SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention
By alternating the use of multi-view processing and channel segmentation attention modules, the problem of information loss in SAR image super-resolution reconstruction after speckle noise suppression is solved, and high-fidelity restoration of high-resolution images is achieved.
Patent Information
- Application Number
- CN202511359292.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-23
- Publication Date
- 2026-02-10
AI Technical Summary
Existing technologies struggle to effectively recover lost information during super-resolution reconstruction of SAR images after speckle noise suppression, and often reduce the spatial resolution of the images.
A SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention is adopted. By alternately connecting the multi-view processing downsampling module and the channel segmentation attention upsampling module, speckle noise is suppressed and spatial resolution and high-frequency feature information are restored.
It achieves high-fidelity restoration of high-resolution SAR images while suppressing speckle noise, and has low computational complexity.
Smart Images
Figure CN121504723A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the field of image super-resolution reconstruction, in particular to a SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention. BACKGROUND
[0002] Synthetic aperture radar (SAR) is widely used on satellite platforms for all-weather and all-time observation tasks, and the corresponding image products (SAR images) have rich applications in various industries. However, due to the imaging characteristics of SAR images, they inevitably suffer from severe speckle noise, which seriously affects the subsequent application of SAR images.
[0003] To suppress the speckle noise, a multi-view processing method is usually used on the original SAR image data. This method can effectively suppress the speckle noise, but it usually reduces the spatial resolution of the image, causing the loss of some small spatial targets and the loss of their own spatial information. Therefore, SAR image speckle noise suppression and SAR image super-resolution reconstruction are important and basic research directions in the field of SAR image processing. Researchers usually study these two problems one by one without considering both at the same time. It is difficult to effectively recover the lost information in the speckle noise suppression process when performing super-resolution reconstruction on the image after speckle noise suppression. SUMMARY
[0004] The purpose of the present application is to provide a SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention, which solves the technical problem of difficulty in effectively recovering the lost information when performing super-resolution reconstruction on the image after existing speckle noise suppression.
[0005] To solve the above problems, the present application provides a SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention, specifically a SAR image super-resolution reconstruction method based on multi-view processing down-sampling and channel segmentation attention up-sampling, including the following steps: S1, obtaining a low-resolution SAR image; S2, constructing a SAR image super-resolution reconstruction model; the SAR image super-resolution reconstruction model includes a plurality of groups of alternately connected multi-view processing down-sampling modules and channel segmentation attention up-sampling modules; S3, inputting the low-resolution SAR image into a shallow feature extraction module to convert the low-resolution SAR image to a high-dimensional vector domain and obtain a shallow high-dimensional feature vector; S4, inputting the shallow high-dimensional feature vector into the SAR image super-resolution reconstruction model to obtain a final high-dimensional feature vector that sufficiently suppresses the speckle noise and does not lose the spatial resolution information and high-frequency feature information of the image itself. S5, input the final high-dimensional feature vector in step S4 into a separate channel segmentation attention up-sampling module, and map it from the vector domain back to the image domain through a convolutional layer to obtain a reconstructed high-resolution SAR image.
[0006] A storage device stores instructions and data for implementing a SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention.
[0007] A SAR image super-resolution reconstruction device based on multi-view processing and channel segmentation attention includes a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention.
[0008] The method uses the multi-view processing down-sampling module and the channel segmentation attention up-sampling module to suppress the coherent speckle noise and recover the spatial resolution information and the corresponding high-frequency feature information, and finally uses the separate channel segmentation attention up-sampling module to restore the spatial resolution required by the super-resolution reconstruction of the method, thereby effectively solving the problem of information loss caused by the coherent speckle noise suppression task and the subsequent super-resolution reconstruction, and realizing the restoration of high-resolution SAR images with small calculation complexity and high fidelity. BRIEF DESCRIPTION OF DRAWINGS
[0009] Figure 1 is the overall flowchart of the method of the present application; Figure 2 is a structural schematic diagram of the multi-view processing down-sampling module of the present application; Figure 3 is a structural schematic diagram of the channel segmentation attention up-sampling module of the present application; Figure 4 is a schematic diagram of the SAR image super-resolution reconstruction result of the present application; Figure 5 is a hardware device working schematic diagram of the present application. DETAILED DESCRIPTION
[0010] To make the purpose, technical scheme and advantages of the present application clearer, the embodiments of the present application will be further described below with reference to the drawings.
