Remote sensing image super-resolution reconstruction method and system based on multi-scale enhancement module

Abstract

The application provides a remote sensing image super-resolution reconstruction method and system based on a multi-scale enhancement module, which comprises the following steps: down-sampling a high-resolution remote sensing image to a target low-resolution remote sensing image, and performing data enhancement to expand a sample library; taking the low-resolution remote sensing image as input, and extracting a remote sensing image rough feature map through a convolution layer; inputting the remote sensing image rough feature map into a first multi-scale enhancement module, using three different convolution kernel size combinations of parallel sub-networks in the multi-scale enhancement module, and realizing efficient fusion through a feature fusion layer; obtaining a fine remote sensing image feature map by connecting multiple multi-scale enhancement modules in series, fusing the remote sensing image rough feature map, and updating a feature map of the target remote sensing image to generate a high-resolution remote sensing image result. The scheme of the application is superior to other remote sensing image super-resolution methods in qualitative and quantitative evaluation, and can produce a higher-precision remote sensing image high-resolution reconstruction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and relates to a remote sensing image super-resolution reconstruction method and system based on a multi-scale enhancement module. Background Art

[0002] Image super-resolution (SR) is a classic problem in the field of computer vision. The goal of image SR is to reconstruct a high-resolution (HR) image with better visual quality and more accurate detail information from a single or multiple low-resolution (LR) images. Remote sensing images can observe ground objects from outer space, providing rich information for monitoring the earth's surface. In recent years, remote sensing images have been widely used in environmental monitoring, resource exploration, land cover classification and other fields. However, due to hardware and environmental factors, the quality and resolution of remote sensing images cannot meet the requirements of actual analysis. Therefore, how to improve the resolution and image quality of remote sensing images has aroused the interest of many researchers, and remote sensing image SR reconstruction technology has become a research hotspot.

[0003] Recently, due to the rapid development of convolutional neural networks (CNNs), deep learning-based SR methods have gradually entered the mainstream. Dong et al. proposed the first CNN-based image SR method, which directly established an end-to-end mapping relationship between LR and HR images. Shi et al. constructed a sub-pixel CNN, which provides a new method to directly and effectively learn the mapping function from LR to HR images. Lei et al. proposed a multi-branch structure to extract local and global features of remote sensing images and achieved good reconstruction results. Xu et al. used a deep memory connection method to combine the image details of remote sensing images with environmental information. Jiang et al. designed an ultra-dense residual network, which utilized the rich long and short line connections in the network to enhance the network's ability to extract remote sensing image features. Dong et al. used enhanced residual blocks and residual channel attention groups to obtain multi-level remote sensing feature information. Lu et al. used the extracted blocks of different scales as multi-scale information input into the network to fuse high-frequency information of different scales to reconstruct HR remote sensing images.

[0004] The aforementioned super-resolution of remote sensing images often overlooks the issue of redundant feature information extracted by deep learning networks. Furthermore, due to complex factors such as image degradation and content diversity, adaptive multi-scale feature information extraction and fusion are more beneficial for super-resolution of remote sensing images. This paper provides a method and system based on a multi-scale enhancement network, utilizing a multi-scale enhancement module to extract multi-scale feature information from remote sensing images, thereby refining this multi-scale feature information to obtain remote sensing images of higher visual quality. Summary of the Invention

[0005] In view of the shortcomings of existing remote sensing image super-resolution reconstruction technology, the purpose of the present invention is to provide a new end-to-end remote sensing image super-resolution solution.

[0006] The present invention provides a remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module, comprising the following steps:

[0007] Step a: downsampling the high-resolution remote sensing image to a target low-resolution remote sensing image, and performing data enhancement on the high-resolution and low-resolution remote sensing image pairs to expand the sample library;

[0008] Step b: using the low-resolution remote sensing image as the input of the deep network, and extracting the rough feature map of the remote sensing image through the convolutional layer;

[0009] Step c: Input the extracted remote sensing image rough feature map into the first multi-scale enhancement module. The multi-scale enhancement module uses a parallel sub-network with three different convolution kernel size combinations to achieve efficient fusion of multi-scale remote sensing image features through a feature fusion layer.

