Remote sensing image super-resolution reconstruction method based on degradation mechanism

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a remote sensing image super-resolution reconstruction method based on a degradation mechanism. Comprising the following steps: S1, constructing a degradation model, inputting a high-resolution remote sensing image into the degradation model, and obtaining a degraded image corresponding to the high-resolution remote sensing image; s2, constructing a super-resolution network, and inputting the degraded image into the super-resolution network to obtain a reconstructed image; s3, training the super-resolution network by using the joint loss function to obtain a trained super-resolution network; and S4, inputting a degraded image to be reconstructed into the trained super-resolution network to obtain a reconstructed super-resolution image. According to the method, the resolution of the remote sensing image is improved, the quality and processing efficiency of the image are remarkably improved, and the method has high practical application value.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a remote sensing image super-resolution reconstruction method based on a degradation mechanism. Background Art

[0002] Super-resolution reconstruction technology of remote sensing images has received widespread attention in recent years. Its main purpose is to improve the detail performance of remote sensing images by improving the resolution of images and enhance the accuracy of target detection, classification and recognition. At present, there are many methods applied to super-resolution reconstruction of remote sensing images, mainly including interpolation methods, reconstruction methods based on signal processing and deep learning methods.

[0003] Interpolation is a traditional super-resolution reconstruction method, and typical techniques include bilinear interpolation and bicubic interpolation. These difference methods fill in the high-resolution image by interpolating between the pixels of the original low-resolution image. This method is simple and computationally efficient, but it cannot effectively restore the high-frequency details of the image, resulting in the reconstruction result being blurred and unclear, especially when dealing with complex textures or edge information.

[0004] The super-resolution reconstruction method based on signal processing uses the characteristics of the image in the spatial domain or frequency domain to achieve super-resolution reconstruction of degraded images by solving specific optimization problems. In this type of reconstruction method, some studies use regularization techniques to constrain the reconstruction results of the image, thereby reducing the influence of noise and artifacts. However, when the computational complexity is high and the processed images are real-world remote sensing images, the super-resolution reconstruction method based on signal processing often has limited reconstruction quality due to the lack of high-frequency information in the image.

[0005] In recent years, deep learning technology has made significant progress in the field of image super-resolution reconstruction. In particular, deep models such as convolutional neural networks (CNNs) have been widely used in remote sensing image super-resolution reconstruction due to their powerful feature extraction and learning capabilities. Super-resolution methods based on deep learning usually obtain higher quality reconstructed images by training deep neural networks and using paired high-resolution and low-resolution images for supervised learning. However, most of the existing deep learning methods are based on idealized degradation models, which do not fully match the complex degradation mechanisms in the actual remote sensing imaging process, resulting in the quality of reconstruction results still needing to be improved when processing real remote sensing images.

[0006] Existing super-resolution reconstruction methods have their own advantages and disadvantages. Traditional interpolation methods are simple to calculate, but their ability to restore high-frequency details is insufficient. Although signal processing-based reconstruction methods can improve reconstruction effects through mathematical models, they are complex to calculate and have poor performance when processing real remote sensing images. Deep learning methods have excellent performance in reconstruction quality, but most current methods rely on idealized degradation models and find it difficult to accurately simulate the complex degradation process of remote sensing imaging. Existing degradation models are significantly different from the actual remote sensing imaging process, resulting in low quality of reconstructed images and the inability to fully restore high-frequency details and texture information in the image, affecting their performance in practical applications.

[0007] Most existing methods cannot effectively handle multiple degradation factors in the complex imaging chain when processing remote sensing images, such as atmospheric scattering, blur in the optical system, noise interference, etc. Although the super-resolution method based on deep learning has improved the reconstruction effect to a certain extent, its network structure and loss function design are still imperfect, making it difficult to take into account the extraction and utilization of local and global feature information, resulting in insufficient visual effects and information fidelity of the reconstructed image. Summary of the invention

[0008] In view of this, the invention aims to provide a remote sensing image super-resolution reconstruction method based on degradation mechanism, so as to solve the problem that the degradation model in the existing super-resolution reconstruction method is quite different from the actual remote sensing imaging process, resulting in low quality of reconstructed images, inability to fully restore high-frequency details and texture information in the image, and most existing methods cannot effectively handle multiple degradation factors in the complex imaging chain, such as atmospheric scattering, blur in the optical system, noise interference, etc. While improving the resolution of remote sensing images, the present invention significantly improves the image quality and processing efficiency, and has high practical application value.

[0009] To achieve the above object, the technical solution created by the present invention is implemented as follows: A remote sensing image super-resolution reconstruction method based on degradation mechanism specifically comprises the following steps: S1: construct a degradation model, input the high-resolution remote sensing image into the degradation model, and obtain a degraded image corresponding to the high-resolution remote sensing image; S2: construct a super-resolution network, input the degraded image into the super-resolution network, and obtain a reconstructed image; The super-resolution network includes a spatial domain branch and a frequency domain branch. The spatial domain branch extracts the local multi-scale feature information of the degraded image through multiple levels of convolution operations to obtain spatial domain features. The frequency domain branch analyzes and processes the global frequency features of the degraded image to obtain frequency domain features. The spatial domain features and frequency domain features are fused to obtain a reconstructed image. S3: Use the joint loss function to train the super-resolution network to obtain a trained super-resolution network; The joint loss function includes pixel-level loss function, multi-scale feature loss function and global feature loss function; S4: Input the degraded image to be reconstructed into the trained super-resolution network to obtain a reconstructed super-resolution image.

[0010] Furthermore, in step S1, the degradation model includes a spatial domain filtering module, a frequency domain filtering module and a degradation factor random control module. The high-resolution remote sensing image is input into the spatial domain filtering module for processing to obtain a first spatial domain image; the high-resolution remote sensing image is Fourier transformed and then input into the frequency domain filtering module, the degradation factor random control module controls the frequency domain filtering module to randomly introduce degradation factors into the high-resolution remote sensing image after Fourier transform, and the high-resolution remote sensing image after the degradation factors are introduced is inverse Fourier transformed to obtain a second spatial domain image; the first spatial domain image and the second spatial domain image are synthesized, and then down-sampling operations and noise introduction operations are performed to obtain a degraded image.

