Remote sensing image super-resolution reconstruction method and device for optical satellite ground processing system

Abstract

The application discloses a remote sensing image super-resolution reconstruction method and device for an optical satellite ground processing system. The method acquires multi-temporal optical satellite images and performs azimuth correction, constructs a prior knowledge base containing boundary constraints, regional homogeneity constraints and perspective transformation matrices based on gradient response and image pixel-level segmentation; performs radiometric calibration and atmospheric correction on the corrected images, constructs a degradation model to generate training paired data; modulates shallow features by boundary constraints, clusters and integrates features by regional homogeneity constraints, and outputs multi-scale fusion feature tensors through channel re-calibration; performs hierarchical up-sampling on the feature tensors, introduces boundary preservation cost and geometric deformation penalty double regularization to generate an initial high-resolution image; adjusts the output radiation fidelity image channel by channel according to the radiation statistics of the corrected image; and dynamically adjusts the cost coefficient by comprehensively considering the no-reference and full-reference quality scores and cyclic feedback optimization until the score reaches the threshold to output the final reconstructed image. The geometric fidelity and radiation consistency are improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and specifically to a method and apparatus for super-resolution reconstruction of remote sensing images for optical satellite ground processing systems. Background Technology

[0002] With the rapid development of high-resolution Earth observation technology, optical satellite remote sensing imagery has been widely used in fields such as land resource surveys, urban planning, environmental monitoring, and disaster assessment. However, limited by factors such as the cutoff frequency of satellite sensor optical systems, detector pixel size, and atmospheric turbulence, the spatial resolution of existing optical satellite multispectral imagery is mostly in the range of meters to tens of meters, which is insufficient to meet the requirements for high-precision ground feature identification. Super-resolution reconstruction technology is an effective way to enhance the application value of remote sensing data.

[0003] Traditional super-resolution methods for remote sensing images include interpolation, reconstruction modeling, and example learning. Interpolation is computationally efficient but results in blurred edges; reconstruction modeling has limited adaptability to complex terrains; and example learning relies on external sample libraries and lacks cross-sensor generalization ability. In recent years, while super-resolution methods based on deep convolutional neural networks (such as SRCNN, VDSR, EDSR, and RCAN) have made some progress, their direct application to optical satellite remote sensing images still suffers from the following drawbacks: First, they lack effective utilization of prior geometric-structural information such as the continuity of remote sensing image edges and regional homogeneity, leading to jagged or broken edges in reconstructed images. Second, they lack explicit modeling of imaging physical processes such as sensor point spread function and atmospheric scattering, resulting in poor compatibility between different satellite platforms. Third, they neglect radiometric fidelity constraints in the super-resolution reconstruction process, leading to radiometric drift in the reconstructed results and difficulty in integrating them into the standardized production processes of optical satellite ground processing systems. Fourth, they lack a quality evaluation and feedback iteration mechanism that dynamically adjusts the optimization direction based on intermediate reconstruction quality states, resulting in slow network convergence and unstable performance.

[0004] To address the aforementioned issues, a super-resolution reconstruction method for remote sensing images oriented towards optical satellite ground processing systems needs to be proposed. This method achieves high geometric fidelity and high radiometric consistency in remote sensing image super-resolution reconstruction by constructing a multi-level geometric-structure prior field, establishing a physical degradation model, introducing dual regularization constraints, and combining radiometric fidelity correction and quality evaluation feedback iteration. Summary of the Invention

[0005] This invention addresses the problems in existing optical satellite remote sensing image super-resolution reconstruction technologies, such as insufficient prior utilization of geometric structure, lack of modeling of imaging physical processes, neglect of radiometric fidelity constraints, and lack of quality evaluation feedback mechanisms. It provides a method and apparatus for remote sensing image super-resolution reconstruction for optical satellite ground processing systems to improve the geometric fidelity and radiometric consistency of reconstructed images.

[0006] In a first aspect, embodiments of this application provide a method for super-resolution reconstruction of remote sensing images for optical satellite ground processing systems, the method comprising:

[0007] Acquire multi-temporal optical satellite images and perform orientation correction. Construct a prior knowledge base containing boundary constraints, regional homogeneity constraints, and perspective transformation matrices based on gradient response and image pixel-level segmentation. Output the corrected image and the prior knowledge base.

[0008] The corrected image is radiometrically calibrated and atmospherically corrected. A degradation model is constructed based on the sensor diffusion function to generate training pair data. The corrected image, the training pair data, and the prior knowledge base are synchronously fed into the network.

[0009] The shallow features modulated by the boundary constraints in the prior knowledge base and the clustering integration features of the region homogeneity constraints are used to output a multi-scale feature tensor after channel recalibration and to pass the perspective transformation matrix.

[0010] The multi-scale feature tensor is subjected to hierarchical upsampling, and dual regularization is introduced, which includes a boundary preservation cost derived from the boundary constraints and a geometric deformation penalty derived from the perspective transformation matrix, to generate an initial high-resolution image.

[0011] The initial high-resolution image is linearly adjusted channel by channel based on the radiometric statistics of the corrected image to output a radiometrically faithful image;

[0012] The no-reference quality score and the full-reference quality score relative to the reference image of the radiometric image are calculated. The coefficients of the boundary preservation cost and the geometric deformation penalty are dynamically adjusted based on the weighted result of the two, and the results are iteratively fed back to the network parameter optimization until the score reaches the threshold, and the final high-resolution reconstructed image is output.