[0010] In step d, a fine remote sensing image feature map is obtained by connecting multiple multi-scale enhancement modules in series. Residual connections are used outside some multi-scale enhancement modules to reduce model complexity to reduce overfitting. The obtained fine remote sensing image feature map is merged with the coarse feature map of the remote sensing image extracted in step b above to update the feature map of the target remote sensing image to produce a high-resolution remote sensing image result.

[0011] Moreover, in step a, the high-resolution remote sensing image is downsampled by bicubic interpolation to generate the corresponding target low-resolution remote sensing image, and then the high-resolution and low-resolution remote sensing image pairs obtained above are enhanced by three enhancement methods: rotation, scaling, and flipping to expand the sample library.

[0012] Moreover, in step b, the low-resolution remote sensing image I of the training set is LR Input to the first 1×1 convolutional layer of the deep network, through which the coarse feature map X of the remote sensing image is extracted.

[0013] Moreover, in step c, the extracted remote sensing image rough feature map X is input into the multi-scale enhancement module. The multi-scale enhancement module uses three parallel sub-networks with different convolution kernel size combinations. Different sub-networks are combined using convolution layers with different convolution kernel sizes to extract features of remote sensing images at different scales. Then, the feature fusion layer is used to achieve efficient fusion of multi-scale remote sensing image features. The feature extraction process of the multi-scale enhancement module is expressed as:

[0014] Conv-a(X)=Conv1×1 (Conv 1×1 (X))+Conv 3×3 (Conv 1×1 (X))

[0015] Conv-b(X)=Conv 3×3 (Conv 1×1 (X))+Conv 5×5 (Conv 1×1 (X))

[0016] Conv-c(X)=Conv 5×5 (Conv 1×1 (X))+Conv 1×1 (Conv 1×1 (X))

[0017] M n (X)=Conv 1×1 (Concat(Conv-a(X),Conv-b(X),Conv-c(X)))

[0018] where Conv-a(), Conv-b(), Conv-c() represent three parallel sub-networks in the multi-scale enhancement module, Conv 1×1 (), Conv 3×3 () and Conv 5×5 () represent 1x1 convolution layer, 3x3 convolution layer and 5x5 convolution layer respectively, Concat() represents feature fusion layer, and M n () represents the output of the nth multi-scale enhancement module.

[0019] Moreover, the implementation of the three parallel sub-networks Conv-a(), Conv-b(), Conv-c() in the multi-scale enhancement module is as follows,

[0020] In the sub-network Conv-a(), the input enters the 1x1 convolution layer first, then enters the 1x1 convolution layer and the 3x3 convolution layer respectively, and then the point-by-point addition layer to get the output;

[0021] In the sub-network Conv-b(), the input enters the 1x1 convolution layer first, then enters the 3x3 convolution layer and the 5x5 convolution layer respectively, and then the point-by-point addition layer to get the output;

[0022] In the sub-network Conv-c(), the input enters the 1x1 convolution layer first, then enters the 1x1 convolution layer and the 5x5 convolution layer respectively, and then the point-by-point addition layer to get the output.

[0023] Moreover, in step d, a fine remote sensing image feature map Y is obtained by connecting multiple multi-scale enhancement modules in series, and it is fused with the coarse feature map X of the remote sensing image extracted in step b above through a point-by-point addition layer to update the feature map of the target remote sensing image to generate a high-resolution remote sensing image result. The L of the training process is SR The loss function is expressed as

[0024]

[0025] Where F() is the mapping relationship of deep learning network training, is the original high-resolution image of the remote sensing image data pair, and || ||1 represents the norm of the matrix.

[0026] On the other hand, the present invention also provides a remote sensing image super-resolution reconstruction system based on a multi-scale enhancement module, which is used to implement the remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module as described above.

