Satellite image super-resolution reconstruction method, system, device and readable storage medium

Abstract

The application discloses a satellite image super-resolution reconstruction method, system, device and readable storage medium, and relates to the technical field of computer vision.The method comprises the following steps: acquiring a high-resolution satellite image and converting the high-resolution satellite image into a low-resolution image; performing edge extraction on the low-resolution image to obtain a low-resolution edge image; performing up-sampling processing on the low-resolution edge image to obtain a high-resolution edge feature image; performing texture feature extraction on the low-resolution edge image through an edge branch network to obtain a reconstructed high-resolution edge image; performing structure feature extraction on the low-resolution image through a super-resolution branch network to obtain a reconstructed high-resolution image; determining a recovery image according to the high-resolution edge feature image and the reconstructed high-resolution edge image; performing image correction on the recovery image to obtain a reconstructed edge prior image; fusing the reconstructed edge prior image and the reconstructed high-resolution image to obtain a fused image; and performing dimension reduction processing on the fused image to obtain a reconstructed super-resolution image.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and specifically to a method, system, device, and readable storage medium for super-resolution reconstruction of satellite images. Background Technology

[0002] Satellite image super-resolution (SR) algorithms based on convolutional neural networks (CNNs) predict detail information through prior information in training samples, greatly improving the reconstruction performance of satellite images. In recent years, image super-resolution reconstruction (ISRR) based on edge priors has been greatly developed. How to fully explore and effectively utilize the prior information of images to obtain better reconstruction performance has become a research hotspot in ISRR.

[0003] As is well known, images can be divided into low-frequency and high-frequency components. The low-frequency component refers to areas where image intensity changes smoothly, while the high-frequency component refers to areas where image intensity changes drastically, i.e., the edges of the image. Numerous edge-assisted and edge-guided image processing methods have verified the feasibility and necessity of image edge priors. However, these methods still have some limitations in achieving ISRR.

[0004] For example, Yang et al. proposed Deep Edge-Guided Recurrent Residual Learning (DEGREE) for image super-resolution. This method constructs an edge-guided recurrent residual network. By introducing image edge priors into this recurrent residual network, the recurrent residual learning recovers the differences between low-resolution and high-resolution images, further enhancing the edge preservation ability of the SR process. It also accelerates the training convergence speed by constructing bypass connections across multiple degrees, thereby achieving the recovery of high-frequency details in the image. However, using this method to achieve ISRR has the drawback of poor image reconstruction performance, mainly for three reasons: First, DEGREE uses low-resolution (LR) images preprocessed by bicubic interpolation as input, which introduces additional noise and produces artifacts; second, DEGREE only applies off-the-shelf edge detectors to obtain image edges on the preprocessed LR images, which introduces additional noise and causes image edge blurring; third, DEGREE directly adds the learned image edge features to the low-resolution image, which is essentially a residual learning and does not maximize the use of image edge prior information.

[0005] The Edge Enhancement GAN for Remote Sensing Image Super-Resolution (EEGAN) constructs a Ultra-Dense Subnetwork (UDSN) and an Edge Enhancement Subnetwork (EESN). This method first generates a high-resolution image eroded by artifacts and noise in the UDSN, then performs masking on the high-resolution image in the EESN to extract and enhance image contours, obtaining a restored intermediate image and enhanced edge images. Finally, the restored intermediate image and enhanced edge images are combined to generate a super-resolution image. However, the enhanced edge images obtained in the intermediate process deviate from the underlying high-frequency information of the low-resolution image and introduce additional artifacts and distortions into the reconstructed satellite image at a small scale.

[0006] Supporting incremental evolution and generational transformation, the novel network architecture SeaNet, characterized by its on-site, resilient, and autonomous nature, uses original high-resolution (HR) edge images and designs a soft-edge-assisted network to reconstruct high-quality super-resolution images with the help of soft edges in the image. Similar to EEGAN, this method does not consider the potential high-frequency information of the image in the low-resolution space.

[0007] Existing edge-prior-based image super-resolution reconstruction methods, such as EEGAN (a single-path network), only consider edge information in the high-resolution space, neglecting the potential high-frequency information of the original LR edge image and failing to maximize the utilization of image edge prior information. Dual-path networks, such as DEGREE, only consider edge information in the low-resolution space, resulting in coarse and inaccurate edge maps and poor reconstruction performance. Therefore, although these methods introduce convolutional neural networks to process images, accurately extracting image priors and effectively using this information to assist image reconstruction remain problems that need further exploration in this field. Summary of the Invention

[0008] The technical problem this invention aims to solve is that existing edge-prior-based image super-resolution reconstruction methods, when performing super-resolution reconstruction of satellite images, only consider edge information in high-resolution space or only consider edge information in low-resolution space, failing to fully utilize prior edge information, resulting in poor visual effects and low quality of the generated super-resolution reconstructed images. To solve this problem, this invention provides a satellite image super-resolution reconstruction method, system, device, and readable storage medium.

[0009] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0010] A method for super-resolution reconstruction of satellite images, characterized by comprising:

[0011] Step S1: Acquire a high-resolution satellite image, convert the high-resolution satellite image into a low-resolution image, and perform edge extraction on the low-resolution image to obtain a low-resolution edge image;

[0012] Step S2: Upsample the low-resolution edge image to obtain a high-resolution edge feature image;

[0013] Step S3: Extract texture features from the low-resolution edge image using a pre-established edge branching network to obtain a reconstructed high-resolution edge image; extract structural features from the low-resolution image using a pre-established super-resolution branching network to obtain a reconstructed high-resolution image;

[0014] Step S4: Based on the high-resolution edge feature image and the reconstructed high-resolution edge image, determine the restored image, perform image correction on the restored image, and obtain the reconstructed edge prior image;

[0015] Step S5: Fuse the reconstructed edge prior image and the reconstructed high-resolution image to obtain a fused image, and perform dimensionality reduction processing on the fused image to obtain a reconstructed super-resolution image.

[0016] The beneficial effects of this invention are as follows: Low-resolution edge images are introduced and edge branch networks are designed for them, converting low-resolution edge images into high-resolution edge feature images and reconstructing high-resolution edge images to assist in the reconstruction of super-resolution images; by integrating the reconstructed edge prior images obtained based on high-resolution edge feature images and reconstructed high-resolution edge images into the super-resolution branch network, a prior structure is provided for reconstructing high-resolution edge images and reconstructing high-resolution images; the reconstructed edge prior images reveal the sharpness of each local region in the image, and correction is performed based on high-frequency information in the low-resolution edge images to obtain more accurate reconstructed edges; for the edge branch network and the super-resolution branch network, a closed loop of low-resolution image-edge branch network-super-resolution branch network-reconstructed super-resolution image is constructed. By introducing the edge branch network, the edge branch network and the super-resolution branch network jointly provide supervision for the reconstruction of super-resolution images, thereby better preserving the structure and obtaining reconstructed super-resolution images with high perceptual quality and low distortion, solving the problem of artifacts and distortions generated during edge image processing.

[0017] Based on the above technical solution, the present invention can be further improved as follows.

[0018] Furthermore, in step S1, converting the high-resolution satellite image into a low-resolution image includes:

[0019] The high-resolution satellite image is divided into blocks to obtain a block-processed high-resolution satellite image.

