Space-wavelet domain collaborative lightweight SAR image speckle suppression method and system

Abstract

The invention discloses a space-wavelet domain collaborative lightweight SAR image speckle suppression method and system, and the method comprises the steps: firstly obtaining a single SAR image containing speckle noise, carrying out the frequency domain decomposition of the image through a learnable wavelet transformation module, and obtaining four sub-band coefficients; secondly, utilizing a wavelet gating and spatial domain fusion module to respectively match the four sub-band coefficients with the resolution and map the four sub-band coefficients into space-wavelet domain collaborative joint representation corresponding to each layer of the spatial domain network; and then, a multi-attention fusion module is utilized, weighted fusion is performed after alignment based on space-frequency collaborative joint representation, image reconstruction is completed through up-sampling, and suppression of speckles is realized. And finally, training by adopting a joint loss function, and storing an optimal training weight for suppressing speckle noise. According to the method, complementary enhancement of local textures and a global scattering mode is realized, edge details are successfully kept, and the speckle suppression capability of the SAR image is improved.

Description

Technical Field

[0001] This invention belongs to the field of synthetic aperture radar image processing technology, specifically relating to a lightweight SAR image speckle suppression method and system with space-wavelet domain coordination. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave imaging sensor, plays an indispensable role in civilian fields such as environmental monitoring and ocean observation due to its all-weather, all-day Earth observation capabilities. However, limited by its coherent imaging mechanism, SAR images inevitably generate severe multiplicative speckle noise during the generation process. This not only significantly reduces the visual quality of the images but also poses a serious challenge to subsequent interpretation tasks such as target detection and recognition. Therefore, effectively suppressing speckle noise while preserving the edge and texture details of the image is a key prerequisite for improving the application value of SAR images.

[0003] In recent years, deep learning-based denoising methods have demonstrated great potential. However, existing methods mostly focus on spatial domain feature learning, and the collaborative modeling of interference patterns propagating in the transform domain and image structure is still insufficient. Furthermore, network structures are generally redundant and computationally complex, making it difficult to meet the needs of real-time onboard processing. In addition, pure convolutional neural networks have inherent limitations in modeling long-range dependencies, while pure Transformer architectures, although capable of capturing global context, are insufficient in learning local details and are computationally complex. Although some studies have attempted to incorporate transform domain priors, most treat wavelet transforms as static feature extractors, failing to achieve fully learnable and adaptive basis functions, thus limiting their generalization ability and denoising accuracy in complex SAR scenarios. Summary of the Invention

[0004] Based on the aforementioned shortcomings in the existing technology, the purpose of this invention is to provide a lightweight SAR image speckle suppression method and system that combines space-wavelet domain collaboration. It adopts an architecture that combines learnable wavelet transform with a spatial domain global feature extraction network to achieve effective noise suppression and accurate preservation of image details.

[0005] To achieve the aforementioned objectives, this invention provides a lightweight SAR image speckle suppression method using a space-wavelet domain collaborative approach, comprising the following steps:

[0006] S1. Acquire a single SAR image containing speckle noise.

[0007] S2. Adaptive frequency domain decomposition of SAR images is performed using a learnable wavelet transform module. Wavelet responses in four directions (LL, LH, HL, HH) are obtained through a trainable wavelet filter. The responses in each direction are downsampled to obtain four sub-band coefficients. The wavelet filter parameters are automatically updated during training to complete the adaptive selection of the wavelet basis.

[0008] S3. Using the wavelet gating and spatial domain fusion module, the four sub-band coefficients are matched with the resolution and mapped to joint representations of spatial-wavelet domain collaboration with each layer of the spatial domain backbone network. LL corresponds to the ×2 receptive field level of the backbone network, LH and HL correspond to the ×3 level, and HH corresponds to the ×4 level. Each gating feature map is modulated by element-wise multiplication with the spatial domain features of the corresponding level, and then merged with the original features in a residual manner. The gating strength is updated during training.