[0027] Furthermore, the following modules are included,

[0028] The first module is used to downsample the high-resolution remote sensing image to a target low-resolution remote sensing image, and perform data enhancement on the high-resolution and low-resolution remote sensing image pairs to expand the sample library;

[0029] The second module is used to take the low-resolution remote sensing image as the input of the deep network and extract the rough feature map of the remote sensing image through the convolution layer;

[0030] The third module is used to input the extracted rough feature map of the remote sensing image into the first multi-scale enhancement module. The multi-scale enhancement module uses a parallel sub-network with three different convolution kernel size combinations to achieve efficient fusion of multi-scale remote sensing image features through a feature fusion layer;

[0031] The fourth module is used to obtain a fine remote sensing image feature map by connecting multiple multi-scale enhancement modules in series, using residual connections outside some multi-scale enhancement modules to reduce model complexity to reduce overfitting, and fusing the obtained fine remote sensing image feature map with the coarse feature map of the remote sensing image to update the feature map of the target remote sensing image to produce a high-resolution remote sensing image result.

[0032] Alternatively, it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the remote sensing image super-resolution reconstruction method based on the multi-scale enhancement module as described above.

[0033] Alternatively, it includes a readable storage medium having a computer program stored thereon, and when the computer program is executed, it implements the remote sensing image super-resolution reconstruction method based on the multi-scale enhancement module as described above.

[0034] The present invention uses three convolutional layers with different convolution kernel sizes to extract multi-scale features of remote sensing images based on their multi-scale characteristics. Unlike other invention patents, the present invention combines convolutional layers with different convolution kernel sizes through pairwise convolution to extract multi-scale features, and uses a parallel approach to share feature information. This allows for efficient extraction of multi-scale features of remote sensing images without increasing the number of network parameters. A feature fusion layer is then used to effectively combine the features extracted by the three parallel networks. Furthermore, the present invention uses residual connections outside each multi-scale enhancement module to reduce model complexity and mitigate issues such as gradient explosion and vanishing during network training. Furthermore, the present invention differs from other inventions in its use of techniques. The network in this aspect does not use other attention models to enhance the network's feature extraction capabilities, but only uses a cleverly designed combination of different convolution kernels. This approach does not increase the number of network parameters, and while reducing network computing resources, the present invention's test cases outperform existing state-of-the-art remote sensing super-resolution methods in both subjective and objective results.

[0035] The proposed scheme is superior to other remote sensing image super-resolution methods in both qualitative and quantitative evaluations, and can produce higher-precision high-resolution reconstruction results of remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The figure is a schematic diagram of the remote sensing image super-resolution process based on the multi-scale enhancement module according to an embodiment of the present invention.

[0037] Figure 2 This is a network framework diagram of remote sensing image super-resolution based on a multi-scale enhancement module according to an embodiment of the present invention.

[0038] Figure 3 The following diagrams show experimental results of the embodiments of the present invention on a public dataset, where part (a) is the interpolated (Bicubic) image; part (b) is the original high-resolution image; and part (c) is the experimental result diagram of the present invention. DETAILED DESCRIPTION

[0039] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0040] Reference Figure 1 and Figure 2The embodiment of the application is a remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module, which comprises the following specific process steps:

[0041] Step a, high-resolution remote sensing image I HR is down-sampled to target low-resolution remote sensing image I LR , and the high and low resolution remote sensing image pair is subjected to data enhancement to expand the sample library.

[0042] In the embodiment, the high-resolution remote sensing image is further preferably down-sampled by a bicubic interpolation method to generate a corresponding target low-resolution remote sensing image, and then the high and low resolution remote sensing image pair obtained above is subjected to data enhancement by three enhancement methods of rotation, scaling and flipping to expand the sample library.

[0043] Step b, the low-resolution remote sensing image is input into a deep network as an input, and a rough feature map of the remote sensing image is extracted through a 1x1 convolutional layer.

