A remote sensing image super-resolution method based on context-aware edge enhancement

Abstract

The application discloses a remote sensing image super-resolution method based on context-aware edge enhancement, and aims at the problem that existing super-resolution methods cannot accurately restore object edges, and establishes an edge enhancement module, enhances context semantic features and edge features through a bidirectional weighted fusion method, and uses context features to guide edge features. The edge restoration module aims to fuse multi-scale edge features enhanced by the edge feature enhancement module, and reconstruct a fine edge map. Meanwhile, since existing loss functions only constrain generated images in the pixel domain, an edge loss function is designed to constrain the similarity of generated super-resolution images and high-resolution images in the edge domain. Test results on the UCMerced dataset show that the application achieves the best LPIPS value, can obtain reconstructed images more in line with human visual habits, and meanwhile, the application achieves the best classification accuracy, and has high practical application value.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing, and in particular relates to a remote sensing image super-resolution method based on context-aware edge enhancement. Background Technology

[0002] With the development of remote sensing image application technology, remote sensing images are widely used in target detection, yield prediction, and change detection. However, due to hardware limitations and environmental interference such as clouds and fog during the acquisition process, remote sensing images often suffer severe degradation and are significantly inconsistent with the actual scene. Natural images, on the other hand, do not encounter these problems, thus super-resolution algorithms for natural images perform poorly when processing remote sensing images. Low-quality remote sensing images can seriously affect the results of subsequent practical applications; therefore, super-resolution algorithms for remote sensing images have received widespread attention in recent years.

[0003] In existing technologies, many super-resolution algorithms use neural networks to learn the correspondence between pairs of low-resolution and high-resolution images. However, they only use peak signal-to-noise ratio (PSNR) as the network's optimization objective, which leads to overly smooth generated images that are inconsistent with the needs of machine analysis and human visual perception. Furthermore, some research has introduced generative adversarial networks (GANs) into the super-resolution field to obtain more realistic reconstructed images. However, due to the lack of constraints on high-frequency information and the severe degradation of remote sensing images, GAN-based super-resolution methods often result in high-frequency details that do not accurately reflect reality. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing work and provide a remote sensing image super-resolution method based on context-aware edge enhancement. This method can effectively improve the reconstruction effect of low-resolution remote sensing images and make the super-resolution image suitable for both human eye perception and machine vision applications.

[0005] This invention is achieved using the following technical solution:

[0006] A context-aware edge enhancement-based super-resolution method for remote sensing images includes the following steps:

[0007] 1) The low-resolution image to be processed is input into the initial feature extraction module, and is first processed by the pre-super-resolution module to generate a pre-super-resolution image;

[0008] 2) Input the pre-super-resolution image into the edge and context semantic feature extraction layers respectively to obtain the corresponding two types of multi-scale features;

[0009] 3) Input the two types of multi-scale features into the edge enhancement module. First, the bidirectional edge feature fusion module and the bidirectional context feature fusion module perform bidirectional weighted fusion. At the same time, the context edge information exchange module is used to exchange information between the edge branch and the context branch.

[0010] 4) Input the multi-scale edge features, which have been iteratively enhanced by multiple edge enhancement modules, into the edge reconstruction module for edge aggregation and reconstruction;

[0011] 5) Replace the blurred edges in the pre-super-resolution image with the enhanced edges obtained from the aggregated reconstruction to obtain a super-resolution reconstructed image with fine and accurate edges;

[0012] 6) Use the VGG discriminator to determine the authenticity of the super-resolution reconstructed image;

[0013] 7) Calculate pixel loss, feature loss, adversarial loss and edge loss respectively, and perform back gradient propagation according to the loss function to optimize network parameters.

[0014] A further improvement of the present invention is that, in step 1), a pre-super-resolution module is used to provide accurate information before processing the low-resolution image to generate the super-resolution image.

[0015] A further improvement of the present invention is that, in step 2), a convolutional layer that extracts multi-scale edge features and contextual semantic features is used to adapt to the size of various objects in the remote sensing image; wherein, the contextual semantic features are extracted by the convolutional network on the pre-super-resolution image, and the edge features are extracted by the convolutional network on the edge map obtained by the Laplacian operator operation of the pre-super-resolution image.