[0020] The high-resolution image after block processing is subjected to bicubic interpolation to obtain a low-resolution image;

[0021] In step S1, edge extraction is performed on the low-resolution image to obtain a low-resolution edge image, including:

[0022] The low-resolution image is sharpened using the Laplacian operator to obtain a sharpened image. The boundary positions in the sharpened image are marked to obtain a Laplacian image. The Laplacian image is then superimposed on the low-resolution image to obtain a low-resolution edge image.

[0023] The beneficial effects of adopting the above-mentioned further scheme are as follows: by performing image block processing and bicubic interpolation processing on high-resolution satellite images, the image size is reduced and the image quality is lowered, laying the foundation for subsequent acquisition of multi-scale global structure prior information of the image using basic blocks; by using the Laplacian operator to sharpen the image, the sharpened image is marked to obtain the Laplacian image, and then the Laplacian image is superimposed with the low-resolution image, which not only maintains the effect of sharpening processing, but also restores the background information of the low-resolution image, accurately locates the image edge, and extracts obvious image edge information, resulting in good edge extraction effect.

[0024] Further, in step S3, texture features are extracted from the low-resolution edge image using a pre-established edge branching network to obtain a reconstructed high-resolution edge image, including:

[0025] The low-resolution edge image is used to extract features using a pre-constructed convolutional layer A to obtain a shallow feature image;

[0026] For each of the pre-established basic blocks, deep feature extraction is performed on the shallow feature image through the basic block to obtain a fused feature image;

[0027] Deep feature images are obtained by extracting features from each of the fused feature images through a pre-constructed convolutional layer B.

[0028] The deep feature image is upsampled by a pre-constructed upsampling layer to obtain an upsampled deep feature image.

[0029] By sequentially extracting features from the upsampled deep feature image using pre-constructed convolutional layers G1 and G2, a reconstructed high-resolution edge image is obtained.

[0030] The beneficial effects of adopting the above-mentioned further scheme are as follows: the method of extracting texture features from low-resolution edge images to obtain reconstructed high-resolution edge images is similar to the method of extracting structural features from low-resolution images to obtain reconstructed high-resolution images. By performing shallow feature extraction on low-resolution edge images and low-resolution images, upsampling features of the images are obtained. Then, deep feature extraction is performed on the shallow feature images, so that the shallow feature images are processed through multiple basic blocks to extract deep features of the images. Finally, upsampling and reconstruction processing is performed on the deep feature images to obtain reconstructed high-resolution edge images and reconstructed high-resolution images, which lays the foundation for cross-supervision of images.

[0031] Furthermore, for each of the basic blocks, the basic block includes a multi-scale global prior block and a dense connection module connected in sequence. Each multi-scale global prior block includes multiple upsampling and downsampling units of different scales and a convolutional layer D. The multiple upsampling and downsampling units are parallel. The dense connection module includes multiple dense connection units with the same structure and a convolutional layer F.

[0032] The step of extracting deep features from the shallow feature image using the basic blocks to obtain a fused feature image includes:

[0033] Each of the upsampling and downsampling units in the basic block is used to extract features from the shallow feature image to obtain multiple feature images;

[0034] The first multi-scale image is obtained by adding the multiple feature images element by element.

[0035] The first multi-scale image is reduced in dimensionality by the convolutional layer D in the basic block to obtain the second multi-scale feature image.

[0036] By performing feature fusion on the second multi-scale feature image through each of the dense connection units in the basic block, multiple third multi-scale feature images are obtained;

[0037] The convolutional layer F in the basic block is used to reduce the dimensionality of multiple third multi-scale feature images to obtain dense feature images.

[0038] The second multi-scale feature image and the dense feature image are added element by element to obtain a fused feature image.

[0039] The beneficial effects of adopting the above-mentioned further scheme are: by performing parallel upsampling and downsampling of shallow feature images at different scales through multiple basic blocks, the prior information of the multi-scale global structure of the image can be fully extracted to obtain the multi-scale features of the image. Then, the image is fused using dense connection units to obtain deep feature images, which can effectively improve the quality of image feature extraction.

[0040] Further, in step S4, determining the restored image based on the high-resolution edge feature image and the reconstructed high-resolution edge image includes:

[0041] The reconstructed high-resolution edge images are processed by pre-constructed convolutional layers H and I to extract image features, generating residual images A and B with the same structure.

[0042] The residual image A and the high-resolution edge feature image are added element by element to obtain the restored image;

[0043] The step of performing image correction on the restored image to obtain a reconstructed edge prior image includes:

[0044] The restored image is processed by attention mask feature processing using a pre-constructed convolutional layer J and a preset activation function C to obtain a pixel-level attention mask for each pixel in the restored image.

[0045] The pixel-level attention mask is multiplied element-wise with the residual image B by a pre-constructed convolutional layer K to obtain local image guided attention features;

[0046] The attention features of the local image are normalized using a preset activation function D to obtain a normalized image. The normalized image is then added element-wise to the reconstructed high-resolution edge image to obtain a prior image of the reconstructed edge.

[0047] The beneficial effects of adopting the above-mentioned further scheme are as follows: by establishing a mapping from the edge branch network to the super-resolution branch network, the consistency of the edge branch network and the super-resolution branch network in spatial structure is maintained, which lays the foundation for adding the normalized image and the reconstructed high-resolution edge image element by element to obtain the reconstructed edge prior image; step S4 obtains the reconstructed edge prior image by performing high-frequency information correction on the high-resolution edge feature image and the reconstructed high-resolution edge image, which provides supervision information for the reconstruction of satellite images and lays the foundation for obtaining the reconstructed super-resolution image.

[0048] Further, in step S5, the reconstructed edge prior image and the reconstructed high-resolution image are fused to obtain a fused image, including:

[0049] The reconstructed edge prior image and the reconstructed high-resolution image are stitched together in spatial dimension to obtain a fused image;

[0050] In step S5, the fused image is subjected to dimensionality reduction processing to obtain a reconstructed super-resolution image, including:

[0051] The fused image is spatially reduced using a pre-constructed convolutional layer L to obtain a reconstructed super-resolution image.

[0052] The beneficial effect of adopting the above-mentioned further scheme is that by stitching the reconstructed edge prior image and the reconstructed high-resolution image in spatial dimension, and then performing spatial dimensionality reduction operation on the stitched fused image, the resulting reconstructed super-resolution image has good visual effect and high quality.

[0053] To address the aforementioned technical problems, the present invention also provides a satellite image super-resolution reconstruction system, the system comprising:

[0054] The image preprocessing module is used to acquire high-resolution satellite images, convert the high-resolution satellite images into low-resolution images, and perform edge extraction on the low-resolution images to obtain low-resolution edge images.

[0055] The upsampling reconstruction module is used to upsample the low-resolution edge image to obtain a high-resolution edge feature image.

[0056] The feature extraction module is used to extract texture features from the low-resolution edge image through a pre-established edge branch network to obtain a reconstructed high-resolution edge image; and to extract structural features from the low-resolution image through a pre-established super-resolution branch network to obtain a reconstructed high-resolution image.

[0057] The interactive supervised attention module determines the restored image based on the high-resolution edge feature image and the reconstructed high-resolution edge image, performs image correction on the restored image, and obtains the reconstructed edge prior image.

