band and the second frequency sub-band are:

[0007] in, Represents low frequency. Represents high frequency levels, Represents vertical high frequency. Represents high frequency on the diagonal.

[0008] Optionally, in one embodiment of this application, the first processing network and the second processing network process their respective image features in parallel.

[0009] Optionally, in one embodiment of this application, the first processing network includes dynamic convolutional blocks, residual dense blocks, and a gradient attention module, wherein, The formula for calculating dynamic convolutional blocks is:

[0010] in, This is a fused feature map of a dynamic convolutional block. This represents a 5×5 convolution operation. For learning weights, This refers to the index subscript of the parallel convolution kernel. For convolution operations, For dynamic convolution kernel weights, This is the input feature map for this module; The first residual dense block l The layer output is represented as:

[0011] in, For the first l The output feature map of the layer, It is the ReLU activation function. For the first l Layer weights, Indicates feature concatenation operation; The calculation formula for the gradient attention module: in, This is a Sobel attention weight graph; For activation functions; W sobel For vertical Sobel attention weights, This is to take the absolute value of the result of the convolution operation.

[0012] in, This is the Sobel attention weight graph. For activation function, W sobel For vertical Sobel attention weights, This is to take the absolute value of the result of the convolution operation.

[0013] Optionally, in one embodiment of this application, the second processing network includes sparse convolution kernels, a strip window attention mechanism module, and a cyclic shifting module, wherein, The formula for calculating sparse convolution kernels is: in, The output feature map of sparse convolution. It is a sparse convolution kernel in the horizontal direction. It is a sparse convolution kernel in the vertical direction. For learnable fringe direction attention weight parameters; in, The output feature map of sparse convolution. It is a sparse convolution kernel in the horizontal direction. It is a sparse convolution kernel in the vertical direction. For learnable fringe direction attention weight parameters; The calculation formula for the strip window attention mechanism module is as follows:

[0014] in, This is an attention mechanism function used to weight and aggregate values ​​based on the similarity between the query and the key. The normalized exponential function is used to map attention scores into a probability distribution such that the sum of all weights is 1. It is a strip mask matrix. Q Used to characterize the query requirements of the currently processed pixel or region during the denoising process. K Feature identifiers used to characterize individual pixels or regions in an image. V This represents the original image information or deep feature information to be aggregated. This is the transpose of the key matrix, used in conjunction with the query matrix. Q Perform matrix multiplication to calculate the similarity between each query and key pair. The scaling factor is usually taken as the dimension of the key vector (i.e., the feature length of each key), used to scale the dot product result to a suitable range and avoid gradient saturation; The calculation formula for the cyclic displacement module is as follows:

[0015] in, This is the feature map after cyclic displacement. Roll This is a cyclic displacement sliding operation. F For the feature map, Δ y This is the preset displacement.

[0016] Optionally, in one embodiment of this application, fusing the first feature map and the second feature map to obtain a remote sensing image with stripe noise removed includes: The first and second feature maps are input into a pre-constructed multi-frequency feature fusion network. The multi-frequency feature fusion network outputs a remote sensing image with stripe noise removed. The fusion formula of the multi-frequency feature fusion network is as follows:

[0017] in, This is the merged attention weight map. To merge convolutional kernels, It is the Sigmoid activation function. This is a low-frequency feature map. This is a high-frequency feature map.

[0018] Optionally, in one embodiment of this application, the multi-frequency feature fusion network employs a multi-objective loss function and a dynamic weight adjustment strategy to optimize network parameters, wherein the formula for the multi-objective loss function is:

[0019] in, For the total loss function, The weighting coefficients for the mean squared error loss are: These are the weighting coefficients for the horizontal gradient loss. Weighting coefficients to counteract losses, It is the loss of the mean square error. It is a horizontal gradient loss constraining the loss of the difference between adjacent columns. It is to counteract some of the losses; The formula for the dynamic weight adjustment strategy is:

[0020] in, For dynamic weight adjustment strategy, Control the rate of weight growth. To normalize the training progress (initial stage) Later ).

[0021] A second aspect of this application provides a stripe noise removal device for remote sensing images, comprising: an extraction module for decomposing the remote sensing image into a first frequency sub-band and a second frequency sub-band, extracting a first image feature from the first frequency sub-band and a second image feature from the second frequency sub-band, wherein the first frequency sub-band is a low-frequency sub-band with a frequency lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with a frequency higher than a second preset frequency; a first processing module for inputting the first image features into a pre-constructed first processing network, wherein the first processing network extracts multi-scale features from the first image features and enhances the stripe regions of the multi-scale features based on an attention mechanism to obtain a first feature map; a second processing module for inputting the second image features into a pre-constructed second processing network, wherein the second processing network processes the second image features through sparse convolution and a striped window attention mechanism to process the stripe noise, and reconstructs the denoised image through inverse wavelet transform to obtain a second feature map; and a fusion module for fusing the first feature map and the second feature map to obtain a remote sensing image with stripe noise removed.

[0022] Optionally, in one embodiment of this application, the second frequency sub-band includes at least one of a horizontal high-frequency sub-band, a vertical high-frequency sub-band, and a diagonal high-frequency sub-band, and the expressions for the first frequency sub-band and the second frequency sub-band are:

[0023] in, Represents low frequency. Represents high frequency levels, Represents vertical high frequency. Represents high frequency on the diagonal.

[0024] Optionally, in one embodiment of this application, the first processing network and the second processing network process their respective image features in parallel.

[0025] Optionally, in one embodiment of this application, the first processing network includes dynamic convolutional blocks, residual dense blocks, and a gradient attention module, wherein, The formula for calculating dynamic convolutional blocks is:

[0026] in, This is a fused feature map of a dynamic convolutional block. This represents a 5×5 convolution operation. For learning weights, This refers to the index subscript of the parallel convolution kernel. For convolution operations, For dynamic convolution kernel weights, This is the input feature map for this module; The first residual dense block lThe layer output is represented as:

[0027] in, For the first l The output feature map of the layer, It is the ReLU activation function. For the first l Layer weights, Indicates feature concatenation operation; The calculation formula for the gradient attention module: in, This is a Sobel attention weight graph; For activation functions; W sobel For vertical Sobel attention weights, This is to take the absolute value of the result of the convolution operation.

