Image denoising method and device fusing global-local adaptive attention

Abstract

The invention relates to an image denoising method and device fusing global-local adaptive attention. The method comprises the following steps: acquiring an original remote sensing image; the original remote sensing image is input into a pre-trained denoising model to obtain a denoised image, and the denoising model is an image denoising model adopting an encoder-decoder architecture; the encoder adopts a four-level symmetrical hierarchical structure; the decoder adopts a four-level symmetrical structure which is symmetrical with the encoder; in the encoding stage, the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer are used for carrying out shallow-to-deep layer-by-layer feature extraction on an original remote sensing image in sequence; and in a decoding stage, decoding the fourth feature layer by layer through a fourth decoding layer, a third decoding layer, a second decoding layer and a first decoding layer to obtain a denoised image. By means of the method, optimization balance is achieved between noise reduction and detail keeping, and the image quality and the reliability of follow-up visual tasks are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, and more specifically, to an image denoising method and apparatus that integrates global-local adaptive attention. Background Technology

[0002] Remote sensing image denoising technology is a crucial step in improving image quality and ensuring the accuracy of subsequent interpretation, and it is widely used in fields such as land cover classification, target recognition, and change detection. During the acquisition process, remote sensing images are often affected by factors such as sensor noise, atmospheric scattering, and changes in illumination, resulting in Gaussian, impulse, or mixed noise, which reduces image clarity and spectral consistency.

[0003] Traditional denoising methods include spatial domain filtering (such as mean filtering, median filtering, and bilateral filtering), transform domain denoising (such as wavelet, Curvelet, and Contourlet transforms), and model-driven methods based on sparse representation and low-rank decomposition. Although these methods can effectively smooth noise, they often lead to the loss of texture details and are not adaptable to complex noise. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an image denoising method and apparatus that integrates global-local adaptive attention, aiming to solve at least one of the above-mentioned technical problems.

[0005] In a first aspect, the technical solution of the present invention to solve the above-mentioned technical problems is as follows: an image denoising method integrating global-local adaptive attention, the method comprising: Acquire raw remote sensing images; The original remote sensing image is input into a pre-trained denoising model to obtain a denoised image. The denoising model is an image denoising model with an encoder-decoder architecture. The encoder adopts a four-level symmetrical hierarchical structure, including a first coding layer, a second coding layer, a third coding layer, and a fourth coding layer; the decoder adopts a four-level symmetrical structure that is symmetrical to the encoder, including a fourth decoding layer, a third decoding layer, a second decoding layer, and a first decoding layer. During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer in sequence, to obtain the first feature output by the first encoding layer, the second feature output by the second encoding layer, the third feature output by the third encoding layer and the fourth feature output by the fourth encoding layer. During the decoding stage, the fourth feature is decoded layer by layer through the fourth decoding layer, the third decoding layer, the second decoding layer, and the first decoding layer to obtain the denoised image.

[0006] The beneficial effects of this invention are as follows: By adopting an encoder-decoder architecture and setting a four-level symmetrical hierarchical encoder and a symmetrical decoder, the first to fourth features are obtained by sequentially extracting features from the original remote sensing image layer by layer from shallow to deep through the four-level encoding layers during the encoding stage. During the decoding stage, the fourth feature is decoded layer by layer through the four-level decoding layers. This architecture realizes multi-scale feature extraction from shallow details to deep semantics through hierarchical encoding. The symmetrical decoder structure can form skip connections with the encoder to fully preserve the detailed information extracted by each level of encoding layer, avoiding the texture loss problem caused by single-scale processing in traditional methods. The four-level hierarchical design ensures sufficient receptive field to capture the global context of large-size remote sensing images while realizing spatial resolution restoration through hierarchical decoding. The end-to-end model architecture avoids the complex process of multi-stage processing in traditional methods and significantly improves the denoising efficiency of large-size remote sensing images. The multi-level feature extraction and fusion mechanism enhances the adaptability and discrimination ability to complex noise patterns, thereby achieving an optimized balance between noise reduction and detail preservation, significantly improving image quality and the reliability of subsequent visual tasks.

[0007] Based on the above technical solution, the present invention can be further improved as follows.

[0008] Furthermore, each decoding layer of the decoder includes a GLAE fusion module; For each GLAE fusion module, the output of the previous decoding layer of the decoding layer where the GLAE fusion module is located, and the output of the corresponding encoding layer of the decoding layer where the GLAE fusion module is located, are enhanced and fused to obtain the enhanced fused features, which are then used as the input of the next decoding layer of the decoding layer where the GLAE fusion module is located.

[0009] Furthermore, each GLAE fusion module includes a multi-scale convolutional feature extraction unit, a statistical fusion unit, and an adaptive attention weighting unit connected in sequence; The multi-scale convolutional feature extraction unit is used to extract multi-resolution spatial features from the input features of the GLAE fusion module through multi-scale convolution operations to obtain multi-scale feature maps. The statistical fusion unit is used to take multi-scale feature maps as input, perform multi-scale information fusion through statistical feature calculation, and obtain statistical fusion features. The adaptive attention weighting unit is used to take the multi-scale convolutional features and statistical fusion features as global context information, and the input features of the GLAE fusion module as local detail information. It calculates the attention weight between the query matrix generated by the local detail information and the key matrix generated by the global context information in the spatial dimension, and performs weighted fusion on the value matrix generated by the global context information to obtain the output features of the corresponding encoding layer.

[0010] Furthermore, each coding layer of the encoder described above corresponds to a Transformer block, and each Transformer block uses a different number of attention heads; During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first, second, third, and fourth encoding layers, including: During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the Transformer blocks corresponding to the first, second, third, and fourth encoding layers.

[0011] Furthermore, each Transformer block consists of a first-layer normalization unit, a multi-depth convolution head transpose attention unit, a first residual connection unit, a second-layer normalization unit, a gated depthwise convolution feedforward network, and a second residual connection unit. The first normalization unit is used to standardize the input features of its corresponding coding layer according to the channel dimension, so as to obtain the first-level normalized features. The multi-depth convolutional head transposed attention unit is used to aggregate channel context information through a 1×1 convolution to obtain channel aggregated features. Local spatial features in the channel aggregated features are extracted through a 3×3 depth convolution to obtain three sets of feature maps: query, key, and value. Based on the three sets of feature maps, the transposed attention map is calculated in the channel dimension to obtain the channel attention weight. The value feature map is dynamically fused through the channel attention weight to obtain the denoising and enhancement features. The first residual connection unit is used to take the first standardized feature as the input of the identity mapping branch in the first residual connection unit, take the denoising and enhancement feature as the input of the residual mapping branch in the first residual connection unit, and add the identity mapping branch input and the residual mapping branch input element by element to obtain the residual fusion feature. The second normalization unit is used to take the residual fusion features as input and perform normalization through the channel dimension to obtain the secondary normalized features. A gated deep convolutional feedforward network is used to take the quadratically normalized features as input. Local spatial features are extracted through 1×1 convolution and 3×3 deep convolution in the first branch of the gated deep convolutional feedforward network to obtain spatial enhancement features. A gate signal is generated by 1×1 convolution, 3×3 deep convolution and GELU activation in the second branch of the gated deep convolutional feedforward network. The spatial enhancement features and the gate signal are subjected to element-wise multiplication and gating to obtain gated filtering features. The gated filtering features are then mapped to the original channel dimension through 1×1 convolution to obtain gated enhancement features. The second residual connection unit is used to take the quadratic normalized features as the input of the identity mapping branch in the second residual connection unit, take the gated enhancement features as the input of the residual mapping branch in the second residual connection unit, and add the identity mapping branch input and the residual mapping branch input in the second residual connection unit element by element to obtain the output features of the coding layer to which the Transformer block belongs.