[0058] The fusion module is used to fuse images to obtain a fused image;

[0059] The spatial dimensionality reduction module is used to reduce the dimensionality of the image to obtain a reconstructed super-resolution image;

[0060] The image preprocessing module, feature extraction module, interactive supervised attention module, fusion module, and spatial dimensionality reduction module are connected in sequence. The input of the upsampling reconstruction module is connected to the output of the image preprocessing module, and the output of the upsampling reconstruction module is connected to the input of the interactive supervised attention module.

[0061] The beneficial effects of this invention are: this system can realize super-resolution reconstruction of satellite images, and the reconstructed super-resolution images have good visual effects and high quality. It solves the problem that existing edge-prior-based image super-resolution reconstruction methods only consider the edge information of the image in high-resolution space or only consider the edge information of the image in low-resolution space when performing super-resolution reconstruction of satellite images, and do not make full use of the image edge prior information, resulting in poor visual effects and low quality of the generated super-resolution reconstructed images.

[0062] Furthermore, the pre-established edge branch network and the pre-established super-resolution branch network have the same structure, and the edge branch network runs in parallel with the super-resolution branch network; the pre-established edge branch network includes a first shallow feature extraction module, a first deep feature extraction module, and a first upsampling reconstruction module, used to extract texture features from the low-resolution edge image to obtain a reconstructed high-resolution edge image; the pre-established super-resolution branch includes a second shallow feature extraction module, a second deep feature extraction module, and a second upsampling reconstruction module, used to extract structural features from the low-resolution image to obtain a reconstructed high-resolution image;

[0063] Both the first shallow feature extraction module and the second shallow feature extraction module include a convolutional layer A, which is used to extract shallow features from the image;

[0064] Both the first deep feature extraction module and the second deep feature extraction module include multiple basic blocks and a convolutional layer B, which are used to extract deep features from the image.

[0065] Each of the basic blocks has the same structure, and each of the basic blocks is connected sequentially. Each basic block includes a multi-scale global prior block and a densely connected module connected sequentially.

[0066] Each of the multi-scale global prior blocks includes multiple upsampling and downsampling units of different scales and a convolutional layer D. The multiple upsampling and downsampling units contained in each multi-scale global prior block are used in parallel to acquire image features at different scales.

[0067] Each dense connection module includes multiple structurally identical dense connection units and a convolutional layer F. Each dense connection unit is sequentially connected to the convolutional layer F. The output of the convolutional layer F is connected to the input of the convolutional layer B. Each dense connection unit includes a dense connection block, a convolutional layer E, and an activation function A, all connected in sequence. The inputs and outputs of each dense connection unit within each dense connection module, the outputs of preceding dense connection units and the inputs of subsequent dense connection units, and the outputs of preceding dense connection units and subsequent dense connection units are connected in a skip connection. The input of each dense connection module is also connected in a skip connection to the outputs of each dense connection unit it contains.

[0068] Each dense connection block includes multiple dense blocks with identical structures and a convolutional layer e1. Each dense block is sequentially connected to the convolutional layer e1. Each dense block includes a convolutional layer e2 and an activation function B. The inputs and outputs of each dense block contained in each dense connection block, the outputs of preceding dense blocks and the inputs of subsequent dense blocks, and the outputs of preceding dense blocks and subsequent dense blocks are skipped. The inputs of each dense connection block are skipped to the outputs of each of its contained dense blocks.

[0069] Both the first upsampling reconstruction module and the second upsampling reconstruction module include an upsampling layer, a convolutional layer G1, and a convolutional layer G2 connected in sequence, which are used to perform upsampling reconstruction processing on the image.

[0070] The beneficial effects of adopting the above-mentioned further scheme are as follows: the deep feature image is used as the input of the first upsampling reconstruction module and the second upsampling reconstruction module. The image features are upsampled and feature detail information is extracted by passing through the upsampling layer, convolutional layer G1 and convolutional layer G2 in sequence to obtain the reconstructed high-resolution edge image and the reconstructed high-resolution image. The reconstructed high-resolution edge image provides supervision for the reconstructed high-resolution image to assist in the reconstruction of the super-resolution image.

[0071] Furthermore, the interactive supervised attention module includes:

[0072] The image restoration unit includes a convolutional layer H and a convolutional layer I, used to extract image features from the reconstructed high-resolution edge image through the convolutional layer H and the convolutional layer I to generate residual images A and B with the same structure; and to add the residual image A and the high-resolution edge feature image element by element to obtain the restored image;

[0073] The attention generation unit includes a convolutional layer J, a convolutional layer K, and an activation function C. It is used to perform attention mask feature processing on the restored image through the convolutional layer J and the activation function C to obtain a pixel-level attention mask corresponding to each pixel in the restored image; and to multiply the pixel-level attention mask with the residual image B element-wise through the convolutional layer K to obtain local image guided attention features.

[0074] The image correction unit includes an activation function D, which is used to normalize the attention-guided features of the local image through the activation function D to obtain a normalized image. The normalized image is then added element-wise to the matrix elements at the same position of the reconstructed high-resolution edge image to obtain a reconstructed edge prior image.

[0075] The beneficial effect of adopting the above-mentioned further scheme is that by performing high-frequency information correction on the edge branch network through the interactive supervised attention module, a more accurate reconstructed edge prior image can be obtained.

[0076] To address the aforementioned technical problems, the present invention also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the satellite image super-resolution reconstruction method as described above.

[0077] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the satellite image super-resolution reconstruction method as described above.

[0078] It should be noted that, in this invention, the images obtained after feature extraction all refer to feature maps. Attached Figure Description

[0079] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0080] Figure 2 This is a schematic diagram of the system structure of the present invention;

[0081] Figure 3 This is a schematic diagram of the system flow of the present invention;

[0082] Figure 4 This is a schematic diagram of the structure of the multi-scale global prior block in this invention;

[0083] Figure 5 This is a schematic diagram of the upsampling and downsampling unit in this invention;

[0084] Figure 6 This is a schematic diagram of the dense connection module in this invention;

[0085] Figure 7 This is a schematic diagram of the interactive supervised attention module in this invention;

[0086] Figure 8 This diagram illustrates the results of image reconstruction using the method of the present invention and existing image super-resolution reconstruction methods on image data A in the public satellite image dataset Draper.

[0087] Figure 9 This diagram illustrates the results of image reconstruction using the method of the present invention and existing image super-resolution reconstruction methods on image data B in the public satellite image dataset Draper.

[0088] Figure 10This diagram illustrates the results of image reconstruction using the method of the present invention and existing image super-resolution reconstruction methods on image data A in the public satellite image dataset NWPU-RESISC45.

[0089] Figure 11 This diagram illustrates the results of image reconstruction using the method of the present invention and existing image super-resolution reconstruction methods on image data B in the public satellite image dataset NWPU-RESISC45.