[0011] Before formally describing the present application, the scheme of the present application is first described generally for easy understanding.
[0012] Please refer to Figures 1-3The application provides a SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention, which comprises the following steps: S1, acquiring a low-resolution SAR image; S2, constructing a SAR image super-resolution reconstruction model; the SAR image super-resolution reconstruction model comprises a plurality of groups of alternately connected multi-view processing down-sampling modules and channel segmentation attention up-sampling modules; S3, inputting the low-resolution SAR image into a shallow feature extraction module, converting the low-resolution SAR image to a high-dimensional vector domain, and obtaining a shallow high-dimensional feature vector; S4, inputting the shallow high-dimensional feature vector into the SAR image super-resolution reconstruction model, and obtaining a final high-dimensional feature vector which sufficiently suppresses coherent noise and does not lose spatial resolution information and high-frequency feature information of itself; S5, inputting the final high-dimensional feature vector in step S4 into a separate channel segmentation attention up-sampling module, and then mapping it from the vector domain to the image domain through a convolution layer to obtain a reconstructed high-resolution SAR image.
[0013] It should be noted that step S3 is specifically as follows:
[0014] Wherein, , are a low-resolution SAR image with a size of 1xHxW and a shallow high-dimensional feature vector with a size of 128xHxW extracted respectively, and conv1 is a convolution operation. It should be noted that in step S4, after the shallow high-dimensional feature vector is input into the SAR image super-resolution reconstruction model, it first enters the multi-view processing down-sampling module; the processing process of the multi-view processing down-sampling module is as follows: S41, the shallow high-dimensional feature vector is subjected to layer normalization processing to obtain a layer normalized feature vector:
[0015] Wherein, is the layer normalized feature vector, and LayerNorm is a layer normalization operation. S42, performing different size step convolution operations on the layer normalized feature vector to realize coherent speckle noise suppression:
[0016]
[0017]
[0018] Wherein, are high-dimensional feature vectors with the size of 128xH / 2xW / 2 obtained by using the conv2, conv3, and conv4 convolution kernels for spatial multi-view processing, respectively; S43, the high-dimensional feature vectors obtained by different multi-view processing strategies are fused to obtain a high-dimensional feature vector after information fusion processing:
[0019] wherein, is a high-dimensional feature vector after information fusion processing of multiple high-dimensional feature vectors obtained by different multi-view processing strategies, is an operation of concatenating feature vectors in the channel dimension; S43, the high-dimensional feature vectors obtained by different multi-view processing strategies are fused to obtain a high-dimensional feature vector after information fusion processing:
[0020] wherein, is a high-dimensional feature vector after information fusion processing, and conv6 is a 1x1 convolution operation; S44, the high-dimensional feature vector after information fusion is input into three convolution layers with different convolution sizes to perform multi-scale convolution feature extraction:
[0021]
[0022]
[0023] wherein, are high-dimensional feature vectors of different scales obtained by using conv7, conv8, and conv9, respectively, and conv7, conv8, and conv9 are convolution layers with different sizes; S45, the high-dimensional feature vectors of different scales are input into a convolution operation and an activation function to extract high-frequency feature information and obtain a high-frequency feature residual vector:
[0024] wherein, is a high-frequency feature residual vector extracted by multiple scale convolution operations, is an operation of concatenating feature vectors in the channel dimension, and relu is a torch.nn.relu activation function in pytorch, conv 10 is a convolution operation; S46, the high-frequency feature residual vector is input into a 1x1 convolution kernel to obtain a high-dimensional feature vector after sufficient fusion of different multi-view processing: M 10 After a convolution operation and M Adding these 5 together yields the high-dimensional feature vector output by the multi-view processing downsampling module. F 2.
