Remote sensing image super-resolution method and system based on residual separation attention mechanism and medium

Abstract

This application relates to an image super-resolution method, system, and medium based on a residual separation attention mechanism. The method includes: downsampling a high-resolution remote sensing satellite image to a target low-resolution remote sensing satellite image; extracting a coarse feature map; constructing an upsampling module; constructing a residual separation attention module; constructing a residual separation attention group to further extract fine features; fusing the coarse feature map as the output of the global residual learning and the final residual attention group; finally, inputting the fine remote sensing satellite feature maps obtained at each stage into a multi-level feature fusion module for fusion; and performing ordinary convolution on the fused final fine remote sensing satellite feature map to reconstruct the target high-resolution remote sensing satellite image. The network proposed in this invention outperforms other state-of-the-art remote sensing satellite image super-resolution algorithms, generating higher-quality satellite images, especially at high sampling rates.

Description

Technical Field

[0001] This application belongs to the field of computer vision remote sensing image super-resolution technology, specifically relating to a remote sensing image super-resolution method, system, and medium based on residual separation attention mechanism. Background Technology

[0002] Remote sensing satellite imagery, as a crucial means of Earth observation, is widely used in various fields, including mineral resource development, environmental monitoring, public safety, and military applications. Spatial resolution is considered a key indicator of satellite image quality. However, spaceborne imaging systems are often affected by complex imaging environments, resulting in images with relatively low spatial resolution. Therefore, improving the spatial resolution and image quality of remote sensing satellite images is an important and challenging task in the field of remote sensing image processing.

[0003] Given the unique value of high-quality remote sensing satellite imagery in various application scenarios, Single-Image Super-Resolution (SISR) methods have attracted widespread attention in remote sensing satellite image processing in recent years. Besides leveraging more advanced hardware to improve image quality, Super-Resolution (SR) technology also provides a convenient and effective strategy for enhancing image quality. With the rapid development of satellite photogrammetry technology, the demand for developing efficient and high-precision remote sensing satellite image super-resolution (SR) methods is increasing. To obtain high-quality images, many researchers have adopted convolutional neural network-based techniques to learn the mapping function between low-resolution (LR) and high-resolution (HR) images and have established various SR models, effectively improving image resolution and quality. Liebel et al. first applied Single Image Super-Resolution Using Deep Learning (SRCNN) to satellite image super-resolution. Considering that satellite image SR methods cannot be directly trained using natural images, they used SENTINEL-2 images to generate a remote sensing dataset to relearn the mapping relationship. Lei et al. introduced a multi-branch structural framework (Super-resolution for remote sensing images via a local–global combined network) to learn the network's multi-scale representation capabilities. This framework combines shallow and deep feature maps to facilitate interaction between feature information, thereby better guiding network reconstruction. Qin et al. designed a multi-scale network (Remote sensing image super-resolution using a multi-scale convolutional neural network) based on GoogLeNet. This network uses multi-scale kernels to extract image features and obtains more comprehensive deep features after connecting the features of each channel, thus improving SR performance.

[0004] While deep learning-based super-resolution methods for remote sensing satellite images have achieved some satisfactory results, most current research focuses on improving remote sensing image resolution at smaller sampling factors, primarily by adding an upsampling layer to the network. However, research on reconstruction at high sampling factors remains limited. For example, Pan et al. utilized a backprojection strategy to more comprehensively address the dependency between LR and HR (Super-resolution of a single remote sensing image based on residual dense backprojection networks, RDBPN). Dong et al. proposed a dense sampling framework (Remote sensing image super-resolution using novel dense-sampling networks, DSSR), which reuses upsampling mechanisms to upsample low-dimensional features to study large-scale factor SR reconstruction. Although existing RDBPN and DSSR methods attempt to solve remote sensing satellite image SR at high sampling factors (e.g., ×8) by densely connecting features through upsampling modules, these methods all use the same weight size to learn remote sensing feature maps of different regions, lacking the ability to represent features across feature channels. Furthermore, due to the limited prior information in LR images, predicting local features becomes more challenging under high sampling rates. Additionally, the extra prior information translates to higher computational costs and increased training difficulty, further negatively impacting the subsequent reconstruction process. Therefore, extracting reliable and accurate fine-grained features under high sampling rates and enhancing the expressive power of cross-channel features are particularly important. Summary of the Invention