[0012] Furthermore, the multi-depth convolutional head transposed attention unit in each Transforme block consists of a third-layer normalization unit, a first convolutional unit, a second convolutional unit, a transposed attention mapping unit, and a third residual connection unit; The third normalization unit is used to normalize the first-normalized features by the channel dimension to obtain MDTA (Multi-Dconv Head Transposed Attention) normalized features. The first convolutional unit is used to take the MDTA-normalized features as input and perform cross-channel context aggregation through 1×1 pointwise convolution to obtain channel aggregated features. The second convolutional unit takes the channel aggregation feature as input and extracts local spatial features through a 3×3 depth convolution to obtain three sets of feature maps: query, key, and value. The transposed attention mapping unit is used to calculate the transposed attention map in the channel dimension based on the three sets of feature maps: query, key, and value, to obtain the channel attention weights. The value feature map is then dynamically fused using the channel attention weights to obtain the attention-weighted features. The third residual connection unit is used to take the first-standardized features as the input of the identity mapping branch in the third residual connection unit, take the attention-weighted features as the input of the residual mapping branch in the third residual connection unit, and add the input of the identity mapping branch in the third residual connection unit to the residual mapping branch element by element to obtain the denoising and enhancement features.

[0013] Furthermore, the gated depthwise convolutional feedforward network in each Transformer block consists of a fourth-layer normalization unit, two sets of parallel third and fourth convolutional units, a GELU activation unit, an element-wise multiplication gate mechanism unit, and a fourth residual connection unit. The fourth normalization unit is used to normalize the input quadratic normalized features by the channel dimension to obtain GDFN (Gated Dconv Feed-Forward Netwoek) normalized features; Two sets of parallel third and fourth convolutional units are used to take GDFN-normalized features as input, extract local spatial features through 1×1 convolution and 3×3 depth convolution of the third convolutional unit to obtain spatially enhanced features, and perform feature transformation through 1×1 convolution and 3×3 depth convolution of the fourth convolutional unit to obtain gated basic features. The GELU activation unit is used to take the spatial augmentation features as input and perform a nonlinear transformation through the GELU activation function to obtain the gated signal; The element-level multiplication gating mechanism unit is used to take the spatial augmentation features and the gating signal as inputs, and perform gating filtering through element-level multiplication to obtain the gating filtering features. The fourth residual connection unit is used to take the quadratic standardized features as the input of the identity mapping branch in the fourth residual connection unit, take the gating and filtering features after 1×1 convolution as the input of the residual mapping branch in the fourth residual connection unit, and add the input of the identity mapping branch in the fourth residual connection unit to the residual mapping branch element by element to obtain the gating enhanced features.

[0014] Secondly, to solve the above-mentioned technical problems, the present invention also provides an image denoising device that integrates global-local adaptive attention, the device comprising: The acquisition module is used to acquire raw remote sensing images; The denoising module is used to input the original remote sensing image into a pre-trained denoising model to obtain a denoised image. The denoising model is an image denoising model with an encoder-decoder architecture. The encoder adopts a four-level symmetrical hierarchical structure, including a first coding layer, a second coding layer, a third coding layer, and a fourth coding layer; the decoder adopts a four-level symmetrical structure that is symmetrical to the encoder, including a fourth decoding layer, a third decoding layer, a second decoding layer, and a first decoding layer. During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer in sequence, to obtain the first feature output by the first encoding layer, the second feature output by the second encoding layer, the third feature output by the third encoding layer and the fourth feature output by the fourth encoding layer. During the decoding stage, the fourth feature is decoded layer by layer through the fourth decoding layer, the third decoding layer, the second decoding layer, and the first decoding layer to obtain the denoised image.

[0015] Thirdly, in order to solve the above-mentioned technical problems, the present invention also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the image denoising method of the present application that integrates global-local adaptive attention.

[0016] Fourthly, in order to solve the above-mentioned technical problems, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the image denoising method of the present application that integrates global-local adaptive attention.

[0017] Additional aspects and advantages of this application will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below.

[0019] Figure 1 This is a flowchart illustrating an image denoising method that integrates global-local adaptive attention, as provided in one embodiment of the present invention. Figure 2 This is a schematic diagram of an original remote sensing image provided in one embodiment of the present invention; Figure 3 A general model architecture diagram provided for one embodiment of the present invention; Figure 4 A schematic diagram of a gated depthwise convolutional feedforward network unit provided in one embodiment of the present invention; Figure 5 This is a schematic diagram of a transposed attention unit for a multi-depth convolutional head according to an embodiment of the present invention; Figure 6 A schematic diagram of a GLAE fusion module provided in one embodiment of the present invention; Figure 7 This is a schematic diagram of a denoised image obtained by an image denoising method according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an image denoising device that integrates global-local adaptive attention according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0020] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0021] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0022] The solution provided in this invention can be applied to any application scenario requiring denoising of remote sensing images. The solution provided in this invention can be executed by any electronic device, such as a user's terminal device, including at least one of the following: smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, smart TV, or smart in-vehicle device.

[0023] This invention provides a possible implementation, such as... Figure 1 The diagram shows a flowchart of an image denoising method that integrates global-local adaptive attention. This method can be executed by any electronic device, such as a terminal device, or jointly executed by a terminal device and a server. For ease of description, the method provided in this embodiment will be described below using a terminal device as the execution subject as an example. Figure 1 The flowchart shown indicates that the method may include the following steps: S10, acquire raw remote sensing images; S20, input the original remote sensing image into the pre-trained denoising model to obtain the denoised image, wherein the denoising model is an image denoising model with an encoder-decoder architecture; The encoder adopts a four-level symmetrical hierarchical structure, including a first coding layer, a second coding layer, a third coding layer, and a fourth coding layer; the decoder adopts a four-level symmetrical structure that is symmetrical to the encoder, including a fourth decoding layer, a third decoding layer, a second decoding layer, and a first decoding layer. During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer in sequence, to obtain the first feature output by the first encoding layer, the second feature output by the second encoding layer, the third feature output by the third encoding layer and the fourth feature output by the fourth encoding layer. During the decoding stage, the fourth feature is decoded layer by layer through the fourth decoding layer, the third decoding layer, the second decoding layer, and the first decoding layer to obtain the denoised image.