[0044] In the embodiment, the low-resolution remote sensing image I LR of the training set is further preferably input into the first 1x1 convolutional layer of the deep network, and a rough feature map X of the remote sensing image is extracted through the convolutional layer, that is, the feature map extraction process can be represented as:

[0045] X = Conv 1×1 (I LR ) (1)

[0046] Wherein, Conv 1×1 () represents a 1x1 convolutional layer.

[0047] Step c, the extracted rough feature map of the remote sensing image is input into the first multi-scale enhancement module, a parallel subnetwork combined by three different sizes of convolution kernels and activation layers is used in the multi-scale enhancement module, multi-scale remote sensing image features are efficiently fused through a feature fusion layer, and a residual connection mode is used outside the multi-scale enhancement module to reduce the model complexity and reduce overfitting.

[0048] In the embodiment, it is further preferred to first input the extracted remote sensing image rough feature map X into the first multi-scale enhancement module, and the subsequent step d will also connect multiple multi-scale enhancement modules in series. Each multi-scale enhancement module has the same structure. Three parallel sub-networks Conv-a(), Conv-b(), and Conv-c() with different convolution kernel size combinations are used in a multi-scale enhancement module. Different sub-networks are combined using convolution layers with different convolution kernel sizes to achieve feature extraction of remote sensing images at different scales. The remote sensing image features extracted by Conv-a(), Conv-b(), and Conv-c() are respectively enhanced through corresponding nonlinear activation functions to enhance the representation ability of the network. Then, the feature fusion layer is used to achieve efficient fusion of multi-scale remote sensing image features. The extracted multi-scale information is output through a 1×1 convolution layer, for example, input to the next multi-scale enhancement module. That is, the feature extraction process of the multi-scale enhancement module can be expressed as:

[0049] Conv-a(X)=Conv 1×1 (Conv 1×1 (X))+Conv 3×3 (Conv 1×1 (X)) (2)

[0050] Conv-b(X)=Conv 3×3 (Conv 1×1 (X))+Conv 5×5 (Conv 1×1 (X)) (3)

[0051] Conv-c(X)=Conv 5×5 (Conv 1×1 (X))+Conv 1×1 (Conv 1×1 (X)) (4)

[0052] M n (X) = Conv 1×1 (Concat(Conv-a(X),Conv-b(X),Conv-c(X))) (5)

[0053] Among them, Conv-a(), Conv-b(), and Conv-c() represent three parallel sub-networks in the multi-scale enhancement module. 1×1 (), Conv 3×3 () and Conv 5×5 () represent 1×1 convolution layer, 3×3 convolution layer and 5×5 convolution layer respectively, Concat() represents feature fusion layer, M n () represents the output of the nth multi-scale enhancement module.

[0054] The implementation of the three parallel sub-networks Conv-a(), Conv-b(), and Conv-c() in the three multi-scale enhancement modules is as follows:

[0055] In the sub-network Conv-a(), after the input enters the 1×1 convolution layer, it enters the 1×1 convolution layer and the 3×3 convolution layer respectively, and then passes through the point-by-point addition layer to obtain the output;

[0056] In the sub-network Conv-b(), the input enters the 1×1 convolution layer, then enters the 3×3 convolution layer and the 5×5 convolution layer, and then passes through the point-by-point addition layer to obtain the output;

[0057] In the sub-network Conv-c(), after the input enters the 1×1 convolution layer, it enters the 1×1 convolution layer and the 5×5 convolution layer respectively, and then gets the output through the point-by-point addition layer.

[0058] In step d, after the first multi-scale enhancement module, a fine remote sensing image feature map is obtained by connecting multiple multi-scale enhancement modules in series, which is then fused with the coarse feature map of the remote sensing image extracted in the above step. The feature map of the updated target remote sensing image is then passed through the convolution layer, the pixel reconstruction layer, and the convolution layer in sequence to produce a high-resolution remote sensing image result.