[0016] A further improvement of this invention is that, in step 3), the bidirectional edge feature fusion module and the bidirectional context feature fusion module combine top-down and bottom-up methods to introduce positional information into the fused features, so that the fused features simultaneously possess semantic and positional information; the edge features output by the (i+1)th bidirectional edge feature fusion module are in The edge features output by the i-th bidirectional edge feature fusion module are those with the smallest size. Represented as:

[0017]

[0018] Here, Resize(·) indicates that interpolation is used to keep the tensor dimension consistent; w1 and w2 are two learnable weight vectors, representing the importance of features at two different scales; ∈ is a numerical value used to increase the stability of model training, and its value is set to 1×10.-6 LReLU(·) represents the LeakyReLU activation function.

[0019] A further improvement of this invention is that, in step 3), the process of information exchange between contextual semantic features and edge features is represented as follows:

[0020]

[0021]

[0022] in, and These represent the contextual semantic features and edge features prior to information exchange, respectively. and These represent the contextual semantic features and edge features after information exchange, respectively. · represents element-wise multiplication, Conv(·) represents 3×3 convolution, and LReLU(·) represents the LeakyReLU activation function.

[0023] A further improvement of this invention is that, in step 4), the edge reconstruction module aggregates multi-scale edge features and reconstructs the enhanced edge image through three parts: feature integration, coordinate attention mechanism, and feature fusion; let F R =[f r1 f r2 f r3 ] represents the output features of residual connections with different scales; the multi-scale features are first adjusted to the same shape as the target super-resolution image, and then concatenated to form a unified feature F. M This process can be represented as:

[0024]

[0025] F M =Concat(f r1 , R2(f r2 ), R4(f r3 ))

[0026] Here, Concat(·) represents a tensor concatenation operation along the channel dimension; R2(·) and R4(·) represent 2× and 4× upsampling operations, respectively. After the scaling operation, the three features will remain in the same feature dimension, i.e., f. r1 , R2(f r2 ), R4(f r3 )∈R 4H×4W×C Where H and W represent the length and width of the low-resolution image, respectively, and C represents the number of output channels for the feature; the output feature tensor F M ∈R4H×4W×3C First, the input is fed into the coordinate attention mechanism to more accurately locate and identify the target of interest, enabling the model to adaptively learn the importance information of different channels for reconstructing the object, as well as the ability to perceive the spatial coordinates of the object in the entire image. The output tensor of the coordinate attention mechanism is then passed through a feature fusion network, which consists of three convolutional layers with kernel sizes of 1×1, 3×3, and 1×1, respectively. For the output features of each convolutional layer, a LeakyReLU activation layer is applied to improve the model's non-linear fitting ability. The edge reconstruction module fuses the enhanced multi-scale features into one and generates an enhanced edge map E. * The calculation process is described as follows:

[0027] E * =FS(CA(F M ))

[0028] Where CA(·) represents the channel attention module and FS(·) represents the feature fusion module.

[0029] A further improvement of the present invention is that, in step 5), the noisy edges I in the pre-super-resolution image are... Pre Replace with the reconstructed edge map E * This process enables the final super-resolution image to reflect the true edges of objects at different scales and of different categories. The process can be represented as:

[0030] I SR =I Pre +E * -L(I Pre ).

[0031] A further improvement of the present invention is that, in step 7), pixel loss... Feature loss and combat losses They are represented as follows:

[0032]

[0033]

[0034]

[0035] Among them, I SR The output super-resolution image of the network, I HR φ represents the actual corresponding high-resolution reference image. l (·) represents the l-th layer of the VGG19 network, ω lLet G(·) represent the weight coefficients corresponding to the l-th layer; G(·) and D(·) represent the generator network and discriminator network, respectively. D(·) is used to calculate the probability that the input is a real high-resolution image or a super-resolution image generated by the super-resolution network; to strengthen the constraint on high-frequency components during image generation, thereby enhancing the consistency of the edges between the reconstructed image and the real image, the edge loss function is used. Represented as:

[0036]

[0037] Finally, the total loss function of the generator network is given by the following formula:

[0038]

[0039] Here, α, β, and γ represent weight parameters designed for each loss function, which are used to balance the magnitudes of the various loss functions.