[0024] The method of this invention employs an encoder-decoder architecture with a four-level symmetrical hierarchical encoder and decoder. In the encoding stage, the four-level encoding layers sequentially extract features from the original remote sensing image, from shallow to deep, to obtain the first to fourth features. In the decoding stage, the four-level decoding layers sequentially decode the fourth feature. This architecture achieves multi-scale feature extraction from shallow details to deep semantics through hierarchical encoding. The symmetrical decoder structure can form skip connections with the encoder to fully preserve the detailed information extracted at each encoding layer, avoiding the texture loss problem caused by single-scale processing in traditional methods. The four-level hierarchical design ensures sufficient receptive field to capture the global context of large-size remote sensing images while restoring spatial resolution through sequential decoding. The end-to-end model architecture avoids the complex multi-stage processing of traditional methods, significantly improving the denoising efficiency for large-size remote sensing images. The multi-level feature extraction and fusion mechanism enhances adaptability and discrimination ability to complex noise patterns, thus achieving an optimized balance between noise reduction and detail preservation, significantly improving image quality and the reliability of subsequent visual tasks.

[0025] The following specific embodiments further illustrate the solution of the present invention. In recent years, deep learning has driven the development of denoising technology, with end-to-end methods based on CNN (Convolutional Neural Network) and Transformer becoming mainstream. CNN can automatically extract features but has a limited receptive field, while Transformer relies on attention mechanisms to achieve global feature modeling, better preserving details while suppressing noise. Combining strategies such as multi-scale feature fusion, adaptive attention, and generative adversarial networks (GANs) significantly improves denoising performance and robustness. Future research will focus on multi-source data fusion, lightweight networks, and model interpretability optimization to improve efficiency and application breadth while ensuring accuracy.

[0026] Based on this, this invention proposes an image denoising method integrating global-local adaptive attention. This method employs an encoder-decoder architecture, with a Transformer block at its core, combining Multi-Depth Convolutional Head Transposed Attention (MDTA) and Gated Deep Convolutional Feedforward Network (GDFN) to achieve efficient feature modeling. This method balances global dependencies and local details with linear complexity, utilizes gating mechanisms to filter effective features, and suppresses noise, thereby improving the denoising efficiency of large-size remote sensing images while maintaining the integrity of ground structure and spectral details. The decoder section introduces a multi-scale fusion global-local adaptive attention fusion enhancement module (GLAE, Global-Local-Aware Encoder). This module enhances the stability and discriminative power of features through multi-scale feature extraction and statistical fusion, and dynamically coordinates global and local information through adaptive attention. During image denoising, GLAE effectively distinguishes noise from effective signals, strengthens key structures, weakens irrelevant responses, accurately suppresses noise, and preserves edge and texture details, thus achieving a balance between noise reduction and detail preservation, significantly improving image quality and enhancing the reliability of subsequent visual tasks.

[0027] This invention proposes an image denoising method that integrates global-local adaptive attention. It achieves efficient feature modeling by using a Transformer block as the core and fusing MDTA and GDFN. At the same time, it introduces a GLAE module to model global correlations and local details through multi-scale feature fusion and adaptive attention, thereby achieving an optimized balance between denoising and detail preservation, and thus significantly improving image quality and interpretation reliability.

[0028] Based on the above, in this embodiment, an image denoising method that integrates global-local adaptive attention may include the following steps: S10, acquire raw remote sensing images; The original remote sensing images can be acquired by remote sensing sensors such as optical sensors and SAR (Synthetic Aperture Radar) sensors mounted on satellites, aircraft (such as airplanes and drones) or ground remote sensing platforms. These sensors receive electromagnetic wave signals reflected or emitted by objects on the ground, and generate image data containing the spectral and spatial information of ground objects through signal conversion and data recording, including noise.

[0029] As an example, see Figure 2 The diagram shows the original remote sensing image.

[0030] S20: Input the original remote sensing image into the pre-trained denoising model to obtain the denoised image. The denoising model is an image denoising model with an encoder-decoder architecture. The noise in the original remote sensing image is effectively suppressed, and the denoised image is a high-quality image with clear details and structure, which can provide a reliable data foundation for subsequent tasks such as land cover classification and target detection.

[0031] See Figure 3 The encoder employs a four-level symmetrical hierarchical structure, including a first encoding layer, a second encoding layer, a third encoding layer, and a fourth encoding layer. The decoder employs a four-level symmetrical structure, symmetrical to the encoder, including a fourth decoding layer, a third decoding layer, a second decoding layer, and a first decoding layer. The four-level symmetrical hierarchical structure of the encoder corresponds to the four-level symmetrical hierarchical structure of the decoder; that is, the first encoding layer corresponds to the first decoding layer, the second encoding layer corresponds to the second decoding layer, the third encoding layer corresponds to the third decoding layer, and the fourth encoding layer corresponds to the fourth decoding layer. In other words, the output of the first encoding layer serves as the input of the first decoding layer, the output of the second encoding layer serves as the input of the second decoding layer, the output of the third encoding layer serves as the input of the third decoding layer, and the output of the fourth encoding layer serves as the input of the fourth decoding layer.

[0032] During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer in sequence, to obtain the first feature output by the first encoding layer, the second feature output by the second encoding layer, the third feature output by the third encoding layer and the fourth feature output by the fourth encoding layer. Specifically, performing layer-by-layer feature extraction from the original remote sensing image through the first coding layer (Level 1), the second coding layer (Level 2), the third coding layer (Level 3), and the fourth coding layer (Level 4) means that feature extraction is performed on the original remote sensing image through the first coding layer to obtain the first feature; further feature extraction is performed on the first feature through the second coding layer to obtain the second feature; further feature extraction is performed on the second feature through the third coding layer to obtain the third feature; and further feature extraction is performed on the third feature through the fourth coding layer to obtain the fourth feature.

[0033] The original remote sensing image can be represented as The first feature can be represented as The second feature can be represented as The third feature can be represented as The fourth feature can be represented as .

[0034] During the decoding stage, the fourth feature is decoded layer by layer through the fourth decoding layer, the third decoding layer, the second decoding layer, and the first decoding layer to obtain the denoised image.

[0035] The decoder's four-level symmetric structure follows a fixed process of "upsampling restoration → channel reduction → GLAE fusion → Transformer refinement," gradually restoring spatial resolution and refining feature representation. Finally, the high-resolution denoised features output by the model are convolutionally mapped to generate clear and realistic denoised images.

[0036] Specifically, the above-mentioned layer-by-layer decoding process of the fourth feature through the fourth decoding layer, the third decoding layer, the second decoding layer, and the first decoding layer refers to decoding the fourth feature through the fourth decoding layer to obtain the first decoded feature, decoding the first decoded feature through the third decoding layer to obtain the second decoded feature, decoding the second decoded feature through the second decoding layer to obtain the third decoded feature, and decoding the third decoded feature through the first decoding layer to obtain the denoised image.