[0028] in, This is the Sobel attention weight graph. For activation function, W sobel For vertical Sobel attention weights, This is to take the absolute value of the result of the convolution operation.

[0029] Optionally, in one embodiment of this application, the second processing network includes sparse convolution kernels, a strip window attention mechanism module, and a cyclic shifting module, wherein, The formula for calculating sparse convolution kernels is:

[0030] in, The output feature map of sparse convolution. It is a sparse convolution kernel in the horizontal direction. It is a sparse convolution kernel in the vertical direction. For learnable fringe direction attention weight parameters; The calculation formula for the strip window attention mechanism module is as follows:

[0031] in, This is an attention mechanism function used to weight and aggregate values ​​based on the similarity between the query and the key. The normalized exponential function is used to map attention scores into a probability distribution such that the sum of all weights is 1. It is a strip mask matrix. Q Used to characterize the query requirements of the currently processed pixel or region during the denoising process. KFeature identifiers used to characterize individual pixels or regions in an image. V This represents the original image information or deep feature information to be aggregated. This is the transpose of the key matrix, used in conjunction with the query matrix. Q Perform matrix multiplication to calculate the similarity between each query and key pair. The scaling factor is usually taken as the dimension of the key vector (i.e., the feature length of each key), used to scale the dot product result to a suitable range and avoid gradient saturation; The calculation formula for the cyclic displacement module is as follows:

[0032] in, This is the feature map after cyclic displacement. Roll This is a cyclic displacement sliding operation. F For the feature map, Δ y This is the preset displacement.

[0033] Optionally, in one embodiment of this application, the fusion module is further configured to input the first feature map and the second feature map into a pre-constructed multi-frequency feature fusion network, the multi-frequency feature fusion network outputting a remote sensing image with stripe noise removed, and the fusion formula of the multi-frequency feature fusion network is:

[0034] in, This is the merged attention weight map. To merge convolutional kernels, It is the Sigmoid activation function. This is a low-frequency feature map. This is a high-frequency feature map.

[0035] Optionally, in one embodiment of this application, the multi-frequency feature fusion network employs a multi-objective loss function and a dynamic weight adjustment strategy to optimize network parameters, wherein the formula for the multi-objective loss function is:

[0036] in, For the total loss function, The weighting coefficients for the mean squared error loss are: These are the weighting coefficients for the horizontal gradient loss. Weighting coefficients to counteract losses, It is the loss of the mean square error. It is a horizontal gradient loss constraining the loss of the difference between adjacent columns. It is to counteract some of the losses; The formula for the dynamic weight adjustment strategy is:

[0037] in, For dynamic weight adjustment strategy, Control the rate of weight growth. To normalize the training progress (initial stage) Later ).

[0038] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the strip noise removal method for remote sensing images as described in the above embodiments.

[0039] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the strip noise removal method for remote sensing images as described in the above embodiments.

[0040] Therefore, this application has the following beneficial effects: First, the remote sensing image is decomposed into a first frequency sub-band and a second frequency sub-band. First image features are extracted from the first frequency sub-band, and second image features are extracted from the second frequency sub-band. The first frequency sub-band is a low-frequency sub-band with frequencies lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with frequencies higher than a second preset frequency. This achieves preliminary separation of high and low frequency information, providing clear frequency domain features for subsequent fine processing. Second, the first image features are input into a pre-constructed first processing network. The first processing network extracts multi-scale features from the first image features and enhances the striped regions based on an attention mechanism. An enhanced first feature map is generated to efficiently extract and enhance low-frequency structural information and global texture features. Then, the second image feature map is input into a pre-constructed second processing network. This network processes stripe noise through sparse convolution and a striped window attention mechanism, and reconstructs the denoised image using inverse wavelet transform to obtain the second feature map. This achieves precise suppression of high-frequency details and stripe noise while maintaining the integrity of image details. Finally, the first and second feature maps are fused to obtain a remote sensing image with stripe noise removed, achieving complementary fusion of low-frequency structure and high-frequency details, significantly improving the visual quality and denoising effect of the image. This solves the problem that related technologies rely on manual parameter adjustment or modeling for specific noise, making it difficult to preserve image details and effectively suppress stripe noise.

[0041] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0042] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for stripe noise removal from remote sensing images according to an embodiment of this application; Figure 2 This is a schematic diagram of a multi-frequency feature fusion network structure according to an embodiment of this application; Figure 3 This is a schematic diagram of the multi-frequency domain stripe noise removal process according to an embodiment of this application; Figure 4 This is an example diagram of a stripe noise removal apparatus for remote sensing images according to an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0043] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0044] The following description, with reference to the accompanying drawings, describes a method, apparatus, device, and medium for stripe noise removal from remote sensing images according to embodiments of this application. Addressing the problems mentioned in the background art, this application provides a method for stripe noise removal from remote sensing images. In this method, firstly, the remote sensing image is decomposed into a first frequency sub-band and a second frequency sub-band. First image features are extracted from the first frequency sub-band, and second image features are extracted from the second frequency sub-band, respectively. The first frequency sub-band is a low-frequency sub-band with a frequency lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with a frequency higher than a second preset frequency, thereby achieving preliminary separation of high and low frequency information and providing clear frequency domain features for subsequent fine processing. Secondly, the first image features are input into a pre-constructed first processing network, which extracts multiple... The system extracts scale features and enhances stripe regions using an attention mechanism to generate an enhanced first feature map, achieving efficient extraction and enhancement of low-frequency structural information and global texture features. Then, a second image feature is input into a pre-constructed second processing network. This network processes stripe noise through sparse convolution and a striped window attention mechanism, and reconstructs the denoised image using inverse wavelet transform to obtain a second feature map. This achieves precise suppression of high-frequency details and stripe noise while maintaining the integrity of image details. Finally, the first and second feature maps are fused to obtain a remote sensing image with stripe noise removed, achieving complementary fusion of low-frequency structure and high-frequency details, significantly improving the visual quality and denoising effect of the image. This solves the problem that related technologies rely on manual parameter adjustment or modeling for specific noise, making it difficult to preserve image details and effectively suppress stripe noise.