[0005] The purpose of this application is to provide an image super-resolution method, system, and medium based on a residual separation attention mechanism, thereby solving the technical problem that existing remote sensing satellite super-resolution reconstruction algorithms have certain limitations in network reconstruction performance and image quality under high sampling multiple conditions.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] In a first aspect, embodiments of this application provide an image super-resolution method based on a residual separation attention mechanism, including:

[0008] S1: Downsample the high-resolution remote sensing satellite image to the target low-resolution remote sensing satellite image;

[0009] S2: The target low-resolution remote sensing image is divided into blocks. After separating overlapping image blocks, the coarse feature extraction module performs a depth separation convolution operation on the low-resolution remote sensing image to extract the coarse feature map F. C ;

[0010] S3: Construct an upsampling module and introduce methods to remove artifacts to help the network reduce excessive artifacts in remote sensing satellite images.

[0011] S4: Construct a residual separation attention module to enable the network to generate local attention and enhance the interaction of global structural information of remote sensing satellite images at the channel level;

[0012] S5: Construct a residual separation attention group to further extract fine features, and then convert the extracted coarse feature map F... C The input residual separation attention group obtains a fine remote sensing feature map. The residual separation attention network contains several residual separation attention groups. Each group contains an upsampling module and a residual separation attention module. The upsampling module considers hierarchical prior information, assists the network in the high-resolution feature space, and predicts high-frequency local detail information. The residual separation attention module is used to adaptively explore and learn global structural information in the low-resolution feature space.

[0013] S6: Transfer the coarse feature map F C The output of the global residual learning and the last residual attention group are fused together, and finally the fine remote sensing satellite feature maps obtained from each stage are input into the multi-level feature fusion module for fusion.

[0014] S7: Perform ordinary convolution on the final refined remote sensing satellite feature map obtained by fusion to reconstruct a high-resolution remote sensing satellite image of the target.

[0015] In some alternative implementations, in step S2, a coarse feature extractor is constructed using an inverse residual module and a 3x3 convolutional layer, and the coarse feature extractor is used to extract a coarse feature map F. C The inverse residual block consists of a 1*1 point convolutional layer Pwconv, a 3*3 depthwise separable convolutional layer Dwconv, a batch normalization (BN) layer, and an activation function PReLU. The coarse feature map F_C is represented as: F C =H C (I LR ), where H C (·) represents the coarse feature extraction operation, I LR This represents the input low-resolution remote sensing satellite image.

[0016] In some alternative implementations, step S3 includes:

[0017] The coarse feature map FC Input an 8x8 deconvolutional layer Deconv and an inverse residual module to obtain a low-resolution feature map L1, followed by a coarse feature map F. C After element-wise subtraction with the low-resolution feature map L1, the image is input again into an 8*8 deconvolution layer Deconv. The high-resolution feature maps obtained from the two deconvolution layers are then fused to obtain the final upsampling module output.

[0018] In some alternative implementations, step S4 includes:

[0019] The obtained high-resolution fused feature map is input into a 3*3 depthwise separable convolutional layer, and then divided into n splits through each channel. The number of output channels is set to c. The value of each part's output channel is... Then, these n parts are merged using an element-wise adder: F Splits =H Sum (Split1, Split2,...,Split n Split n H represents the nth split after the partitioning operation. Sum This indicates an element-wise summation operation, followed by F. Splits After passing through an adaptive average pooling layer and two point convolutional layers, the data is further divided into n splits by a softmax function. Each current split is multiplied by the previous split. Finally, an element-wise adder is used to fuse the multiplied features as the output of the residual separation attention module. The residual separation attention is represented as: F RSA =H Sum (H Ep (Split1, Split′1),...,H Ep (Split n Split′ n H Sum H represents the second element-wise addition operation. Ep This indicates element-wise multiplication.

[0020] In some optional implementations, the residual separation attention group comprises: each residual separation attention group includes an upsampling module and a residual separation attention module, which extracts the deep feature map of the m-th residual separation attention group. This can be represented as: F1 = H RSAG，m (F C m = 1, m = 2, ..., M, where H RSAG，m(·) represents the m-th residual separation attention group operation. The number of residual separation attention groups is determined by the ablation experiment, and the optimal value is m = 10.

[0021] In some optional implementations, the multi-level feature fusion module is configured to: input high-resolution feature maps output from different residual separation attention groups into the multi-level feature fusion module for feature fusion. Where U1 = H Up，1 (F C ), H Up，1 This represents the first upsampling operation, and U1 is the output of the first upsampling operation. U2 represents the output of the second upsampling operation. m = 3, ..., M, where This is the feature fusion of the outputs of the (m-1)th and (m-2)th residual attention modules, where concat is the feature fusion operation.