[0037] Optionally, each decoding layer of the decoder includes a GLAE fusion module (see [link to decoder]). Figure 6 ); For each GLAE fusion module, the output of the previous decoding layer of the decoding layer where the GLAE fusion module is located, and the output of the corresponding encoding layer of the decoding layer where the GLAE fusion module is located, are enhanced and fused to obtain the enhanced fused features, which are then used as the input of the next decoding layer of the decoding layer where the GLAE fusion module is located.

[0038] If it is a GLAE fusion module corresponding to the first decoding layer, then the input of the GLAE fusion module only includes the output of the encoding layer corresponding to the decoding layer where the GLAE fusion module is located.

[0039] As an example, if it is a GLAE fusion module corresponding to the second decoding layer, the input of the GLAE fusion module includes the output of the third decoding layer (second decoding feature) and the output of the second encoding layer (second feature).

[0040] Optionally, each GLAE fusion module includes a multi-scale convolutional feature extraction unit, a statistical fusion unit, and an adaptive attention weighting unit connected in sequence; The multi-scale convolutional feature extraction unit is used to extract multi-resolution spatial features from the input features of the GLAE fusion module (including the output of the previous decoding layer of the decoding layer where the GLAE fusion module is located, and the output of the encoding layer corresponding to the decoding layer where the GLAE fusion module is located) through multi-scale convolution operations, and obtain multi-scale feature maps. The statistical fusion unit is used to take multi-scale feature maps as input and perform multi-scale information fusion (e.g., 1×1 convolution processing) through statistical feature calculation (specifically, the minimum, maximum, sum and mean values ​​of the channel dimensions of the multi-scale feature maps can be calculated separately) to obtain statistical fusion features; The adaptive attention weighting unit is used to take the multi-scale convolutional features and statistical fusion features as global context information, and the input features of the GLAE fusion module as local detail information. It calculates the attention weight between the query matrix generated by the local detail information and the key matrix generated by the global context information in the spatial dimension, and performs weighted fusion on the value matrix generated by the global context information to obtain the output features of the corresponding encoding layer.

[0041] Using statistical fusion features as input, the global and local information are dynamically weighted through an adaptive attention mechanism to obtain the output features of the corresponding coding layer. The global information is the output feature of the coding layer to which the corresponding GLAE fusion module belongs, and the local information is the output feature of the coding layer above the coding layer to which the corresponding GLAE fusion module belongs.

[0042] Optionally, the aforementioned adaptive attention weighting unit is specifically used to generate a query matrix Q by performing a 1×1 convolution on the input features X of the GLAE fusion module, and to convert the multi-scale convolutional features... Features of statistical fusion After concatenation along the channel dimension, the matrix Q is used as the key matrix K and the value matrix V. Attention weights for Q and K are calculated in the spatial dimension and normalized by SoftMax. Finally, V is weighted and fused using these attention weights to obtain the global-local adaptive enhancement features, which are the output features of the corresponding coding layer.

[0043] Each GLAE fusion module achieves dynamic coordination between global semantics and local details through "multi-scale convolutional feature extraction—statistical fusion—adaptive attention weighting". Each GLAE fusion module captures multi-scale features using convolutional kernels with different receptive fields, and fuses multi-scale information through statistical operations such as min, max, sum, and mean to enhance feature stability and discriminativity. Subsequently, an adaptive attention mechanism dynamically allocates weights in the spatial and channel dimensions to strengthen key structures and suppress redundant responses. The relevant calculation formulas for the Global-Local Adaptive Attention Fusion Enhancement (GLAE) module are as follows, where the input features of the GLAE fusion module are... : ; ; ; ; ; in, The following describes the convolutional feature extraction at different scales: a multi-scale feature extraction module is used, which includes 1×1 convolution, 3×3 convolution with different dilation rates (6, 12, 18), and 5×5 convolution, to capture local and multi-scale contextual information simultaneously. Represents multi-scale feature maps; The method calculates the minimum, maximum, elemental sum, and mean of image features in the spatial dimension, achieving aggregation of multiple statistical features to comprehensively represent global extrema, energy, and average distribution characteristics. Indicates statistical fusion characteristics, This indicates the splicing of tensor channel dimensions. This represents a global-local adaptive enhancement feature.

[0044] The Global-Local Adaptive Attention Fusion Enhancement (GLAE) module designed using this invention can effectively distinguish between noise and real signals, accurately preserve the texture, edges and spectral details of ground features and suppress noise in a multi-scale space, avoid over-smoothing, and thus significantly improve image quality and subsequent interpretation accuracy.

[0045] Optionally, each coding layer of the encoder described above corresponds to a Transformer block, and each Transformer block uses a different number of attention heads; During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first, second, third, and fourth encoding layers, including: During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the Transformer blocks corresponding to the first, second, third, and fourth encoding layers.

[0046] Optionally, the Transformer block corresponding to the first coding layer uses 1 attention head, the Transformer block corresponding to the second coding layer uses 2 attention heads, the Transformer block corresponding to the third coding layer uses 4 attention heads, and the Transformer block corresponding to the fourth coding layer uses 8 attention heads.

[0047] Specifically, Level 1 in the encoder receives the raw remote sensing imagery, also known as the initial features. The basic structural information at high resolution is extracted through multiple Transformer blocks with one attention head each, and the resolution is reduced to [the desired value]. The channels are expanded to 96 to generate the first feature. This is simultaneously transmitted to Level 2 in the encoder and the corresponding second decoding layer. Level 2 in the encoder... Based on this, a Transformer block with 2 attention heads is used to enhance the mid-layer texture and edge features, and the second feature is obtained through downsampling and channel expansion. This information is then passed to Level 3 in the encoder and the corresponding third decoding layer. Level 3 in the encoder further captures global dependencies at low resolution using a Transformer block with four attention heads, outputting the third feature. Level 4 in the encoder is the deepest level, consisting of Transformer blocks with 8 attention heads. High-level semantic compression and feature abstraction are performed to generate the encoder's final latent features, i.e., the fourth feature. This provides the core representation for high-resolution feature reconstruction of the decoder.