[0045] Specifically, Figure 1 This is a flowchart illustrating a method for stripe noise removal from remote sensing images, provided as an embodiment of this application.

[0046] like Figure 1 As shown, the method for removing stripe noise from this remote sensing image includes the following steps: In step S101, the remote sensing image is decomposed into a first frequency sub-band and a second frequency sub-band, and a first image feature is extracted from the first frequency sub-band and a second image feature is extracted from the second frequency sub-band, respectively. The first frequency sub-band is a low-frequency sub-band with a frequency lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with a frequency higher than a second preset frequency.

[0047] In this application, remote sensing images are large-scale images of the Earth's surface taken from the air by sensors such as satellites and drones. The first frequency sub-band is the low-frequency portion (below a first preset frequency) after frequency domain decomposition; the second frequency sub-band is the high-frequency portion (above a second preset frequency) after frequency domain decomposition. Frequency domain decomposition is the process of converting an image from the spatial domain to the frequency domain and separating different frequency components; in this application, it represents the basic operation for extracting low-frequency and high-frequency features. The first preset frequency is the threshold used to divide the low-frequency sub-band; in this application, it represents the highest frequency boundary of the low-frequency sub-band. The second preset frequency is the threshold used to divide the high-frequency sub-band; in this application, it represents the lowest frequency boundary of the high-frequency sub-band. The low-frequency sub-band is the portion of the image decomposition with frequencies below the first preset frequency; the high-frequency sub-band is the portion of the image decomposition with frequencies above the second preset frequency. The first image feature is the feature information extracted from the first frequency sub-band; the second image feature is the feature information extracted from the second frequency sub-band.

[0048] Understandably, the first frequency sub-band preserves the main structure and brightness distribution of ground features, which is beneficial for maintaining the texture and information of the image; the second frequency sub-band contains high-frequency detail information and stripe noise, allowing high-frequency noise to be independently identified and suppressed. By separating low-frequency and high-frequency processing, the accuracy of stripe noise removal is improved, interference and information aliasing during processing are reduced, and clearer and more targeted input features are provided for subsequent multi-scale feature extraction and attention mechanism enhancement. By decomposing remote sensing images into low-frequency and high-frequency sub-bands, effective separation of different frequency components of the image is achieved, thereby enabling targeted processing of the overall structure and detail information of the image.

[0049] In one embodiment of this application, the second frequency sub-band includes at least one of a horizontal high-frequency sub-band, a vertical high-frequency sub-band, and a diagonal high-frequency sub-band, and the expressions for the first and second frequency sub-bands are as follows:

[0050] in, For wavelet basis functions, This represents the convolution operation. Represents low frequency. Represents high frequency levels, Represents vertical high frequency. Represents high frequency on the diagonal.

[0051] It is understood that the embodiments of this application use wavelet decomposition to clearly divide the image into low-frequency sub-bands. ) and multi-directional high-frequency subbands ( , , It can accurately distinguish the overall structural information of an image from high-frequency details and stripe noise in different directions, giving a clear target for subsequent processing of horizontal, vertical or diagonal stripes.

[0052] In this application, a low-frequency branch Flow and a high-frequency branch Fhigh are introduced to separately model the different frequency domain information after wavelet decomposition. The first frequency sub-band obtained by wavelet decomposition... As a low-frequency branch The input, at least one sub-band in the second frequency sub-band (including) , , This is used as the input to the high-frequency branch Fhigh for subsequent processing.

[0053] In step S102, the first image features are input into a pre-constructed first processing network. The first processing network extracts multi-scale features from the first image features and enhances the stripe region of the multi-scale features based on an attention mechanism to obtain a first feature map.

[0054] In this application, the first processing network is a deep neural network designed for the first image features, representing a computational network used for low-frequency feature processing and enhancing structural information; the multi-scale features are image features of different spatial scales extracted by the first processing network, representing a set of features that capture the overall structure of the image and information at different resolutions; the attention mechanism is a computational method used to highlight important features and suppress irrelevant information; the striped region is a specific high-frequency or structurally abnormal part of the image that appears due to stripe noise or stripe effect, representing a region that needs to be focused on by the attention mechanism; and the first feature map is the feature representation output by the first processing network, representing a multi-scale feature map of the striped region after low-frequency subband processing.

[0055] Understandably, by performing multi-scale analysis on the first image features extracted from the low-frequency subband through the first processing network, structural and textural information can be captured simultaneously. The attention mechanism is used to focus on enhancing the stripe region, so that the low-frequency features related to stripe noise are effectively enhanced and distinguished, and the generated first feature map retains the overall structure of the image.

[0056] The first image features are input into a pre-constructed first processing network. Specifically, the first processing network first performs wavelet transform on the first image features to obtain... , , , Sub-bands, among which The subbands preserve the overall structure and main energy of the image; therefore, the low-frequency branch Flow is responsible for feature extraction and representation mapping, which can be formally represented as follows: ′= Correspondingly, , , The high-frequency subband contains more texture, detail, and noise components, and is processed by the high-frequency branch Fhigh. Their mapping relationship can be expressed as follows: ′ = Through this frequency-division processing architecture, the first processing network can acquire complementary multi-scale features from both low-frequency and high-frequency dimensions. Building upon this, the network further introduces an attention mechanism to focus on enhancing the striped regions within the aforementioned multi-scale features, ultimately generating a first feature map that highlights effective texture information.