[0022] In some alternative implementations, the high-resolution remote sensing satellite image of the target is represented as: H Rec The reconstruction operation is represented by a 3x3 convolution, I SR A high-resolution remote sensing satellite image representing the target.

[0023] In some alternative implementations, the loss function L for the entire network H Represented as: L H =λ1L pix +λ2L per +λ3L bec Among them, the weighting coefficients λ1, λ2, and λ3 are used to balance the loss. For pixel domain loss, and This represents the i-th low-resolution remote sensing satellite image and the i-th high-resolution remote sensing satellite image in the dataset; For the loss of the receptive domain, f VGG 54 Here, HRSAN is the feature extraction function of the VGG network, and HRSAN is the output of the residual separation attention network. For binary cross-entropy loss, This represents the result of the output of the residual separation attention network after Sigmoid activation.

[0024] According to another aspect of the present invention, a residual separation attention mechanism is provided.

[0025] Remote sensing satellite super-resolution systems include:

[0026] The downsampling module downsamples high-resolution remote sensing satellite images to the target low-resolution remote sensing satellite image;

[0027] The coarse feature extractor module divides the low-resolution remote sensing satellite image of the target into blocks. After dividing the overlapping image blocks, the inverse residual module is used to extract coarse remote sensing feature maps.

[0028] The residual separation attention group is used to construct the residual separation attention network as a fine feature extractor. The coarse remote sensing image feature map is input into the residual separation attention group to obtain a fine remote sensing image feature map. The residual separation attention network contains several residual separation attention groups, each containing an upsampling module and a residual separation attention module. The upsampling module considers hierarchical prior information, assists the network in the high-resolution feature space, and predicts high-frequency local detail information. The residual separation attention module is used to adaptively explore and learn global structural information in the low-resolution feature space.

[0029] The multi-level feature fusion module gradually fuses feature maps from different levels through dense connections, enabling local attention to interact across feature groups and achieving global attention in the network, thus obtaining more refined remote sensing feature maps.

[0030] The remote sensing image reconstruction module is used to perform convolution operations on the feature-fused remote sensing satellite feature map to reconstruct a high-resolution remote sensing satellite image of the target.

[0031] According to another aspect of the present invention, a computer-readable storage medium is provided thereon storing a computer program that, when executed by a processor, implements the image super-resolution method based on residual separation attention mechanism as described above.

[0032] Compared with the prior art, the beneficial effects of the present invention are: the remote sensing satellite super-resolution method based on the residual split attention mechanism (RSAM) can effectively solve the problems of insufficient texture details and inaccurate structural information in the SR method of single remote sensing satellite images, especially under high sampling rate conditions.

[0033] Furthermore, a residual separation attention mechanism is proposed to help the network focus on regions rich in detail, thereby reducing attention to parts with insufficient information. The network's overall expressive power is enhanced by separating and fusing intermediate residual feature maps from different channel dimensions. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating an image super-resolution method based on a residual separation attention mechanism provided in an embodiment of the present invention.

[0036] Figure 2 This is a network framework diagram of an image super-resolution network based on a residual separation attention mechanism provided in an embodiment of the present invention;

[0037] Figure 3 This is a network framework diagram based on residual separation attention groups provided in an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of a device structure provided in an embodiment of the present invention.

[0039] Figure 5 This is a result image (under 4x downsampling conditions) on the RSCN7 public dataset provided by an embodiment of the present invention. The last two rows of images show the error distribution between the generated image and the original image.

[0040] Figure 6 This is a result image (under 8x downsampling conditions) on the RSCN7 public dataset provided by an embodiment of the present invention. The last two rows of images show the error distribution between the generated image and the original image. Detailed Implementation

[0041] The technical solutions of the embodiments of this application will now be described with reference to the accompanying drawings. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0042] The terms “comprising,” “including,” or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0043] The terms “first,” “second,” etc., are used only to distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance, nor as requiring or implying any such actual relationship or order between these entities or operations.

[0044] The image super-resolution method based on residual separation attention mechanism in this invention embodiment, such as... Figure 1 As shown, it includes the following steps:

[0045] S1: Downsample the high-resolution remote sensing satellite image to the target low-resolution remote sensing satellite image;

[0046] In this embodiment of the invention, bicubic interpolation can be used to downsample a high-resolution remote sensing satellite image to a target low-resolution remote sensing satellite image.