[0048] Optionally, each Transformer block consists of a first-layer normalization unit, a multi-depth convolutional head transposed attention (MDTA) unit (see [link to relevant documentation]). Figure 5 The network consists of a first residual connection unit, a second normalization unit, and a gated depthwise feedforward network (GDFN, see [link]). Figure 4 It consists of the second residual connection unit; The first normalization unit is used to standardize the input features of its corresponding coding layer according to the channel dimension, so as to obtain the first-level normalized features. The multi-depth convolutional head transposed attention unit is used to aggregate channel context information through a 1×1 convolution to obtain channel aggregated features. Local spatial features in the channel aggregated features are extracted through a 3×3 depth convolution to obtain three sets of feature maps: query, key, and value. Based on the three sets of feature maps, the transposed attention map is calculated in the channel dimension to obtain the channel attention weight. The value feature map is dynamically fused through the channel attention weight to obtain the denoising and enhancement features. The first residual connection unit is used to take the first standardized feature as the input of the identity mapping branch in the first residual connection unit, take the denoising and enhancement feature as the input of the residual mapping branch in the first residual connection unit, and add the identity mapping branch input and the residual mapping branch input element by element to obtain the residual fusion feature. The second normalization unit is used to take the residual fusion features as input and perform normalization through the channel dimension to obtain the secondary normalized features. A gated deep convolutional feedforward network is used to take quadratically normalized features as input. Local spatial features are extracted through 1×1 convolution and 3×3 deep convolution in the first branch of the gated deep convolutional feedforward network to obtain spatial enhancement features. A gated signal is generated by 1×1 convolution, 3×3 deep convolution and GELU (Gaussian Error Linear Unit) activation in the second branch of the gated deep convolutional feedforward network. The spatial enhancement features and the gated signal are subjected to element-wise multiplication and gating to obtain gated filtered features. The gated filtered features are then mapped to the original channel dimension through 1×1 convolution to obtain gated enhancement features. The second residual connection unit is used to take the quadratic normalized features as the input of the identity mapping branch in the second residual connection unit, take the gated enhancement features as the input of the residual mapping branch in the second residual connection unit, and add the identity mapping branch input and the residual mapping branch input in the second residual connection unit element by element to obtain the output features of the coding layer to which the Transformer block belongs.

[0049] Optionally, the multi-depth convolutional head transpose attention unit in each Transformer block consists of a third-layer normalization unit, a first convolutional unit, a second convolutional unit, a transpose attention mapping unit, and a third residual connection unit; The third normalization unit is used to normalize the first-normalized features according to the channel dimension to obtain MDTA normalized features. The first convolutional unit is used to take the MDTA-normalized features as input and perform cross-channel context aggregation through 1×1 pointwise convolution to obtain channel aggregated features. The second convolutional unit takes the channel aggregation feature as input and extracts local spatial features through a 3×3 depth convolution to obtain three sets of feature maps: query, key, and value. The transposed attention mapping unit is used to calculate the transposed attention map in the channel dimension based on the three sets of feature maps: query, key, and value, in order to capture global dependencies and obtain channel attention weights. The value feature map is then dynamically fused using the channel attention weights to obtain attention-weighted features. The third residual connection unit is used to take the first-standardized features as the input of the identity mapping branch in the third residual connection unit, take the attention-weighted features as the input of the residual mapping branch in the third residual connection unit, and add the input of the identity mapping branch in the third residual connection unit to the residual mapping branch element by element to obtain the denoising and enhancement features.

[0050] The relevant calculation formulas for the Multi-Depth Convolution Head Transposed Attention (MDTA) unit in each Transformer block are as follows: ; ; ; ; in, The input feature map (first-order normalized features), where H and W are the spatial dimensions, and C is the number of channels. It is a 1×1 convolution weight (without bias). The weights are 3×3 depthwise convolutional weights (without bias). It is a learnable scaling parameter.

[0051] Y represents the MDTA normalized feature, LN(X) represents the channel-dimensional normalization of X, Q represents the query feature map, K represents the key feature map, and V represents the value feature map. , and These represent the query weight matrix, key weight matrix, and value weight matrix under a 1×1 convolution, respectively. , and These represent the query weight matrix, key weight matrix, and value weight matrix under a 3×3 depthwise convolution, respectively. , and These represent the transposes of Q, K, and V, respectively. , and These represent the dimensions of Q, K, and V, respectively. Indicates attention-weighted features, This represents the noise reduction and enhancement features.

[0052] The Multi-Depth Convolution Head Transposed Attention (MDTA) unit designed in this invention can effectively suppress complex noise in remote sensing images, preserve the edges of ground features and spectral details, and achieve efficient and accurate image denoising while maintaining linear computational complexity and taking into account local details and global correlations.

[0053] Optionally, the gated deep convolutional feedforward network (GDFN) in each Transformer block consists of a fourth-layer normalization unit, two sets of parallel third and fourth convolutional units, a GELU activation unit, an element-wise multiplication gate mechanism unit, and a fourth residual connection unit; it is a key structure in the efficient Transformer block.

[0054] The fourth normalization unit is used to normalize the input quadratic normalized features by the channel dimension to obtain GDFN normalized features; Two sets of parallel third and fourth convolutional units are used to take GDFN-normalized features as input, extract local spatial features through 1×1 convolution and 3×3 depth convolution of the third convolutional unit to obtain spatially enhanced features, and perform feature transformation through 1×1 convolution and 3×3 depth convolution of the fourth convolutional unit to obtain gated basic features. The GELU activation unit is used to take the spatial augmentation features as input and perform a nonlinear transformation through the GELU activation function to obtain the gated signal; The element-level multiplication gating mechanism unit is used to take the spatial augmentation features and the gating signal as inputs, and perform gating filtering through element-level multiplication to obtain the gating filtering features. The fourth residual connection unit is used to take the quadratic standardized features as the input of the identity mapping branch in the fourth residual connection unit, take the gating and filtering features after 1×1 convolution as the input of the residual mapping branch in the fourth residual connection unit, and add the input of the identity mapping branch in the fourth residual connection unit to the residual mapping branch element by element to obtain the gating enhanced features.

[0055] The relevant calculation formulas for Gated Depthwise Feedforward Network (GDFN) are as follows: ; ; ; ; in, The input feature map (quadratic normalized features), where H and W are the spatial dimensions, and C is the number of channels. It is a 1×1 convolution weight (without bias). The weights are 3×3 depthwise convolutional weights (without bias). This represents element-wise multiplication. It is a 1×1 pointwise convolution weight matrix.

[0056] Y represents the GDFN normalized feature, Y1 represents the spatial augmentation feature, and Y2 represents the gated basic feature. Gating(Y) represents the gated signal obtained by performing a nonlinear transformation on the spatial augmentation feature Y1 using the GELU activation function, and Gating(Y) represents the gated selection feature obtained by performing element-wise multiplication between the gated signal GELU(Y1) and the gated basic feature Y2. This indicates that the gated feature Gating(Y) is mapped by a 1×1 convolution and then the residuals of the second-normalized feature are added together to obtain the gated enhancement feature.

[0057] The gated deep convolutional feedforward network (GDFN) designed using this invention enhances feature selection capabilities while maintaining linear and efficient computation. It can effectively highlight useful features and suppress redundant noise, providing strong support for structural detail preservation and feature discrimination in image denoising tasks.

[0058] This invention constructs an end-to-end image denoising model, where the decoder input includes two parts: one is the fourth feature from the encoder Level 4. First, it serves as the starting point for decoding, carrying global structural and semantic information, i.e., global information; second, it is the output of each layer of the encoder. , , The data is passed to the corresponding layer of the decoder via the GLAE cross-layer fusion module to supplement high-resolution details and local structures, i.e., local information.

[0059] Optionally, in the decoding stage, the fourth feature is decoded layer by layer through the fourth decoding layer, the third decoding layer, the second decoding layer, and the first decoding layer to obtain the denoised image. The denoised image can be found in [reference needed]. Figure 7 As shown.