[0040] The present invention has at least the following beneficial technical effects:

[0041] The remote sensing image super-resolution method based on context-aware edge enhancement provided by this invention, in its specific operation, firstly inputs the low-resolution image to be processed into the initial feature extraction module, and then processes it using a pre-super-resolution module to generate a pre-super-resolution image. The pre-super-resolution module can reconstruct most of the region except for the edges. The pre-super-resolution image is then input into the edge and context semantic feature extraction layers respectively to obtain two types of multi-scale features. Further, the two types of multi-scale features are input into the edge enhancement module, where they are first bidirectionally weighted and fused by a bidirectional edge feature fusion module and a bidirectional context feature fusion module. Simultaneously, a context edge information exchange module exchanges information between the edge branch and the context branch. The multi-scale edge features, enhanced iteratively by multiple edge enhancement modules, are then input into the edge reconstruction module for edge aggregation and reconstruction. Finally, a VGG discriminator network is used to determine the realism of the super-resolution reconstructed image, and pixel loss, feature loss, adversarial loss, and edge loss are calculated using the super-resolution image and the high-resolution image respectively. Backward gradient propagation is then performed based on the loss function to optimize the network parameters.

[0042] Furthermore, the remote sensing image super-resolution framework in this invention, designed for both human and machine vision analysis, addresses the challenge of handling remote sensing images. Remote sensing images typically contain a large number of objects of varying scales, shapes, and complex spatial relationships, resulting in highly rich contextual semantic features. However, these features are severely impacted during the complex degradation process of remote sensing images. CEEGAN facilitates information exchange between contextual semantic features and edge features, integrating contextual features into the image to recover texture information in the generated super-resolution image as much as possible. These two types of features complement each other to meet the needs of both human visual habits and machine vision applications.

[0043] Furthermore, in this invention, the context-guided edge enhancement module explores and integrates multi-scale edge features and contextual semantic features to enhance images with blurred edge details. High-level contextual semantic features are beneficial for guiding the extraction and enhancement of edge features. Therefore, we maintain the information interaction between contextual semantics and edge branches to help edge branches understand high-level semantic information in the image. In addition, we fuse the multi-scale features enhanced by the edge feature enhancement module and reconstruct the edge map through the edge recovery module.

[0044] Furthermore, one edge loss function in this invention, by adding constraints to the super-resolution image and the high-resolution image in the edge domain, generates an image that extends the model's attention from the pixel domain to the edge domain, making the generated image more realistic and preserving important edge details.

[0045] In summary, this invention provides a context-aware edge enhancement-based super-resolution method for remote sensing images. Addressing the problem that existing super-resolution methods cannot accurately recover object edges, this invention establishes an edge enhancement module that enhances features through bidirectional weighted fusion. Simultaneously, an information exchange module between contextual semantic features and edge features is designed, using contextual semantic features to guide edge feature development. Finally, an edge recovery module fuses the multi-scale edge features enhanced by the edge feature enhancement module to reconstruct a refined edge map. To address the issue that existing loss functions only constrain the generated image in the pixel domain, an edge loss function is designed to constrain the similarity between the generated super-resolution image and the high-resolution image in the edge domain. Overall, this invention can obtain images that better conform to human visual habits and has good practical application value. Attached Figure Description

[0046] Figure 1 : Flowchart of the remote sensing image super-resolution method based on context-aware edge enhancement of this invention;

[0047] Figure 2 : A flowchart of the bidirectional feature fusion module in this invention;

[0048] Figure 3 : A flowchart of the information exchange module between contextual semantic features and edge features in this invention;

[0049] Figure 4 : A flowchart of the edge reconstruction module in this invention;

[0050] Figure 5 : A flowchart for calculating the loss function in this invention. Detailed Implementation

[0051] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0052] like Figure 1 As shown, the present invention provides a remote sensing image super-resolution method based on context-aware edge enhancement, which includes the following steps:

[0053] 1) The low-resolution image to be processed is input into the initial feature extraction module. First, it is processed by the pre-super-resolution module to generate a pre-super-resolution image. The pre-super-resolution module can reconstruct most of the region except for the edges.