[0047] In this embodiment of the invention, the RSCN7 dataset is preferably used, which contains a total of 2800 remote sensing images from 7 scene categories, with 400 images in each category. 2100 and 700 images were selected as the training and test sets, respectively. Each image in the dataset has a pixel size of 400×400. This invention uses a bicubic degradation model to downsample the dataset, with downsampling factors of 4 and 8. Therefore, the sizes of the downsampled low-resolution images are 100×100 and 50×50 pixels, respectively.

[0048] S2: Construct a coarse feature extractor to divide the target low-resolution image into blocks.

[0049] After separating overlapping image patches, a coarse feature extractor is used to extract coarse remote sensing satellite feature maps.

[0050] In this embodiment of the invention, a coarse feature extractor can be constructed using an inverse residual module and a 3*3 convolutional layer, wherein the coarse remote sensing satellite feature map is represented as follows:

[0051] F C =H C (I LR )

[0052] The inverse residual block consists of a 1*1 point convolutional layer Pwconv, a 3*3 depthwise separable convolutional layer Dwconv, a batch normalized (BN) layer, and an activation function PReLU. C (·) represents the coarse feature extraction operation, I LR This represents the input low-resolution remote sensing satellite image.

[0053] S3: Construct an upsampling module and introduce methods to remove artifacts to help the network reduce excessive artifacts from remote sensing satellite images;

[0054] In this embodiment of the invention, the upsampling module's workflow consists of the following steps:

[0055] The obtained coarse remote sensing satellite feature map F C Input an 8x8 deconvolutional layer Deconv and an inverse residual module to obtain a low-resolution feature map L1, followed by a coarse feature map F. C After element-wise subtraction with the low-resolution feature map L1, it is input again into the 8*8 deconvolution layer Deconv. The high-resolution feature maps obtained from the two deconvolution layers are fused to obtain the final output of the upsampling module.

[0056] S4: Construct a Residual Split Attention Module to enhance the interaction of global structural information of remote sensing satellite images at the channel level by increasing the local attention of the network;

[0057] In this embodiment of the invention, the residual separation attention module's workflow consists of the following steps: The high-resolution fused feature map obtained from the upsampling module is input into a 3*3 depthwise separable convolutional layer, and then divided into n splits through each channel. The number of output channels is set to c, and the value of each part's output channel is... Then, these n parts are merged using an element-wise adder:

[0058] F Splits =H Sum (Split1, Split2,...,Split n )

[0059] Among them, Split n H represents the nth split after the partitioning operation. Sum This indicates an element-wise summation operation, followed by F. Splits After passing through an adaptive average pooling layer and two point convolutional layers, the data is further divided into n splits by a softmax function. Each current split is multiplied by the previous split. Finally, an element-wise adder is used to fuse the multiplied features as the output of the residual separation attention module. The residual separation attention is represented as:

[0060] F RSA =H Sum (H Ep (Split1, Split′1),...,H Ep (Splitn Split′ n ))

[0061] Among them, H Sum H represents the second element-wise addition operation. Ep This indicates element-wise multiplication.

[0062] S5: Construct a residual separation attention group to further extract fine features and obtain refined remote sensing data.

[0063] Sensing satellite feature map;

[0064] In this embodiment of the invention, the residual separation attention group workflow consists of the following steps:

[0065] The extracted coarse remote sensing feature map is input into the residual separation attention group to obtain a refined remote sensing feature map. The residual separation attention network comprises several residual separation attention groups, each containing an upsampling module and a residual separation attention module. The upsampling module considers hierarchical prior information, assists the network in the high-resolution feature space, and predicts high-frequency local detail information. The residual separation attention module is used to adaptively explore and learn global structural information in the low-resolution feature space. The deep feature map extracted from the m-th residual separation attention group... Represented as:

[0066] F1 = H RSAG，m (F C m = 1, m = 2, ..., M,

[0067] Where H RSAG，m (·) represents the m-th residual separation attention group operation, residual separation attention

[0068] The number of force groups is determined by the ablation experiment, with the optimal value being m = 10.