[0090] Figure 12 Schematic diagram A shows the results of image reconstruction of SuperView-1 satellite images using the method of the present invention and existing image super-resolution reconstruction methods;

[0091] Figure 13 Schematic diagram B shows the results of image reconstruction of SuperView-1 satellite images using the method of the present invention and existing image super-resolution reconstruction methods;

[0092] Figure 14 This is a schematic diagram of the electronic device structure of the present invention. Detailed Implementation

[0093] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0094] Example 1

[0095] This embodiment provides a method for super-resolution reconstruction of satellite images, such as... Figure 1 As shown, it includes:

[0096] Step S1: Acquire a high-resolution satellite image, convert the high-resolution satellite image into a low-resolution image, and perform edge extraction on the low-resolution image to obtain a low-resolution edge image.

[0097] The conversion of the high-resolution satellite image into a low-resolution image specifically includes:

[0098] The high-resolution satellite image is subjected to image block processing to obtain a block-processed high-resolution satellite image; wherein, block processing refers to dividing the image into multiple sub-images by cropping, that is, dividing the high-resolution satellite image into high-resolution satellite images;

[0099] The high-resolution image after block processing is subjected to bicubic interpolation to obtain a low-resolution image.

[0100] Edge extraction is performed on the low-resolution image to obtain a low-resolution edge image, specifically including:

[0101] The low-resolution image is sharpened using the Laplacian operator to obtain a sharpened image. The boundary positions in the sharpened image are marked to obtain a Laplacian image. The Laplacian image is then superimposed on the low-resolution image to obtain a low-resolution edge image.

[0102] Step S2: Upsample the low-resolution edge image to obtain a high-resolution edge feature image.

[0103] Step S2 specifically includes:

[0104] Feature upsampling is performed on low-resolution edge images using a pre-constructed upsampling layer;

[0105] By sequentially extracting features from the low-resolution edge image after feature upsampling using pre-constructed convolutional layers G1 and G2, a high-resolution edge feature image is obtained. In this embodiment, the size of both convolutional layers G1 and G2 is 3×3.

[0106] Step S3: Extract texture features from the low-resolution edge image using a pre-established edge branch network to obtain a reconstructed high-resolution edge image; extract structural features from the low-resolution image using a pre-established super-resolution branch network to obtain a reconstructed high-resolution image.

[0107] Texture features are extracted from the low-resolution edge image using a pre-established edge branching network to obtain a reconstructed high-resolution edge image, specifically including:

[0108] 1. Shallow Feature Extraction: Features are extracted from the low-resolution edge image using a pre-constructed convolutional layer A to obtain a shallow feature image. In this embodiment, the size of the convolutional layer A is 3×3.

[0109] 2. Deep Feature Extraction:

[0110] 2.1. For each of the pre-established basic blocks, deep feature extraction is performed on the shallow feature image through the basic block to obtain a fused feature image, specifically:

[0111] For each of the basic blocks, the basic block includes a multi-scale global prior block and a dense connection module connected in sequence. Each multi-scale global prior block includes multiple upsampling and downsampling units of different scales and a convolutional layer D. The multiple upsampling and downsampling units are parallel. The dense connection module includes multiple dense connection units with the same structure and a convolutional layer F. In this embodiment, the number of basic blocks is 6, the number of upsampling and downsampling units in each multi-scale global prior block is 3, the number of dense connection units in each dense connection module is 3, and the size of the convolutional layer F is 1×1.

[0112] The step of extracting deep features from the shallow feature image using the basic blocks to obtain a fused feature image includes:

[0113] The shallow feature image is extracted by each of the upsampling and downsampling units in the basic block to obtain three feature images;

[0114] The three feature images are added element by element to obtain the first multi-scale image;

[0115] The first multi-scale image is reduced in dimensionality by the convolutional layer D in the basic block to obtain the second multi-scale feature image; in this embodiment, the size of the convolutional layer D is 1×1.

[0116] By performing feature fusion on the second multi-scale feature image through each of the dense connection units in the basic block, six third multi-scale feature images are obtained;

[0117] The six third multi-scale feature images are dimensionality-reduced by the convolutional layer F in the basic block to obtain dense feature images;

[0118] The second multi-scale feature image and the dense feature image are added element by element to obtain a fused feature image.

[0119] 2.2. Feature extraction is performed on each of the fused feature images using a pre-constructed convolutional layer B to obtain a deep feature image. In this embodiment, the size of the convolutional layer B is 3×3.

[0120] 3. Upsampling and reconstruction processing:

[0121] 3.1. The deep feature image is upsampled by a pre-constructed upsampling layer to obtain the upsampled deep feature image;

[0122] 3.2. The deep feature image after feature upsampling is sequentially extracted using pre-constructed convolutional layers G1 and G2 to obtain a reconstructed high-resolution edge image; in this embodiment, the size of both convolutional layer G1 and convolutional layer G2 is 3×3.

[0123] The low-resolution image is reconstructed by extracting structural features from a pre-established super-resolution branch network. Specifically, this involves sequentially performing shallow feature extraction, deep feature extraction, and upsampling reconstruction on the low-resolution image. The principle of reconstructing a high-resolution image from a low-resolution image is the same as that of reconstructing a high-resolution edge image from a low-resolution edge image; the similarities will not be elaborated further.

[0124] Step S4: Based on the high-resolution edge feature image and the reconstructed high-resolution edge image, determine the restored image, perform image correction on the restored image, and obtain the reconstructed edge prior image.

[0125] Specifically, determining the restored image based on the high-resolution edge feature image and the reconstructed high-resolution edge image includes:

[0126] The reconstructed high-resolution edge images are processed by pre-constructed convolutional layers H and I to extract image features, generating residual images A and B with the same structure.

[0127] The residual image A and the high-resolution edge feature image are added element by element to obtain the restored image.

[0128] The step of performing image correction on the restored image to obtain a reconstructed edge prior image specifically includes:

[0129] The restored image is processed by attention mask feature processing using a pre-constructed convolutional layer J and a preset activation function C to obtain a pixel-level attention mask for each pixel in the restored image.

[0130] The pixel-level attention mask is multiplied element-wise with the residual image by a pre-constructed convolutional layer K to obtain local image guided attention features;

[0131] The local image is guided by an activation function D to map attention features to the interval (0, 1) to obtain a normalized image. This normalized image is then element-wise added to the reconstructed high-resolution edge image to obtain a prior image of the reconstructed edge. In this embodiment, the sizes of convolutional layers H, I, J, and K are all 3×3; both activation functions C and D are Sigmoid activation functions.

[0132] Step S5: Fuse the reconstructed edge prior image and the reconstructed high-resolution image to obtain a fused image, and perform dimensionality reduction processing on the fused image to obtain a reconstructed super-resolution image.

[0133] Specifically, fusing the reconstructed edge prior image and the reconstructed high-resolution image to obtain a fused image includes: stitching the reconstructed edge prior image and the reconstructed high-resolution image together in the spatial dimension to obtain a fused image.

[0134] The fused image is subjected to dimensionality reduction processing to obtain a reconstructed super-resolution image. Specifically, this includes: using a pre-constructed convolutional layer L to perform spatial dimensionality reduction on the fused image to obtain the reconstructed super-resolution image. The size of the convolutional layer L is 1×1.

[0135] Simulation Experiment 1:

[0136] In the Draper public satellite image dataset, 1000 public satellite images were selected for training and 200 public satellite images were selected for testing. The resolution of the public satellite images used for training and testing was 3099×2329.