[0060] It should be noted that, Figures 3 to 6 In this context, Norm represents layer normalization, and 3×3 Dconv represents depthwise convolution. This indicates Downsample. Indicates Upsample; This indicates Element-wise Addition, which adds elements one by one. This represents Element-wise Multiplication, which multiplies elements one by one. This represents Matrix Multiplication. Indicates concatenation, which joins multiple tensors into a single tensor along a specified dimension. Reshape means changing the shape (dimensions) of the tensor. This represents GELU Activation, the GELU activation function.

[0061] To better illustrate and understand the principle of the method provided by this invention, the following description uses an optional specific embodiment to illustrate the solution of this invention. It should be noted that the specific implementation of each step in this specific embodiment should not be construed as a limitation of the solution of this invention. Other implementations that can be conceived by those skilled in the art based on the principle of the solution provided by this invention should also be considered within the scope of protection of this invention.

[0062] This invention proposes an image denoising method that integrates global-local adaptive attention. It achieves efficient feature modeling by using a Transformer block as the core and fusing MDTA and GDFN. At the same time, it introduces a GLAE module to model global correlations and local details through multi-scale feature fusion and adaptive attention, thereby achieving an optimized balance between denoising and detail preservation, and thus significantly improving image quality and interpretation reliability.

[0063] This method employs an encoder-decoder architecture: in the encoding stage, the encoder uses a four-level symmetrical hierarchical structure (Level 1-Level 4), progressively deepening and compressing features from top to bottom. Level 1 receives the initial features. The basic structural information at high resolution is extracted through multiple Transformer blocks with one attention head each, and the resolution is reduced to [the desired value]. The channels are expanded to 96, generating This is simultaneously transmitted to Level 2 and the corresponding decoding layer. Level 2 in Based on this, a Transformer block with 2 attention heads is used to enhance the mid-layer texture and edge features, which are then downsampled and expanded to obtain... This information is then passed to Level 3 and the corresponding decoding layer. Level 3 further captures global dependencies at low resolution using Transformer blocks with 4 attention heads, and outputs the results. Level 4, being the deepest level, uses Transformer blocks with 8 attention heads. High-level semantic compression and feature abstraction are performed to generate the encoder's final latent features. This provides the core representation for high-resolution feature reconstruction in the decoder. The decoder input consists of two parts: one is the latent features from the encoder's Level 4. First, it serves as the starting point for decoding, carrying global structural and semantic information; second, it is the output of each layer of the encoder. , , The denoised features are passed to the corresponding level of the decoder via the GLAE cross-layer fusion module to supplement high-resolution details and local structures. The decoder's four-level symmetrical structure (Level 4-Level) follows a fixed process of "upsampling restoration → channel reduction → GLAE fusion → Transformer refinement," gradually restoring spatial resolution and refining feature representation. Finally, the high-resolution denoised features output by the model are convolutionally mapped to generate clear and realistic denoised images.

[0064] Based on and Figure 1 Based on the same principle as the method shown, this embodiment of the invention also provides an image denoising device 20 that fuses global-local adaptive attention, such as... Figure 8 As shown, the image denoising device 20 that integrates global-local adaptive attention may include an acquisition module 210 and a denoising module 220, wherein: The acquisition module 210 is used to acquire raw remote sensing images; The denoising module 220 is used to input the original remote sensing image into a pre-trained denoising model to obtain a denoised image. The denoising model is an image denoising model with an encoder-decoder architecture. The encoder adopts a four-level symmetrical hierarchical structure, including a first coding layer, a second coding layer, a third coding layer, and a fourth coding layer; the decoder adopts a four-level symmetrical structure that is symmetrical to the encoder, including a fourth decoding layer, a third decoding layer, a second decoding layer, and a first decoding layer. During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer in sequence, to obtain the first feature output by the first encoding layer, the second feature output by the second encoding layer, the third feature output by the third encoding layer and the fourth feature output by the fourth encoding layer. During the decoding stage, the fourth feature is decoded layer by layer through the fourth decoding layer, the third decoding layer, the second decoding layer, and the first decoding layer to obtain the denoised image.

[0065] Optionally, each decoding layer of the decoder includes a GLAE fusion module; For each GLAE fusion module, the output of the previous decoding layer of the decoding layer where the GLAE fusion module is located, and the output of the corresponding encoding layer of the decoding layer where the GLAE fusion module is located, are enhanced and fused to obtain the enhanced fused features, which are then used as the input of the next decoding layer of the decoding layer where the GLAE fusion module is located.

[0066] Optionally, each GLAE fusion module includes a multi-scale convolutional feature extraction unit, a statistical fusion unit, and an adaptive attention weighting unit connected in sequence; The multi-scale convolutional feature extraction unit is used to extract multi-resolution spatial features from the input features of the GLAE fusion module through multi-scale convolution operations to obtain multi-scale feature maps. The statistical fusion unit is used to take multi-scale feature maps as input, perform multi-scale information fusion through statistical feature calculation, and obtain statistical fusion features. The adaptive attention weighting unit is used to take the multi-scale convolutional features and statistical fusion features as global context information, and the input features of the GLAE fusion module as local detail information. It calculates the attention weight between the query matrix generated by the local detail information and the key matrix generated by the global context information in the spatial dimension, and performs weighted fusion on the value matrix generated by the global context information to obtain the output features of the corresponding encoding layer.

[0067] Optionally, each coding layer of the encoder corresponds to a Transformer block, and each Transformer block uses a different number of attention heads; During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first, second, third, and fourth encoding layers, including: During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the Transformer blocks corresponding to the first, second, third, and fourth encoding layers.

[0068] Optionally, each Transformer block consists of a first-layer normalization unit, a multi-depth convolution head transpose attention unit, a first residual connection unit, a second-layer normalization unit, a gated depthwise convolutional feedforward network, and a second residual connection unit. The first normalization unit is used to standardize the input features of its corresponding coding layer according to the channel dimension, so as to obtain the first-level normalized features. The multi-depth convolutional head transposed attention unit is used to aggregate channel context information through a 1×1 convolution to obtain channel aggregated features. Local spatial features in the channel aggregated features are extracted through a 3×3 depth convolution to obtain three sets of feature maps: query, key, and value. Based on the three sets of feature maps, the transposed attention map is calculated in the channel dimension to obtain the channel attention weight. The value feature map is dynamically fused through the channel attention weight to obtain the denoising and enhancement features. The first residual connection unit is used to take the first standardized feature as the input of the identity mapping branch in the first residual connection unit, take the denoising and enhancement feature as the input of the residual mapping branch in the first residual connection unit, and add the identity mapping branch input and the residual mapping branch input element by element to obtain the residual fusion feature. The second normalization unit is used to take the residual fusion features as input and perform normalization through the channel dimension to obtain the secondary normalized features. A gated deep convolutional feedforward network is used to take the quadratically normalized features as input. Local spatial features are extracted through 1×1 convolution and 3×3 deep convolution in the first branch of the gated deep convolutional feedforward network to obtain spatial enhancement features. A gate signal is generated by 1×1 convolution, 3×3 deep convolution and GELU activation in the second branch of the gated deep convolutional feedforward network. The spatial enhancement features and the gate signal are subjected to element-wise multiplication and gating to obtain gated filtering features. The gated filtering features are then mapped to the original channel dimension through 1×1 convolution to obtain gated enhancement features. The second residual connection unit is used to take the quadratic normalized features as the input of the identity mapping branch in the second residual connection unit, take the gated enhancement features as the input of the residual mapping branch in the second residual connection unit, and add the identity mapping branch input and the residual mapping branch input in the second residual connection unit element by element to obtain the output features of the coding layer to which the Transformer block belongs.