[0059] In step d of the embodiment, a refined remote sensing image feature map Y is obtained by connecting multiple identical multi-scale enhancement modules in series. Assuming that there are N multi-scale enhancement modules in total, residual connections are used outside the 2nd, 3rd...N-1st multi-scale enhancement modules to reduce the complexity of the model and reduce overfitting. After experiments, it is preferably recommended that N be 7. In step c, the output of the first multi-scale enhancement module is used as the input of the second multi-scale enhancement module, the input of the second multi-scale enhancement module is added to the output of the second multi-scale enhancement module, and then input to the third multi-scale enhancement module; the input of the third multi-scale enhancement module is added to the output of the third multi-scale enhancement module, and then input to the fourth multi-scale enhancement module; ... until the input of the N-1th multi-scale enhancement module is added to the output of the N-1th multi-scale enhancement module, and the result is input to the Nth multi-scale enhancement module. The output of the Nth multi-scale enhancement module is fused with the rough feature map X of the remote sensing image extracted in step b above through a point-by-point addition layer. Then, the feature map of the updated target remote sensing image is sequentially passed through a 1×1 convolution layer, a pixel reassembly layer, and a 1×1 convolution layer to generate a high-resolution remote sensing image result I. SR . The training process L SR The loss function is expressed as

[0060]

[0061] Where F() is the mapping relationship of deep learning network training, is the original high-resolution image of the remote sensing image data pair, and || ||1 represents the norm of the matrix.

[0062] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0063] In some possible embodiments, a remote sensing image super-resolution reconstruction system based on a multi-scale enhancement module is provided, comprising the following modules:

[0064] The first module is used to downsample the high-resolution remote sensing image to a target low-resolution remote sensing image, and perform data enhancement on the high-resolution and low-resolution remote sensing image pairs to expand the sample library;

[0065] The second module is used to take the low-resolution remote sensing image as the input of the deep network and extract the rough feature map of the remote sensing image through the convolution layer;

[0066] The third module is used to input the extracted rough feature map of the remote sensing image into the first multi-scale enhancement module. The multi-scale enhancement module uses a parallel sub-network with three different convolution kernel size combinations to achieve efficient fusion of multi-scale remote sensing image features through a feature fusion layer;

[0067] The fourth module is used to obtain a fine remote sensing image feature map by connecting multiple multi-scale enhancement modules in series, using residual connections outside some multi-scale enhancement modules to reduce model complexity to reduce overfitting, and fusing the obtained fine remote sensing image feature map with the coarse feature map of the remote sensing image to update the feature map of the target remote sensing image to produce a high-resolution remote sensing image result.

[0068] In some possible embodiments, a remote sensing image super-resolution reconstruction system based on a multi-scale enhancement module is provided, including a processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute a remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module as described above.

[0069] In some possible embodiments, a remote sensing image super-resolution reconstruction system based on a multi-scale enhancement module is provided, including a readable storage medium, on which a computer program is stored. When the computer program is executed, a remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module as described above is implemented.

[0070] Reference Figure 3 , the application test experiment of the embodiment of the present invention is as follows:

[0071] The experiment adopts a public data set NWPU-RESISC45, and 90 remote sensing images are selected as a test set, each remote sensing image has a size of 256*256 pixels, and a corresponding low-resolution remote sensing image has a size of 64*64 pixels after down-sampling four times.

[0072] Compared with other remote sensing image super-resolution reconstruction methods, the effectiveness of the present application is expressed by providing experimental data, and the evaluation indexes of the comparison experiment results are shown in Table 1, and the experimental results are shown in Figure 1. Figure 3 (a) is a Bicubic image, (b) is an original high-resolution image, and (c) is an experimental result of the present application.

[0073] Table 1

[0074] Algorithm Bicubic HSEnet DLGNN The present invention PSNR 27.25 28.81 28.69 29.04 SSIM 0.6786 0.7582 0.7519 0.7588 VIF 0.7807 0.8444 0.8379 0.8446

[0075] As can be seen from the above table experimental data, the present application method obtains a high score compared with the comparison method, that is, is better than the comparison method.