[0054] 2) Input the pre-super-resolution image into the edge and context semantic feature extraction layers respectively to obtain the corresponding two types of multi-scale features;

[0055] 3) Input the two types of multi-scale features into the edge enhancement module. First, the bidirectional edge feature fusion module and the bidirectional context feature fusion module perform bidirectional weighted fusion. At the same time, the context edge information exchange module is used to exchange information between the edge branch and the context branch.

[0056] 4) Input the multi-scale edge features, which have been iteratively enhanced by multiple edge enhancement modules, into the edge reconstruction module for edge aggregation and reconstruction;

[0057] 5) Replace the blurred edges in the pre-super-resolution image with the enhanced edges obtained from the aggregated reconstruction to obtain a super-resolution reconstructed image with fine and accurate edges;

[0058] 6) Use the VGG discriminator to determine the authenticity of the super-resolution reconstructed image;

[0059] 7) Calculate pixel loss, feature loss, adversarial loss and edge loss respectively, and perform back gradient propagation according to the loss function to optimize network parameters.

[0060] Low-resolution images undergo extremely complex degradation processes, often losing important high-frequency information. Directly performing super-resolution processing on low-resolution images rarely yields satisfactory edges. Therefore, a pre-super-resolution module is used to provide accurate information before processing low-resolution images to generate super-resolution images.

[0061] Objects in remote sensing images typically span a wide range of scales, making it difficult to capture the diverse object information required for super-resolution reconstruction using only single-scale features. Therefore, we propose a convolutional layer capable of extracting multi-scale edge features and contextual semantic features, adapting it to the varying sizes of objects in remote sensing images. Specifically, the contextual semantic features are extracted from the pre-super-resolution image by the convolutional network, while the edge features are extracted from the edge map obtained by applying the Laplacian operator to the pre-super-resolution image.

[0062] Bidirectional feature fusion module, such as Figure 2 As shown, remote sensing images undergo complex degradation processes, and natural factors such as clouds and fog often interfere with the shape of targets and blur the edge features extracted from them. Furthermore, the complex background features of remote sensing images introduce a large amount of noise during feature extraction. These problems lead to insufficient initial edge feature representation capabilities. With the continuous stacking of convolutions, deeper convolutional layers extract higher-level features considered to have stronger semantic information, while shallower convolutional layers extract shallower features considered to have stronger positional information. Top-down feature fusion methods can obtain feature maps with higher semantic information for reconstruction, improving accuracy to some extent, but are limited by unidirectional information flow. Bottom-up feature fusion structures can obtain richer semantic information from the fused features but also lose a lot of positional information. This paper combines top-down and bottom-up methods in the bidirectional edge feature fusion module and the bidirectional context feature fusion module, incorporating positional information into the fused features. This allows the fused features to possess both high semantic and positional information, significantly improving the accuracy of image super-resolution reconstruction. The edge features output by the (i+1)th bidirectional edge feature fusion module are... in To have the smallest feature size, the edge features output by the i-th bidirectional edge feature fusion module can be... Represented as:

[0063]

[0064] Here, `Resize(·)` indicates that interpolation is used to maintain the same dimensionality of the tensor. `w1` and `w2` are two learnable weight vectors, representing the importance of features at two different scales. `∈` is a numerical value used to increase the stability of model training, and its value is set to 1 × 10^2. -6 LReLU(·) represents the LeakyReLU activation function.

[0065] The information exchange module between contextual semantic features and edge features, such as Figure 3 As shown, contextual semantic features reflect high-level features in an image. These features help edge branches understand high-level contextual semantic information, enabling the network to learn edges better and thus obtain more accurate edge reconstruction results. The accuracy of edge reconstruction is improved by utilizing the interaction between contextual semantic features in the pixel domain and edge features in the edge domain. The process of information exchange between contextual semantic features and edge features can be represented as follows:

[0066]

[0067]

[0068] in, and These represent the contextual semantic features and edge features prior to information exchange, respectively. and These represent the contextual semantic features and edge features after information exchange, respectively. · represents element-wise multiplication, Conv(·) represents 3×3 convolution, and LReLU(·) represents the LeakyReLU activation function.