[0011] Further, the frequency domain filtering module includes an atmospheric disturbance factor, an optical system disturbance factor, a first platform component disturbance factor, a second platform component disturbance factor, and a third platform component disturbance factor; Atmospheric disturbance factor I is: ; ; in, is the angular spatial frequency, , is the focal length of the optical system, is the spatial frequency, is the refractive index structure constant, is the wavelength, R is the propagation path, a1 is the turbulence effect, a2 is the aerosol effect, is the modulation transfer function of the turbulence effect, is the modulation transfer function of the aerosol effect; ; in, is the scattering coefficient of atmospheric aerosol, is the absorption coefficient of atmospheric aerosol, is the angular spatial cutoff frequency of the aerosol; The optical system disturbance factor II is: ; ; in, for , is the spatial frequency in the horizontal direction, is the spatial cutoff frequency, = , D is the diameter of the optical system, b1 is the diffraction effect, b2 is the defocus effect, b3 is the image shift effect, is the modulation transfer function of the diffraction effect, is the modulation transfer function of the defocus effect, is the modulation transfer function of the image shift effect; ; in, is a first-order Bessel function, , is the defocus amount, F is the focal ratio of the optical system, It is an intermediate parameter and has no physical meaning; ; ; in, is the image shift caused by the change in satellite velocity, is the image shift caused by the satellite yaw, is the image shift caused by the change of satellite attitude angle, is the image shift caused by the rotation of the Earth, Image shift caused by random vibration of satellite and optical remote sensor. is the image shift caused by the jitter of the integration time, It is an intermediate variable and has no physical meaning; The disturbance factor III of the first platform component is: ; ; Where N is the sampling frequency, d is the diffuse spot diameter, For low frequency vibration, It is high frequency vibration; ;

[0012] in, is the zero-order Bessel function, is the platform amplitude, is the spatial frequency; The disturbance factor IV of the second platform component is: ; ; in, is the spatial frequency along the scanning direction, is the CCD pixel size, d1 is the spatial response, d2 is the temporal response, d3 is the CCD response, is the modulation transfer function of the spatial response, is the modulation transfer function of the time response, is the modulation transfer function of the CCD response; ; in, is the time frequency value along the scanning direction, is the 3-dB point frequency value of the detector time response; ; in, is the amount of charge transferred by the CCD, is the charge transfer efficiency, is the sampling frequency of the CCD signal; The third platform component disturbance factor V is: ; ; in, is the frequency value of the -3dB point of the preamplifier response, e1 is the preamplifier response, e2 is the postamplifier response, and e3 is the lift response. is the modulation transfer function of the preamplifier response, is the modulation transfer function of the post-amplifier response, is the modulation transfer function of the lift response; ; in, The frequency value of the -3dB point of the post-amplifier response; ; in, is the frequency value when the boost pressure is maximum, for The magnitude of the value at .

[0013] Furthermore, the formula for the random control module of the degradation factor to control the frequency domain filtering module to randomly introduce the degradation factor into the high-resolution remote sensing image after Fourier transformation is: K= · · ; Among them, U is uniform distribution, X1~X5 are coefficients that obey uniform distribution, n1 is 0, and n2 is 1.

[0014] Further, in step S2, the spatial domain branch includes a first 3×3 unconstrained blueprint convolution layer and n cascaded basic units, each basic unit includes a cascaded first basic module and a second basic module, and the output of the first basic module is the input of the second basic module; The degraded image is copied four times, and three of the degraded images are subjected to RGB three-channel extraction operations, and the remaining degraded image is subjected to Y channel extraction operations, and all the extracted feature maps are subjected to feature splicing operations to obtain a spliced ​​feature map, and the spliced ​​feature map is subjected to a convolution operation by the first 3×3 unconstrained blueprint convolution layer to obtain an intermediate feature map, and the intermediate feature map is input into n cascaded basic units for processing to obtain a first feature map; The frequency domain branch includes a mask, a second 3×3 unconstrained blueprint convolution layer, a third 3×3 unconstrained blueprint convolution layer, a fourth 3×3 unconstrained blueprint convolution layer, a first 3×3 convolution layer, a second 3×3 convolution layer, a third 3×3 convolution layer, a first batch of normalization layers, a second batch of normalization layers, a third batch of normalization layers, a first SiLU activation function, a second SiLU activation function, and a third SiLU activation function; Perform Fourier transform on the degraded image to obtain a frequency domain image, perform point multiplication operation on the frequency domain image and the mask plate to obtain frequency domain images under different masks, and process the frequency domain images under different masks through the attention mechanism to obtain the second feature map, the third feature map and the fourth feature map; The second feature map is input into the second 3×3 unconstrained blueprint convolution layer for processing to obtain the fifth feature map. The fifth feature map is processed by the first 3×3 convolution layer, the first batch of normalization layers and the first SiLU activation function in sequence to obtain the sixth feature map. The fifth feature map and the sixth feature map are feature concatenated to obtain the seventh feature map. The seventh feature map is inverse Fourier transformed to obtain the eighth feature map. The third feature map is input into the third 3×3 unconstrained blueprint convolution layer for processing to obtain the ninth feature map. The ninth feature map is processed by the second 3×3 convolution layer, the second batch of normalization layers and the second SiLU activation function in sequence. After processing, the tenth feature map is obtained, the ninth feature map and the tenth feature map are feature-concatenated to obtain the eleventh feature map, the eleventh feature map is inverse Fourier transformed to obtain the twelfth feature map; the fourth feature map is input into the fourth 3×3 unconstrained blueprint convolution layer for processing to obtain the thirteenth feature map, the thirteenth feature map is sequentially processed by the third 3×3 convolution layer, the third batch normalization layer and the third SiLU activation function to obtain the fourteenth feature map, the thirteenth feature map and the fourteenth feature map are feature-concatenated to obtain the fifteenth feature map, the fifteenth feature map is inverse Fourier transformed to obtain the sixteenth feature map.

[0015] Furthermore, the output features of the first basic module of each basic unit, the first feature map, the eighth feature map, the twelfth feature map and the sixteenth feature map are feature spliced ​​to obtain a seventeenth feature map. The seventeenth feature map is sequentially convolved by the first 1×1 convolution layer and the fifth 3×3 unconstrained blueprint convolution layer to obtain an eighteenth feature map after feature supplementation. The eighteenth feature map and the intermediate feature map are added to obtain a nineteenth feature map. The nineteenth feature map is sequentially processed by the fourth 3×3 convolution layer and upsampling to obtain a reconstructed image.