[0025] It should be noted that in step S4, the output of the multi-view processing downsampling module is further input to the channel segmentation attention upsampling module. The processing procedure of the channel segmentation attention upsampling module is as follows: S47, Yes F After layer normalization, the input is fed into a convolutional layer and activation function, as shown in the following formula:
[0026] in, For the initial reconstruction of the high-dimensional feature vector, LayerNorm is the layer normalization operation, relu is the torch.nn.relu activation function in PyTorch, and conv12 is the convolution operation; S48. After passing C1 through a convolutional layer, perform vector segmentation in the channel dimension to perform subsequent channel attention operations, as shown in the following equation:
[0027] in, The features of each channel are of size 32×H / 2×W / 2; attention requires the feature vectors of the corresponding channels, the remaining feature vectors of the channel segmentation of size (128-32)×H / 2×W / 2, and conv13 is the convolution operation; The channel segmentation operation splits the input high-dimensional feature vector into a 32-dimensional feature vector and a feature vector of the remaining dimension in the channel dimension. S49, will C 2. The input is fed into a global average pooling layer to integrate all spatial information within the channel, as shown in the following formula:
[0028] in, This is the feature vector after the global pooling layer operation. This is a global average pooling layer operation; S410, to We use two branches to construct attention-weighted convolutional kernels that focus on features at different scales, as shown below:
[0029]
[0030] in, Convolutional kernels with attention weights that focus on features at different scales are constructed respectively. ReLU is the torch.nn.reLU activation function in PyTorch. conv14, conv15, conv16 and conv17 are all convolution operations. reshape1 and reshape2 are vector size adjustment operations. S411. Using the two attention-weighted convolutional kernels obtained above that focus on features at different scales, we can apply them to the high-dimensional feature vector. Perform a convolution operation, then sum the results of the convolutions to obtain the feature vector after channel segmentation and attention processing, as shown below:
[0031] in, These are the feature vectors processed by two attention-weighted convolutional kernels that focus on features at different scales. This is a convolution operation; S412, will and The channel dimensions are concatenated to restore the feature vector, resulting in a feature vector with the same dimensions as before segmentation, as shown below:
[0032] in, This is the feature vector after concatenation and restoration along the channel dimension. This is an operation for concatenating vectors along the channel dimension; S413, will The input is fed into two consecutive convolutional layers and then added to the input feature vector. The feature vector obtained after restoring the high-frequency feature information is shown below:
[0033] in, conv18 and conv19 are the feature vectors after restoring high-frequency feature information, and convolution operations are respectively. S414, to Upsampling is performed to restore the spatial resolution lost by the multi-view processing downsampling module, as shown in the following equation:
[0034] in, The high-dimensional feature vector after restoring the high-frequency features output by the channel segmentation attention upsampling module, i.e. The operation is applied at twice the original value.
[0035] It should be noted that, since the SAR image super-resolution reconstruction model in step S2 includes multiple sets of alternately connected multi-look processing downsampling modules and channel segmentation attention upsampling modules, steps S41 to S414 are executed multiple times, except that the input to step S41 in the first pass is a shallow high-dimensional feature vector. In addition, the input to the remaining steps S41 is the output of the previous channel segmentation attention upsampling module, and the high-dimensional feature vector after high-frequency feature restoration is output by the channel segmentation attention upsampling module in the last pass, which is the final high-dimensional feature vector.
[0036] It should be noted that step S5 is specifically as follows:
[0037] in, For the reconstructed high-resolution SAR image, CSABfinal is the last group of channel segmentation attention upsampling modules in the SAR image super-resolution reconstruction model. F final-1 conv20 is the output of the penultimate channel segmentation attention upsampling module in the SAR image super-resolution reconstruction model; conv20 is the convolution operation.
[0038] As one embodiment, the present invention mainly includes the following processing: (1) Shallow feature extraction module First, the low-resolution SAR image is input into the shallow feature extraction module to transform it from the image domain to the high-dimensional vector domain, as shown below: (1) in, , The input vectors are a low-resolution SAR image of size 1×H×W and an extracted shallow high-dimensional feature vector of size 128×H×W. conv1 is a 3×3 convolution operation with 1 input channel and 128 output channels.
[0039] (2) Multi-view processing downsampling module Next, the extracted shallow high-dimensional feature vectors The input is fed to the multi-look processing downsampling module to suppress speckle noise using multi-look processing operations in the spatial domain, and residual connections are used to recover some high-frequency feature information, as shown below: (2) in, , These represent the input high-dimensional feature vector of size 128×H×W to this module, and the output high-dimensional feature vector of size 128×H / 2×W / 2 after suppressing speckle noise, respectively. MPDB1 is the multi-look processing downsampling module proposed in this paper, and the corresponding structural diagram is shown below. Figure 2 As shown below, the specific process will be described in detail below.
[0040] First, the input shallow high-dimensional feature vector Perform layer normalization as follows: (3) in, The vector is the feature vector after layer normalization, and LayerNorm is the layer normalization operation.