[0057] In one embodiment of this application, the first processing network includes dynamic convolutional blocks, residual dense blocks, and gradient attention modules, wherein the calculation formula for the dynamic convolutional blocks is:

[0058] in, This is a fused feature map of a dynamic convolutional block. This represents a 5×5 convolution operation. For learning weights, This refers to the index subscript of the parallel convolution kernel. For convolution operations, For dynamic convolution kernel weights, This is the input feature map for this module; The first residual dense block l The layer output is represented as:

[0059] in, For the first l The output feature map of the layer, It is the ReLU activation function. For the first l Layer weights, Indicates feature concatenation operation; The calculation formula for the gradient attention module:

[0060] in, This is the Sobel attention weight graph. For activation function, W sobel For vertical Sobel attention weights, This is to take the absolute value of the result of the convolution operation.

[0061] It is understood that the embodiments of this application achieve efficient enhancement of low-frequency subband features and accurate highlighting of stripe regions through multi-module collaborative processing within the first processing network. The dynamic convolutional block utilizes learnable multi-path convolution to fuse features of different directions and scales, improving the adaptive expressive ability of low-frequency features; the residual dense block enhances the network's ability to capture complex structural information through cross-layer feature reuse and multi-scale feature accumulation; the gradient attention module enhances the vertical stripe sensitive region based on the Sobel operator, enabling the low-frequency features related to stripe noise to be processed in a focused manner.

[0062] In one embodiment of this application, the first processing network sequentially includes a dynamic convolutional block, a residual dense block, and a gradient attention module, used for multi-scale and multi-directional depth representation of the first image features. First, the dynamic convolutional block (DCB) achieves elastic expansion of the feature space through four parallel convolutional kernels, and its fused output can be expressed as:

[0063] in, For learnable weights and satisfying This structure can capture rich feature patterns in different directions and scales, and achieve effective fusion through adaptive weights, thereby enhancing the dynamic feature extraction capability.

[0064] Building upon this, residual dense blocks are used to further accumulate multi-scale features, employing a five-layer dense connection structure, the first of which... The layer output is:

[0065] in, This is a Sobel attention weight graph; For activation functions; W sobel For vertical Sobel attention weights, This is to take the absolute value of the result of the convolution operation.

[0066] Furthermore, to highlight the stripe structure features and suppress irrelevant region responses, the first processing network also includes a gradient attention module. This module employs a vertical Sobel operator to construct a gradient-sensitive mechanism. The Sobel operator is a classic image gradient detection operator, primarily used to calculate gradient changes in the horizontal or vertical directions of an image, thereby extracting edge or texture structures.

[0067] The formula for calculating the attention map is:

[0068] The expression for the Sobel operator is:

[0069] Used to extract vertical structure gradients, by analyzing low-frequency... Convolution of sub-bands combined with Sigmoid adaptive mapping effectively enhances the response intensity of striped regions. The attention features obtained after processing by this module are further used to highlight structurally significant regions, ultimately forming an enhanced feature map that guides subsequent feature fusion.

[0070] In step S103, the second image features are input into a pre-constructed second processing network. The second processing network processes the second image features through sparse convolution and strip window attention mechanism to handle strip noise, and reconstructs the denoised image through inverse wavelet transform to obtain the second feature map.

[0071] The second processing network is a deep neural network designed for the second image features, representing a computational model for high-frequency feature processing, strip noise suppression, and detail enhancement in this application; sparse convolution is a convolution method that applies convolution operations only in specific directions or positions; the strip window attention mechanism is a local attention mechanism that calculates attention weights in specific strip regions of the image; inverse wavelet transform is the process of recombining the sub-bands after frequency domain decomposition back into the spatial domain image, representing the operation of reconstructing a denoised image from the high-frequency processing results in this application; the denoised image is the image processed by the second processing network and inverse wavelet transform; the second feature map is the feature representation output by the second processing network, representing the feature map for strip noise suppression and detail enhancement after high-frequency sub-band processing in this application.

[0072] Understandably, the second processing network targets the second image features extracted from the high-frequency subbands to achieve efficient suppression of stripe noise and preservation of high-frequency details. Sparse convolution can enhance stripe noise features in the horizontal or vertical directions, accurately capturing the stripe noise; the striped window attention mechanism further highlights the stripe region and suppresses irrelevant high-frequency interference through attention calculation in local areas; inverse wavelet transform reconstructs the processed high-frequency features into a spatial domain image, ensuring the fidelity of high-frequency details after denoising. Overall, this step improves the directional suppression accuracy of stripe noise while preserving image detail information, providing a high-quality high-frequency feature map for the final multi-frequency feature fusion, and enhancing the overall effect of remote sensing image denoising.

[0073] In one embodiment of this application, the second processing network includes sparse convolution kernels, a strip window attention mechanism module, and a cyclic shifting module, wherein, The formula for calculating sparse convolution kernels is:

[0074] in, The output feature map of sparse convolution. It is a sparse convolution kernel in the horizontal direction. It is a sparse convolution kernel in the vertical direction. These are the learnable fringe direction attention weight parameters.

[0075] The calculation formula for the strip window attention mechanism module is as follows:

[0076] in, This is an attention mechanism function used to weight and aggregate values ​​based on the similarity between the query and the key. The normalized exponential function is used to map attention scores into a probability distribution such that the sum of all weights is 1. It is a strip mask matrix. Q Used to characterize the query requirements of the currently processed pixel or region during the denoising process. K Feature identifiers used to characterize individual pixels or regions in an image. V This represents the original image information or deep feature information to be aggregated. This is the transpose of the key matrix, used in conjunction with the query matrix. Q Perform matrix multiplication to calculate the similarity between each query and key pair. The scaling factor is usually taken as the dimension of the key vector (i.e., the feature length of each key), used to scale the dot product result to a suitable range and avoid gradient saturation; The calculation formula for the cyclic displacement module is as follows:

[0077] in, This is the feature map after cyclic displacement. Roll This is a cyclic displacement sliding operation. F For the feature map, Δ y This is the preset displacement.