[0016] Further, in step S3, the output of the joint loss function is the cumulative sum of the pixel-level loss function, the multi-scale feature loss function, and the global feature loss function; The pixel-level loss function is an L1 loss function, and the input of the pixel-level loss function is a degraded image and a reconstructed image; The multi-scale feature loss function is a Huber loss function with a parameter of 1, and the input of the multi-scale feature loss function is the reconstructed stitched image and the high-resolution stitched image; The global feature loss function is a Huber loss function with a parameter of 1.25, and the input of the global feature loss function is the reconstructed stitched image and the high-resolution stitched image.

[0017] Furthermore, the reconstructed stitched image is obtained by: inputting the reconstructed image into the global feature extraction module and the local feature extraction module respectively for feature extraction, and stitching the extracted global features and all local features to obtain the reconstructed stitched image; The method for obtaining a high-resolution stitched image is as follows: the high-resolution remote sensing image is input into the global feature extraction module and the local feature extraction module for feature extraction, and the extracted global features and all local features are stitched to obtain a high-resolution stitched image.

[0018] Further, the global feature extraction module includes a fifth 3×3 convolutional layer, an activation layer submodule, a pooling layer submodule, a fully connected layer submodule, a flattening transpose submodule, a first encoder, a second encoder, a third encoder, a fourth encoder, and a fifth encoder; The degraded image or the reconstructed image is input into the global feature extraction module, and the degraded image or the reconstructed image is processed by the fifth 3×3 convolutional layer, the activation layer submodule, the pooling layer submodule and the fully connected layer submodule in sequence to obtain a feature map A1. The feature map A1 is flattened and transposed by the flattening and transposing submodule to obtain a feature map A2. The feature map A2 is input into the first encoder for encoding to obtain a feature map A3. The feature map A3 is input into the second encoder for encoding to obtain a feature map A4. The feature map A4 is input into the third encoder for encoding to obtain a feature map A5. The feature map A5 is input into the fourth encoder for encoding to obtain a feature map A6. The feature map A6 is input into the fifth encoder for encoding to obtain a feature map A7. The feature maps A3, A4, A5, A6 and A7 are processed by using the attention mechanism to obtain the global features of the degraded image or the reconstructed image.

[0019] Further, the local feature extraction module includes a first basic unit convolution layer, a second basic unit convolution layer, a third basic unit convolution layer, a fourth basic unit convolution layer, a fifth basic unit convolution layer, a first activation layer, a second activation layer, a third activation layer, a fourth activation layer, a fifth activation layer, a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer, a fourth maximum pooling layer and a fifth maximum pooling layer; The degraded image or the reconstructed image is input into the local feature extraction module, the degraded image or the reconstructed image is processed by the first basic unit convolution layer, the first activation layer and the first maximum pooling layer in sequence to obtain a feature map B1, the feature map B1 is processed by the second basic unit convolution layer, the second activation layer and the second maximum pooling layer in sequence to obtain a feature map B2, the feature map B2 is processed by the third basic unit convolution layer, the third activation layer and the third maximum pooling layer in sequence to obtain a feature map B3, the feature map B3 is processed by the fourth basic unit convolution layer, the fourth activation layer and the fourth maximum pooling layer in sequence to obtain a feature map B4, the feature map B4 is processed by the fifth basic unit convolution layer, the fifth activation layer and the fifth maximum pooling layer in sequence to obtain a feature map B5, the feature map B1, the feature map B2, the feature map B3, the feature map B4 and the feature map B5 are processed by the attention mechanism to obtain the local features of the degraded image or the reconstructed image.

[0020] Compared with the prior art, the invention can achieve the following beneficial effects: (1) The remote sensing image super-resolution reconstruction method based on the degradation mechanism created by the present invention can better simulate the degradation process in remote sensing imaging through the degradation model based on the degradation mechanism and the joint space-frequency domain super-resolution network, and effectively improve the quality of the super-resolution image, especially in terms of high-frequency details and texture restoration, so that the quality of the reconstructed image is significantly improved.

[0021] (2) The present invention creates the remote sensing image super-resolution reconstruction method based on the degradation mechanism, which expands the degradation space through the degradation factor random control module, so that the remote sensing image super-resolution reconstruction method of the present invention has stronger adaptability and robustness, can process different types of remote sensing images, and shows better generalization ability in practical applications.

[0022] (3) The present invention creates a remote sensing image super-resolution reconstruction method based on the degradation mechanism. The combined spatial domain and frequency domain super-resolution network enables the present invention to fully extract and utilize the local and global feature information of the degraded image, thereby improving the fineness and overall visual effect of the reconstructed image.

[0023] (4) The remote sensing image super-resolution reconstruction method based on the degradation mechanism created by the present invention optimizes the reconstructed image at three levels: pixel level, global feature and multi-scale feature, through the design of a joint loss function, thereby ensuring high-quality super-resolution output.

[0024] (5) The remote sensing image super-resolution reconstruction method based on the degradation mechanism created by the present invention can be applied to satellite remote sensing, environmental monitoring, urban planning, national defense and security, etc., by improving the detail restoration ability of degraded images, and improving the spatial resolution and quality of reconstructed images. Specific applications include but are not limited to the construction of high-precision geographic information systems (GIS), remote sensing image classification, target detection and recognition, etc. Specifically: 1) Satellite remote sensing: improve the resolution of remote sensing images and enhance the recognition and monitoring capabilities of ground targets. 2) Environmental monitoring: used to track changes in dynamic environments such as pollution spread and natural disasters. 3) Urban planning: by enhancing image resolution, more accurate basic data is provided to support building and road planning. 4) Agriculture and forestry monitoring: used for refined management of agriculture and forestry such as crop health assessment and forest cover monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The drawings constituting part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments and descriptions of the present invention are used to explain the present invention and do not constitute an improper limitation on the present invention. In the drawings: Figure 1 A schematic diagram of a flow chart of a remote sensing image super-resolution reconstruction method based on a degradation mechanism according to an embodiment of the present invention; Figure 2 A structural block diagram of a remote sensing image super-resolution reconstruction method based on a degradation mechanism according to an embodiment of the present invention; Figure 3 A schematic diagram of the network structure of the degradation model described in the embodiment of the present invention; Figure 4A schematic diagram of the network structure of the super-resolution network described in the embodiment of the present invention; Figure 5 A schematic diagram of the network structure of the global feature extraction module described in the embodiment of the present invention; Figure 6 A schematic diagram of the network structure of the local feature extraction module described in the embodiment of the present invention; Figure 7 A schematic diagram of the network structure of the first basic module described in the embodiment of the present invention; Figure 8 A schematic diagram of the network structure of the second basic module described in the embodiment of the present invention; Fig. 9 A reconstruction effect diagram of reconstructing a severely degraded remote sensing image as described in an embodiment of the present invention; Fig.10 A reconstruction effect diagram of reconstructing the first type of remote sensing image with complex textures according to an embodiment of the present invention; Fig.11 This is a reconstruction effect diagram of reconstructing the second type of remote sensing image with complex texture described in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the invention more clear, the invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described here are only used to explain the invention and do not constitute a limitation of the invention.