[0041] Then, the eigenvectors after layer normalization... Different strided convolution operations of varying sizes are performed to achieve a 2x downsampling of the spatial dimension. Multi-scale, multi-view processing in the spatial dimension is then used to suppress speckle noise, as shown below: (4) (5) (6) in, The three convolutional operations are conv2, conv3, and conv4, respectively, which are high-dimensional feature vectors of size 128×H / 2×W / 2 for spatial multi-view processing. conv2 is a 3×3 convolution operation with a stride of 2 and padding of 1. conv3 is a 5×5 convolution operation with a stride of 2 and padding of 2. conv4 is a 7×7 convolution operation with a stride of 2 and padding of 3.
[0042] Next, the high-dimensional feature vectors obtained using different multi-view processing strategies are analyzed. Feature concatenation is performed, and then the data is input into a convolutional layer and activation function to fuse information from multiple high-dimensional feature vectors obtained by different multi-view processing strategies, as shown below: (7) in, This refers to the high-dimensional feature vector obtained after information fusion processing of multiple high-dimensional feature vectors obtained by different multi-view processing strategies. This is an operation that concatenates feature vectors along the channel dimension. ReLU is the torch.nn.reLU activation function in PyTorch, and conv5 is a 3×3 convolution operation with 384 input channels and 128 output channels.
[0043] Based on the above operations, it is believed that different multi-view processing strategies have been used to effectively suppress speckle noise.
[0044] However, some high-frequency feature information is inevitably lost during multi-view processing. The lost high-frequency information is then restored using residual connection operations.
[0045] Specifically, the high-dimensional feature vector after information fusion is first... The input is fed into a 1×1 convolutional layer to fully fuse information from different multi-view processing strategies, as shown below: (8) in, To obtain the fully fused high-dimensional feature vector, conv6 is a 1×1 convolution operation with 128 input and 128 output channels.
[0046] Then, the fully fused high-dimensional feature vectors The input is fed into three convolutional layers with different convolutional sizes for multi-scale convolutional feature extraction, as shown below: (9) (10) (11) in, These are high-dimensional feature vectors of different scales extracted using conv7, conv8, and conv9, respectively. conv7, conv8, and conv9 are convolution operations of size 3×3, 5×5, and 7×7, respectively, with 128 input and output channels.
[0047] Next, the extracted high-dimensional feature vectors at different scales are... After concatenation, the input is fed into a convolutional operation and activation function to extract high-frequency feature information, as shown below: (12) in, The high-frequency feature residual vectors are extracted from convolution operations at multiple scales. This is an operation that concatenates feature vectors along the channel dimension. ReLU is the torch.nn.reLU activation function in PyTorch, and conv10 is a 3×3 convolution operation with 384 input channels and 128 output channels.
[0048] Finally, the high-frequency feature residual vectors extracted from the convolution operations at multiple scales are... After further processing, a high-dimensional feature vector that fully suppresses speckle noise is added. The output of this module is shown below: (13) in, conv11 is the high-dimensional feature vector output by the multi-view processing downsampling module, and conv11 is a 1×1 convolution operation with 128 input and output channels.
[0049] (3) Channel Segmentation Attention Upsampling Module The multi-view downsampling module is considered to have effectively suppressed speckle noise in the high-dimensional feature vectors. However, some high-frequency information in the spatial dimension is inevitably lost during the processing of different multi-view strategies.
[0050] Next, the channel segmentation attention upsampling module is immediately used to recover the spatial resolution information and the lost high-frequency feature information, as shown below: (14) in, The input high-dimensional feature vector of size 128×H / 2×W / 2 is given by this module, and the output high-dimensional feature vector of size 128×H×W after high-frequency feature restoration is given by this module. CSAB1 is the channel segmentation attention upsampling module proposed in this paper, and the corresponding structural diagram is shown below. Figure 3 As shown below, the specific process will be described in detail below.
[0051] First, the high-dimensional feature vector after suppressing the input speckle noise... After layer normalization, the input is fed into a convolutional layer and activation function, as shown below: (15) in, The high-dimensional feature vector is initially restored. LayerNorm is the layer normalization operation, relu is the torch.nn.relu activation function in PyTorch, conv12 is a 3×3 convolution operation, and both the input and output channels are 128.
[0052] Next, the initially restored high-dimensional feature vectors After passing through a convolutional layer, vectors are split along the channel dimension to perform subsequent channel attention operations, as shown below: (16) in, The features vectors for the corresponding channels are 32×H / 2×W / 2, and the remaining features vectors for the channel segmentation are (128-32)×H / 2×W / 2. conv13 is a 1×1 convolution operation with 128 input and output channels. The channel segmentation operation splits the input high-dimensional feature vector into a 32-dimensional feature vector and a feature vector of the remaining dimension in the channel dimension.