[0013] In a second aspect, embodiments of this application provide a remote sensing image super-resolution reconstruction apparatus for an optical satellite ground processing system, applied to the method described in the first aspect, the system comprising:

[0014] The data acquisition and prior knowledge construction module is used to acquire multi-temporal optical satellite images and perform orientation correction. Based on gradient response and image pixel-level segmentation, it constructs a prior knowledge base containing boundary constraints, regional homogeneity constraints and perspective transformation matrices, and outputs the corrected image and the prior knowledge base.

[0015] The radiometric correction and sample generation module is used to perform radiometric calibration and atmospheric correction on the corrected image, construct a degradation model based on the sensor diffusion function to generate training pair data, and synchronously feed the corrected image, the training pair data and the prior knowledge base into the network.

[0016] The feature extraction and fusion module is used to utilize the boundary constraints modulate shallow features and the region homogeneous constraint clustering integration features in the prior knowledge base, output a multi-scale fused feature tensor after channel recalibration, and transmit the perspective transformation matrix;

[0017] The hierarchical upsampling and regularization module is used to perform hierarchical upsampling on the multi-scale fused feature tensor, and introduces dual regularization of boundary preservation cost derived from the boundary constraints and geometric deformation penalty derived from the perspective transformation matrix to generate an initial high-resolution image.

[0018] The radiation fidelity correction module is used to linearly adjust the initial high-resolution image channel by channel based on the radiation statistics of the corrected image, and output a radiation fidelity image.

[0019] The quality evaluation and feedback iteration module is used to calculate the no-reference quality score and the full-reference quality score relative to the reference image of the radiometric image. Based on the weighted result of the two, the coefficients of the boundary preservation cost and the geometric deformation penalty are dynamically adjusted and iteratively fed back to the network parameter optimization until the score reaches the threshold, and the final high-resolution reconstructed image is output.

[0020] Thirdly, embodiments of this application provide an electronic device, including:

[0021] processor;

[0022] Memory used to store processor-executable instructions;

[0023] The processor is configured to implement the remote sensing image super-resolution reconstruction method for optical satellite ground processing systems as described in the first aspect when executing the instructions.

[0024] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to perform the remote sensing image super-resolution reconstruction method for optical satellite ground processing systems as described in the first aspect.

[0025] This invention constructs a multi-level geometric-structural prior knowledge base containing boundary constraints, regional homogeneity constraints, and perspective transformation matrices, and embeds it into the entire feature extraction and upsampling process of a deep network. Combined with a physical degradation model constructed from sensor diffusion functions, dual regularization constraints, and channel-by-channel linear correction based on radiometric statistics, it effectively improves the continuity of ground feature edges and geometric fidelity of reconstructed images, while ensuring the consistency of radiometric distribution with ground processing system standards. Furthermore, through a dynamic weighted feedback iteration mechanism that integrates no-reference and full-reference quality scores, it achieves adaptive adjustment of the network optimization direction, improving reconstruction quality and convergence stability. Attached Figure Description

[0026] Figure 1 This is a schematic flowchart of a method for super-resolution reconstruction of remote sensing images for optical satellite ground processing systems, provided as an embodiment of this application.

[0027] Figure 2 This is an architectural diagram of a remote sensing image super-resolution reconstruction device for an optical satellite ground processing system, provided as an embodiment of this application.

[0028] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0030] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0031] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] Example 1

[0033] Figure 1 This is a schematic flowchart illustrating a method for super-resolution reconstruction of remote sensing images for an optical satellite ground processing system, provided as an embodiment of this application. Figure 1 As shown, a method for super-resolution reconstruction of remote sensing images for optical satellite ground processing systems includes:

[0034] S1. Acquire multi-temporal optical satellite images and perform orientation correction. Based on gradient response and image pixel-level segmentation, construct a prior knowledge base containing boundary constraints, regional homogeneity constraints, and perspective transformation matrices. Output the corrected image and the prior knowledge base.

[0035] This step is used for data preparation and structural prior extraction. Azimuth correction aims to eliminate image geometric distortions caused by different satellite orbital inclination and yaw angles, establishing a unified geometric benchmark for subsequent processing. Gradient response refers to calculating the direction and intensity of the most drastic brightness changes for each pixel in the image, used to capture ground feature edges. Image pixel-level segmentation divides the image into several small regions with similar internal textures. Based on these two processing results, three types of prior information are constructed: boundary constraints to define the location where edges should appear, region homogeneity constraints to ensure consistent features within the same ground feature, and perspective transformation matrices to describe the local geometric mapping relationships between images from different time phases, providing clear spatial guidance for edge enhancement and geometric correction in the subsequent network.

[0036] Specifically, in this embodiment, the construction of a prior knowledge base based on gradient response and pixel-level segmentation, including boundary constraints, region homogeneity constraints, and perspective transformation matrices, includes:

[0037] S1.1 Calculate the gradient magnitude and suppress non-maximum values ​​of the corrected image, extract edge response points and generate an initial edge response map; perform superpixel segmentation on the corrected image to obtain a pixel-level segmentation map.