[0069] Edge reconstruction modules utilizing multi-scale features, such as Figure 4 As shown, features at a single scale cannot fully represent objects with vastly different shapes and sizes in remote sensing images. The edge reconstruction module, through feature integration, coordinate attention mechanism, and feature fusion, aggregates multi-scale edge features and reconstructs an enhanced edge image. Let F... R =[f r1 f r2 f r3 [] represents the output features of residual connections with different scales. The multi-scale features are first resized to the same shape as the target super-resolution image, and then concatenated to form a unified feature F. M This process can be represented as:

[0070]

[0071] R M =Concat(f r1, R2(f r2 ), R4(f r3 ))

[0072] Here, Concat(·) represents a tensor concatenation operation along the channel dimension. R2(·) and R4(·) represent 2× and 4× upsampling operations, respectively. After the scaling operation, the three features will remain in the same feature dimension, i.e., f. r1 , R2(f r2 ), R4(f r3 )∈R 4H×4W×C Where H and W represent the length and width of the low-resolution image, respectively, and C represents the number of output channels for the features. The output feature tensor R M ∈R 4H×4W×3C First, the input is fed into the coordinate attention mechanism to more accurately locate and identify the target of interest, enabling the model to adaptively learn the importance information of different channels for reconstructing the object, as well as its ability to perceive the spatial coordinates of the object in the entire image. The output tensor of the coordinate attention mechanism is then passed through a feature fusion network, which consists of three convolutional layers with kernel sizes of 1×1, 3×3, and 1×1, respectively. For the output features of each convolutional layer, a LeakyReLU activation layer is applied to improve the model's non-linear fitting ability. The edge reconstruction module fuses the enhanced multi-scale features into a single entity and generates an enhanced edge map E. * The calculation process can be described as follows:

[0073] E * =FS(CA(F M ))

[0074] Where CA(·) represents the channel attention module and FS(·) represents the feature fusion module.

[0075] Next, the noisy edges I in the pre-super-resolution image are... Pre Replace with the reconstructed edge map E * This makes the final super-resolution image I SR It can reflect the true edges of objects at different scales and of different categories. This process can be represented as:

[0076] I SR =I Pre +E * -L(I Pre )

[0077] This invention provides a constraint-generated edge loss function, although it uses pixel loss. Feature loss and combat losses The combination of these elements can increase the realism of the generated images, but it can also introduce artifacts and inaccurate details into the generated images. and They are represented as follows:

[0078]

[0079]

[0080]

[0081] Among them, I SR The output super-resolution image of the network, I HR φ represents the actual corresponding high-resolution reference image. l (·) represents the l-th layer of the VGG19 network, ω l Let G(·) represent the weight coefficients corresponding to the l-th layer. G(·) and D(·) represent the generator network and discriminator network, respectively. D(·) is used to calculate the probability that the input is a real high-resolution image or a super-resolution image generated by the super-resolution network. To strengthen the constraint on high-frequency components during image generation, thereby enhancing the consistency of edges between the reconstructed image and the real image, an edge loss function is used. Represented as:

[0082]

[0083] Finally, the total loss function of the generator network is given by the following formula:

[0084]

[0085] Here, α, β, and γ represent the weight parameters designed for each loss function. Their purpose is to balance the magnitude of each loss function and prevent the training process of the entire model from being hindered by the excessively large values ​​of some loss functions.

[0086] The experimental setup and results analysis are given below.