[0137] This method divides the selected public satellite images into blocks: the public satellite images are cropped to a resolution of 192×192 to obtain high-resolution satellite images, and image enhancement processing is performed on the high-resolution satellite images: the high-resolution satellite images are downsampled to a resolution of 48×48 using bicubic interpolation to obtain low-resolution images.

[0138] Under the premise of using the same training and testing data, the hyperparameters (i.e., the parameters in the first column of Table 1) of this method (DCLN) and the existing methods (Bicubic interpolation, Enhanced Deep Super-Resolution Network, Generative Adversarial Network for Image Super-Resolution (SR) (SRResNet), Deep Residual Channel Attention Network (RCAN), Multi-Scale Residual Network (MSRN), Innovative Hybrid High-Order Attention Network for SR (MHAN), Edge Enhancement Subnetwork (EEGAN), and Soft Edge Auxiliary Network (SeaNet)) are adjusted, and the same data distribution is used to optimize the algorithm performance of each method. The corresponding values ​​of this method and the aforementioned existing methods on different parameters are shown in Table 1 below.

[0139] Table 1 - Different parameters and their corresponding values ​​for different methods on the Draper public satellite image dataset.

[0140]

[0141] In Table 1, Param represents the test parameter, in Param / M; FLOPs represents floating-point operations per second, in FLOPs / M; PSNR represents peak signal-to-noise ratio; SSIM represents structural similarity; FISM represents feature similarity; VIF represents visual information fidelity; ERGAS represents relative global dimensionless error; LPIPS represents image perceptual similarity; the method corresponding to superscript A is the one among the aforementioned 9 methods that has the best performance under this parameter; the method corresponding to superscript B is the one among the aforementioned 9 methods that has relatively good performance under this parameter.

[0142] As shown in Table 1, our method outperforms the other eight methods in terms of PSNR, SSIM, FISM, VIF, ERGAS, and LPIPS. In particular, for PSNR, our method achieves a PSNR value of 34.08 dB, which is 0.59 dB higher than the second-best method RCAN's PSNR value of 33.62 dB, indicating that the reconstructed super-resolution image obtained by our method is a high-quality image.

[0143] like Figure 8 and Figure 9 As shown, the boxes in the image represent the key regions. Comparing the results of super-resolution reconstruction of image data from the public satellite image dataset Draper using this method with existing methods, it is evident that the reconstructed super-resolution image obtained by this method is closest to the original image data (i.e., HR in the image). Therefore, it can be concluded that the reconstructed super-resolution image obtained by this method for satellite image super-resolution reconstruction has the advantages of excellent visual effects and high quality. Figure 8 and Figure 9 It can be seen that Bicubic cannot generate additional details. For CNN-based image super-resolution reconstruction methods, such as EDSR, SRResNet, MSRN, and EEGAN, some texture details can be inferred, but due to the global optimization scheme and low feature utilization of these methods, the image contours are blurred. Some attention-based image super-resolution reconstruction methods, such as RCAN and MHAN, produce overly smooth images with artifacts at the edges, and the texture details are also blurred. Comparing the reconstructed super-resolution image obtained using this method with the original image and reconstructed super-resolution images obtained by other methods, it can be seen that this method can reconstruct an image with more realistic texture details.

[0144] Simulation Experiment 2:

[0145] In the NWPU-RESISC45 public satellite image dataset, 1000 public satellite images were selected for training and 200 public satellite images were selected for testing. The resolution of the public satellite images used for training and testing was 3099×2329.

[0146] This method crops the selected public satellite image to a resolution of 192×192 to obtain a high-resolution satellite image, and then performs bicubic interpolation downsampling on the high-resolution satellite image to a resolution of 48×48 to obtain a low-resolution image.

[0147] Under the premise of using the same training and testing data, the hyperparameters of this method and the existing Bicubic, EDSR, SRResNet, RCAN, MSRN, MHAN, EEGAN and SeaNet (i.e. the parameters in the first column of Table 2) are adjusted, and the same data distribution is used to optimize the algorithm performance of each method. The corresponding values ​​of this method and the aforementioned existing methods on different parameters are shown in Table 2 below.

[0148] Table 2 - Different parameters and their corresponding values ​​for different methods on the public satellite imagery dataset NWPU-RESISC45

[0149]

[0150]

[0151] In Table 2, the meanings of the superscripts for each parameter, its unit, and its value are the same as those for each parameter, its unit, and its value in Table 1. The similarities will not be repeated here.

[0152] As shown in Table 2, our method outperforms the other eight methods in terms of PSNR, SSIM, FISM, VIF, and ERGAS. In particular, for PSNR, our method achieves a PSNR value of 28.93 dB, which is 0.28 dB higher than the second-best method RCAN's PSNR value of 28.65 dB, indicating that the reconstructed super-resolution image obtained by our method is a high-quality image.

[0153] like Figure 10 and Figure 11 As shown, the magnified area in the image is the key region. Comparing the results of super-resolution reconstruction of image data from the public satellite image dataset NWPU-RESISC45 using this method with existing methods, it can be seen that the reconstructed super-resolution image obtained by this method is closest to the original image data (i.e., HR in the figure). Therefore, it can be concluded that the reconstructed super-resolution image obtained by this method for satellite image super-resolution reconstruction has the advantages of good visual effect and high quality. Figure 10 and Figure 11 As can be seen, in the magnified area of ​​the image, other methods besides this one produce obvious jagged edges and background blur, and also introduce artifacts and distortion effects. However, by processing satellite images with this method, not only are texture details better preserved and enhanced, but texture detail information is also effectively preserved and enhanced, and sharp edges and clear image content are displayed.

[0154] Simulation Experiment 3:

[0155] This method is compared with existing methods such as Bicubic, EEGAN, and MHAN to analyze two SuperView-1 satellite images (i.e., Figure 12 and Figure 13 Satellite image super-resolution reconstruction is performed on the image on the left side of the image, such as... Figure 12 and Figure 13 As shown, the areas marked with rectangles in the SuperView-1 satellite image are key areas. Figure 12 and Figure 13 The DCLN in the image represents the final reconstructed super-resolution image obtained by downsampling the key region at a scale of 4 using this method. Comparing the super-resolution reconstruction results of SuperView-1 satellite images using this method with those of existing methods, it is evident that Bicubic, EEGAN, and MHAN reconstruction results contain obvious artifacts. In contrast, the reconstructed super-resolution image obtained using this method displays fine-grained texture details at the image edges, and the image content is clearer. This demonstrates that this method has a powerful artifact removal capability.

[0156] Example 2

[0157] This embodiment provides a satellite image super-resolution system, such as Figure 2 and Figure 3 As shown, it includes:

[0158] (1) Image preprocessing module, used to acquire high-resolution satellite images, convert the high-resolution satellite images into low-resolution images, extract edges from the low-resolution images, and obtain low-resolution edge images.

[0159] The image preprocessing module specifically includes:

[0160] Image acquisition unit, used to acquire high-resolution satellite images;

[0161] The block processing unit is used to perform image block processing on the high-resolution satellite image to obtain a block-processed high-resolution satellite image.

[0162] An image enhancement unit is used to perform bicubic interpolation on the high-resolution image after block processing to obtain a low-resolution image.