[0009] S4. Using the multi-attention fusion module, after aligning the joint representation based on spatial-wavelet domain collaboration, weighted fusion is performed to generate a fused feature map. The spatial size of the input image is then restored layer by layer through upsampling to complete image reconstruction and achieve SAR speckle noise suppression.

[0010] S5. The network is trained using a joint loss function, and the optimal training weights are saved for the suppression of speckle noise in SAR images.

[0011] As a preferred embodiment, in step S2, the learnable wavelet transform module adaptively generates the optimal wavelet basis through end-to-end training, the core of which is a set of trainable one-dimensional filter kernels that satisfy the orthogonal mirror relationship constraint.

[0012] As a preferred embodiment, in step S3:

[0013] Considering the input SAR image I, after forward propagation through the RepViT backbone network, a multi-resolution spatial domain feature set f1, f2, f3 is obtained. Simultaneously, the input SAR image undergoes learnable wavelet transform (DWT) decomposition to generate four sets of sub-band coefficients:

[0014] DWT θ (I) = LL, {LH, HL, HH}

[0015] DWT θ The term represents a learnable wavelet decomposition layer with parameter θ, where LL is the low-frequency approximate subband, and LH, HL, and HH are the high-frequency detail subbands in the horizontal, vertical, and diagonal directions, respectively.

[0016] For the spatial domain feature f1, the LL subband is used as the gate signal, and structural enhancement is achieved through the following modulation process:

[0017]

[0018] Where u represents the bilinear upsampling operation, adjusting the LL subband resolution to match f1; F LL This is a fusion network consisting of two convolutional layers, responsible for converting single-channel subbands into spatial attention maps; σ is the Sigmoid activation function; γ LL represents the learnable gating strength parameter corresponding to the LL subband; ⊙ denotes successive multiplication.

[0019] For the spatial domain feature f2, design a dual-gating mechanism:

[0020]

[0021] ⊕ represents channel splicing operation, γ LH-HL For the learnable gated strength parameters corresponding to the LH and HL subbands, 1 represents a 1-matrix, F HL-LH Process the concatenated dual-channel input and generate a composite attention map through convolution operations;

[0022] For the spatial domain feature f3, the HH subband is used as a gating signal, and high-frequency injection is used to enhance the model's ability to identify salient structures in the macro-context:

[0023]

[0024] Where γ HH F is the learnable gating strength parameter corresponding to the HH subband. HH Indicates the convolution operation;

[0025] The feature set G = {g} after wavelet-gated modulation LL ,g LH-HL ,g HH This forms a joint representation that combines space and wavelet domain.

[0026] As a preferred embodiment, in step S4, firstly, feature g... LL ,g LH-HL ,g HH The features F are concatenated, and then the first layer of channel attention is designed using the SE module. Channel statistics are obtained through global average pooling, and then the channel importance weights M are learned through a fully connected layer with a bottleneck structure. c M obtained from this c This is a weighted coefficient map for the channel dimension, with a size consistent with the number of channels. The input feature F is modulated by multiplying each channel element-wise to obtain feature F. c .

[0027] Next, a two-layer convolutional network is used to learn the spatial weight distribution to construct the second spatial attention layer M. s And calculate Fs =F c ⊙M s .

[0028] Finally, the relative contributions of features at each scale are learned through global pooling and fully connected layers, and a third-level inter-scale attention mechanism M is constructed. scale ;

[0029] F fused =F s ⊙M scale

[0030] F fused Upsampling restores the spatial size of the input SAR image, resulting in the reconstructed output image P with speckle suppression.

[0031] In another aspect, the present invention provides a lightweight SAR image speckle suppression system with space-wavelet domain coordination, comprising the following modules:

[0032] The SAR image module is used to acquire single SAR images containing speckle noise.

[0033] The sub-band coefficient module is used to perform frequency domain decomposition on SAR images using a learnable wavelet transform module to obtain four sub-band coefficients.