[0078] Understandably, by introducing sparse convolutional kernels, a striped window attention mechanism module, and a cyclic shift module into the second processing network, this application can significantly improve the extraction and suppression of strip noise in high-frequency subbands. Specifically, the sparse convolutional kernels employ an asymmetric convolutional structure in both the horizontal and vertical directions; the striped window attention mechanism utilizes a striped mask matrix to limit the attention range, making the attention calculation more focused on the long strip-shaped dependencies of the striped region and reducing interference from irrelevant locations; and the cyclic shift module performs periodic shifts on the features, enabling the striped window to form continuous contextual associations at different positions. The synergy of these three components allows the second processing network to more accurately capture the directional characteristics and local variations of striped noise.

[0079] In one embodiment of this application, the high-frequency attention optimization model achieves efficient enhancement processing of high-frequency features, especially strip noise, through a two-stage structure combining window attention and global attention. This application as a whole consists of sparse convolution kernels, a strip window attention mechanism, and a cyclic shift module, as follows: First, the high-frequency processing pathway adopts a two-stage architecture. First, sparse convolution kernels are used to enhance the representation of structural features in the horizontal direction. Then, a strip window attention mechanism module and a cyclic shift module are introduced to expand the receptive field through cyclic shift. Finally, high-frequency information is reconstructed by inverse wavelet transform.

[0080] In the sparse convolution stage, 7×1 and 1×7 sparse convolution kernels are used for the high-frequency subbands. The computational relationship for enhancing orientation perception is as follows:

[0081] in, The output feature map of sparse convolution. It is a sparse convolution kernel in the horizontal direction. It is a sparse convolution kernel in the vertical direction. For learnable fringe direction attention weight parameters; Subsequently, a 4×32 strip window structure is introduced, and local strip attention modeling is completed through attention weights, as shown in the following formula:

[0082] in, This is an attention mechanism function used to weight and aggregate values ​​based on the similarity between the query and the key. The normalized exponential function is used to map attention scores into a probability distribution such that the sum of all weights is 1. It is a strip mask matrix. Q Used to characterize the query requirements of the currently processed pixel or region during the denoising process. K Feature identifiers used to characterize individual pixels or regions in an image. V This represents the original image information or deep feature information to be aggregated. This is the transpose of the key matrix, used in conjunction with the query matrix. Q Perform matrix multiplication to calculate the similarity between each query and key pair. The scaling factor is usually taken as the dimension of the key vector (i.e., the feature length of each key), used to scale the dot product result to a suitable range and avoid gradient saturation; Furthermore, this application introduces a cyclic displacement module consisting of vertical shift, horizontal shift, and cross-modal attention, which performs horizontal displacement of the feature map using the following cyclic displacement operator:

[0083] This expands the receptive field across the window boundary, with particular emphasis on horizontal shift attention for targeted optimization against vertical strip noise.

[0084] Finally, the dual outputs are dimensionality reduced by 1×1 convolution and then reconstructed using inverse wavelet transform to obtain high-quality denoising results. This effectively reduces the overall computational complexity from O(n²) of traditional window attention to O(n²). It significantly improves computational efficiency while maintaining high-precision reasoning.

[0085] In one embodiment of this application, a first processing network and a second processing network process their respective image features in parallel.

[0086] Understandably, the first and second processing networks process their respective image features in parallel, thereby achieving efficient extraction and optimization of multi-frequency features from remote sensing images. Specifically, the low-frequency branch of the first processing network extracts global structural features, while the high-frequency branch performs detail enhancement and noise suppression; meanwhile, the second processing network further optimizes high-frequency strip features through sparse convolution kernels, a strip window attention mechanism module, and a cyclic shift module.

[0087] In step S104, the first feature map and the second feature map are fused to obtain a remote sensing image with strip noise removed.

[0088] The first feature map is the feature map after extracting, transforming and enhancing the stripe response of the first image features by the low-frequency branch and high-frequency branch output by the first processing network; the second feature map is the feature map obtained by processing the high-frequency sub-bands of the high-frequency attention optimization result output by the second processing network through sparse convolution, strip window attention mechanism module and cyclic shift module.

[0089] Understandably, the step of fusing the first and second feature maps can fully leverage the complementary advantages of low-frequency structural features and high-frequency texture features, thereby enhancing detail while preserving the overall image structure. Through the fusion operation, this application can more effectively suppress stripe noise, reduce residual stripes and artifacts, and improve the quality and visual effect of the denoised image.

[0090] In one embodiment of this application, fusing a first feature map and a second feature map to obtain a remote sensing image with stripe noise removed includes: inputting the first feature map and the second feature map into a pre-constructed multi-frequency feature fusion network, the multi-frequency feature fusion network outputting the remote sensing image with stripe noise removed, and the fusion formula of the multi-frequency feature fusion network being:

[0091] in, This is the merged attention weight map. To merge convolutional kernels, It is the Sigmoid activation function. This is a low-frequency feature map. This is a high-frequency feature map.

[0092] Understandably, this application inputs the first feature map and the second feature map into a pre-constructed multi-frequency feature fusion network for processing. It utilizes fusion convolution kernels to weightedly combine features from different frequency bands, enhancing texture details and the response of striped regions while maintaining the overall image structure. By fully leveraging the complementarity of low-frequency and high-frequency features, it achieves efficient suppression of striped noise.

[0093] The multi-frequency feature fusion network architecture of this application embodiment is shown in the figure. Figure 2 As shown, the input remote sensing image is first decomposed into low-frequency sub-bands through discrete wavelet transform. as well as , , Three high-frequency sub-bands are used to characterize the overall structure and detailed texture of the image, respectively. The low-frequency sub-band enters the low-frequency processing branch, where low-frequency variation features are adaptively extracted through dynamic convolutional blocks, and important regions are further enhanced by residual dense blocks and a visual attention module to obtain low-frequency reconstructed features. Simultaneously, the three high-frequency sub-bands are merged into the high-frequency processing branch, where directional texture and stripe information are extracted by a feature enhancement module, and high-frequency reconstructed features are generated after fine optimization by residual blocks. Finally, the low-frequency and high-frequency reconstruction results are fused in the spatial domain through inverse wavelet transform to obtain a high-quality remote sensing image after stripe noise removal.