[0027] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0028] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0029] In the description of the invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installation", "connection" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the invention can be understood according to specific circumstances.

[0030] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0031] like Figure 1 to Figure 2 As shown, a remote sensing image super-resolution reconstruction method based on degradation mechanism specifically includes the following steps: S1: construct a degradation model, input the high-resolution remote sensing image into the degradation model, and obtain a degraded image corresponding to the high-resolution remote sensing image; S2: construct a super-resolution network, input the degraded image into the super-resolution network, and obtain a reconstructed image; The super-resolution network includes a spatial domain branch and a frequency domain branch. The spatial domain branch extracts the local multi-scale feature information of the degraded image through multiple levels of convolution operations to obtain spatial domain features. The frequency domain branch analyzes and processes the global frequency features of the degraded image to obtain frequency domain features. The spatial domain features and frequency domain features are fused to obtain a reconstructed image. S3: Use the joint loss function to train the super-resolution network to obtain a trained super-resolution network; The joint loss function includes pixel-level loss function, multi-scale feature loss function and global feature loss function; S4: Input the degraded image to be reconstructed into the trained super-resolution network to obtain a reconstructed super-resolution image.

[0032] This paper designs a degradation model based on the remote sensing imaging chain, which simulates various degradation factors that may exist in the remote sensing imaging process, including atmospheric disturbances, optical system blur, sensor noise, etc. By introducing a random control module for degradation factors, the expression capacity of the degradation space is expanded, making the model more generalized and robust, ensuring that the model can adapt to different degradation situations.

[0033] The following is a detailed introduction to the network structure of the degradation model. Figure 3As shown in the figure, the degradation model includes a spatial domain filtering module, a frequency domain filtering module and a degradation factor random control module. The high-resolution remote sensing image is input into the spatial domain filtering module for processing to obtain the first spatial domain image; the high-resolution remote sensing image is Fourier transformed (FFT) and then input into the frequency domain filtering module. The degradation factor random control module controls the frequency domain filtering module to randomly introduce degradation factors into the high-resolution remote sensing image after Fourier transformation, and the high-resolution remote sensing image after the introduction of degradation factors is subjected to inverse Fourier transformation (iFFT) to obtain the second spatial domain image; the first spatial domain image and the second spatial domain image are synthesized and then down-sampled and noise introduced to obtain a degraded image. This degradation model not only enhances the adaptability to complex degradation scenes, but also effectively improves the accuracy of super-resolution reconstruction.

[0034] The frequency domain filtering module includes atmospheric disturbance factors, optical system disturbance factors, first platform component disturbance factors, second platform component disturbance factors and third platform component disturbance factors (the first platform component disturbance factors, the second platform component disturbance factors and the third platform component disturbance factors are sensor noise factors).

[0035] Atmospheric disturbance factor I is: ; ; in, is the angular spatial frequency, , is the focal length of the optical system, is the spatial frequency, is the refractive index structure constant, is the wavelength, R is the propagation path, a1 is the turbulence effect, a2 is the aerosol effect, is the modulation transfer function of the turbulence effect, is the modulation transfer function of the aerosol effect; ; in, is the scattering coefficient of atmospheric aerosol, is the absorption coefficient of atmospheric aerosol, is the angular spatial cutoff frequency of the aerosol; The optical system disturbance factor II is: ; ; in, for , is the spatial frequency in the horizontal direction, is the spatial cutoff frequency, = , D is the diameter of the optical system, b1 is the diffraction effect, b2 is the defocus effect, b3 is the image shift effect, is the modulation transfer function of the diffraction effect, is the modulation transfer function of the defocus effect, is the modulation transfer function of the image shift effect; ; in, is a first-order Bessel function, , is the defocus amount, F is the focal ratio of the optical system, It is an intermediate parameter and has no physical meaning; ; ; in, is the image shift caused by the change in satellite velocity, is the image shift caused by the satellite yaw, is the image shift caused by the change of satellite attitude angle, is the image shift caused by the rotation of the Earth, Image shift caused by random vibration of satellite and optical remote sensor. is the image shift caused by the jitter of the integration time, It is an intermediate variable and has no physical meaning; The disturbance factor III of the first platform component is: ; ; Where N is the sampling frequency, d is the diffuse spot diameter, For low frequency vibration, It is high frequency vibration; ;

[0036] in, is the zero-order Bessel function, is the platform amplitude, is the spatial frequency; The disturbance factor IV of the second platform component is: ; ; in, is the spatial frequency along the scanning direction, is the CCD pixel size, d1 is the spatial response, d2 is the temporal response, d3 is the CCD response, is the modulation transfer function of the spatial response, is the modulation transfer function of the time response, is the modulation transfer function of the CCD response; ; in, is the time frequency value along the scanning direction, is the 3-dB point frequency value of the detector time response; ; in, is the amount of charge transferred by the CCD, is the charge transfer efficiency, is the sampling frequency of the CCD signal; The third platform component disturbance factor V is: ; ; in, is the frequency value of the -3dB point of the preamplifier response, e1 is the preamplifier response, e2 is the postamplifier response, and e3 is the lift response. is the modulation transfer function of the preamplifier response, is the modulation transfer function of the post-amplifier response, is the modulation transfer function of the lift response; ; in, The frequency value of the -3dB point of the post-amplifier response; ; in, is the frequency value when the boost pressure is maximum, for The magnitude of the value at .

[0037] The formula for the random control module of degradation factors to control the frequency domain filtering module to randomly introduce degradation factors into the high-resolution remote sensing image after Fourier transformation is: K= · · ; Among them, U is uniform distribution, X1~X5 are coefficients that obey uniform distribution, n1 is 0, and n2 is 1.