[0053] Then, the high-dimensional feature vectors obtained by segmenting in the channel dimension... An attention mechanism is used to apply additional attention to high-frequency features to reconstruct the corresponding features. Specifically, the high-dimensional feature vectors are... The input is fed into a global average pooling layer to integrate all spatial information within the channel, as shown below: (17) in, The feature vector after the global pooling layer operation is 32×1×1 in size. This is a global average pooling layer operation.
[0054] Next, the feature vector after the global pooling layer operation is processed. We use two branches to construct attention-weighted convolutional kernels that focus on features at different scales, as shown below: (18) (19) in, Convolutional kernels with attention weights of size 3×3×1×32 and 5×5×1×32 are constructed to focus on features at different scales. ReLU is the torch.nn.reLU activation function in PyTorch. conv14, conv15, conv16, and conv17 are all 1×1 convolutional operations, but conv14 has 32 input channels and 128 output channels, conv15 has 128 input channels and 288 output channels, conv16 has 32 input channels and 320 output channels, and conv17 has 320 input channels and 800 output channels. reshape1 and reshape2 are vector resizing operations. reshape1 resizes a 288×1×1 vector to 3×3×1×32, and reshape2 resizes an 800×1×1 vector to 5×5×1×32.
[0055] Then, the two attention-weighted convolutional kernels obtained above, which focus on features at different scales, are used to process the high-dimensional feature vector. Perform a convolution operation, then sum the results of the convolutions to obtain the feature vector after channel segmentation and attention processing, as shown below: (20) in, The vectors are the result of two attention-weighted convolutional kernels that focus on features at different scales, with × representing the convolution operation.
[0056] Through the above operations, the high-dimensional feature vector after channel dimension segmentation is completed. The attention mechanism operation will then process the partial channel feature vectors. The feature vectors of the remaining dimensions are split by channels. The channel dimensions are concatenated to restore the feature vector, resulting in a feature vector with the same dimensions as before segmentation, as shown below: (twenty one) in, This is the feature vector after concatenation and restoration along the channel dimension, with a size of 128×H / 2×W / 2. This is an operation for concatenating vectors along the channel dimension.
[0057] Next, the feature vectors after channel dimension restoration are... The input is fed into two consecutive convolutional layers and then combined with the feature vector input from this module. The feature vector obtained after restoring the high-frequency feature information is shown below: (twenty two) in, The feature vector is 128×H / 2×W / 2 after the high-frequency feature information is restored. conv18 and conv19 are convolution operations of size 1×1 and 3×3, respectively, with 128 input and output channels.
[0058] Finally, the feature vectors restored from the high-frequency feature information Upsampling is performed to restore the spatial resolution lost by the multi-view processing downsampling module, as shown below: (twenty three) in, This module outputs a high-dimensional feature vector of size 128×H×W after high-frequency feature restoration. The operation is applied at twice the original value.
[0059] (4) Alternately execute the multi-view processing downsampling module and the channel segmentation attention upsampling module. By operating the above two modules, the goal of effectively suppressing speckle noise without sacrificing spatial resolution and high-frequency details is achieved. Next, the two modules will be processed alternately twice to fully and accurately accomplish the above objective, as shown below: (twenty four) (25) in, The two modules are the feature vectors processed alternately by the first and second multi-view processing downsampling modules and the channel segmentation attention upsampling module, respectively. MPDB2 and MPDB3 are multi-view processing downsampling modules, and their detailed structures are the same as MPDB1, so they will not be described in detail here. CSAB2 and CSAB3 are channel segmentation attention upsampling modules, and their detailed structures are the same as CSAB1, so they will not be described in detail here either.
[0060] (5) Super-resolution reconstruction By using alternating multi-view processing downsampling modules and channel segmentation attention upsampling modules, high-dimensional feature vectors are obtained that effectively suppress coherent noise without losing their spatial resolution and high-frequency feature information. Finally, it is fed into a separate channel segmentation attention upsampling module to restore the spatial resolution required for the super-resolution reconstruction of this method, and then mapped back from the vector domain to the image domain through a convolutional layer, as shown below: (26) in, The image is a SAR image reconstructed by super-resolution at 2x. conv20 is a 3×3 convolution operation with 128 input channels and 1 output channel. CSAB4 is a channel segmentation attention upsampling module, whose detailed structure is the same as CSAB1, and will not be described in detail here.