[0038] This sub-step is used for the initial extraction of edges and regions. Gradient magnitude calculation involves determining the rate of change of brightness for each pixel in the horizontal and vertical directions. The square root of the sum of the squares of these two values ​​yields the gradient intensity, reflecting the probability that the pixel belongs to an edge. Non-maximum suppression is a thinning process that retains only the pixels with the largest local gradients in the gradient direction, suppressing non-maximum responses and compressing the edge line width to the single-pixel level, thus obtaining a discrete set of edge response points, forming the initial edge response map. Superpixel segmentation divides the image into several small regions with consistent internal brightness and texture but without overlap. Each region is called a superpixel, serving as the basic unit for subsequent region analysis, rather than a single pixel.

[0039] S1.2. Using the edge response points as seed points, perform boundary shifting along the gradient direction to spread the edge response points into continuous edge lines, and clip and correct the continuous edge lines according to the superpixel boundaries in the image pixel-level segmentation map to generate the boundary constraints.

[0040] This sub-step connects discrete edge points into complete edge lines. Single-pixel-level edge response points often contain breaks or isolated noise, making them unsuitable as direct effective constraints. Using each edge response point as a seed, the process extends along both sides of the gradient direction of that point, connecting neighboring response points with the same direction to form a continuous edge line segment. However, during this extension process, the edge line may cross the boundaries of different land cover types, resulting in deviations. Therefore, it is necessary to trim the edge line according to the boundaries of the superpixel segmentation map, ensuring the edge line terminates at the superpixel boundary. Simultaneously, the direction of the edge line is fine-tuned and corrected based on the superpixel boundary position, ultimately forming precise spatial edge position constraints.

[0041] S1.3. Using each superpixel region in the pixel-level segmentation map as the basic unit, the grayscale mean and texture response of pixels in each region are statistically analyzed to establish a region homogeneous descriptor and construct an adjacency relationship graph between adjacent superpixels. Based on the confidence propagation algorithm, the region attribute information is iteratively transmitted on the adjacency relationship graph. The transmission cost is used to quantify the difficulty of merging adjacent superpixel regions, realize the merging and optimization of homogeneous regions, and generate the region homogeneous constraint.

[0042] This sub-step is used to merge overly fine superpixel segments into larger, semantically meaningful regions. The grayscale mean reflects the average brightness level of the region, while the texture response reflects the regularity of pixel grayscale variations within the region. Together, they constitute a region homogeneous descriptor, used to measure whether two adjacent superpixels belong to the same land cover type. Each superpixel is treated as a node, and adjacency relationships as connections, forming an adjacency graph. Confidence propagation is an algorithm for information transmission on a graph structure. Each node transmits its class judgment belief to its neighbors and receives beliefs from its neighbors. After multiple iterations, the beliefs across the entire graph tend to converge, ultimately merging adjacent superpixels with similar descriptors into larger homogeneous regions, forming region homogeneity constraints.

[0043] S1.4 Extract stable feature points within each superpixel region, perform feature point matching across temporal images, fit a local perspective transformation model to the matching point pairs, and generate the perspective transformation matrix.

[0044] This sub-step establishes the spatial geometric correspondence between multi-temporal images. Feature points invariant to changes in brightness, scale, and rotation are selected within the superpixel region to ensure stable detection of the same ground location across different temporal images. Cross-temporal matching pairs feature points in the current image with those in another temporal image, based on the similarity of texture descriptors in the local neighborhood of each feature point. After obtaining a sufficient number of matching point pairs, a perspective transformation model is fitted individually to each local region. This model describes the positional mapping of a planar region across different temporal images after translation, rotation, scaling, and tilting, represented in matrix form, and serves as the basis for subsequent geometric deformation penalties.

[0045] Furthermore, in this embodiment, the confidence-based propagation algorithm iteratively transmits region attribute information on the adjacency graph. The transmission cost is used to quantify the difficulty of merging adjacent superpixel regions, thereby achieving the merging and optimization of homogeneous regions. Specifically, this includes:

[0046] S1.3.1. Treat each superpixel region in the adjacency graph as a node, define the node state as its corresponding land cover category label, and define the region merging cost between adjacent nodes as a weighted sum of spectral feature differences and texture feature differences. This sub-step is used to establish a cost metric for region merging. The adjacency graph is an undirected graph structure where each node represents a superpixel region, and the lines connecting nodes indicate that two superpixels are spatially adjacent. The land cover category label refers to the type of land cover each superpixel is classified as, such as vegetation, water bodies, buildings, bare soil, etc. The merging cost is a numerical value used to evaluate whether it is reasonable to merge two adjacent superpixels into the same category. Spectral feature differences refer to the difference between the average brightness values ​​of two regions across multiple bands, and texture feature differences refer to the difference between the pixel arrangement patterns within two regions. The two are weighted and added together; the smaller the difference, the lower the merging cost, meaning that the two are more likely to belong to the same category.

[0047] To quantify the rationality of merging adjacent superpixel regions, this invention defines the following region merging cost function with joint spectral and texture weights. Compared to the traditional merging criterion that only uses spectral difference, this formula introduces texture structure consistency constraints, which can effectively prevent heterogeneous objects from being incorrectly merged due to similar spectra:

[0048] ,

[0049] in, For adjacent superpixel nodes and The cost of merging regions is in the range of [0, +∞), and the smaller the value, the easier it is to merge the two regions. , Superpixels , The spectral feature vector, where B is the number of image bands. The Each component is a superpixel In the The average gray level of the band; This is an L2 norm operator used to calculate the Euclidean distance of spectral vectors, quantifying differences in spectral features; , Superpixels , The texture response matrix, where L is the side length of the local texture window, and the matrix elements are the gray-level co-occurrence matrix statistics (such as contrast, entropy, correlation) within the corresponding window. Matrix dot product operation is used to calculate the texture similarity between two superpixels; These are the spectral difference weighting coefficients and the texture difference weighting coefficients, respectively, satisfying... It can adaptively adjust according to the type of terrain features, for example, taking urban areas. To emphasize texture constraints, vegetation areas are selected To highlight spectral constraints.