[0038] In the design of super-resolution network, the present invention proposes a joint spatial domain and frequency domain super-resolution network (JSF-SRNet). Among them, the spatial domain branch is responsible for extracting the local multi-scale feature information of the degraded image, and the frequency domain branch is used to capture the global frequency characteristics of the degraded image. The two branches realize the information complementarity between the spatial domain and the frequency domain through a specific fusion mechanism, and make full use of the multi-dimensional feature information of the image to improve the reconstruction effect. Specifically, the spatial domain branch processes the texture and details of the degraded image at different scales, and the frequency domain branch focuses on the recovery of high-frequency information through frequency separation technology, ensuring the clarity and structural integrity of the reconstructed image.

[0039] like Figure 4 As shown, the spatial domain branch includes a first 3×3 unconstrained blueprint convolution layer and n cascaded basic units, each basic unit includes a cascaded first basic module and a second basic module, and the output of the first basic module is the input of the second basic module; The degraded image is copied four times, and the RGB three-channel extraction operation is performed on three of the degraded images, and the Y channel extraction operation is performed on the remaining degraded image. All the extracted feature maps are subjected to feature splicing operation to obtain a spliced ​​feature map, and the spliced ​​feature map is convolved through the first 3×3 unconstrained blueprint convolution layer (the unconstrained blueprint convolution layer is BSConv-U (Unconstrained Blueprint Convolution, unconstrained blueprint separable convolution), which has the advantages of small number of parameters (lightweight), fast operation speed, and enhanced intra-core correlation of the convolution kernel) to obtain an intermediate feature map, which is input into n cascaded basic units for processing to obtain the first feature map.

[0040] The frequency domain branch includes a mask (specifically a low-pass filter, a high-pass filter or a velvet filter), a second 3×3 unconstrained blueprint convolution layer, a third 3×3 unconstrained blueprint convolution layer, a fourth 3×3 unconstrained blueprint convolution layer, a first 3×3 convolution layer, a second 3×3 convolution layer, a third 3×3 convolution layer, a first batch of normalization layers, a second batch of normalization layers, a third batch of normalization layers, a first SiLU activation function, a second SiLU activation function and a third SiLU activation function; The degraded image is subjected to fast Fourier transform ( Figure 4 F is used to obtain a frequency domain image, perform a dot multiplication operation on the frequency domain image and the mask plate to obtain frequency domain images under different masks, and after processing the frequency domain images under different masks through the attention mechanism, obtain the second feature map, the third feature map and the fourth feature map; The second feature map is input into the second 3×3 unconstrained blueprint convolution layer for processing to obtain the fifth feature map. The fifth feature map is processed by the first 3×3 convolution layer, the first batch of normalization layers and the first SiLU activation function in sequence to obtain the sixth feature map. The fifth feature map and the sixth feature map are feature concatenated ( Figure 4 Indicated by C in the figure), obtain the seventh feature map, and perform inverse fast Fourier transform on the seventh feature map ( Figure 4 denoted by iF in the figure), obtaining an eighth feature map; inputting the third feature map into the third 3×3 unconstrained blueprint convolution layer for processing to obtain a ninth feature map, processing the ninth feature map sequentially through the second 3×3 convolution layer, the second batch normalization layer, and the second SiLU activation function to obtain a tenth feature map, concatenating the ninth feature map with the tenth feature map to obtain an eleventh feature map, and performing inverse Fourier transform on the eleventh feature map to obtain a twelfth feature map; inputting the fourth feature map into the fourth 3×3 unconstrained blueprint convolution layer for processing to obtain a thirteenth feature map, and processing the thirteenth feature map sequentially through the third 3×3 convolution layer, the third batch normalization layer, and the third SiLU activation function to obtain a fourteenth feature map, concatenating the thirteenth feature map with the fourteenth feature map to obtain a fifteenth feature map, and performing inverse Fourier transform on the fifteenth feature map to obtain a sixteenth feature map.

[0041] The output features of the first basic module of each basic unit, the first feature map, the eighth feature map, the twelfth feature map and the sixteenth feature map are concatenated to obtain the seventeenth feature map. The seventeenth feature map is convolved by the first 1×1 convolution layer and the fifth 3×3 unconstrained blueprint convolution layer in sequence to obtain the eighteenth feature map after feature supplementation. The eighteenth feature map and the intermediate feature map are added ( Figure 4 After the nineteenth feature map is obtained, the nineteenth feature map is processed by the fourth 3×3 convolution layer and upsampling in sequence to obtain a reconstructed image.

[0042] The present invention also proposes a joint loss function (JLF-PMG), which optimizes the network from three dimensions: pixel level, global feature level, and multi-scale feature level, to ensure that the super-resolution image has higher visual quality and structural integrity while retaining details. The pixel-level loss function calculates the pixel difference between the reconstructed image and the high-resolution remote sensing image through the L1 norm; the multi-scale feature loss function extracts local features of different scales through a convolutional neural network and calculates the difference between the reconstructed image features and the high-resolution remote sensing image features; the global feature loss function extracts the global context information of the degraded image through a conversion network (such as the ViT model) to ensure the consistency of the overall structure of the degraded image.

[0043] The output of the joint loss function is the cumulative sum of the pixel-level loss function, the multi-scale feature loss function, and the global feature loss function; The pixel-level loss function is an L1 loss function, and the input of the pixel-level loss function is a degraded image and a reconstructed image; The multi-scale feature loss function is a Huber loss function with a parameter of 1, and the input of the multi-scale feature loss function is the reconstructed stitched image and the high-resolution stitched image; The global feature loss function is a Huber loss function with a parameter of 1.25, and the input of the global feature loss function is the reconstructed stitched image and the high-resolution stitched image.

[0044] The reconstructed stitched image is obtained by: inputting the reconstructed image into the global feature extraction module and the local feature extraction module respectively for feature extraction, and stitching the extracted global features and all local features to obtain the reconstructed stitched image; The method for obtaining a high-resolution stitched image is as follows: the high-resolution remote sensing image is input into the global feature extraction module and the local feature extraction module for feature extraction, and the extracted global features and all local features are stitched to obtain a high-resolution stitched image.

[0045] like Figure 5 As shown, the global feature extraction module includes the fifth 3×3 convolutional layer, an activation layer sub-module, a pooling layer sub-module, a fully connected layer sub-module, a flattened transpose sub-module, a first encoder, a second encoder, a third encoder, a fourth encoder and a fifth encoder; the structure of the first encoder, the second encoder, the third encoder, the fourth encoder and the fifth encoder are the same, which is consistent with the encoder structure used in the paper "Translating Images into Maps" published on the website arXiv in October 2021. The encoder in the paper is matched with the super-resolution network for use.