[0094] In one embodiment of this application, the multi-frequency feature fusion network employs a multi-objective loss function and a dynamic weight adjustment strategy to optimize network parameters. The formula for the multi-objective loss function is:

[0095] in, For the total loss function, The weighting coefficients for the mean squared error loss are: These are the weighting coefficients for the horizontal gradient loss. Weighting coefficients to counteract losses, It is the loss of the mean square error. It is a horizontal gradient loss constraining the loss of the difference between adjacent columns. It is to counteract some of the losses; The formula for the dynamic weight adjustment strategy is:

[0096] in, For dynamic weight adjustment strategy, Control the rate of weight growth. To normalize the training progress (initial stage) Later ).

[0097] Understandably, the multi-objective loss function integrates mean squared error loss, high-frequency gradient loss, and adversarial loss, enabling the network to enhance high-frequency texture details and suppress strip noise while preserving the overall image structure. The dynamic weight adjustment strategy adaptively adjusts the weights of each loss term according to the normalized training progress, ultimately achieving coordinated optimization between low-frequency structure and high-frequency details.

[0098] In one embodiment of this application, the multi-frequency feature fusion network uses a joint loss function to guide network training, as shown in the formula: The joint loss function takes into account three key optimization objectives: pixel-level reconstruction, horizontal gradient constraint, and adversarial training.

[0099] Among them, the pixel-level mean square error loss is:

[0100] Used to ensure accurate reconstruction of the overall structure and brightness information of the image; The horizontal gradient loss is:

[0101] Used to enhance the texture details of an image, especially the structural information of the stripe direction; The losses from the confrontation are:

[0102] Adversarial training enhances the realism and naturalness of high-frequency features.

[0103] During training, a dynamic weight adjustment strategy is used to achieve incremental optimization. The weight formula is as follows:

[0104] The dynamic weight adjustment strategy enables network training to transition from focusing on pixel-level reconstruction in the early stages to strengthening high-frequency details and adversarial optimization in the later stages, thereby achieving smooth and progressive optimization from global structure to local texture.

[0105] The strip noise removal process for remote sensing images in this application embodiment is as follows: Figure 3 As shown, In step S201, the original image is input as the basic data for subsequent frequency domain processing. In step S202, a frequency domain decomposition network is constructed to decompose the original image in the frequency domain, obtaining multi-frequency components containing low-frequency and high-frequency information. In step S203, dynamic convolution is applied to the low-frequency components to extract low-frequency structural features. In step S204, an attention mechanism is introduced to the high-frequency components to enhance the expression of high-frequency detail features. In step S205, the low-frequency features processed by dynamic convolution are fused with the high-frequency features enhanced by the attention mechanism to obtain multi-frequency fusion features. In step S206, an output image is generated based on the multi-frequency fusion features to optimize image quality.

[0106] According to the strip noise removal method for remote sensing images proposed in this application, firstly, the remote sensing image is decomposed into a first frequency sub-band and a second frequency sub-band. First image features are extracted from the first frequency sub-band, and second image features are extracted from the second frequency sub-band, respectively. The first frequency sub-band is a low-frequency sub-band with a frequency lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with a frequency higher than a second preset frequency. This achieves preliminary separation of high and low frequency information, providing clear frequency domain features for subsequent fine processing. Secondly, the first image features are input into a pre-constructed first processing network. The first processing network extracts multi-scale features from the first image features and, based on annotation... An attention mechanism enhances the stripe region, generating an enhanced first feature map that efficiently extracts and enhances low-frequency structural information and global texture features. Then, the second image feature is input into a pre-constructed second processing network. This network processes stripe noise through sparse convolution and a striped window attention mechanism, and reconstructs the denoised image using inverse wavelet transform to obtain the second feature map. This achieves precise suppression of high-frequency details and stripe noise while maintaining the integrity of image details. Finally, the first and second feature maps are fused to obtain a remote sensing image with stripe noise removed, achieving complementary fusion of low-frequency structure and high-frequency details, significantly improving the visual quality and denoising effect of the image. This solves the problem that related technologies rely on manual parameter adjustment or modeling for specific noise, making it difficult to preserve image details and effectively suppress stripe noise.

[0107] Next, referring to the accompanying drawings, a strip noise removal apparatus for remote sensing images according to an embodiment of this application is described.

[0108] Figure 4 A block diagram of a strip noise removal device for remote sensing images according to an embodiment of this application.

[0109] like Figure 4 As shown, the strip noise removal device 10 for the remote sensing image includes: an extraction module 100, a first processing module 200, a second processing module 300, and a fusion module 400.

[0110] The system includes the following components: an extraction module 100, which decomposes the remote sensing image into a first frequency sub-band and a second frequency sub-band, extracts first image features from the first frequency sub-band and second image features from the second frequency sub-band, respectively. The first frequency sub-band is a low-frequency sub-band with a frequency lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with a frequency higher than a second preset frequency. A first processing module 200, which inputs the first image features into a pre-constructed first processing network, extracts multi-scale features from the first image features, and enhances the stripe regions of the multi-scale features based on an attention mechanism to obtain a first feature map. A second processing module 300, which inputs the second image features into a pre-constructed second processing network, processes the second image features through sparse convolution and a striped window attention mechanism to handle stripe noise, and reconstructs the denoised image through inverse wavelet transform to obtain a second feature map. A fusion module 400, which fuses the first feature map and the second feature map to obtain a remote sensing image with stripe noise removed.