[0050] In step S1.3.1, before each iteration, the spectral mean and texture matrix of each superpixel are calculated, and then substituted into the formula to calculate the mean of all adjacent superpixel pairs. As a message transmission cost for confidence propagation, only when Merging of two types of superpixels is only allowed when the value is less than the preset merging threshold.

[0051] S1.3.2 Initialize category label confidence at each node. Forward message passing is performed from the leaf nodes of the adjacency graph to the root node. Each node updates its own label confidence vector based on the received messages from neighboring nodes. This sub-step is used to propagate category information from bottom to top. A leaf node is a superpixel located at the edge of the adjacency graph with only one neighboring node. The root node is a pre-selected center or top node. Confidence is a probability value representing the degree of certainty that a node belongs to a certain category label. Initially, an initial confidence value is assigned based on the spectral and texture features of each node. Forward message passing refers to sending messages layer by layer from the leaf nodes towards the root node. The content sent by each node to its parent node is an estimate of the probability of each category of the parent node based on its own observations and information transmitted from its child nodes. After receiving messages from its child nodes, each node corrects its own confidence vector for each category accordingly, so that the category judgment of the lower-level region gradually affects the upper-level region.

[0052] S1.3.3 After completing the forward pass, a backward message backtracking is performed from the root node to the leaf nodes. Each node calculates the marginal posterior probability by combining the forward and backward messages, and selects the category label corresponding to the highest posterior probability as the final category label for that node. This sub-step is used to correct the category determination from top to bottom. The forward pass only realizes the convergence of information from the local to the global level, but the reverse influence of the upper layer on the lower layer has not yet been established. The backward message backtracking starts from the root node and transmits the category determination results of the higher level back to each level of child nodes. After receiving the backtracking message from the parent node, each node combines the child node messages received during the forward pass with the parent node messages received during the backward backtracking to calculate the marginal posterior probability of each possible category label, that is, the final probability value of the node belonging to a certain category after considering the information of all adjacent nodes. The category with the highest probability is selected as the final decision result for that node.

[0053] S1.3.4. Based on the final category labels of each node, adjacent superpixel regions with the same category label are merged, and the region homogeneity descriptor is recalculated using the merged region unit. The above message passing and region merging process is iteratively executed until the region label change rate in two adjacent iterations is lower than the preset convergence threshold, and the optimized region homogeneity constraint is output. This sub-step is used for iterative merging and convergence determination. After obtaining the final category label of each node, spatially adjacent superpixels with the same label are merged into a larger region unit. The number of regions is reduced after merging, and the homogeneity descriptor of each region needs to be recalculated because the mean and texture features of pixels in the merged region may change. Then, the adjacency graph is reconstructed using the merged region as the new node, and message passing and label determination are executed again, and so on iteratively. The region label change rate refers to the proportion of nodes whose category has changed compared to the previous round in the current iteration to the total number of nodes. When this proportion is lower than the preset convergence threshold, the merging result is considered to have stabilized, the iteration stops, and the final region homogeneity constraint is output.

[0054] S2. Perform radiometric calibration and atmospheric correction on the corrected image, construct a degradation model based on the sensor diffusion function to generate training pair data, and synchronously feed the corrected image, the training pair data, and the prior knowledge base into the network.

[0055] This step is used for radiometric normalization and training sample preparation. Radiometric calibration converts the raw digital quantization values ​​recorded by the sensor into physically meaningful radiance values, while atmospheric correction eliminates the absorption and scattering effects of atmospheric molecules and aerosols on the surface reflection signal, ensuring that the image accurately reflects the surface reflection characteristics. The sensor diffusion function describes the response distribution of the optical satellite imaging system to an ideal point light source, reflecting the blurring effect of the optical system. Based on this function, a degradation model is constructed to simulate the complete process of high-resolution images becoming low-resolution images after optical blurring, spatial downsampling, and detector noise. This generates a large number of high- and low-resolution image pairs for network training, enabling the network to learn reconstruction capabilities that match the degradation characteristics of the target satellite platform.

[0056] This step is used to achieve radiometric normalization and prepare training samples that conform to the physical characteristics of satellite imaging. It specifically includes the following sub-steps:

[0057] S2.1 Radiometric Calibration and Atmospheric Correction. First, read the raw digital quantization values ​​of the corrected image. The radiation is converted into the radiance of the top of the atmosphere using the sensor's radiation calibration coefficient. :

[0058] ,

[0059] in, The sensor calibration gain and offset for band b were obtained from the satellite ephemeris parameter file. Subsequently, atmospheric correction was performed using the 6S radiative transfer model to eliminate the effects of scattering and absorption by atmospheric molecules and aerosols, yielding the true surface reflectance. This serves as the reference radiation value for subsequent processing.