[0034] The space-frequency co-representation module is used to match and map the four sub-band coefficients to the resolution using wavelet gating and spatial domain fusion modules, respectively, into space-frequency co-representations corresponding to each layer of the spatial domain backbone network.

[0035] The speckle noise suppression and reconstruction module utilizes a multi-attention fusion module, performs weighted fusion based on spatial-frequency collaborative joint representation alignment, and completes image reconstruction through upsampling to achieve SAR speckle noise suppression.

[0036] The training module uses a joint loss function for training and saves the optimal training weights for suppressing speckle noise in SAR images.

[0037] The beneficial effects of this invention are as follows:

[0038] (1) This invention breaks through the limitations of traditional methods in processing in the spatial domain or fixed transform domain. It realizes data-driven adaptive frequency domain analysis through learnable wavelet transform, and can generate the optimal sparse frequency domain representation for specific SAR image content.

[0039] (2) The wavelet subband is innovatively used as a dynamic gate signal to hierarchically modulate the multi-scale features extracted by the spatial domain backbone network, establishes a precise correspondence between frequency band characteristics and semantic hierarchy, and realizes complementary enhancement of local texture and global scattering mode.

[0040] (3) A channel-space-scale triple attention module was introduced, which systematically solved the weight allocation problem in multi-source and multi-scale feature fusion, successfully preserving edge details and improving the ability of SAR image speckle suppression. Attached Figure Description

[0041] Figure 1 This is a flowchart of a lightweight SAR image speckle suppression method based on space-small wavelet domain coordination according to an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the experimental results of the embodiments and comparative methods of the present invention. Detailed Implementation

[0043] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0044] In one aspect, this invention provides a lightweight SAR image speckle suppression method based on space-wavelet domain collaboration, which can be summarized as follows: acquiring SAR image data and constructing a dataset containing simulated and real scenes; constructing a denoising network composed of a learnable wavelet transform module, a wavelet gating and spatial domain fusion module, and a multi-attention fusion module; training the network using a joint loss function that combines reconstruction loss and physical constraints and retaining the optimal model; and finally using the optimal model to denoise the SAR image.

[0045] Example 1:

[0046] like Figure 1 As shown, the lightweight SAR image speckle suppression method based on space-wavelet domain coordination in this embodiment includes the following steps:

[0047] (1) Acquire a single SAR image containing speckle noise, and use a learnable wavelet transform module to perform adaptive frequency domain decomposition on the input SAR image to obtain four wavelet sub-bands as frequency domain feature representations.

[0048] The learnable wavelet transform module adaptively generates the optimal wavelet basis through end-to-end training. Its core is a set of trainable one-dimensional filter kernels that satisfy orthogonal mirror relationship constraints.

[0049] Four wavelet subbands: LL∈R B×1×H / 2×W / 2 For the low-frequency approximate subband, LH, HL, HH∈R B×1×H / 2×W / 2 These are high-frequency detail subbands in the horizontal, vertical, and diagonal directions, respectively.

[0050] (2) Feature extraction and interaction are performed using a learnable wavelet transform module and the RepViT spatial domain backbone network, and fusion is achieved through a wavelet gating mechanism. The execution process of the wavelet gating and spatial domain fusion module at each stage can be represented by the following formula:

[0051] Consider the input SAR image I∈R B×1×H×W After forward propagation through the RepViT backbone network, a multi-resolution spatial domain feature set f1, f2, f3 is obtained, where These feature maps correspond to downsampling scales of 1 / 2, 1 / 3, and 1 / 4 of the original resolution, respectively. Simultaneously, the input SAR image undergoes learnable wavelet transform (DWT) decomposition to generate four sets of sub-band coefficients.