[0069] S6: Construct a multi-level feature fusion module to fuse high-resolution feature maps output by different residual separation attention groups:

[0070]

[0071] Where U1 = H Up，1 (F C ), H Up，1 This represents the first upsampling operation, and U1 is the output of the first upsampling operation. U2 represents the output of the second upsampling operation. m = 3, ..., M, where This is the feature fusion of the outputs of the (m-1)th and (m-2)th residual attention modules, where concat is the feature fusion operation.

[0072] S7: Construct a remote sensing satellite image reconstruction module, which performs ordinary convolution on the final refined remote sensing satellite feature map obtained by fusion to reconstruct a high-resolution remote sensing satellite image of the target.

[0073] In this embodiment of the invention, the remote sensing satellite image reconstruction module is represented as follows:

[0074]

[0075] Among them, H Rec The reconstruction operation is represented by a 3x3 convolution, I SR A high-resolution remote sensing satellite image representing the target.

[0076] S8: Loss function L of the entire network H Represented as:

[0077] L H =λ1L pix +λ2L per +λ3L bec

[0078] Among them, the weighting coefficients λ1, λ2, and λ3 are used to balance the loss. For pixel domain loss, and This represents the i-th low-resolution remote sensing satellite image and the i-th high-resolution remote sensing satellite image in the dataset; For the loss of the receptive domain, f VGG 54 H is the feature extraction function for VGG networks. RSAN This is the output of the residual separation attention network; For binary cross-entropy loss, This represents the result of the output of the residual separation attention network after Sigmoid activation.

[0079] This invention also provides a remote sensing satellite super-resolution method based on a residual separation attention mechanism.

[0080] Rate systems, such as Figure 3 As shown, it includes:

[0081] The downsampling module 101 downsamples the high-resolution remote sensing satellite image to the target low-resolution remote sensing satellite image;

[0082] The coarse feature extractor module 102 divides the target low-resolution remote sensing satellite image into blocks, and after dividing the overlapping image blocks, it uses the inverse residual module to extract coarse remote sensing feature maps.

[0083] The residual separation attention group 103 is used to construct a residual separation attention network as a fine feature extractor. The coarse remote sensing image feature map is input into the residual separation attention group to obtain a fine remote sensing image feature map. The residual separation attention network contains several residual separation attention groups, each containing an upsampling module and a residual separation attention module. The upsampling module considers hierarchical prior information, assists the network in the high-resolution feature space, and predicts high-frequency local detail information. The residual separation attention module is used to adaptively explore and learn global structural information in the low-resolution feature space.

[0084] The multi-level feature fusion module 104 gradually fuses feature maps from different levels through dense connections, enabling local attention to interact across feature groups, achieving global attention in the network, and obtaining more refined remote sensing feature maps.

[0085] The remote sensing image reconstruction module 105 is used to perform convolution operations on the feature-fused remote sensing satellite feature map to reconstruct a high-resolution remote sensing satellite image of the target.

[0086] The specific implementation methods of each module can be referred to the description of the above method embodiments, and the embodiments of the present invention will not be repeated.

[0087] The present invention also provides a computer storage medium storing a computer program executable by a computer processor, the computer program performing the above-described image super-resolution method based on residual separation attention mechanism.

[0088] Finally, this invention provides a test example. The invention uses the RSCN7 dataset, which contains a total of 2800 remote sensing images from 7 scene categories, with 400 images per category. 2100 and 700 images were selected as the training and test sets, respectively. Each image in the dataset has a pixel size of 400×400. This invention uses a bicubic degradation model to downsample the dataset, with downsampling factors of 4 and 8. Therefore, the sizes of the downsampled low-resolution images are 100×100 and 50×50 pixels, respectively. It should be noted that all training and testing are based on the luminance channel in the YCbCr color space, and sampling factors of 4x and 8x are used for training and testing. Four evaluation metrics—Peak signal-to-noise ratio (PSNR), Structural similarity (SSIM), Visual Information Fidelity (VIF), and Erreur Relative Globale Adimensionnelle de Synthèse (ERGAS)—were used to evaluate the SR reconstruction results and test the performance of SR reconstruction under the luminance channel. In each iteration of training, 16 32×32 LR blocks were extracted as input. The model of this invention was trained using the Adam optimizer, β... 1 =0.9, β 2 =0.999, and ∈ =10 -8 Learning rate initialization □ = 10 -4 The size is reduced to one-tenth of the original size every 500 cycles. The model of this invention was trained using the PyTorch platform and a GTX 3090 graphics card. Table 1 shows the comparison results under the conditions of reconstruction magnification of 4 and 8 using the above four evaluation metrics. Figure 5 Figure 6 These are comparison images showing the reconstruction results of 4x and 8x remote sensing satellite images. D374 and E327 represent the original high-resolution remote sensing satellite baseline images.