[0060] ,

[0061] in, Atmospheric path radiation, This refers to atmospheric upward transmittance. Atmospheric downward transmittance, The solar zenith angle is given above; all parameters were calculated by the 6S model based on image imaging time, geographical location, and atmospheric model parameters. The final radiation output of the corrected image is denoted as... Where H and W are the height and width of the corrected image, and B is the number of bands.

[0062] S2.2 Degradation Model Construction Based on Sensor Spread Function. The point spread effect of optical satellite sensors is described by the sensor spread function (SSF). This invention uses a Gaussian spread function to simulate the blurring effect of the optical system:

[0063] ,

[0064] in Let be the coordinates of the diffusion function in the spatial domain. The diffusion coefficient corresponding to band b is related to the focal length, pixel size, and orbital altitude of the sensor optical system and can be found in the satellite payload parameter manual.

[0065] To simulate the degradation process in real imaging, this invention constructs the following physics-driven fourth-order degradation model, which uses high-resolution surface reflectance... (That is, the label data for subsequent training can be obtained by fusing high-resolution images from historical archives or multi-temporal images from step S1) and mapped to low-resolution simulated images. :

[0066] ,

[0067] in, 1 is a two-dimensional convolution operator, representing the convolution operation between the high-resolution image and the sensor spread function, simulating the spatial blurring effect of the optical system; ↓s is a downsampling operator, where the sampling factor s is the target magnification factor of the super-resolution reconstruction (e.g., s=2,4), and bicubic interpolation downsampling is used to simulate the spatial integral effect of the detector pixel. The term represents electronic noise and follows a Gaussian distribution. As a quantization operator, floating-point radiometric values ​​are mapped to DN quantization levels consistent with the original corrected image, simulating the detector's analog-to-digital conversion process. This degradation model allows for the batch generation of training paired data that conforms to the imaging characteristics of the target satellite. , where N is the number of samples, and each sample pair corresponds to a low-resolution simulation image and a high-resolution labeled image of the same scene.

[0068] S2.3 Synchronous Feed-in Strategy for Multi-Source Data. The generated corrected image... Training pairing data The prior knowledge base output in step S1 (in For boundary weight graph, For regional homogeneous constraint sets, After spatiotemporal alignment of the perspective transformation matrix, it is synchronously input into a deep convolutional neural network: Spatial alignment: All spatial dimensions of the training paired data and prior knowledge base are uniformly resampled to match the corrected image. Consistency ensures a one-to-one correspondence at the pixel level; Channel alignment: Normalize the data of each band of the multi-band image to the [0,1] interval in the same order to avoid the impact of differences in the units of different data sources on the stability of network training; Feed method: ... As the basic input to the network, the training pair data is used for loss calculation during the supervised training phase. The prior knowledge base serves as an auxiliary input for feature modulation and regularization constraints, but does not participate in network parameter updates; it only acts as a spatial prior guidance signal. The training samples generated by the aforementioned physics-driven degradation model enable the network to learn reconstruction capabilities that highly match the degradation characteristics of the target satellite platform, solving the problems of traditional deep learning methods that rely on external sample libraries and lack cross-sensor generalization ability.

[0069] S3. Utilize the boundary constraints in the prior knowledge base to modulate shallow features and the region homogeneity constraints to cluster and integrate features. Output a multi-scale feature tensor after channel recalibration and pass the perspective transformation matrix. This step is used for prior-guided deep feature extraction and fusion. Shallow features are detailed information extracted by the front-end convolutional layers of the network, such as edges, corners, and textures. They have high resolution but weak semantics. Boundary constraints play a modulation role here, applying higher response weights to pixels near the edges, making the network more sensitive to texture changes in that area. Mid-layer features have stronger structural semantics. Region homogeneity constraints are used to cluster and integrate features in this layer, forcibly bringing feature vectors within the same geographic area closer together, making the feature expressions of similar areas more consistent. Channel recalibration is an attention mechanism that automatically judges the importance of each feature channel, enhances the response of valuable channels, suppresses the response of redundant channels, outputs a multi-scale feature tensor that fuses multi-level information, and simultaneously passes the perspective transformation matrix to subsequent steps as geometric constraints.

[0070] Specifically, in this embodiment, the step of utilizing the boundary constraints to modulate shallow features and the region homogeneous constraint clustering integration features in the prior knowledge base, outputting a multi-scale fused feature tensor after channel recalibration, and transmitting the perspective transformation matrix specifically includes:

[0071] S3.1. The corrected image and the training paired data are input into the front end of a deep convolutional neural network. Shallow texture features, mid-level structural features, and deep semantic features under different receptive fields are extracted through initial convolutional layers, denoted as first-level features, second-level features, and third-level features, respectively. This sub-step is used for the initial extraction of multi-level features. The receptive field refers to the area covered by a pixel on the output feature map of a certain layer of the network, corresponding to the input image. The larger the receptive field, the richer the contextual information contained in the feature. Shallow texture features are extracted through fewer convolutional layers, resulting in a smaller receptive field. They mainly capture detailed texture patterns such as edges, corners, and stripes in the image, and are denoted as first-level features. Mid-level structural features are extracted through convolutional layers of moderate depth, resulting in a moderate receptive field. They reflect structural information such as the shape and component composition of ground objects, and are denoted as second-level features. Deep semantic features are extracted through more convolutional layers, resulting in a larger receptive field. They encode the category classification and global contextual semantic information of ground objects, and are denoted as third-level features. These three types of features represent a progressive abstraction from details to semantics.