[0069] Optionally, the multi-depth convolutional head transpose attention unit in each Transformer block consists of a third-layer normalization unit, a first convolutional unit, a second convolutional unit, a transpose attention mapping unit, and a third residual connection unit; The third normalization unit is used to normalize the first-normalized features according to the channel dimension to obtain MDTA normalized features. The first convolutional unit is used to take the MDTA-normalized features as input and perform cross-channel context aggregation through 1×1 pointwise convolution to obtain channel aggregated features. The second convolutional unit takes the channel aggregation feature as input and extracts local spatial features through a 3×3 depth convolution to obtain three sets of feature maps: query, key, and value. The transposed attention mapping unit is used to calculate the transposed attention map in the channel dimension based on the three sets of feature maps: query, key, and value, to obtain the channel attention weights. The value feature map is then dynamically fused using the channel attention weights to obtain the attention-weighted features. The third residual connection unit is used to take the first-standardized features as the input of the identity mapping branch in the third residual connection unit, take the attention-weighted features as the input of the residual mapping branch in the third residual connection unit, and add the input of the identity mapping branch in the third residual connection unit to the residual mapping branch element by element to obtain the denoising and enhancement features.

[0070] Optionally, the gated depthwise convolutional feedforward network in each Transformer block consists of a fourth normalization unit, two sets of parallel third and fourth convolutional units, a GELU activation unit, an element-wise multiplication gate mechanism unit, and a fourth residual connection unit. The fourth normalization unit is used to normalize the input quadratic normalized features by the channel dimension to obtain GDFN normalized features; Two sets of parallel third and fourth convolutional units are used to take GDFN-normalized features as input, extract local spatial features through 1×1 convolution and 3×3 depth convolution of the third convolutional unit to obtain spatially enhanced features, and perform feature transformation through 1×1 convolution and 3×3 depth convolution of the fourth convolutional unit to obtain gated basic features. The GELU activation unit is used to take the spatial augmentation features as input and perform a nonlinear transformation through the GELU activation function to obtain the gated signal; The element-level multiplication gating mechanism unit is used to take the spatial augmentation features and the gating signal as inputs, and perform gating filtering through element-level multiplication to obtain the gating filtering features. The fourth residual connection unit is used to take the quadratic standardized features as the input of the identity mapping branch in the fourth residual connection unit, take the gating and filtering features after 1×1 convolution as the input of the residual mapping branch in the fourth residual connection unit, and add the input of the identity mapping branch in the fourth residual connection unit to the residual mapping branch element by element to obtain the gating enhanced features.

[0071] The image denoising device integrating global-local adaptive attention in this embodiment of the invention can execute the image denoising method integrating global-local adaptive attention provided in this embodiment of the invention. The implementation principle is similar. The actions performed by each module and unit in the image denoising device integrating global-local adaptive attention in each embodiment of the invention correspond to the steps in the image denoising method integrating global-local adaptive attention in each embodiment of the invention. For detailed functional descriptions of each module of the image denoising device integrating global-local adaptive attention, please refer to the descriptions in the corresponding image denoising method integrating global-local adaptive attention shown above, which will not be repeated here.

[0072] The aforementioned image denoising device that integrates global-local adaptive attention can be a computer program (including program code) running on a computer device. For example, the image denoising device that integrates global-local adaptive attention is an application software. The device can be used to execute the corresponding steps in the method provided in the embodiments of the present invention.

[0073] In some embodiments, the image denoising device integrating global-local adaptive attention provided in this invention can be implemented using a combination of hardware and software. As an example, the image denoising device integrating global-local adaptive attention provided in this invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the image denoising method integrating global-local adaptive attention provided in this invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0074] In other embodiments, the image denoising device merging global-local adaptive attention provided in this invention can be implemented in software. Figure 8 An image denoising apparatus for fusing global-local adaptive attention, stored in memory, is shown. It may be software in the form of programs and plug-ins, and includes a series of modules, including an acquisition module 210 and a denoising module 220, for implementing the image denoising method for fusing global-local adaptive attention provided in the embodiments of the present invention.

[0075] The modules described in the embodiments of the present invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0076] Based on the same principles as the methods shown in the embodiments of the present invention, the embodiments of the present invention also provide an electronic device, which may include, but is not limited to: a processor and a memory; the memory for storing computer programs; and the processor for executing the methods shown in any embodiment of the present invention by invoking the computer programs.

[0077] In one alternative embodiment, an electronic device is provided, such as Figure 9 As shown, Figure 9The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0078] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0079] Bus 4002 may include a pathway for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 9 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.

[0080] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0081] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0082] Among these, electronic devices can also be terminal devices. Figure 9 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0083] This invention provides a computer-readable storage medium storing a computer program that, when run on a computer, enables the computer to execute the corresponding content in the aforementioned method embodiments.

[0084] According to another aspect of the present invention, a computer program product or computer program is also provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various embodiments described above.

[0085] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0086] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0087] The computer-readable storage medium provided in this invention can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0088] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0089] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. An image denoising method of fusing global-local adaptive attention, characterized in that, include: Acquire raw remote sensing images; The original remote sensing image is input into a pre-trained denoising model to obtain a denoised image, wherein the denoising model is an image denoising model using an encoder-decoder architecture; The encoder adopts a four-level symmetrical hierarchical structure, including a first coding layer, a second coding layer, a third coding layer, and a fourth coding layer; the decoder adopts a four-level symmetrical structure that is symmetrical to the encoder, including a fourth decoding layer, a third decoding layer, a second decoding layer, and a first decoding layer. During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer in sequence, to obtain the first feature output by the first encoding layer, the second feature output by the second encoding layer, the third feature output by the third encoding layer and the fourth feature output by the fourth encoding layer. During the decoding stage, the fourth feature is decoded layer by layer through the fourth decoding layer, the third decoding layer, the second decoding layer, and the first decoding layer to obtain the denoised image.

2. The method of claim 1, wherein, Each decoding layer of the decoder includes a GLAE fusion module; For each GLAE fusion module, the output of the previous decoding layer of the decoding layer where the GLAE fusion module is located, and the output of the corresponding encoding layer of the decoding layer where the GLAE fusion module is located, are enhanced and fused to obtain the enhanced fused features, which are then used as the input of the next decoding layer of the decoding layer where the GLAE fusion module is located.