[0076] In summary, the remote sensing image super-resolution reconstruction method and system based on a multi-scale enhancement module provided by the present application firstly down-samples a high-resolution remote sensing image to a target low-resolution remote sensing image, and performs data enhancement on the high and low resolution remote sensing image pair to expand the sample library; then the low-resolution remote sensing image is used as the input of the deep network, the rough feature map of the remote sensing image is extracted through a 1*1 convolution layer, the extracted rough feature map of the remote sensing image is input into the multi-scale enhancement module, three different convolution kernel size combination parallel sub-networks are used in the multi-scale enhancement module, the multi-scale remote sensing image features are efficiently fused through a feature fusion layer, and the model complexity is reduced by using a residual connection mode outside the multi-scale enhancement module to reduce overfitting; finally, a fine remote sensing image feature map is obtained by connecting a plurality of multi-scale enhancement modules in series, the fine remote sensing image feature map is fused with the rough feature map of the remote sensing image extracted in the above step, and then the feature map of the updated target remote sensing image is recombined through a pixel recombination layer to generate a high-resolution remote sensing image result. The network provided by the present application is better than other remote sensing image super-resolution methods in qualitative and quantitative evaluation, and can produce a higher-precision remote sensing image high-resolution reconstruction result.

[0077] Compared with the prior art, the advantages and positive effects of the present application are that the remote sensing image super-resolution reconstruction method and system based on a multi-scale enhancement module not only considers how to finely extract multi-scale features of remote sensing images, but also solves the redundancy problem of the extracted remote sensing image feature information in the prior art, and reduces overfitting by using a residual connection mode to reduce the model complexity. The advantage of the present application is that the multi-scale enhancement network learns the multi-scale feature information of the remote sensing image from a large number of samples, so that the super-resolution reconstruction accuracy of the remote sensing image is higher.

[0078] The specific embodiments described herein are merely illustrative of the spirit of the application. Various modifications or changes in the specific embodiments described herein can occur to those skilled in the art to which the application pertains without departing from the spirit of the application, and it is understood that such modifications or changes are to be considered as within the scope of the application as defined by the appended claims.

Claims

1. A remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module, characterized in that: The following steps are involved: Step a: downsampling the high-resolution remote sensing image to the target low-resolution remote sensing image, and performing data enhancement on the obtained high-resolution and low-resolution remote sensing image pairs to expand the sample library; Step b: using the low-resolution remote sensing image as the input of the deep network, and extracting the rough feature map of the remote sensing image through the convolutional layer; Step c: Input the extracted remote sensing image rough feature map into the first multi-scale enhancement module. The multi-scale enhancement module uses a parallel sub-network with three different convolution kernel size combinations to achieve efficient fusion of multi-scale remote sensing image features through a feature fusion layer. The implementation method is to input the extracted remote sensing image rough feature map X into the multi-scale enhancement module. The multi-scale enhancement module uses three parallel sub-networks with different convolution kernel size combinations. Different sub-networks are combined with convolution layers with different convolution kernel sizes to extract features of remote sensing images at different scales. Then, the feature fusion layer is used to achieve efficient fusion of multi-scale remote sensing image features. The feature extraction process of the multi-scale enhancement module is expressed as: Conv-a(X)=Conv 1×1 (Conv 1×1 (X))+Conv 3×3 (Conv 1×1 (X)) Conv-b(X)=Conv 3×3 (Conv 1×1 (X))+Conv 5×5 (Conv 1×1 (X)) Conv-c(X)=Conv 5×5 (Conv 1×1 (X))+Conv 1×1 (Conv 1×1 (X)) M n (X)=Conv 1×1 (Concat(Conv-a(X),Conv-b(X),Conv-c(X))) Among them, Conv-a(), Conv-b(), and Conv-c() represent three parallel sub-networks in the multi-scale enhancement module. 1×1 (), Conv 3×3 () and Conv 5×5 () represent 1×1 convolution layer, 3×3 convolution layer and 5×5 convolution layer respectively, Concat() represents feature fusion layer, M n () represents the output of the nth multi-scale enhancement module; In step d, a fine remote sensing image feature map is obtained by connecting multiple multi-scale enhancement modules in series. Residual connections are used outside some multi-scale enhancement modules to reduce model complexity to reduce overfitting. The obtained fine remote sensing image feature map is merged with the coarse feature map of the remote sensing image extracted in step b above to update the feature map of the target remote sensing image to produce a high-resolution remote sensing image result.