[0052] DWT θ (I) = LL, {LH, HL, HH}

[0053] For the spatial domain feature f1, its core function is to delineate the main outline and basic structure of the image, which is highly synergistic with the stable structural support provided by the low-frequency global features carried by the LL subband. Therefore, using the LL subband as the gate signal, structural enhancement is achieved through the following modulation process:

[0054]

[0055] in This indicates a bilinear upsampling operation, adjusting the LL subband resolution to match f1; F LL This is a fusion network consisting of two convolutional layers, responsible for converting single-channel subbands into spatial attention maps; σ is the Sigmoid activation function; γ LL represents the learnable gating strength parameter corresponding to the LL subband; ⊙ denotes successive multiplication.

[0056] For the spatial domain feature f2, its core function is to characterize the fine contours and significant textures of the subject. This is highly synergistic with the mid-frequency edge and structural information dominated by the LH and HL subbands. Therefore, a dual-gating mechanism is designed:

[0057]

[0058] Where ⊕ represents the channel splicing operation, γ LH-HL For the learnable gated strength parameters corresponding to the LH and HL subbands, 1 represents a 1-matrix, F HL-LH The concatenated dual-channel input is processed, and a composite attention map is generated through convolution operations.

[0059] For the spatial domain feature f3, its core function lies in modeling the global context and dominant scattering patterns of the image. The diagonal high-frequency details (such as point targets and fine textures) precisely located by the HH subband provide crucial clues for distinguishing strong scatterers from noise. Therefore, the HH subband is used as a gating signal, and high-frequency injection is employed to enhance the model's ability to identify salient structures in a macroscopic context.

[0060]

[0061] Where γ HH F is the learnable gating strength parameter corresponding to the HH subband. HH Indicates the convolution operation;

[0062] The feature set G = {g} after wavelet-gated modulation LL ,g LH-HL ,g HH This forms a joint representation that combines space and wavelet domain.

[0063] (3) Input the modulated multi-scale features into the multi-attention fusion module. This module first concatenates the features and then sequentially passes them through channel attention, spatial attention, and inter-scale attention modules. The learning process of the inter-scale attention weights can be expressed by the following formula:

[0064] First, consider feature g. LL ,g LH-HL ,g HH The features F are concatenated, and then the first layer of channel attention is designed using the SE module. Channel statistics are obtained through global average pooling, and then the channel importance weights are learned through a fully connected layer with a bottleneck structure.

[0065] M c =σ(W2(δ(W1(GAP(F)))))∈R B×4C×1×1

[0066] Where F represents the spatial feature of the input, GAP(·) is the operation of performing global average pooling on the feature in the spatial dimension to obtain channel statistics; W1 and W2 are linear transformation parameters for first reducing and then increasing the dimensionality of the pooling result, and the dimensionality reduction ratio is controlled by the reduction ratio r; δ(·) is the ReLU activation function, from which M is obtained. c This is a weighted coefficient map for the channel dimension, with a size consistent with the number of channels. The input feature F is modulated by multiplying each channel element by element.

[0067] Next, a two-layer convolutional network is used to learn the spatial weight distribution to construct a second layer of spatial attention:

[0068]

[0069] Fs =F c ⊙M s

[0070] Finally, the relative contributions of features at each scale are learned through global pooling and fully connected layers to construct a third level of inter-scale attention:

[0071] M scale =σ(W4(ReLU(W3(GAP(F)) s )))))∈R B×4C×1×1

[0072] F fused =F s ⊙M scale

[0073] F fused Upsampling restores the spatial size of the input SAR image, resulting in the reconstructed output image P with speckle suppression.

[0074] (4) Use the joint loss function to train the network.

[0075] The joint loss function is composed of the following weighted components: L1 norm reconstruction loss L rec Total variational regularization loss L tv Orthogonality constraint loss L for learnable wavelet filters orth With energy normalization constraint loss L norm The overall loss function is as follows:

[0076] L total =L rec +λ1L TV +λ2L orth +λ3L norm

[0077] First, to suppress local oscillations caused by speckle noise and maintain the continuity of the edge structure, a total variational regularization term is introduced. Its expression is as follows:

[0078]

[0079] Where P i,j This represents the pixel value of image P at coordinates (i,j).