[0072] S3.2. The boundary constraints are converted into a boundary weight map with the same spatial size as the first-level feature. The boundary weight map is then used to perform point-by-point weighted modulation on each pixel position of the first-level feature, enhancing the texture response of edge regions and suppressing the texture response of non-edge regions, outputting the modulated shallow edge enhancement feature. This sub-step is used to enhance the detail representation of edge regions. The generation method of the boundary weight map is similar to that of the aforementioned spatial weight mask, mapping the edge line coordinates in the boundary constraints to the same grid size as the first-level feature. Point-by-point weighted modulation means applying a corresponding weight coefficient to each pixel position on the feature map. Edge positions are assigned a coefficient greater than one to amplify their texture response, while non-edge positions are assigned a coefficient less than one to suppress their response. After modulation, the feature response near the edges is more prominent, and the network will naturally focus more attention on the edge regions in subsequent processing, helping to maintain edge sharpness during reconstruction.

[0073] S3.3. Map the homogeneous region labels in the homogeneous region constraint to the spatial locations of the second-level features. Perform mean pooling on the feature vectors of all pixel locations belonging to the same homogeneous region label to obtain the region prototype vectors of each homogeneous region. Then, fill the region prototype vectors back into all pixel locations of the corresponding region to make the features within the same region more consistent, and output the clustered and integrated mid-level region aggregation features. This sub-step is used to make the feature expressions within the same ground feature converge. The homogeneous region constraint has divided the image into several large regions with consistent internal structures, each region having a unique label. Project these labels onto the second-level feature map to mark the region affiliation of each pixel. For each region, perform mean pooling on the feature vectors of all pixel locations within the region, i.e., average channel by channel, to obtain a prototype vector representing the overall features of the region. Then, copy and fill this prototype vector back into each pixel location within the region to make the features of all locations within the region completely consistent. This operation eliminates the feature differences caused by random texture variations within the same region, allowing the network to focus on the common features of the region as a whole.

[0074] S3.4. The modulated shallow edge enhancement features, the clustered and integrated mid-level region aggregation features, and the third-level features are concatenated along the channel dimension and input into the channel attention module. The channel attention module performs global average pooling and global max pooling on each channel feature, generates activation weights for each channel through a shared multilayer perceptron, and recalibrates each concatenated channel using the activation weights to output the multi-scale fusion feature tensor.

[0075] This sub-step is used to fuse features from three different levels and adaptively recalibrate channel importance. Concatenation along the channel dimension merges the three sets of features in the direction of channel count, forming a fused feature with more channels, preserving complete information from details to semantics. The channel attention module is a dynamic channel selection mechanism. Global average pooling compresses the overall response of each channel into a single value reflecting the average activation level of the channel, while global max pooling compresses the maximum response of each channel into a single value reflecting the peak activation level of the channel. Both are used in parallel to extract channel information from different aspects. The shared multilayer perceptron is a small, fully connected network that receives these two values, comprehensively judges the importance of each channel, and outputs activation weights. Larger values ​​indicate that the channel is more critical to the current reconstruction task, while smaller values ​​indicate limited contribution. Multiplying the activation weights by the original channel features enhances important channels and suppresses secondary channels, outputting a refined multi-scale fused feature tensor.

[0076] S3.5. The perspective transformation matrix is ​​associated with the multi-scale fusion feature tensor as geometric constraint information and then synchronously transmitted to subsequent steps.

[0077] This sub-step ensures that spatial geometric constraint information is transmitted synchronously with the features. The perspective transformation matrix represents the coordinate mapping relationship of a local image region and serves as the basis for subsequent geometric deformation penalty calculations. By binding the perspective transformation matrix with the multi-scale fusion feature tensor as additional parameters and transmitting it synchronously in the data stream, it ensures that the matrix can be invoked at any time during subsequent upsampling and loss calculation stages without needing to be extracted or matched again from the original image, thus guaranteeing the consistency of geometric constraint information throughout the entire process.

[0078] Furthermore, in this embodiment, converting the boundary constraints into a boundary weight map with the same size as the first-level feature space specifically includes:

[0079] S3.2.1 Obtain the set of continuous edge line coordinates in the boundary constraint. Using the set of continuous edge line coordinates as an index, in a blank weight map with the same size as the first-level feature space, set the pixel value at the position corresponding to the edge line to the first preset value, and set the pixel value at the position not on the edge line to the second preset value, thereby generating a binarized boundary weight map.

[0080] This sub-step converts the edge line coordinates into a binary mesh map. The continuous edge line coordinate set records the pixel row and column positions traversed by all edge lines. First, a blank weight map with the exact same size as the first-level feature map is created, with all pixels initially set to zero. Then, using the edge line coordinate set as an index, the pixel values ​​at these corresponding positions in the weight map are set to a first preset value, typically one, as a marker value, indicating that these positions belong to edge regions; other unindexed positions retain a second preset value, typically zero, indicating that these positions belong to non-edge regions, thus generating the initial binary boundary weight map.

[0081] S3.2.2 Perform Gaussian kernel convolution diffusion processing on the binarized boundary weight map, so that the weight values ​​at the edge line positions decay in a Gaussian distribution to both sides along the direction perpendicular to the edge line, and generate a diffused boundary weight map.