[0163] Edge extraction unit (i.e.) Figure 3 The low-resolution image (EEM) is used to sharpen the low-resolution image using the Laplacian operator to obtain a sharpened image. The boundary positions in the sharpened image are marked to obtain a Laplacian image. The Laplacian image is then superimposed on the low-resolution image to obtain a low-resolution edge image.

[0164] (2) Upsampling reconstruction module (i.e.) Figure 3The bottommost Upscale is used to upsample the low-resolution edge image to obtain a high-resolution edge feature image.

[0165] The upsampling reconstruction module includes an upsampling layer, a convolutional layer G1, and a convolutional layer G2 connected in sequence, used for upsampling and reconstruction processing of the image. In this embodiment, the size of both the convolutional layer G1 and the convolutional layer G2 is 3×3.

[0166] (3) Feature extraction module, used to extract texture features from the low-resolution edge image through a pre-established edge branch network to obtain a reconstructed high-resolution edge image; and to extract structural features from the low-resolution image through a pre-established super-resolution branch network to obtain a reconstructed high-resolution image.

[0167] Among them, the pre-established edge branch network (i.e. Figure 3 The Edge branch in the network and the pre-established super-resolution branch network (i.e. Figure 3 The edge branch network has the same structure as the super-resolution branch network (SRbranch) in the image, and the edge branch network runs in parallel with the super-resolution branch network. The pre-established edge branch network includes a first shallow feature extraction module, a first deep feature extraction module, and a first upsampling reconstruction module, which are used to extract texture features from the low-resolution edge image to obtain a reconstructed high-resolution edge image. The pre-established super-resolution branch includes a second shallow feature extraction module, a second deep feature extraction module, and a second upsampling reconstruction module, which are used to extract structural features from the low-resolution image to obtain a reconstructed high-resolution image.

[0168] The first shallow feature extraction module (i.e. Figure 3 The lower-middle SFEM), the second shallow feature extraction module (i.e. Figure 3 The upper and middle SFEMs each include a convolutional layer A for shallow feature extraction of the image; in this embodiment, the size of the convolutional layer A is 3×3.

[0169] The first deep feature extraction module (i.e. Figure 3 The lower-middle DFEM), the second deep feature extraction module (i.e. Figure 3 The upper and middle DFEMs each consist of multiple basic blocks and a convolutional layer B, used for deep feature extraction of the image; in this embodiment, the basic blocks (i.e. Figure 3 The number of Basic Blocks in the convolutional layer is 6, and the size of the convolutional layer B is 3×3.

[0170] Each of the basic blocks has the same structure, and each of the basic blocks is connected sequentially. Each basic block includes a multi-scale global prior block and a densely connected module connected sequentially.

[0171] like Figure 4 and Figure 5 As shown, each of the multi-scale global prior blocks (MGPB) includes multiple upsampling and downsampling units (UDBs) of different scales and a convolutional layer D. The multiple upsampling and downsampling units contained in each MGPB are parallel and used to acquire image features at different scales. Each upsampling and downsampling unit includes a convolutional layer C1, an upsampling sub-unit, a convolutional layer C2, a downsampling sub-unit, and a convolutional layer C3 connected in sequence. The upsampling and downsampling sub-units contained in each upsampling and downsampling unit have the same scale. The input of each upsampling and downsampling unit in the edge branch network is the output of the first shallow feature extraction module, and the output of each upsampling and downsampling unit in the edge branch network is connected to the convolutional layer D. In this embodiment, each MGPB contains 3 upsampling and downsampling units. The size of the convolutional layers C1, C2, and C3 is 3×3, and the size of the convolutional layer D is 1×1.

[0172] like Figure 6 As shown, each dense connection module includes multiple densely connected units (DenseBlocks) with identical structures and a convolutional layer F. Each densely connected unit is sequentially connected to the convolutional layer F. The output of the convolutional layer F is connected to the input of the convolutional layer B. Each densely connected unit includes a densely connected block, a convolutional layer E, and an activation function A, all connected sequentially. The input and output of each densely connected unit in each dense connection module, the output of a preceding densely connected unit and the input of a subsequent densely connected unit, and the output of a preceding densely connected unit and the output of a subsequent densely connected unit are connected in a skip connection. The input of each densely connected module is also connected in a skip connection to the output of each densely connected unit it contains. In this embodiment, each dense connection module contains 3 densely connected units, the size of the convolutional layer E is 3×3, the size of the convolutional layer F is 1×1, and the activation function A is the LReLU activation function.

[0173] Each dense connection block includes multiple structurally identical dense blocks and a convolutional layer e1, with each dense block sequentially connected to the convolutional layer e1. Each dense block includes a convolutional layer e2 and an activation function B, with the convolutional layer e2 connected to the activation function B. The inputs and outputs of each dense block within each dense connection block, the outputs of preceding dense blocks and the inputs of subsequent dense blocks, and the outputs of preceding and subsequent dense blocks are skip-connected. The input of each dense connection block is skip-connected to the outputs of each of its contained dense blocks. In this embodiment, each dense connection block contains four dense blocks; the size of convolutional layer e1 is 1×1, the size of convolutional layer e2 is 3×3, and the activation function B is the LReLU activation function.

[0174] After the shallow feature image enters the multi-scale global prior block, it first enters three parallel upsampling and downsampling units of different scales to extract image features at different scales: After entering each upsampling and downsampling unit, image features are first extracted through the convolutional layer C1, then the image features extracted by the convolutional layer C1 are upsampled through the upsampling subunit, then the upsampled features are extracted through the convolutional layer C2, then the features are downsampled through the downsampling subunit with the same scale as the upsampling subunit, and finally the image features are extracted through the convolutional layer C3 to output the feature image. The scales of the upsampling and downsampling subunits can be set by the user. In this embodiment, the scales of the upsampling subunits and the downsampling subunits included in the three upsampling and downsampling units are 2×, 4×, and 8×, respectively. Then, the feature images output by the three upsampling and downsampling units are added element-wise to obtain the first multi-scale image, realizing multi-scale feature fusion (i.e., Figure 4 The first multi-scale image is then reduced in dimensionality using the convolutional layer D to obtain the second multi-scale feature image.

[0175] The second multi-scale feature image is input into the dense connection module. After entering the dense connection module, the image is fed into each of the dense connection units in parallel. Each dense connection unit uses skip connections to perform feature fusion on the second multi-scale feature image to obtain six third multi-scale feature images. The convolutional layer e1 and the convolutional layer F are used to perform dimensionality reduction on the image, and the convolutional layer e2 and the convolutional layer E are used to extract features from the image. The third multi-scale feature images output by each of the dense connection units are subjected to dimensionality reduction by the convolutional layer F to obtain dense feature images.

[0176] The second multi-scale feature image and the dense feature image are added element-wise to obtain a fused feature image. The fused feature image output from each of the basic blocks is input into the convolutional layer B. Feature extraction is performed through the convolutional layer B to obtain a deep feature image.

[0177] The first upsampling reconstruction module (i.e. Figure 3 The URM located in the Edge branch region), and the second upsampling reconstruction module (i.e. Figure 3 The URM located in the SRbranch region includes an upsampling layer, a convolutional layer G1, and a convolutional layer G2 connected in sequence, which are used to perform upsampling reconstruction processing on the image.