3. The method of claim 2, wherein, Each of the GLAE fusion modules includes a multi-scale convolutional feature extraction unit, a statistical fusion unit, and an adaptive attention weighting unit connected in sequence. The multi-scale convolutional feature extraction unit is used to extract multi-resolution spatial features from the input features of the GLAE fusion module through multi-scale convolution operations to obtain a multi-scale feature map. The statistical fusion unit is used to take the multi-scale feature map as input, perform multi-scale information fusion through statistical feature calculation, and obtain statistical fusion features. The adaptive attention weighting unit is used to take the multi-scale convolutional features and the statistical fusion features as global context information, and the input features of the GLAE fusion module as local detail information. It calculates the attention weight between the query matrix generated by the local detail information and the key matrix generated by the global context information in the spatial dimension, and performs weighted fusion on the value matrix generated by the global context information to obtain the output features of the corresponding encoding layer.

4. The method according to any one of claims 1 to 3, characterized in that, Each coding layer of the encoder corresponds to a Transformer block, and each Transformer block uses a different number of attention heads; In the encoding stage, the original remote sensing image is sequentially processed through the first encoding layer, the second encoding layer, the third encoding layer, and the fourth encoding layer, extracting features layer by layer from shallow to deep, including: During the encoding stage, the original remote sensing image is sequentially extracted layer by layer from shallow to deep using the Transformer blocks corresponding to the first, second, third, and fourth encoding layers.

5. The method of claim 4, wherein, Each Transformer block consists of a first-layer normalization unit, a multi-depth convolution head transpose attention unit, a first residual connection unit, a second-layer normalization unit, a gated depthwise convolution feedforward network, and a second residual connection unit. The first normalization unit is used to perform channel-dimensional standardization on the input features of its corresponding coding layer to obtain first-level standardized features; The multi-depth convolution head transposed attention unit is used to aggregate the first-normalized features with channel context information through a 1×1 convolution to obtain channel aggregated features, extract local spatial features from the channel aggregated features through a 3×3 depth convolution to obtain three sets of feature maps: query, key, and value. Based on the three sets of feature maps, the transposed attention map is calculated in the channel dimension to obtain the channel attention weight. The value feature map is dynamically fused using the channel attention weight to obtain denoising and enhancement features. The first residual connection unit is used to take the first-normalized feature as the input of the identity mapping branch in the first residual connection unit, take the denoising and enhancement feature as the input of the residual mapping branch in the first residual connection unit, and add the identity mapping branch input and the residual mapping branch input element by element to obtain the residual fusion feature. The second-layer normalization unit is used to take the residual fusion features as input and perform normalization processing through layer normalization in the channel dimension to obtain secondary normalized features; The gated deep convolutional feedforward network is used to take the secondary normalized features as input, extract local spatial features through 1×1 convolution and 3×3 deep convolution in the first branch of the gated deep convolutional feedforward network to obtain spatial enhancement features, generate a gated signal through 1×1 convolution, 3×3 deep convolution and GELU activation in the second branch of the gated deep convolutional feedforward network, perform element-wise multiplication gated filtering on the spatial enhancement features and the gated signal to obtain gated filtering features, and map the gated filtering features to the original channel dimension through 1×1 convolution to obtain gated enhancement features; The second residual connection unit is used to take the quadratic normalized feature as the input of the identity mapping branch in the second residual connection unit, take the gated enhancement feature as the input of the residual mapping branch in the second residual connection unit, and add the identity mapping branch input and the residual mapping branch input in the second residual connection unit element by element to obtain the output feature of the coding layer to which the Transformer block belongs.

6. The method of claim 5, wherein, The multi-depth convolution head transposed attention unit in each Transformer block consists of a third-layer normalization unit, a first convolution unit, a second convolution unit, a transposed attention mapping unit, and a third residual connection unit; The third-layer normalization unit is used to normalize the first-standardized features according to the channel dimension to obtain MDTA-standardized features. The first convolutional unit is used to take the MDTA-normalized features as input and perform cross-channel context aggregation through 1×1 pointwise convolution to obtain channel aggregated features. The second convolutional unit is used to take the channel aggregation features as input and extract local spatial features through 3×3 depth convolution to obtain three sets of feature maps: query, key, and value. The transposed attention mapping unit is used to calculate the transposed attention map in the channel dimension based on the three sets of feature maps of query, key and value, to obtain the channel attention weight, and to dynamically fuse the value feature map through the channel attention weight to obtain the attention-weighted feature. The third residual connection unit is used to take the first-normalized feature as the input of the identity mapping branch in the third residual connection unit, take the attention-weighted feature as the input of the residual mapping branch in the third residual connection unit, and add the input of the identity mapping branch in the third residual connection unit and the input of the residual mapping branch element by element to obtain the denoising enhancement feature.

7. The method of claim 5, wherein, Each gated deep convolutional feedforward network in the Transformer block consists of a fourth-layer normalization unit, two sets of parallel third and fourth convolutional units, a GELU activation unit, an element-wise multiplication gate mechanism unit, and a fourth residual connection unit. The fourth-layer normalization unit is used to perform channel-dimensional normalization on the input secondary normalized features to obtain GDFN normalized features. The two sets of parallel third and fourth convolutional units are used to take the GDFN normalized features as input, extract local spatial features through the 1×1 convolution and 3×3 depth convolution of the third convolutional unit to obtain spatial enhancement features, and perform feature transformation through the 1×1 convolution and 3×3 depth convolution of the fourth convolutional unit to obtain gated basic features. The GELU activation unit is used to take the spatial enhancement features as input and perform nonlinear transformation through the GELU activation function to obtain the gated signal; The element-level multiplication gating mechanism unit is used to take the spatial enhancement feature and the gating signal as input, and perform gating filtering through element-level multiplication to obtain gating filtering features. The fourth residual connection unit is used to take the quadratic standardized features as the input of the identity mapping branch in the fourth residual connection unit, take the features of the gated filtering features after 1×1 convolution as the input of the residual mapping branch in the fourth residual connection unit, and add the identity mapping branch input in the fourth residual connection unit to the residual mapping branch input element by element to obtain the gated enhancement features.

8. An image denoising device integrating global-local adaptive attention, characterized in that, include: The acquisition module is used to acquire raw remote sensing images; The denoising module is used to input the original remote sensing image into a pre-trained denoising model to obtain a denoised image, wherein the denoising model is an image denoising model using an encoder-decoder architecture. The encoder adopts a four-level symmetrical hierarchical structure, including a first coding layer, a second coding layer, a third coding layer, and a fourth coding layer; the decoder adopts a four-level symmetrical structure that is symmetrical to the encoder, including a fourth decoding layer, a third decoding layer, a second decoding layer, and a first decoding layer. During the encoding stage, the original remote sensing image is subjected to layer-by-layer feature extraction from shallow to deep through the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer in sequence, to obtain the first feature output by the first encoding layer, the second feature output by the second encoding layer, the third feature output by the third encoding layer and the fourth feature output by the fourth encoding layer. During the decoding stage, the fourth feature is decoded layer by layer through the fourth decoding layer, the third decoding layer, the second decoding layer, and the first decoding layer to obtain the denoised image.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1-7.

Images