[0080] The reconstruction loss is measured using the L1 norm as follows:

[0081] L rec =||PT||1

[0082] To ensure that the learnable wavelet filters h and g maintain the mathematical properties of the wavelet transform throughout the optimization process, additional regularization constraints based on the physical properties of the filters are applied. Orthogonality constraints force the low-pass and high-pass filters to remain orthogonal in the vector space:

[0083] L orth =|<h,g> |

[0084] h is the low-pass filter learned by the network, g is the high-pass filter learned by the network, <> represents the vector dot product, and || represents the absolute value.

[0085] The norm constraint of the learnable wavelet filter maintains energy conservation in the decomposition process by limiting the L2 norm of the filter.

[0086] L norm =(||h||2-1) 2

[0087] ||h||2 is the L2 norm of filter h.

[0088] Wherein, λ1, λ2, and λ3 are hyperparameters for balancing the weights of each item, and in this embodiment, they are set to 0.1, 0.01, and 0.01, respectively.

[0089] By using the backpropagation algorithm and the AdamW optimizer, the loss function is minimized to update the network parameters. The learning rate is adjusted using a cosine annealing strategy. The model is iteratively trained for 300 epochs until it converges. The optimal model is then saved for SAR image denoising.

[0090] In another aspect, the present invention provides a lightweight SAR image speckle suppression system with space-wavelet domain coordination, comprising the following modules:

[0091] The SAR image module is used to acquire single SAR images containing speckle noise.

[0092] The sub-band coefficient module is used to perform frequency domain decomposition on SAR images using a learnable wavelet transform module to obtain four sub-band coefficients.

[0093] The space-frequency co-representation module is used to match and map the four sub-band coefficients to the resolution using wavelet gating and spatial domain fusion modules, respectively, into space-frequency co-representations corresponding to each layer of the spatial domain backbone network.

[0094] The speckle noise suppression and reconstruction module utilizes a multi-attention fusion module, performs weighted fusion based on spatial-frequency collaborative joint representation alignment, and completes image reconstruction through upsampling to achieve SAR speckle noise suppression.

[0095] The training module uses a joint loss function for training and saves the optimal training weights for suppressing speckle noise in SAR images.

[0096] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

[0097] To verify the effectiveness and synergistic effect of the proposed learnable wavelet transform, gating fusion mechanism, and multi-scale attention fusion module, ablation experiments were designed. All ablation experiments were conducted on the same NWPU VHR-10 dataset to ensure fair comparability of results.

[0098] The experiment began with the construction of a baseline model (Scheme 1) based entirely on the RepViT architecture. This model performs feature extraction and denoising entirely in the spatial domain, and its performance was used as the evaluation benchmark. As the results show, this baseline model has a certain denoising capability, but it is difficult to achieve the optimal balance between noise suppression and detail preservation.

[0099] Based on the baseline model, a learnable wavelet transform module (Scheme 2) was first introduced. This module aims to adaptively learn the optimal wavelet basis in a data-driven manner to replace the traditional fixed basis functions. Experimental results show that even without a specially designed fusion mechanism, the model's performance, especially the evaluation metrics sensitive to structural information, has been initially improved simply by introducing learnable frequency domain analysis. This demonstrates the potential value of adaptive frequency domain representation for SAR image denoising tasks.

[0100] To explore more effective methods for frequency-space feature interaction, a combination of fixed wavelet basis and gated fusion was tested in Scheme 3. This scheme retains the traditional discrete wavelet transform but innovatively uses its subbands as gate signals to modulate the global features extracted by RepViT. The significant improvement in results strongly confirms the effectiveness of the core architecture of "wavelet gating," indicating that using frequency band information to guide spatial feature modulation is a highly promising direction.