2. The remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module according to claim 1, characterized in that: In step a, the high-resolution remote sensing image is downsampled using the bicubic interpolation method to generate the corresponding target low-resolution remote sensing image. Then, the high-resolution and low-resolution remote sensing image pairs obtained above are enhanced by rotation, scaling, and flipping to expand the sample library.

3. The remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module according to claim 1, characterized in that: In step b, the low-resolution remote sensing image I of the training set is LR Input to the first 1×1 convolutional layer of the deep network, through which the coarse feature map X of the remote sensing image is extracted.

4. The remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module according to claim 3, characterized in that: The implementation of the three parallel sub-networks Conv-a(), Conv-b(), and Conv-c() in the multi-scale enhancement module is as follows: In the sub-network Conv-a(), after the input enters the 1×1 convolution layer, it enters the 1×1 convolution layer and the 3×3 convolution layer respectively, and then passes through the point-by-point addition layer to obtain the output; In the sub-network Conv-b(), the input enters the 1×1 convolution layer, then enters the 3×3 convolution layer and the 5×5 convolution layer, and then passes through the point-by-point addition layer to obtain the output; In the sub-network Conv-c(), after the input enters the 1×1 convolution layer, it enters the 1×1 convolution layer and the 5×5 convolution layer respectively, and then gets the output through the point-by-point addition layer.

5. The remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module according to claim 1, 2, 3 or 4, characterized in that: In step d, a fine remote sensing image feature map Y is obtained by connecting multiple multi-scale enhancement modules in series, and it is fused with the coarse feature map X of the remote sensing image extracted in step b above through a point-by-point addition layer to update the feature map of the target remote sensing image to produce a high-resolution remote sensing image result. The training process L SR The loss function is expressed as Where F() is the mapping relationship of deep learning network training, is the original high-resolution image of the remote sensing image data pair, and || ||1 represents the norm of the matrix.

6. A remote sensing image super-resolution reconstruction system based on a multi-scale enhancement module, characterized by: Used to implement a remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module as described in any one of claims 1-5.

7. The remote sensing image super-resolution reconstruction system based on the multi-scale enhancement module according to claim 6, characterized in that: Includes the following modules, The first module is used to downsample the high-resolution remote sensing image to a target low-resolution remote sensing image, and perform data enhancement on the high-resolution and low-resolution remote sensing image pairs to expand the sample library; The second module is used to take the low-resolution remote sensing image as the input of the deep network and extract the rough feature map of the remote sensing image through the convolution layer; The third module is used to input the extracted rough feature map of the remote sensing image into the first multi-scale enhancement module. The multi-scale enhancement module uses a parallel sub-network with three different convolution kernel size combinations to achieve efficient fusion of multi-scale remote sensing image features through a feature fusion layer; The fourth module is used to obtain a fine remote sensing image feature map by connecting multiple multi-scale enhancement modules in series, using residual connections outside some multi-scale enhancement modules to reduce model complexity to reduce overfitting, and fusing the obtained fine remote sensing image feature map with the coarse feature map of the remote sensing image to update the feature map of the target remote sensing image to produce a high-resolution remote sensing image result.

8. The remote sensing image super-resolution reconstruction system based on the multi-scale enhancement module according to claim 6, characterized in that: The invention comprises a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute a remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module as described in any one of claims 1 to 5.

9. The remote sensing image super-resolution reconstruction system based on the multi-scale enhancement module according to claim 6, characterized in that: It includes a readable storage medium, on which a computer program is stored. When the computer program is executed, it implements a remote sensing image super-resolution reconstruction method based on a multi-scale enhancement module as described in any one of claims 1 to 5.

Images