[0087] This invention was trained and tested on the UCMerced dataset, which contains 21 scene categories, including churches, coastlines, dense residential areas, and highways. Each category has 100 images, all with a resolution of 256×256 pixels and a spatial resolution of 0.3 meters per pixel. 80% of the entire dataset was randomly selected as the training set, and the remaining 20% ​​was used as the validation set. During network training, the input images were cropped into 128×128 pixel patches. Adam was used as the optimizer to train the generator and discriminator networks separately. The batch size was set to 16, the total training epochs were 2000, and the initial learning rate for both the generator and discriminator networks was set to 1×10⁻⁶. -4 During training, the learning rate is reduced to 0.9 times its initial value every 500 epochs, until the learning rate is reduced to less than 1×10⁻⁶. -5 Scaling was stopped at the specified point. The weight parameters α, β, and γ in the generator loss function were set to 0.1, 1, and 0.5, respectively. After training, the model was saved and tested. The test metrics included learning-based perceptual similarity measure (LPIPS) and classification accuracy. This experiment used an Nvidia 2080 Ti GPU for training and testing. The model was implemented using Python 3.8 on the PyTorch 1.12.0 framework under Ubuntu 18.04 and CUDA 11.2.

[0088] Experimental Results and Analysis

[0089] In the experimental results, the comparison algorithms were SRGAN, ESRGAN, EEGAN, and SwinIR. The algorithm corresponding to this invention is the Context-Aware Edge Enhancement-Based Remote Sensing Image Super-Resolution Method (CEEGAN).

[0090] The first set of experiments compares the LPIPS (Percentage Points Per Second) of different super-resolution methods on the UCMerced dataset. LPIPS, by learning perceptual factors, reflects the perceptual distance between two images in the feature domain. The calculation method of LPIPS aligns with human perception of image similarity; a lower LPIPS value indicates that the two images are closer. Due to the varying complexity of different scenes in remote sensing images, the learning difficulty of the models varies, resulting in significant fluctuations in evaluation metrics across different scenarios. For a fair comparison, we calculated LPIPS for each category separately and also calculated the average data for the entire dataset. Table 1 shows the comparison results of different methods at a 4x magnification on the UCMerced dataset.

[0091] The second set of experiments: Comparison of classification accuracy of different super-resolution methods on the UCMerced dataset. To evaluate the performance of super-resolution methods in real-world scenarios, we conducted image classification experiments on the UCMerced dataset. First, a ResNet-34 image classification network was trained on the original images in the dataset. Then, bicubic interpolation and Gaussian blur downsampling were performed on the 256×256 pixel high-resolution images to generate 64×64 pixel low-resolution images, and different super-resolution methods were used to obtain the corresponding reconstructed images. Finally, the reconstructed images were input into the image classification network to obtain the classification accuracy, and the results are shown in Table 2. The super-resolution reconstructed images obtained by CEEGAN performed best on the classification task and could obtain results close to those of the original high-resolution images. This shows that our method can effectively preserve important image features and generate highly realistic images, proving the effectiveness of this invention in practical applications of remote sensing images.

[0092] Table 1 Comparison of LPIPS results on the UCMerced dataset

[0093]

[0094]

[0095] Table 2 Comparison of classification accuracy on the UCMerced dataset

[0096]

[0097] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A remote sensing image super-resolution method based on context-aware edge enhancement, characterized in that, Includes the following steps: 1) The low-resolution image to be processed is input into the initial feature extraction module, and is first processed by the pre-super-resolution module to generate a pre-super-resolution image; 2) Input the pre-super-resolution image into the edge and context semantic feature extraction layers respectively to obtain the corresponding two types of multi-scale features; 3) Input the two types of multi-scale features into the edge enhancement module. First, the bidirectional edge feature fusion module and the bidirectional context feature fusion module perform bidirectional weighted fusion. At the same time, the context edge information exchange module is used to exchange information between the edge branch and the context branch. 4) Input the multi-scale edge features, which have been iteratively enhanced by multiple edge enhancement modules, into the edge reconstruction module for edge aggregation and reconstruction; 5) Replace the blurred edges in the pre-super-resolution image with the enhanced edges obtained from the aggregated reconstruction to obtain a super-resolution reconstructed image with fine and accurate edges; 6) Use the VGG discriminator to determine the authenticity of the super-resolution reconstructed image; 7) Calculate the pixel loss, feature loss, adversarial loss, and edge loss respectively, and perform backpropagation based on the loss function to optimize the network parameters; pixel loss Feature loss and combat losses They are represented as follows: in, The output super-resolution image represents the network's output. This represents the actual corresponding high-resolution reference image. This indicates the VGG19 network's... layer, Indicates the first The weight coefficients corresponding to the layers; and These represent the generator network and the discriminator network, respectively. This is used to calculate the probability that the input is a real high-resolution image or a super-resolution image generated by a super-resolution network; to strengthen the constraint on high-frequency components during image generation, thereby enhancing the consistency of edges between the reconstructed image and the real image, an edge loss function is used. Represented as: Finally, the total loss function of the generator network is given by the following formula: in, This represents the weight parameters designed for each loss function, which are used to balance the magnitudes of the various loss functions.