[0101] Next, combining the advantages of the first two approaches, Scheme 4, namely "learnable wavelet transform + gating fusion," was constructed. This scheme enables the network to not only adaptively find the optimal analysis basis but also dynamically utilize these basis functions to perform precise feature modulation. Experimental results showed a significant performance leap, clearly demonstrating a powerful synergistic effect between learnable adaptive capabilities and gating architecture.

[0102] Ultimately, a complete WRV-Net model (Scheme 5) was constructed, introducing a multi-scale attention fusion module based on Scheme 4. This module is responsible for intelligently integrating wavelet-gated features from different scales. The complete model achieved optimality on all evaluation metrics, fully demonstrating the completeness of the overall design. This module, through a triple attention mechanism of channel, space, and scale, ensures that local details, global context, and high-frequency scattering information can be collaboratively optimized and reconstructed, making it an indispensable part of the entire denoising pipeline.

[0103] Table 1 clearly shows the positive trend brought about by the "addition-based" approach: starting from the pure spatial domain baseline, both PSNR and SSIM increase simultaneously after introducing only learnable wavelets; further superimposing wavelet-gated fusion results in a significant increase in both metrics again; finally, completing the multi-scale attention fusion module leads to the highest levels of both PSNR and SSIM, exhibiting a monotonically increasing trend overall, verifying that each component can continuously contribute to the gain. Actual results are as follows... Figure 2 As shown

[0104] Table 1

[0105]

Claims

1. A lightweight SAR image speckle suppression method based on space-wavelet domain synergy, characterized in that, Includes the following steps: S1. Acquire a single SAR image containing speckle noise; S2. Use the learnable wavelet transform module to perform frequency domain decomposition on the SAR image to obtain four sub-band coefficients; S3. Using the wavelet gating and spatial domain fusion module, the coefficients of the four sub-bands are matched with the resolution and mapped to a joint representation of spatial-wavelet domain coordination with each layer of the spatial domain backbone network. S4. Using a multi-attention fusion module, weighted fusion is performed after joint representation alignment based on space-wavelet domain collaboration, and image reconstruction is completed through upsampling to suppress SAR speckle noise. S5. Use a joint loss function for training and save the optimal training weights for suppressing speckle noise in SAR images.

2. The lightweight SAR image speckle suppression method based on space-wavelet domain coordination according to claim 1, characterized in that, The specific implementation of step S2 is as follows: the SAR image is adaptively decomposed using a learnable wavelet transform module, and wavelet responses in four directions (LL, LH, HL, HH) are obtained through a trainable wavelet filter. The responses in each direction are downsampled to obtain four sub-band coefficients. The wavelet filter parameters are automatically updated during training to complete the adaptive selection of the wavelet basis.

3. The lightweight SAR image speckle suppression method based on space-wavelet domain coordination according to claim 2, characterized in that, The specific implementation process of step S3 is as follows: The input SAR image I is forward-propagated through the RepViT backbone network to obtain a multi-resolution spatial domain feature set f1, f2, f3. Simultaneously, the input SAR image undergoes learnable wavelet transform (DWT) decomposition to generate four sets of sub-band coefficients. DWT θ (I)=LL,{LH,HL,HH} DWT θ The term represents a learnable wavelet decomposition layer with parameter θ, where LL is the low-frequency approximate subband, and LH, HL, and HH are the high-frequency detail subbands in the horizontal, vertical, and diagonal directions, respectively. For the spatial domain feature f1, the LL subband is used as the gate signal, and structural enhancement is achieved through the following modulation process: Where u represents the bilinear upsampling operation, adjusting the LL subband resolution to match f1; F LL The network is a fusion network composed of convolutions; σ is the Sigmoid activation function; γ LL represents the learnable gating strength parameter corresponding to the LL subband; ⊙ denotes successive multiplication; For the spatial domain feature f2, design a dual-gating mechanism: in Indicates channel splicing operation, γ LH-HL For the learnable gated strength parameters corresponding to the LH and HL subbands, 1 represents a 1-matrix, F HL-LH Process the concatenated dual-channel input and generate a composite attention map through convolution operations; For spatial domain feature f3, the HH subband is used as a gate signal, and high-frequency injection is used to enhance the ability to identify salient structures in the macroscopic context. Where γ HH F is the learnable gating strength parameter corresponding to the HH subband. HH Indicates the convolution operation; The feature set G = {g} after wavelet-gated modulation LL ,g LH-HL ,g HH This forms a joint representation that integrates the space and wavelet domains.