[0111] In one embodiment of this application, the second frequency sub-band includes at least one of a horizontal high-frequency sub-band, a vertical high-frequency sub-band, and a diagonal high-frequency sub-band, and the expressions for the first and second frequency sub-bands are as follows:

[0112] in, Represents low frequency. Represents high frequency levels, Represents vertical high frequency. Represents high frequency on the diagonal.

[0113] In one embodiment of this application, a first processing network and a second processing network process their respective image features in parallel.

[0114] In one embodiment of this application, the first processing network includes dynamic convolutional blocks, residual dense blocks, and a gradient attention module, wherein, The formula for calculating dynamic convolutional blocks is:

[0115] in, This is a fused feature map of a dynamic convolutional block. This represents a 5×5 convolution operation. For learning weights, This refers to the index subscript of the parallel convolution kernel. For convolution operations, For dynamic convolution kernel weights, This is the input feature map for this module; The first residual dense block l The layer output is represented as:

[0116] in, For the first l The output feature map of the layer, It is the ReLU activation function. For the first l Layer weights, Indicates feature concatenation operation; The calculation formula for the gradient attention module:

[0117] in, This is the Sobel attention weight graph. For activation function, W sobel For vertical Sobel attention weights, This is to take the absolute value of the result of the convolution operation.

[0118] In one embodiment of this application, the second processing network includes sparse convolution kernels, a strip window attention mechanism module, and a cyclic shifting module, wherein, The formula for calculating sparse convolution kernels is:

[0119] in, The output feature map of sparse convolution. It is a sparse convolution kernel in the horizontal direction. It is a sparse convolution kernel in the vertical direction. For learnable fringe direction attention weight parameters; The calculation formula for the strip window attention mechanism module is as follows:

[0120] in, This is an attention mechanism function used to weight and aggregate values ​​based on the similarity between the query and the key. The normalized exponential function is used to map attention scores into a probability distribution such that the sum of all weights is 1. It is a strip mask matrix. Q Used to characterize the query requirements of the currently processed pixel or region during the denoising process. K Feature identifiers used to characterize individual pixels or regions in an image. V This represents the original image information or deep feature information to be aggregated. This is the transpose of the key matrix, used in conjunction with the query matrix. Q Perform matrix multiplication to calculate the similarity between each query and key pair. The scaling factor is usually taken as the dimension of the key vector (i.e., the feature length of each key), used to scale the dot product result to a suitable range and avoid gradient saturation; The calculation formula for the cyclic displacement module is as follows:

[0121] in, This is the feature map after cyclic displacement. Roll This is a cyclic displacement sliding operation. F For the feature map, Δ y This is the preset displacement.

[0122] In one embodiment of this application, the fusion module 400 is further configured to input the first feature map and the second feature map into a pre-constructed multi-frequency feature fusion network, and the multi-frequency feature fusion network outputs a remote sensing image with stripe noise removed. The fusion formula of the multi-frequency feature fusion network is as follows:

[0123] in, This is the merged attention weight map. To merge convolutional kernels, It is the Sigmoid activation function. This is a low-frequency feature map. This is a high-frequency feature map.

[0124] In one embodiment of this application, the multi-frequency feature fusion network employs a multi-objective loss function and a dynamic weight adjustment strategy to optimize network parameters. The formula for the multi-objective loss function is:

[0125] in, For the total loss function, The weighting coefficients for the mean squared error loss are: These are the weighting coefficients for the horizontal gradient loss. Weighting coefficients to counteract losses, It is the loss of the mean square error. It is a horizontal gradient loss constraining the loss of the difference between adjacent columns. It is to counteract some of the losses; The formula for the dynamic weight adjustment strategy is:

[0126] in, For dynamic weight adjustment strategy, Control the rate of weight growth. To normalize the training progress (initial stage) Later ).

[0127] It should be noted that the foregoing explanation of the method embodiment for strip noise removal of remote sensing images also applies to the strip noise removal device for remote sensing images in this embodiment, and will not be repeated here.

[0128] According to the strip noise removal device for remote sensing images proposed in the embodiments of this application, firstly, the remote sensing image is decomposed into a first frequency sub-band and a second frequency sub-band. First image features are extracted from the first frequency sub-band and second image features are extracted from the second frequency sub-band, respectively. The first frequency sub-band is a low-frequency sub-band with a frequency lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with a frequency higher than a second preset frequency. This achieves preliminary separation of high and low frequency information, providing clear frequency domain features for subsequent fine processing. Secondly, the first image features are input into a pre-constructed first processing network. The first processing network extracts multi-scale features from the first image features and, based on annotation... An attention mechanism enhances the stripe region, generating an enhanced first feature map that efficiently extracts and enhances low-frequency structural information and global texture features. Then, the second image feature is input into a pre-constructed second processing network. This network processes stripe noise through sparse convolution and a striped window attention mechanism, and reconstructs the denoised image using inverse wavelet transform to obtain the second feature map. This achieves precise suppression of high-frequency details and stripe noise while maintaining the integrity of image details. Finally, the first and second feature maps are fused to obtain a remote sensing image with stripe noise removed, achieving complementary fusion of low-frequency structure and high-frequency details, significantly improving the visual quality and denoising effect of the image. This solves the problem that related technologies rely on manual parameter adjustment or modeling for specific noise, making it difficult to preserve image details and effectively suppress stripe noise.

[0129] Figure 5 This application provides a schematic diagram of the structure of an electronic device. The electronic device may include: The memory 501, the processor 502, and the computer program stored on the memory 501 and capable of running on the processor 502.

[0130] When the processor 502 executes the program, it implements the strip noise removal method for remote sensing images provided in the above embodiments.

[0131] Furthermore, electronic devices also include: Communication interface 503 is used for communication between memory 501 and processor 502.

[0132] The memory 501 is used to store computer programs that can run on the processor 502.

[0133] The memory 501 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0134] If the memory 501, processor 502, and communication interface 503 are implemented independently, then the communication interface 503, memory 501, and processor 502 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0135] Optionally, in a specific implementation, if the memory 501, processor 502, and communication interface 503 are integrated on a single chip, then the memory 501, processor 502, and communication interface 503 can communicate with each other through an internal interface.