[0046] The degraded image or the reconstructed image is input into the global feature extraction module, and the degraded image or the reconstructed image is processed by the fifth 3×3 convolutional layer, the activation layer submodule, the pooling layer submodule and the fully connected layer submodule in sequence to obtain a feature map A1. The feature map A1 is flattened and transposed by the flattening and transposing submodule to obtain a feature map A2. The feature map A2 is input into the first encoder for encoding to obtain a feature map A3. The feature map A3 is input into the second encoder for encoding to obtain a feature map A4. The feature map A4 is input into the third encoder for encoding to obtain a feature map A5. The feature map A5 is input into the fourth encoder for encoding to obtain a feature map A6. The feature map A6 is input into the fifth encoder for encoding to obtain a feature map A7. The feature maps A3, A4, A5, A6 and A7 are processed by using the attention mechanism to obtain the global features of the degraded image or the reconstructed image.

[0047] like Figure 6 As shown, the local feature extraction module includes a first basic unit convolution layer, a second basic unit convolution layer, a third basic unit convolution layer, a fourth basic unit convolution layer, a fifth basic unit convolution layer, a first activation layer, a second activation layer, a third activation layer, a fourth activation layer, a fifth activation layer, a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer, a fourth maximum pooling layer and a fifth maximum pooling layer; the structures of the first basic unit convolution layer, the second basic unit convolution layer, the third basic unit convolution layer, the fourth basic unit convolution layer and the fifth basic unit convolution layer are the same, which is consistent with the basic unit convolution layer used in the paper "Rethinking Depthwise Separable Convolutions: How Intra-Kernel Correlations Lead to Improved MobileNets" published in the 2020 CVPR conference.

[0048] The degraded image or the reconstructed image is input into the local feature extraction module, the degraded image or the reconstructed image is processed by the first basic unit convolution layer, the first activation layer and the first maximum pooling layer in sequence to obtain a feature map B1, the feature map B1 is processed by the second basic unit convolution layer, the second activation layer and the second maximum pooling layer in sequence to obtain a feature map B2, the feature map B2 is processed by the third basic unit convolution layer, the third activation layer and the third maximum pooling layer in sequence to obtain a feature map B3, the feature map B3 is processed by the fourth basic unit convolution layer, the fourth activation layer and the fourth maximum pooling layer in sequence to obtain a feature map B4, the feature map B4 is processed by the fifth basic unit convolution layer, the fifth activation layer and the fifth maximum pooling layer in sequence to obtain a feature map B5, the feature map B1, the feature map B2, the feature map B3, the feature map B4 and the feature map B5 are processed by the attention mechanism to obtain the local features of the degraded image or the reconstructed image.

[0049] like Figure 7 As shown, the first basic module includes a sixth 3×3 unconstrained blueprint convolution layer, a seventh 3×3 unconstrained blueprint convolution layer, an eighth 3×3 unconstrained blueprint convolution layer, a second 1×1 convolution layer, a third 1×1 convolution layer, a fourth 1×1 convolution layer, a fifth 1×1 convolution layer, a first maximum pooling submodule and a first upsampling layer. The feature map C input to the first basic module is processed by the sixth 3×3 unconstrained blueprint convolution layer and the second 1×1 convolution layer, respectively, to obtain feature maps C1 and C2, and the feature map C1 is processed by the seventh 3×3 unconstrained blueprint convolution layer and the first upsampling layer, respectively. Correspondingly, feature maps C3 and C4 are obtained. After feature map C3 is processed by the first maximum pooling sub-module, feature map C5 is obtained. After feature map C5 is processed by the eighth 3×3 unconstrained blueprint convolution layer and the third 1×1 convolution layer respectively, feature maps C6 and C7 are obtained accordingly. After feature map C6 is processed by the fourth 1×1 convolution layer, feature map C8 is obtained. Feature maps C, C2, C4 and C8 are concatenated to obtain feature map C9. After feature map C9 is processed by the fifth 1×1 convolution layer, feature map C10 which is input to the second basic module is obtained.

[0050] like Figure 8As shown, the second basic module includes the sixth 1×1 convolution layer, the seventh 1×1 convolution layer, the eighth 1×1 convolution layer, the ninth 1×1 convolution layer, the ninth 3×3 unconstrained blueprint convolution layer, the first 5×5 unconstrained blueprint convolution layer, the first 7×7 unconstrained blueprint convolution layer, the second maximum pooling submodule, the second upsampling layer, and the first Sigmoid activation layer. After the feature map C10 is processed by the sixth 1×1 convolution layer, the feature map D1 is obtained. After the feature map D1 is processed by the seventh 1×1 convolution layer and the second maximum pooling submodule, the feature map D2 and the feature map D3 are obtained respectively. The feature map D3 is processed by the ninth 3×3 unconstrained blueprint convolution layer, After the first 5×5 unconstrained blueprint convolution layer and the first 7×7 unconstrained blueprint convolution layer are processed, feature maps D4, D5 and D6 are obtained respectively. Feature maps D4, D5 and D6 are input to the eighth 1×1 convolution layer for processing to obtain feature map D7. Feature map D7 is processed by the second upsampling layer to obtain feature map D8. Feature map D2 and feature map D8 are added to obtain feature map D9. Feature map D9 is processed by the ninth 1×1 convolution layer and the first Sigmoid activation layer in turn to obtain feature map D10. Channel attention operation is performed on feature map D10 and feature map C10 ( Figure 8 denoted by T in the figure), and obtain the output features of the second basic module.

[0051] like Fig. 9 As shown in the figure, based on the EuroSAT dataset, Origin is the original image; C-DM is the classic degradation model; RS-DM: remote sensing degradation model (degradation model of the present invention). For severely degraded remote sensing images (such as satellite images with strong atmospheric interference), the degradation model of the present invention effectively simulates the influence of atmospheric scattering through the frequency domain filtering module, and reconstructs the image using a joint spatial domain-frequency domain super-resolution network. The experimental results show that the reconstructed image has significant improvements in detail restoration and texture preservation compared to the existing technology.

[0052] like Figure 10-11 As shown, based on the dataset WHU-RS19 and the dataset PatternNet, For remote sensing images with complex textures (such as remote sensing images of urban buildings), the spatial domain branch of the present invention can effectively capture the detailed features of different scales in the image, while the frequency domain branch enhances the ability to restore high-frequency information. Experimental results show that the present invention performs well in the reconstruction of building edges and road structures, with clear details and accurate structures.