4. The lightweight SAR image speckle suppression method based on space-wavelet domain coordination according to claim 3, characterized in that, The specific implementation process of step S4 is as follows: First, consider feature g. LL ,g LH-HL ,g HH The features F are concatenated, and then the first layer of channel attention is designed using the SE module. Channel statistics are obtained through global average pooling, and then the channel importance weights are learned through a fully connected layer with a bottleneck structure. M c =σ(W2(δ(W1(GAP(F))))) Where F represents the spatial feature of the input, GAP(·) is the operation of performing global average pooling on the feature in the spatial dimension to obtain channel statistics; W1 and W2 are linear transformation parameters for first reducing and then increasing the dimensionality of the pooling result, and the dimensionality reduction ratio is controlled by the reduction ratio r; δ(·) is the ReLU activation function, from which M is obtained. c This is a weighted coefficient map for the channel dimension, with a size consistent with the number of channels. The input feature F is modulated by multiplying each channel element-wise to obtain feature F. c ; We employ convolutional networks to learn the spatial weight distribution and construct a second layer of spatial attention: M s =σ(Conv(ReLU(Conv(F c )))) F s =F c ⊙M s Finally, the relative contributions of features at each scale are learned through global pooling and fully connected layers to construct a third level of inter-scale attention: M scale =σ(W4(ReLU(W3(GAP(F s )))))∈R B×4C×1×1 F fused =F s ⊙M scale F fused Upsampling restores the spatial size of the input SAR image, resulting in the reconstructed output image P with speckle suppression.

5. The lightweight SAR image speckle suppression method based on space-wavelet domain coordination according to claim 4, characterized in that, The specific implementation process of step S5 is as follows: The joint loss function combines the MSE reconstruction loss, the total variation regularization term, and soft orthogonality and norm constraints on the learnable wavelet filter, as detailed below: L total =L rec +λ1L TV +λ2L orth +λ3L norm Where λ1, λ2, and λ3 are the balance coefficients of each regularization term; Given a reconstructed image P∈R B×1×H×W and the corresponding real target image T∈R B×1×H×W Reconstruction loss L rec The L1 norm is used as the metric. The total variational regularization term is expressed as follows: Where P i,j This represents the pixel value of image P at coordinates (i,j); Based on the regularization constraints and orthogonality constraints of the filter's physical properties, the low-pass and high-pass filters are forced to remain orthogonal in the vector space, resulting in the soft orthogonality constraint L for the learnable wavelet filter. orth ; The norm constraint of the learnable wavelet filter maintains the energy conservation of the decomposition process by limiting the L2 norm of the filter.

6. A lightweight SAR image speckle suppression system based on space-wavelet domain coordination, used to implement the lightweight SAR image speckle suppression method according to any one of claims 1 to 5, characterized in that, Includes the following modules: The SAR image module is used to acquire a single SAR image containing speckle noise; The sub-band coefficient module is used to perform frequency domain decomposition on SAR images using a learnable wavelet transform module to obtain four sub-band coefficients. The space-frequency co-representation module is used to match the four sub-band coefficients with the resolution and map them into space-frequency co-representations corresponding to each layer of the spatial domain backbone network by using wavelet gating and spatial domain fusion module. The speckle noise suppression and reconstruction module utilizes a multi-attention fusion module, performs weighted fusion based on joint representation alignment using spatial-frequency coordination, and completes image reconstruction through upsampling to achieve SAR speckle noise suppression. The training module uses a joint loss function for training and saves the optimal training weights for suppressing speckle noise in SAR images.

Images