[0089] The remote sensing satellite image SR methods selected for comparison include: Bicubic, EDSR, DBPN, RDN, RCAN, SAN, and MHAN. Bicubic is a classic image interpolation algorithm; EDSR is a classic deep residual network SR method; DBPN is an SR method based on a deep backprojection network; RDN is an SR method based on a deep residual dense connection network; RCAN is an SR method based on a deep residual channel attention network; SAN is an SR method based on a second-order attention network; and MHAN is an SR method based on a hybrid high-order attention mechanism. Figure 5 D374 is the original high-resolution remote sensing satellite baseline image. Figure 6 E327 is the original high-resolution remote sensing satellite reference image, and RSAN is the high-resolution remote sensing satellite image generated by this invention. It can be seen from this that this invention has achieved the best visual effect in the visual results.

[0090] Table 1 Comparison results of the present invention with seven excellent algorithms.

[0091] method Scale Param / M PSNR↑ SSIM↑ VIF↑ ERGAS↓ Bicubic ×4 - 30.62 0.7899 0.4490 1.4738 EDSR ×4 40.09 32.71 0.8471 0.5064 1.1786 DBPN ×4 10.43 32.73 0.8480 0.5072 1.1756 RDN ×4 1.15 32.21 0.8341 0.4903 1.2488 RCAN ×4 15.59 32.78 0.8487 0.5091 1.1724 SAN ×4 15.82 32.76 0.8483 0.5083 1.1717 MHAN ×4 11.35 32.65 0.8454 0.5050 1.1894 This invention ×4 8.21 32.87 0.8494 0.5128 1.1619 Bicubic ×8 - 25.77 0.5762 0.2289 3.0669 EDSR ×8 83.22 28.72 0.7158 0.3184 1.8114 DBPN ×8 23.21 28.74 0.7189 0.3195 1.8028 RDN ×8 1.29 28.50 0.7073 0.3079 1.8412 RCAN ×8 15.74 28.73 0.7162 0.3119 1.8032 SAN ×8 15.97 28.77 0.7189 0.3170 1.7957 MHAN ×8 11.51 28.56 0.7131 0.3179 1.8447 This invention ×8 9.72 28.84 0.7206 0.3247 1.7832

[0092] As can be seen from the experimental results in the table above, this invention has significant advantages over the other seven methods.

[0093] All parts not described in this specification are prior art or common knowledge. This embodiment is only for illustrating the invention and is not intended to limit the scope of the invention. Equivalent substitutions and modifications made to the invention by those skilled in the art are considered to fall within the protection scope of the claims.

[0094] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A remote sensing image super-resolution method based on residual separation attention mechanism, characterized in that, include: S1: Downsample the high-resolution remote sensing satellite image to the target low-resolution remote sensing satellite image; S2: The target low-resolution remote sensing satellite image is divided into blocks. After separating overlapping image blocks, the coarse feature extraction module performs a depthwise separation convolution operation on the low-resolution remote sensing satellite image to extract coarse feature maps. ; S3: Construct an upsampling module and introduce methods to remove artifacts to help the network reduce excessive artifacts in remote sensing satellite images; S4: Construct a residual separation attention module to enable the network to generate local attention and enhance the interaction of global structural information of remote sensing satellite images at the channel level; S5: Construct a residual separation attention group to further extract fine features, and then convert the extracted coarse feature map... The input residual separation attention group is used to obtain a fine remote sensing feature map. The residual separation attention network contains several residual separation attention groups. Each group contains an upsampling module and a residual separation attention module. The upsampling module considers hierarchical prior information, assists the network in the high-resolution feature space and predicts high-frequency local detail information. The residual separation attention module is used to adaptively explore and learn global structural information in the low-resolution feature space. S6: Transform the coarse feature map The output of the global residual learning and the last residual attention group are fused together, and finally the fine remote sensing satellite feature maps obtained from each stage are input into the multi-level feature fusion module for fusion. S7: Perform ordinary convolution on the final refined remote sensing satellite feature map obtained by fusion to reconstruct a high-resolution remote sensing satellite image of the target; Step S4 includes: inputting the high-resolution fused feature map obtained from the above operations into a 3*3 depth separable convolutional layer, and then segmenting it through each channel into... indivual And the number of output channels is set to Then the value of the output channel of each part is Then, this is processed by an element-wise adder. The parts are merged: Represents the first after the partitioning operation indivual , This indicates an element-wise summation operation, followed by After passing through an adaptive average pooling layer and two point convolutional layers, and then through a... The function is split into indivual Each current Multiply by the corresponding previous ones respectively Finally, an element-wise adder is used to fuse the multiplied features as the output of the residual separation attention module. The residual separation attention is represented as: , This indicates the second element-wise addition operation. This indicates element-wise multiplication. Step S5 includes: for a residual separation attention network, which contains several residual separation attention groups, for the first... Deep feature maps extracted by group residual separation attention groups Represented as: in Indicates the first The number of residual separation attention groups is determined by the ablation experiment; the optimal value is [value missing]. ; The high-resolution feature maps output from the different residual separation attention groups obtained in step S5 are input into a multi-level feature fusion module for feature fusion. ,in , This indicates the first upsampling operation. This is the output of the first upsampling operation. , This indicates the output result of the second upsampling operation. ,in For the first The and the first Feature fusion of the outputs of each residual attention module This is a feature fusion operation.