2. The remote sensing image super-resolution method based on context-aware edge enhancement according to claim 1, characterized in that, In step 1), a pre-super-resolution module is used to provide accurate information before processing the low-resolution image to generate the super-resolution image.

3. The remote sensing image super-resolution method based on context-aware edge enhancement according to claim 2, characterized in that, In step 2), a convolutional layer is used to extract multi-scale edge features and contextual semantic features so that it can adapt to the size of various objects in the remote sensing image. The contextual semantic features are extracted by the convolutional network on the pre-super-resolution image, and the edge features are extracted by the convolutional network on the edge map obtained by the Laplacian operator operation on the pre-super-resolution image.

4. The remote sensing image super-resolution method based on context-aware edge enhancement according to claim 3, characterized in that, In step 3), the bidirectional edge feature fusion module and the bidirectional context feature fusion module combine top-down and bottom-up methods to incorporate location information into the fused features, enabling the fused features to possess both semantic and location information; The edge features output by the bidirectional edge feature fusion module are: ,in For features with the smallest size, by the first The edge features output by the bidirectional edge feature fusion module Represented as: in, This indicates that interpolation is used to keep the dimensions of the tensor consistent. and These are two learnable weight vectors, representing the importance of features at two different scales; This is a numerical value used to increase the stability of model training; its value is set to... , This represents the LeakyReLU activation function.

5. The remote sensing image super-resolution method based on context-aware edge enhancement according to claim 4, characterized in that, In step 3), the process of information exchange between contextual semantic features and edge features is represented as follows: in, and These represent the contextual semantic features and edge features prior to information exchange, respectively. and These represent the contextual semantic features and edge features after information exchange, respectively. This represents element-wise multiplication. represent Convolution operation, This represents the LeakyReLU activation function.

6. The remote sensing image super-resolution method based on context-aware edge enhancement according to claim 5, characterized in that, In step 4), the edge reconstruction module aggregates multi-scale edge features and reconstructs the enhanced edge image through three parts: feature integration, coordinate attention mechanism, and feature fusion; let's assume... This represents the output features of residual connections with different scales; Multi-scale features are first reshaped to match the target super-resolution image, and then stitched together to form a unified feature set. This process can be represented as: in, This indicates a tensor concatenation operation performed along the channel dimension; and They represent and After the upsampling operation and the scaling operation, the three features will remain in the same feature dimension. , ,in and These represent the length and width of the low-resolution image, respectively. The number of output channels representing the features; the output feature tensor The input is first fed into the coordinate attention mechanism to more accurately locate and identify the target of interest, enabling the model to adaptively learn the importance information of different channels for reconstructing the object, as well as the ability to perceive the spatial coordinates of the object in the entire image. The output tensor of the coordinate attention mechanism will then be passed through a feature fusion network, which consists of three convolutional layers, each with a kernel size of [missing information]. The convolutional layers, for each output feature, are followed by a LeakyReLU activation layer to improve the model's non-linear fitting ability; the edge reconstruction module fuses the enhanced multi-scale features into one and generates an enhanced edge map. The calculation process is described as follows: in, This indicates the channel attention module. This indicates the feature fusion module.

7. The remote sensing image super-resolution method based on context-aware edge enhancement according to claim 6, characterized in that, In step 5), noisy edges in the pre-super-resolution image are removed. Replace with the reconstructed edge map This process enables the final super-resolution image to reflect the true edges of objects at different scales and of different categories. The process can be represented as: .

Images