[0053] Experiments show that the method proposed in the present invention can significantly improve the image reconstruction effect in various degradation scenarios, especially has obvious advantages in high-frequency detail and texture restoration.

[0054] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the disclosure of the present invention can be performed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0055] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A remote sensing image super-resolution reconstruction method based on degradation mechanism, characterized by: The specific steps include: S1: construct a degradation model, input the high-resolution remote sensing image into the degradation model, and obtain a degraded image corresponding to the high-resolution remote sensing image; S2: construct a super-resolution network, input the degraded image into the super-resolution network, and obtain a reconstructed image; The super-resolution network includes a spatial domain branch and a frequency domain branch. The spatial domain branch extracts local multi-scale feature information of the degraded image through multiple levels of convolution operations to obtain spatial domain features. The frequency domain branch analyzes and processes the global frequency features of the degraded image to obtain frequency domain features. The spatial domain features and the frequency domain features are fused to obtain a reconstructed image; S3: Use the joint loss function to train the super-resolution network to obtain a trained super-resolution network; The joint loss function includes a pixel-level loss function, a multi-scale feature loss function and a global feature loss function; S4: Input the degraded image to be reconstructed into the trained super-resolution network to obtain a reconstructed super-resolution image.

2. The remote sensing image super-resolution reconstruction method based on degradation mechanism according to claim 1, characterized in that: In step S1, the degradation model includes a spatial domain filtering module, a frequency domain filtering module and a degradation factor random control module, the high-resolution remote sensing image is input into the spatial domain filtering module for processing to obtain a first spatial domain image; the high-resolution remote sensing image is Fourier transformed and then input into the frequency domain filtering module, the degradation factor random control module controls the frequency domain filtering module to randomly introduce degradation factors into the high-resolution remote sensing image after Fourier transformation, and the high-resolution remote sensing image after the degradation factor is introduced is inverse Fourier transformed to obtain a second spatial domain image; The first spatial domain image and the second spatial domain image are synthesized, and then a downsampling operation and a noise introduction operation are performed to obtain a degraded image.

3. The remote sensing image super-resolution reconstruction method based on degradation mechanism according to claim 2, characterized in that: The frequency domain filtering module includes an atmospheric disturbance factor, an optical system disturbance factor, a first platform component disturbance factor, a second platform component disturbance factor, and a third platform component disturbance factor; Atmospheric disturbance factor I is: ; ； in, is the angular spatial frequency, , is the focal length of the optical system, is the spatial frequency, is the refractive index structure constant, is the wavelength, R is the propagation path, a1 is the turbulence effect, a2 is the aerosol effect, is the modulation transfer function of the turbulence effect, is the modulation transfer function of the aerosol effect; ； in, is the scattering coefficient of atmospheric aerosol, is the atmospheric aerosol absorption coefficient, is the angular spatial cutoff frequency of the aerosol; The optical system disturbance factor II is: ; ； in, for , is the spatial frequency in the horizontal direction, is the spatial cutoff frequency, = , D is the diameter of the optical system, b1 is the diffraction effect, b2 is the defocus effect, b3 is the image shift effect, is the modulation transfer function of the diffraction effect, is the modulation transfer function of the defocus effect, is the modulation transfer function of the image shift effect; ； in, is a first-order Bessel function, , is the defocus amount, F is the focal ratio of the optical system, It is an intermediate parameter and has no physical meaning; ； ； in, is the image shift caused by the change in satellite velocity, is the image shift caused by the satellite yaw, is the image shift caused by the change of satellite attitude angle, is the image shift caused by the rotation of the Earth, Image shift caused by random vibration of satellite and optical remote sensor. is the image shift caused by the jitter of the integration time, It is an intermediate variable and has no physical meaning; The disturbance factor III of the first platform component is: ; ； Where N is the sampling frequency, d is the diffuse spot diameter, For low frequency vibration, It is high frequency vibration; ； in, is the zero-order Bessel function, is the platform amplitude, is the spatial frequency; The disturbance factor IV of the second platform component is: ； ； in, is the spatial frequency along the scanning direction, is the CCD pixel size, d1 is the spatial response, d2 is the temporal response, d3 is the CCD response, is the modulation transfer function of the spatial response, is the modulation transfer function of the time response, is the modulation transfer function of the CCD response; ； in, is the time frequency value along the scanning direction, is the 3-dB point frequency value of the detector time response; ； in, is the amount of charge transferred by the CCD, is the charge transfer efficiency, is the sampling frequency of the CCD signal; The third platform component disturbance factor V is: ； ； in, is the frequency value of the -3dB point of the preamplifier response, e1 is the preamplifier response, e2 is the postamplifier response, and e3 is the lift response. is the modulation transfer function of the preamplifier response, is the modulation transfer function of the post-amplifier response, is the modulation transfer function of the lift response; ； in, The frequency value of the -3dB point of the post-amplifier response; ； in, is the frequency value when the boost pressure is maximum, for The magnitude of the value at .

4. The remote sensing image super-resolution reconstruction method based on degradation mechanism according to claim 3 is characterized in that: The formula for the random control module of the degradation factor to control the frequency domain filtering module to randomly introduce the degradation factor into the high-resolution remote sensing image after Fourier transformation is: K= · · ； Among them, U is uniform distribution, X1~X5 are coefficients that obey uniform distribution, n1 is 0, and n2 is 1.