2. The remote sensing image super-resolution method based on residual separation attention mechanism according to claim 1, characterized in that, In step S2, a coarse feature extractor is constructed using an inverse residual module and a 3*3 convolutional layer, and the coarse feature extractor is used to extract a coarse feature map. The inverse residual block consists of a 1*1 point convolutional layer. 3x3 depth separable convolutional layer Batch normalization Layers, activation functions Composition, the coarse feature map Represented as: ,in It is a coarse feature extraction operation. This represents the input low-resolution remote sensing satellite image.

3. The remote sensing image super-resolution method based on residual separation attention mechanism according to claim 1, characterized in that, Step S3 includes: The coarse feature map obtained from the above operations Input 8x8 deconvolution layer A reverse residual module to obtain low-resolution feature maps Subsequently, coarse feature map Compared with low-resolution feature maps After performing element-wise subtraction, input the data into an 8x8 deconvolution layer again. The high-resolution feature maps obtained from the two deconvolution layers are fused to obtain the final upsampling module output.

4. The remote sensing image super-resolution method based on residual separation attention mechanism according to claim 1, characterized in that, The high-resolution remote sensing satellite image of the target is represented as follows: , The reconstruction operation is implemented using a 3x3 convolution. A high-resolution remote sensing satellite image representing the target.

5. A remote sensing image super-resolution method based on residual separation attention mechanism according to any one of claims 1 to 4, characterized in that, Loss function of the entire network Represented as: Among them, the weighting coefficient Used to balance losses For pixel domain loss, and Represents the first in the dataset Zhang low-resolution remote sensing satellite image and the first Zhang high-resolution remote sensing satellite image; For the loss of the perceptual domain, This is the feature extraction function for the VGG network. This is the output of the residual separation attention network; For binary cross-entropy loss, This represents the result of the output of the residual separation attention network after Sigmoid activation.

6. A remote sensing image super-resolution system based on a residual separation attention mechanism, used to implement the method of claim 1, characterized in that, include: The downsampling module downsamples high-resolution remote sensing satellite images to the target low-resolution remote sensing satellite image; The coarse feature extractor module divides the low-resolution remote sensing satellite image of the target into blocks. After separating the overlapping image blocks, the inverse residual module is used to extract coarse feature maps. ; Residual separation attention group is used to construct a residual separation attention network as a fine feature extractor, which extracts the coarse feature map. The residual separation attention group is input to obtain a fine remote sensing image feature map. The residual separation attention network contains several residual separation attention groups. Each group contains an upsampling module and a residual separation attention module. The upsampling module considers hierarchical prior information, assists the network in the high-resolution feature space, and predicts high-frequency local detail information. The residual separation attention module is used to adaptively explore and learn global structural information in the low-resolution feature space. The multi-level feature fusion module gradually fuses feature maps from different levels through dense connections, enabling local attention to interact across feature groups and achieving global attention in the network, thus obtaining more refined remote sensing feature maps. The remote sensing image reconstruction module is used to perform convolution operations on the feature-fused remote sensing satellite feature map to reconstruct a high-resolution remote sensing satellite image of the target.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the remote sensing image super-resolution method based on residual separation attention mechanism as described in any one of claims 1 to 4.

Images