[0061] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the SAR image super-resolution reconstruction results of the present invention. Figure 4 (a) in the image is the input low-resolution remote sensing image; Figure 4 Image (b) shows the result of super-resolution reconstruction at 4x. It can be seen that after super-resolution reconstruction, the present invention can effectively solve the problem of information loss caused by super-resolution reconstruction after speckle noise suppression, and realize the restoration of high-resolution SAR images with low computational complexity and high fidelity.
[0062] Please see Figure 5 , Figure 5This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a SAR image super-resolution reconstruction device 401 based on multi-view processing and channel segmentation attention, a processor 402, and a storage device 403.
[0063] A SAR image super-resolution reconstruction device 401 based on multi-view processing and channel segmentation attention: The SAR image super-resolution reconstruction device 401 based on multi-view processing and channel segmentation attention implements the SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention.
[0064] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention.
[0065] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention.
[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A SAR image super-resolution reconstruction method based on multi-look processing and channel segmentation attention, characterized in that: The method includes the following steps: S1. Acquire low-resolution SAR images; S2. Construct a SAR image super-resolution reconstruction model; the SAR image super-resolution reconstruction model includes: multiple sets of alternating connected multi-look processing downsampling modules and channel segmentation attention upsampling modules; S3. Input the low-resolution SAR image into a shallow feature extraction module to convert the low-resolution SAR image into a high-dimensional vector domain and obtain shallow high-dimensional feature vectors. S4. Input the shallow high-dimensional feature vector into the SAR image super-resolution reconstruction model to obtain the final high-dimensional feature vector that fully suppresses coherent noise and does not lose its own spatial resolution information and high-frequency feature information. S5. Input the final high-dimensional feature vector from step S4 into a separate channel segmentation attention upsampling module, and then pass it through a convolutional layer to map it from the vector domain back to the image domain to obtain the reconstructed high-resolution SAR image.
2. The SAR image super-resolution reconstruction method based on multi-look processing and channel segmentation attention as described in claim 1, characterized in that: Step S3 is as follows: in, , The two vectors are a low-resolution SAR image of size 1×H×W and an extracted shallow high-dimensional feature vector of size 128×H×W, respectively, with conv1 being the convolution operation.
3. The SAR image super-resolution reconstruction method based on multi-look processing and channel segmentation attention as described in claim 2, characterized in that: In step S4, after the shallow high-dimensional feature vector is input into the SAR image super-resolution reconstruction model, it first enters the multi-look processing downsampling module; the processing procedure of the multi-look processing downsampling module is as follows: S41. Transform shallow high-dimensional feature vectors Perform layer normalization to obtain the layer-normalized feature vector: in, The feature vectors are normalized by the layer, and LayerNorm is the layer normalization operation; S42. Perform stride convolution operations of different sizes on the normalized feature vectors of the layers to achieve speckle noise suppression: in, These are high-dimensional feature vectors of size 128×H / 2×W / 2, which are processed using convolutional kernels conv2, conv3, and conv4 respectively for spatial multi-view processing. S43, to Feature concatenation is performed to obtain a high-dimensional feature vector after information fusion processing: in, This refers to the high-dimensional feature vector obtained after information fusion processing of multiple high-dimensional feature vectors obtained by different multi-view processing strategies. This is an operation that concatenates feature vectors along the channel dimension; S43, will The input is fed into a 1×1 convolutional kernel, and different multi-view processing is fully fused to obtain a fully fused high-dimensional feature vector: in For the fully fused high-dimensional feature vector, conv6 is a 1×1 convolution operation; S44. Input the fully fused high-dimensional feature vector into three convolutional layers with different convolutional sizes to perform multi-scale convolutional feature extraction: in, These are high-dimensional feature vectors of different scales extracted using conv7, conv8, and conv9, respectively, where conv7, conv8, and conv9 are convolutional layers of different sizes. S45. Input high-dimensional feature vectors of different scales into a convolution operation and activation function to extract high-frequency feature information and obtain high-frequency feature residual vectors: in, The high-frequency feature residual vectors are extracted from convolution operations at multiple scales. To perform feature vector concatenation along the channel dimension, ReLU is the torch.nn.reLU activation function in PyTorch. conv 10 represents the convolution operation; S46. Transform the high-frequency feature residual vector M 10 After a convolution operation and M Adding these 5 together yields the high-dimensional feature vector output by the multi-view processing downsampling module. F 2.