[0178] (4) Interactive supervised attention module (i.e.) Figure 3 The ISAM (In the high-resolution edge feature image) determines the restored image based on the high-resolution edge feature image and the reconstructed high-resolution edge image, performs image correction on the restored image, and obtains the reconstructed edge prior image.

[0179] like Figure 7 As shown, the interactive supervised attention module specifically includes:

[0180] The image restoration unit includes a convolutional layer H and a convolutional layer I, used to reconstruct the high-resolution edge image (i.e., ...) through the convolutional layer H and the convolutional layer I. Figure 7 Image feature extraction is performed on Img1 in the image to generate residual images A and B with the same structure; and the residual image A and the high-resolution edge feature image (i.e. Figure 7 The convolutional layer H and the convolutional layer I are added element by element to obtain the restored image; in this embodiment, the size of the convolutional layer H and the convolutional layer I are both 3×3.

[0181] The attention generation unit, comprising a convolutional layer J, a convolutional layer K, and an activation function C, is used to perform attention mask feature processing on the restored image through the convolutional layer J and the activation function C to obtain a pixel-level attention mask (i.e., ...) for each pixel in the restored image. Figure 7 M1); and the pixel-level attention mask is multiplied element-wise with the residual image B through the convolutional layer K to obtain local image guided attention features; in this embodiment, the size of the convolutional layer J and the convolutional layer K are both 3×3, and the activation function C is the Sigmoid activation function;

[0182] The image correction unit includes an activation function D, used to normalize the attention-guided features of the local image through the activation function D to obtain a normalized image. The normalized image is then added element-wise to the matrix elements at the same positions in the reconstructed high-resolution edge image to obtain a reconstructed edge prior image (i.e.,...). Figure 3 ReconstructedImage in Figure 7 In this embodiment, the activation function D is the Sigmoid activation function.

[0183] (5) Fusion module (i.e.) Figure 3 The MM in the image is used to fuse images to obtain a fused image.

[0184] The fusion module stitches the reconstructed edge prior image and the reconstructed high-resolution image together in the spatial dimension to obtain a fused image.

[0185] (6) Spatial dimensionality reduction module (i.e. Figure 3 The SDRM (Simplified Degradation Reconstruction Mechanism) is used to reduce the dimensionality of an image and obtain a reconstructed super-resolution image.

[0186] The spatial dimensionality reduction module uses a pre-constructed convolutional layer L to perform spatial dimensionality reduction on the fused image to obtain a reconstructed super-resolution image.

[0187] The image preprocessing module, feature extraction module, interactive supervised attention module, fusion module, and spatial dimensionality reduction module are connected in sequence. The input of the upsampling reconstruction module is connected to the output of the image preprocessing module, and the output of the upsampling reconstruction module is connected to the input of the interactive supervised attention module.

[0188] The working principle of this system is as follows: First, a high-resolution satellite image is acquired using the image acquisition unit. The high-resolution satellite image is then segmented using the block processing unit. Next, the high-resolution image after block processing is subjected to bicubic interpolation using the image enhancement unit to obtain a low-resolution image. Then, the low-resolution image is divided into two paths. One path is first fed into the edge extraction module for edge extraction processing to obtain a low-resolution edge image, which is then fed into the edge branch network and the upsampling reconstruction module for image processing. The other path is directly fed into the super-resolution branch network for image processing. The edge branch network and the super-resolution branch network have the same structure. The edge branch network processes the low-resolution edge image in parallel, and the super-resolution branch network processes the low-resolution image in parallel, i.e., the low-resolution edge image is processed separately by the edge branch network and the super-resolution branch network. The edge image and the low-resolution image undergo shallow feature extraction, deep feature extraction, and upsampling reconstruction processing to obtain corresponding reconstructed high-resolution edge images and reconstructed high-resolution images. The upsampling reconstruction module performs upsampling reconstruction processing on the low-resolution edge image to obtain a high-resolution edge feature image. Then, the reconstructed high-resolution edge image and the high-resolution edge feature image are input into the interactive supervised attention module for processing to obtain a reconstructed edge prior image. In this system, the reconstruction process of the edge branch network provides supervision for the reconstruction of the super-resolution branch network. The obtained reconstructed edge prior image and the reconstructed high-resolution image obtained by the super-resolution branch network are then input into the fusion module for processing to obtain a fused image. Finally, the fused image is input into the spatial dimensionality reduction module for processing to obtain a reconstructed super-resolution image.

[0189] Example 3

[0190] This embodiment provides an electronic device, such as... Figure 14 As shown, the device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the satellite image super-resolution reconstruction method as described in Embodiment 1.

[0191] Example 4

[0192] This embodiment provides a computer-readable storage medium, such as... Figure 14 As shown, a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements the satellite image super-resolution reconstruction method as described in Embodiment 1.

[0193] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for super-resolution reconstruction of satellite images, characterized in that, include: Step S1: Acquire a high-resolution satellite image, convert the high-resolution satellite image into a low-resolution image, and perform edge extraction on the low-resolution image to obtain a low-resolution edge image; Step S2: Upsample the low-resolution edge image to obtain a high-resolution edge feature image; Step S3: Extract texture features from the low-resolution edge image using a pre-established edge branching network to obtain a reconstructed high-resolution edge image; extract structural features from the low-resolution image using a pre-established super-resolution branching network to obtain a reconstructed high-resolution image; Step S4: Based on the high-resolution edge feature image and the reconstructed high-resolution edge image, determine the restored image, perform image correction on the restored image, and obtain the reconstructed edge prior image; Step S5: Fuse the reconstructed edge prior image and the reconstructed high-resolution image to obtain a fused image, and perform dimensionality reduction processing on the fused image to obtain a reconstructed super-resolution image.

2. The satellite image super-resolution reconstruction method according to claim 1, characterized in that, In step S1, converting the high-resolution satellite image into a low-resolution image includes: The high-resolution satellite image is divided into blocks to obtain a block-processed high-resolution satellite image. The high-resolution image after block processing is subjected to bicubic interpolation to obtain a low-resolution image; In step S1, edge extraction is performed on the low-resolution image to obtain a low-resolution edge image, including: The low-resolution image is sharpened using the Laplacian operator to obtain a sharpened image. The boundary positions in the sharpened image are marked to obtain a Laplacian image. The Laplacian image is then superimposed on the low-resolution image to obtain a low-resolution edge image.

3. The satellite image super-resolution reconstruction method according to claim 1, characterized in that, In step S3, texture features are extracted from the low-resolution edge image using a pre-established edge branching network to obtain a reconstructed high-resolution edge image, including: The low-resolution edge image is used to extract features using a pre-constructed convolutional layer A to obtain a shallow feature image. The size of the convolutional layer A is 3×3. For each of the pre-established basic blocks, deep feature extraction is performed on the shallow feature image through the basic block to obtain a fused feature image; The deep feature image is obtained by extracting features from each of the fused feature images through a pre-constructed convolutional layer B, wherein the size of the convolutional layer B is 3×3. The deep feature image is upsampled by a pre-constructed upsampling layer to obtain an upsampled deep feature image. By sequentially extracting features from the upsampled deep feature image using pre-constructed convolutional layers G1 and G2, a reconstructed high-resolution edge image is obtained. The size of both convolutional layers G1 and G2 is 3×3.