[0082] This sub-step smooths sharp binary edges into a gradual, banded weight distribution. The edges of the binarized boundary weight map are hard boundaries; the weight values ​​abruptly change from one to zero from the inside to the outside of the edge, lacking a transition zone. The Gaussian kernel is a two-dimensional bell-shaped distribution matrix, with the largest value at the center and gradually decreasing towards the edges. Convolving the binarized boundary weight map with the Gaussian kernel is equivalent to superimposing a bell-shaped distribution at each edge position, causing the weight values ​​at the edges to gradually decay from their maximum value along a direction perpendicular to the edge line, forming a gradually decreasing banded region of a certain width, rather than a hard edge of single pixel width. This allows for a smooth transition near the edges during subsequent weighted modulation.

[0083] S3.2.3. Normalize the diffusion boundary weight map to distribute the weight values ​​across the entire image within a preset numerical range, and output a normalized boundary weight map. This sub-step is used to unify the weight values ​​to a standardized numerical range. After Gaussian kernel convolution, the magnitude of the weight values ​​at each location in the image depends on the density of the edge lines and the convolution parameter settings. The absolute values ​​of the weights in different regions may vary significantly, and direct use will lead to uneven modulation intensity. Normalization scales the weight values ​​of all pixels in the entire image proportionally to a preset numerical range. Typically, the maximum value is scaled to the upper limit and the minimum value to the lower limit, standardizing the weight distribution. This facilitates uniform modulation application subsequently, ensuring that the modulation intensity is comparable and consistent across the entire image.

[0084] S3.2.4. Using the normalized boundary weight map as weight coefficients, the first-level features are multiplied and weighted channel-wise and pixel-wise. Edge regions receive higher weight coefficients than flat regions, thus modulating the shallow texture features by boundary constraints. This sub-step is used to actually apply the weight coefficients to the feature map for modulation. The first-level features are three-dimensional tensors, containing three dimensions: spatial height, spatial width, and number of channels. The values ​​of all channels at the same spatial location constitute the feature vector at that location. Channel-wise refers to operating on each feature channel individually, and pixel-wise refers to processing each location on the feature map separately. The weight value at each location in the normalized boundary weight map is multiplied by the feature values ​​of all channels at that location. Locations closer to the edge receive a larger multiplier, enhancing the feature response, while locations farther from the edge receive a smaller multiplier, weakening the feature response, thereby enhancing the expression of edge information in shallow features.

[0085] To avoid edge artifacts caused by traditional boundary-weighted methods that only perform binary enhancement at edge locations, this invention designs the following continuous boundary modulation formula. Through the coupling effect of Gaussian diffusion weights and feature gradients, it achieves smooth enhancement of the texture response in edge regions:

[0086] ,

[0087] in, This is the first-level shallow feature tensor before modulation, with dimension . , The feature map is defined by its height and width. This represents the number of first-level feature channels. For spatial coordinates, For channel indexing; For the modulated shallow edge enhancement feature tensor, the spatial size is... Maintain consistency; Normalized boundary weights in The weight value at each point is output by step S3.2.3, and its value range is [0,1]. The value in the edge region is close to 1, and the value in the flat region is close to 0. For first-level features The response value along the gradient direction is calculated by the Sobel operator and reflects the degree of drastic change in the characteristics at that location; This is the modulation intensity coefficient, used to control the edge enhancement amplitude. The default value is 0.5, and it can be adaptively adjusted according to the image edge sharpness.

[0088] For each channel and each spatial location of the first-level feature map, the feature gradient at that location is first calculated. Then, the normalized boundary weight map is substituted into the above formula to calculate the modulated feature tensor point by point, thereby achieving an adaptive modulation effect where the clearer the edge, the higher the weight, and the greater the enhancement.

[0089] S4. The multi-scale feature tensor is subjected to hierarchical upsampling, and dual regularization is introduced, which includes the boundary preservation cost derived from the boundary constraints and the geometric deformation penalty derived from the perspective transformation matrix, to generate an initial high-resolution image.

[0090] This step is used for spatial resolution restoration and structural constraints. Hierarchical upsampling refers to gradually increasing the resolution of the feature map, for example, first doubling it, then doubling it again, rather than doing it all at once. This reduces learning difficulty and artifacts. Subpixel convolution is the specific operation for achieving upsampling, expanding the spatial size of the feature map by rearranging pixels between channels. Boundary preservation cost serves as a regularization constraint, measuring whether the reconstruction result maintains the expected brightness abruptness at edge locations; if edges are smoothed, a significant penalty is imposed. Geometric deformation penalty, based on the perspective transformation matrix, constrains the relative spatial relationships between feature points during upsampling to avoid distortion; if the same feature exhibits stretching or compression deformation at different locations, a penalty is applied. These two penalties jointly constrain the optimization direction of the network, ensuring both edge sharpness and spatial geometric accuracy of the reconstructed image.

[0091] Specifically, in this embodiment, the step of hierarchically upsampling the multi-scale fused feature tensor, introducing dual regularization of boundary preservation cost derived from the boundary constraints and geometric deformation penalty derived from the perspective transformation matrix, to generate an initial high-resolution image includes:

[0092] S4.1. The multi-scale fused feature tensor is used as the initial input and processed through multiple upsampling modules in sequence. In each upsampling module, the spatial resolution of the feature map is first doubled through a sub-pixel convolutional layer, and then the channel dimension feature is refined through a residual dense connection block to output the high-resolution feature map of that level.