5. The remote sensing image super-resolution reconstruction method based on degradation mechanism according to claim 1, characterized in that: In step S2, the spatial domain branch includes a first 3×3 unconstrained blueprint convolution layer and n cascaded basic units, each basic unit includes a cascaded first basic module and a second basic module, and the output of the first basic module is the input of the second basic module; The degraded image is copied four times, and three of the degraded images are subjected to RGB three-channel extraction operations, and the remaining degraded image is subjected to Y channel extraction operations, and all the extracted feature maps are subjected to feature splicing operations to obtain a spliced ​​feature map, and the spliced ​​feature map is subjected to a convolution operation by a first 3×3 unconstrained blueprint convolution layer to obtain an intermediate feature map, and the intermediate feature map is input into n cascaded basic units for processing to obtain a first feature map; The frequency domain branch includes a mask, a second 3×3 unconstrained blueprint convolution layer, a third 3×3 unconstrained blueprint convolution layer, a fourth 3×3 unconstrained blueprint convolution layer, a first 3×3 convolution layer, a second 3×3 convolution layer, a third 3×3 convolution layer, a first batch of normalization layers, a second batch of normalization layers, a third batch of normalization layers, a first SiLU activation function, a second SiLU activation function, and a third SiLU activation function; Performing Fourier transformation on the degraded image to obtain a frequency domain image, performing a dot multiplication operation on the frequency domain image and the mask to obtain frequency domain images under different masks, and processing the frequency domain images under different masks through an attention mechanism to obtain a second feature map, a third feature map, and a fourth feature map; The second feature map is input into the second 3×3 unconstrained blueprint convolution layer for processing to obtain the fifth feature map. The fifth feature map is processed by the first 3×3 convolution layer, the first batch of normalization layers and the first SiLU activation function in sequence to obtain the sixth feature map. The fifth feature map and the sixth feature map are feature concatenated to obtain the seventh feature map. The seventh feature map is inverse Fourier transformed to obtain the eighth feature map. The third feature map is input into the third 3×3 unconstrained blueprint convolution layer for processing to obtain the ninth feature map. The ninth feature map is processed by the second 3×3 convolution layer, the second batch of normalization layers and the second SiLU activation function in sequence. After processing, the tenth feature map is obtained, the ninth feature map and the tenth feature map are feature-concatenated to obtain the eleventh feature map, the eleventh feature map is inverse Fourier transformed to obtain the twelfth feature map; the fourth feature map is input into the fourth 3×3 unconstrained blueprint convolution layer for processing to obtain the thirteenth feature map, the thirteenth feature map is sequentially processed by the third 3×3 convolution layer, the third batch normalization layer and the third SiLU activation function to obtain the fourteenth feature map, the thirteenth feature map and the fourteenth feature map are feature-concatenated to obtain the fifteenth feature map, the fifteenth feature map is inverse Fourier transformed to obtain the sixteenth feature map.

6. The remote sensing image super-resolution reconstruction method based on degradation mechanism according to claim 5, characterized in that: The output features of the first basic module of each basic unit, the first feature map, the eighth feature map, the twelfth feature map and the sixteenth feature map are feature-concatenated to obtain a seventeenth feature map. The seventeenth feature map is sequentially convolved by the first 1×1 convolution layer and the fifth 3×3 unconstrained blueprint convolution layer to obtain an eighteenth feature map after feature supplementation. The eighteenth feature map and the intermediate feature map are added to obtain a nineteenth feature map. The nineteenth feature map is sequentially processed by the fourth 3×3 convolution layer and upsampling to obtain a reconstructed image.

7. The remote sensing image super-resolution reconstruction method based on degradation mechanism according to claim 1, characterized in that: In step S3, the output of the joint loss function is the cumulative sum of the pixel-level loss function, the multi-scale feature loss function and the global feature loss function; The pixel-level loss function is an L1 loss function, and the input of the pixel-level loss function is a degraded image and a reconstructed image; The multi-scale feature loss function is a Huber loss function with a parameter of 1, and the input of the multi-scale feature loss function is the reconstructed stitched image and the high-resolution stitched image; The global feature loss function is a Huber loss function with a parameter of 1.25, and the input of the global feature loss function is the reconstructed stitched image and the high-resolution stitched image.

8. The remote sensing image super-resolution reconstruction method based on degradation mechanism according to claim 7, characterized in that: The reconstructed stitched image is obtained by: inputting the reconstructed image into the global feature extraction module and the local feature extraction module respectively for feature extraction, and stitching the extracted global features and all local features to obtain the reconstructed stitched image; The high-resolution stitched image is obtained by inputting the high-resolution remote sensing image into the global feature extraction module and the local feature extraction module for feature extraction, and stitching the extracted global features and all local features to obtain a high-resolution stitched image.

9. The remote sensing image super-resolution reconstruction method based on degradation mechanism according to claim 8, characterized in that: The global feature extraction module includes a fifth 3×3 convolutional layer, an activation layer submodule, a pooling layer submodule, a fully connected layer submodule, a flattening transposition submodule, a first encoder, a second encoder, a third encoder, a fourth encoder, and a fifth encoder; The degraded image or the reconstructed image is input into the global feature extraction module, and the degraded image or the reconstructed image is processed by the fifth 3×3 convolutional layer, the activation layer submodule, the pooling layer submodule and the fully connected layer submodule in sequence to obtain a feature map A1. The feature map A1 is flattened and transposed by the flattening and transposing submodule to obtain a feature map A2. The feature map A2 is input into the first encoder for encoding to obtain a feature map A3. The feature map A3 is input into the second encoder for encoding to obtain a feature map A4. The feature map A4 is input into the third encoder for encoding to obtain a feature map A5. The feature map A5 is input into the fourth encoder for encoding to obtain a feature map A6. The feature map A6 is input into the fifth encoder for encoding to obtain a feature map A7. The feature maps A3, A4, A5, A6 and A7 are processed by using the attention mechanism to obtain the global features of the degraded image or the reconstructed image.

10. The remote sensing image super-resolution reconstruction method based on degradation mechanism according to claim 8, characterized in that: The local feature extraction module includes a first basic unit convolution layer, a second basic unit convolution layer, a third basic unit convolution layer, a fourth basic unit convolution layer, a fifth basic unit convolution layer, a first activation layer, a second activation layer, a third activation layer, a fourth activation layer, a fifth activation layer, a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer, a fourth maximum pooling layer and a fifth maximum pooling layer; The degraded image or the reconstructed image is input into the local feature extraction module, and the degraded image or the reconstructed image is processed by the first basic unit convolution layer, the first activation layer and the first maximum pooling layer in sequence to obtain a feature map B1, and the feature map B1 is processed by the second basic unit convolution layer, the second activation layer and the second maximum pooling layer in sequence to obtain a feature map B2, and the feature map B2 is processed by the third basic unit convolution layer, the third activation layer and the third maximum pooling layer in sequence to obtain a feature map B3, and the feature map B3 is processed by the fourth basic unit convolution layer, the fourth activation layer and the fourth maximum pooling layer in sequence to obtain a feature map B4, and the feature map B4 is processed by the fifth basic unit convolution layer, the fifth activation layer and the fifth maximum pooling layer in sequence to obtain a feature map B5, and the feature maps B1, B2, B3, B4 and B5 are processed by the attention mechanism to obtain the local features of the degraded image or the reconstructed image.

Images