4. The SAR image super-resolution reconstruction method based on multi-look processing and channel segmentation attention as described in claim 3, characterized in that: In step S4, the output of the multi-view processing downsampling module is further input to the channel segmentation attention upsampling module. The processing procedure of the channel segmentation attention upsampling module is as follows: S47, Yes F After layer normalization, the input is fed into a convolutional layer and activation function, as shown in the following formula: in, For the initial reconstruction of the high-dimensional feature vector, LayerNorm is the layer normalization operation, relu is the torch.nn.relu activation function in PyTorch, and conv12 is the convolution operation; S48. After passing C1 through a convolutional layer, perform vector segmentation in the channel dimension to perform subsequent channel attention operations, as shown in the following equation: in, The attention requires the feature vectors of the corresponding channels, the remaining feature vectors of the channel segmentation with a size of (128-32)×H / 2×W / 2, and conv13 is the convolution operation; The channel segmentation operation splits the input high-dimensional feature vector into a 32-dimensional feature vector and a feature vector with the remaining dimension in the channel dimension. S49, will C 2. The input is fed into a global average pooling layer to integrate all spatial information within the channel, as shown in the following formula: in, This is the feature vector after the global pooling layer operation. This is a global average pooling layer operation; S410, to We use two branches to construct attention-weighted convolutional kernels that focus on features at different scale levels, as shown below: in, Convolutional kernels with attention weights that focus on features at different scales are constructed respectively. ReLU is the torch.nn.reLU activation function in PyTorch. conv14, conv15, conv16 and conv17 are all convolution operations. reshape1 and reshape2 are vector size adjustment operations. S411. Use the two attention-weighted convolutional kernels obtained above, which focus on features at different scales, to process the high-dimensional feature vector. Perform a convolution operation, then sum the results of the convolutions to obtain the feature vector after channel segmentation and attention processing, as shown below: in, These are the feature vectors processed by two attention-weighted convolutional kernels that focus on features at different scales. This is a convolution operation; S412, will and The channel dimensions are concatenated to restore the feature vector, resulting in a feature vector with the same dimensions as before segmentation, as shown below: in, This is the feature vector after concatenation and restoration along the channel dimension. This is an operation for concatenating vectors along the channel dimension; S413, will The input is fed into two consecutive convolutional layers and then added to the input feature vector. The feature vector obtained after restoring the high-frequency feature information is shown below: in, conv18 and conv19 are the feature vectors after restoring high-frequency feature information, and convolution operations are respectively. S414, to Upsampling is performed to restore the spatial resolution lost by the multi-view processing downsampling module, as shown in the following equation: in, The high-dimensional feature vector after restoring the high-frequency features output by the channel segmentation attention upsampling module, i.e. The operation is applied at twice the original value.
5. The SAR image super-resolution reconstruction method based on multi-look processing and channel segmentation attention as described in claim 4, characterized in that: Since the SAR image super-resolution reconstruction model in step S2 includes multiple sets of alternately connected multi-look processing downsampling modules and channel segmentation attention upsampling modules, steps S41 to S414 are executed multiple times, except that the input to step S41 in the first pass is a shallow high-dimensional feature vector. In addition, the input to the remaining steps S41 is the output of the previous channel segmentation attention upsampling module, and the high-dimensional feature vector after high-frequency feature restoration is output by the channel segmentation attention upsampling module in the last pass, which is the final high-dimensional feature vector.
6. The SAR image super-resolution reconstruction method based on multi-look processing and channel segmentation attention as described in claim 4, characterized in that: Step S5 is as follows: in, For the reconstructed high-resolution SAR image, CSABfinal is the last group of channel segmentation attention upsampling modules in the SAR image super-resolution reconstruction model. F final-1 conv20 is the output of the penultimate channel segmentation attention upsampling module in the SAR image super-resolution reconstruction model; conv20 is the convolution operation.
7. A storage device, characterized in that: The storage device stores instructions and data for implementing the SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention as described in any one of claims 1 to 5.
8. A SAR image super-resolution reconstruction device based on multi-look processing and channel segmentation attention, characterized in that: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the SAR image super-resolution reconstruction method based on multi-view processing and channel segmentation attention as described in any one of claims 1 to 5.