[0136] The processor 502 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0137] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for stripe noise removal from remote sensing images.

[0138] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0139] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0140] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0141] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0142] Those skilled in the art will understand that all or part of the steps of the methods implementing the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0143] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for stripe noise removal from remote sensing images, characterized in that, Includes the following steps: The remote sensing image is decomposed into a first frequency sub-band and a second frequency sub-band. First image features are extracted from the first frequency sub-band and second image features are extracted from the second frequency sub-band, respectively. The first frequency sub-band is a low-frequency sub-band with a frequency lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with a frequency higher than a second preset frequency. The first image features are input into a pre-constructed first processing network. The first processing network extracts multi-scale features from the first image features and enhances the stripe regions of the multi-scale features based on an attention mechanism to obtain a first feature map. The second image features are input into a pre-constructed second processing network. The second processing network processes the second image features through sparse convolution and a strip window attention mechanism to handle strip noise. The denoised image is reconstructed through inverse wavelet transform to obtain the second feature map. By fusing the first feature map and the second feature map, a remote sensing image with stripe noise removed is obtained.

2. The method for removing strip noise from remote sensing images according to claim 1, characterized in that, The second frequency sub-band includes at least one of a horizontal high-frequency sub-band, a vertical high-frequency sub-band, and a diagonal high-frequency sub-band. The expressions for the first frequency sub-band and the second frequency sub-band are as follows: in, Represents low frequency. Represents high frequency levels, Represents vertical high frequency. Represents high frequency on the diagonal.

3. The method for removing stripe noise from remote sensing images according to claim 1, characterized in that, The first processing network and the second processing network process their respective image features in parallel.

4. The method for strip noise removal from remote sensing images according to claim 1 or 3, characterized in that, The first processing network includes dynamic convolutional blocks, residual dense blocks, and gradient attention modules, wherein, The calculation formula for the dynamic convolutional block is as follows: in, This is a fused feature map of a dynamic convolutional block. This represents a 5×5 convolution operation. For learning weights, This refers to the index of the parallel convolution kernel. For convolution operations, For dynamic convolution kernel weights, This is the input feature map for the dynamic convolutional block; The first of the residual dense blocks l The layer output is represented as: in, For the first l The output feature map of the layer, It is the ReLU activation function. For the first l Layer weights Indicates feature concatenation operation; The calculation formula for the gradient attention module is as follows: in, This is the Sobel attention weight graph. For activation function, W sobel For vertical Sobel attention weights, This is to take the absolute value of the result of the convolution operation.

5. The method for strip noise removal from remote sensing images according to claim 1 or 3, characterized in that, The second processing network includes sparse convolution kernels, a strip window attention mechanism module, and a cyclic shifting module, wherein, The formula for calculating the sparse convolution kernel is: in, The output feature map of sparse convolution. It is a sparse convolution kernel in the horizontal direction. It is a sparse convolution kernel in the vertical direction. For learnable fringe direction attention weight parameters; The calculation formula for the strip window attention mechanism module is as follows: in, For attention mechanism functions, For normalized exponential functions, It is a strip mask matrix. Q Used to characterize the query requirements of the currently processed pixel or region during the denoising process. K Feature identifiers used to characterize individual pixels or regions in an image. V This represents the original image information or deep feature information to be aggregated. This is the transpose of the key matrix. This is the scaling factor; The calculation formula for the cyclic displacement module is as follows: in, This is the feature map after cyclic displacement. Roll This is a cyclic displacement sliding operation. F For the feature map, Δ y This is the preset displacement.

6. The method for removing strip noise from remote sensing images according to claim 1, characterized in that, The process of fusing the first feature map and the second feature map to obtain the remote sensing image with stripe noise removed includes: The first feature map and the second feature map are input into a pre-constructed multi-frequency feature fusion network, which outputs the remote sensing image with strip noise removed. The fusion formula of the multi-frequency feature fusion network is as follows: in, This is the merged attention weight map. To merge convolutional kernels, It is the Sigmoid activation function. This is a low-frequency feature map. This is a high-frequency feature map.

7. The method for removing stripe noise from remote sensing images according to claim 6, characterized in that, The multi-frequency feature fusion network employs a multi-objective loss function and a dynamic weight adjustment strategy to optimize network parameters. The formula for the multi-objective loss function is as follows: in, For the total loss function, The weighting coefficients for the mean squared error loss are: These are the weighting coefficients for the horizontal gradient loss. Weighting coefficients to counteract losses, It is the loss of the mean square error. It is a horizontal gradient loss constraining the loss of the difference between adjacent columns. It is to counteract some of the losses; The formula for the dynamic weight adjustment strategy is: in, For dynamic weight adjustment strategy, Control the rate of weight growth. To normalize the training progress.

8. A stripe noise removal device for remote sensing images, characterized in that, include: The extraction module is used to decompose the remote sensing image into a first frequency sub-band and a second frequency sub-band, and extract a first image feature from the first frequency sub-band and a second image feature from the second frequency sub-band, respectively. The first frequency sub-band is a low-frequency sub-band with a frequency lower than a first preset frequency, and the second frequency sub-band is a high-frequency sub-band with a frequency higher than a second preset frequency. The first processing module is used to input the first image features into a pre-constructed first processing network. The first processing network extracts multi-scale features from the first image features and enhances the stripe region of the multi-scale features based on an attention mechanism to obtain a first feature map. The second processing module is used to input the second image features into a pre-constructed second processing network. The second processing network processes the second image features through sparse convolution and strip window attention mechanism to process strip noise, and reconstructs the denoised image through inverse wavelet transform to obtain the second feature map. The fusion module is used to fuse the first feature map and the second feature map to obtain a remote sensing image with strip noise removed.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the strip noise removal method for remote sensing images according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed, they implement the strip noise removal method for remote sensing images according to any one of claims 1-7.

Images