4. The satellite image super-resolution reconstruction method according to claim 3, characterized in that, For each of the basic blocks, the basic block includes a multi-scale global prior block and a dense connection module connected in sequence. Each multi-scale global prior block includes multiple upsampling and downsampling units of different scales and a convolutional layer D. The multiple upsampling and downsampling units are parallel. The dense connection module includes multiple dense connection units with the same structure and a convolutional layer F. The size of the convolutional layer D is 1×1, and the size of the convolutional layer F is 1×1. The step of extracting deep features from the shallow feature image using the basic blocks to obtain a fused feature image includes: Each of the upsampling and downsampling units in the basic block is used to extract features from the shallow feature image to obtain multiple feature images; The first multi-scale image is obtained by adding the multiple feature images element by element. The first multi-scale image is reduced in dimensionality by the convolutional layer D in the basic block to obtain the second multi-scale feature image. By performing feature fusion on the second multi-scale feature image through each of the dense connection units in the basic block, multiple third multi-scale feature images are obtained; The convolutional layer F in the basic block is used to reduce the dimensionality of multiple third multi-scale feature images to obtain dense feature images. The second multi-scale feature image and the dense feature image are added element by element to obtain a fused feature image.

5. A satellite image super-resolution reconstruction method according to any one of claims 1 to 4, characterized in that, In step S4, determining the restored image based on the high-resolution edge feature image and the reconstructed high-resolution edge image includes: The reconstructed high-resolution edge images are processed by pre-constructed convolutional layers H and I to extract image features, generating residual images A and B with the same structure. The residual image A and the high-resolution edge feature image are added element by element to obtain the restored image; The step of performing image correction on the restored image to obtain a reconstructed edge prior image includes: The restored image is processed by attention mask feature processing using a pre-constructed convolutional layer J and a preset activation function C to obtain a pixel-level attention mask for each pixel in the restored image. The pixel-level attention mask is multiplied element-wise with the residual image B by a pre-constructed convolutional layer K to obtain local image guided attention features; The attention-guided features of the local image are normalized using a preset activation function D to obtain a normalized image. The normalized image is then added element-wise to the reconstructed high-resolution edge image to obtain a prior image of the reconstructed edge. The size of the convolutional layer H, the convolutional layer I, the convolutional layer J, and the convolutional layer K is 3×3. The activation functions C and D are both Sigmoid activation functions.

6. A satellite image super-resolution reconstruction method according to any one of claims 1 to 4, characterized in that, In step S5, the reconstructed edge prior image and the reconstructed high-resolution image are fused to obtain a fused image, including: The reconstructed edge prior image and the reconstructed high-resolution image are stitched together in spatial dimension to obtain a fused image; In step S5, the fused image is subjected to dimensionality reduction processing to obtain a reconstructed super-resolution image, including: The fused image is spatially reduced using a pre-constructed convolutional layer L to obtain a reconstructed super-resolution image. The size of the convolutional layer L is 1×1.

7. A satellite image super-resolution reconstruction system, characterized in that, include: The image preprocessing module is used to acquire high-resolution satellite images, convert the high-resolution satellite images into low-resolution images, and perform edge extraction on the low-resolution images to obtain low-resolution edge images. The upsampling reconstruction module is used to upsample the low-resolution edge image to obtain a high-resolution edge feature image. The feature extraction module is used to extract texture features from the low-resolution edge image through a pre-established edge branch network to obtain a reconstructed high-resolution edge image; and to extract structural features from the low-resolution image through a pre-established super-resolution branch network to obtain a reconstructed high-resolution image. The interactive supervised attention module determines the restored image based on the high-resolution edge feature image and the reconstructed high-resolution edge image, performs image correction on the restored image, and obtains the reconstructed edge prior image. The fusion module is used to fuse images to obtain a fused image; The spatial dimensionality reduction module is used to reduce the dimensionality of the image to obtain a reconstructed super-resolution image; The image preprocessing module, feature extraction module, interactive supervised attention module, fusion module, and spatial dimensionality reduction module are connected in sequence. The input of the upsampling reconstruction module is connected to the output of the image preprocessing module, and the output of the upsampling reconstruction module is connected to the input of the interactive supervised attention module.

8. A satellite image super-resolution reconstruction system according to claim 7, characterized in that, The pre-established edge branch network and the pre-established super-resolution branch network have the same structure, and the edge branch network runs in parallel with the super-resolution branch network. The pre-established edge branch network includes a first shallow feature extraction module, a first deep feature extraction module, and a first upsampling reconstruction module, used to extract texture features from the low-resolution edge image to obtain a reconstructed high-resolution edge image. The pre-established super-resolution branch network includes a second shallow feature extraction module, a second deep feature extraction module, and a second upsampling reconstruction module, used to extract structural features from the low-resolution image to obtain a reconstructed high-resolution image. Both the first shallow feature extraction module and the second shallow feature extraction module include a convolutional layer A, which is used to extract shallow features from the image. The size of the convolutional layer A is 3×3. Both the first deep feature extraction module and the second deep feature extraction module include multiple basic blocks and a convolutional layer B, which are used to extract deep features from the image. The size of the convolutional layer B is 3×3. Each of the basic blocks has the same structure, and each of the basic blocks is connected sequentially. Each basic block includes a multi-scale global prior block and a densely connected module connected sequentially. Each of the multi-scale global prior blocks includes multiple upsampling and downsampling units of different scales and a convolutional layer D. The multiple upsampling and downsampling units contained in each multi-scale global prior block are parallel and used to acquire image features of different scales. The size of the convolutional layer D is 1×1. Each dense connection module includes multiple dense connection units with identical structures and a convolutional layer F. Each dense connection unit is sequentially connected to the convolutional layer F. The output of the convolutional layer F is connected to the input of the convolutional layer B. Each dense connection unit includes a dense connection block, a convolutional layer E, and an activation function A, all connected sequentially. The inputs and outputs of each dense connection unit within each dense connection module, the outputs of preceding dense connection units and the inputs of subsequent dense connection units, and the outputs of preceding and subsequent dense connection units are all skipped. The input of each dense connection module is skipped to the output of each of its dense connection units. The size of the convolutional layer E is 3×3, the size of the convolutional layer F is 1×1, and the activation function A is the LReLU activation function. Each dense connection block includes multiple dense blocks with identical structures and a convolutional layer e1. Each dense block is sequentially connected to the convolutional layer e1. Each dense block includes a convolutional layer e2 and an activation function B. The convolutional layer e2 is connected to the activation function B. The inputs and outputs of each dense block contained in each dense connection block, the outputs of preceding dense blocks and the inputs of subsequent dense blocks, and the outputs of preceding dense blocks and subsequent dense blocks are skipped. The inputs of each dense connection block are skipped to the outputs of each dense block contained within it. The size of the convolutional layer e1 is 1×1, the size of the convolutional layer e2 is 3×3, and the activation function B is the LReLU activation function. Both the first upsampling reconstruction module and the second upsampling reconstruction module include an upsampling layer, a convolutional layer G1, and a convolutional layer G2 connected in sequence, used to perform upsampling reconstruction processing on the image. The size of the convolutional layer G1 and the convolutional layer G2 is 3×3.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the satellite image super-resolution reconstruction method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the satellite image super-resolution reconstruction method as described in any one of claims 1 to 6.

Images