[0093] This sub-step is used to progressively restore the spatial resolution of the image. Multi-level upsampling refers to gradually increasing the resolution of the feature map multiple times, doubling it each time. For example, moving from low resolution to the target resolution may require two to three consecutive upsampling operations. This is easier to learn and results in more stable reconstruction quality than a single multiplication. Subpixel convolution is the specific operation for achieving resolution enhancement. It rearranges the feature map by channel group, transforming the channel dimension information into the spatial dimension, doubling the spatial size while reducing the number of channels. Although the feature map becomes larger after each upsampling, the channel dimension feature representation becomes coarse. Therefore, it is then refined using residual dense connection blocks. This module consists of multiple stacked convolutions. Each layer not only receives the output of the previous layer but also receives the outputs of all previous layers for fusion, ensuring that the features are fully refined before being output to the next level.

[0094] S4.2. At the output of each upsampling module, a boundary preservation cost module is introduced to convert the boundary constraint into a spatial weight mask. The spatial weight mask is used to apply a weighted loss to the high-resolution feature map to constrain the pixel response of the edge region and the pixel response of the background region to maintain a preset contrast difference threshold.

[0095] This sub-step is used to enforce edge sharpness during each level of upsampling. The spatial weight mask is a weight distribution map with the same spatial dimensions as the feature map, assigning larger weights to edge locations and smaller weights to flat areas. Its function is to indicate where the network should focus its attention. Weighted loss calculates the loss by multiplying the error of each pixel by the weight corresponding to that location, giving errors in edge regions a higher penalty weight in network optimization, forcing the network to accurately recover brightness abrupt changes at edge locations. The contrast difference threshold is a pre-set value that requires the brightness difference between pixels on either side of an edge to reach at least this threshold. If it is lower than the threshold, it indicates that the edge has been over-smoothed, requiring a stronger penalty to prevent the network from blurring the edges.

[0096] S4.3 At the output of each upsampling module, a geometric deformation penalty module is introduced to decompose the perspective transformation matrix into rotation components, scaling components, and translation components. The rotation deviation, scaling deviation, and translation deviation of the current feature map spatial coordinates relative to the reference coordinate system are calculated respectively, and the weighted sum of the three is used as the geometric deformation penalty term to constrain the spatial deformation amplitude during the upsampling process.

[0097] This sub-step is used to suppress spatial geometric distortion in each upsampling stage. The perspective transformation matrix is ​​a three-dimensional matrix that describes the coordinate transformation relationship of a local region when mapped from one temporal image to another. It includes rotation components (the angle by which the image is rotated), scaling components (how much the image is stretched or compressed horizontally and vertically, respectively), and translation components (how much the image as a whole is moved horizontally and vertically). The coordinates of each pixel in the feature map output from the current upsampling stage are substituted into this transformation model to calculate the theoretically expected position. This position is then compared with the actual position of each pixel in the feature map to obtain the deviation. The three deviations are squared, multiplied by their respective weights, and then summed as a penalty term. The larger the deviation, the heavier the penalty, thus constraining the upsampling process from producing excessive geometric distortion.

[0098] Existing techniques typically treat boundary preservation and geometric correction as independent loss terms, which can easily lead to conflicts between the two constraints. This invention proposes the following joint regularization loss function, which achieves synergistic optimization of the two constraints by dynamically balancing boundary clarity and geometric consistency through weights:

[0099] ,

[0100] in, The total loss is double-regularized and used to constrain the output features of the hierarchical upsampling process; The boundary preservation cost term is calculated using the following formula: , For high-resolution feature maps in gradient at, To correct the downsampling gradient at the corresponding location in the image, To preset the contrast difference threshold, The spatial weight mask output from step S4.2.3 is used as the loss term to constrain the difference between the edge gradient of the reconstructed image and the edge gradient of the original image to be no less than [value missing]. To avoid blurry edges; This is a geometric deformation penalty term, calculated using the following formula: , It is the Frobenius norm. , , These represent the rotational deviation, scaling deviation, and translation deviation of the current feature map relative to the reference coordinate system. Obtained by perspective transformation matrix decomposition. The weighting coefficients for the three types of biases satisfy the following conditions: ; The cost coefficient is maintained at the boundary and dynamically adjusted by the subsequent feedback iteration module. This is the geometric deformation penalty coefficient, dynamically adjusted by the subsequent feedback iteration module, and satisfies... This achieves a balance between the strengths of the two constraints.

[0101] At the output of each upsampling module, calculate respectively and Substituting into the formula, we get By incorporating this into the network's total loss function and backpropagating to optimize the network parameters, the dual constraints of edge clarity and geometric fidelity are achieved during the upsampling process.

[0102] S4.4 The final output of the multi-level upsampling module is mapped to the image space through a convolutional layer to generate the initial high-resolution image. This sub-step is used to convert the feature representation into a visual image. After multi-level upsampling and refining, the resolution of the feature map has reached the target high resolution, but its data form is still a deeply abstract feature vector, and the values ​​of each channel do not directly correspond to pixel brightness values. Through a convolutional operation, the number of channels of the feature map is compressed to be consistent with the number of bands of the target image, so that the feature vector at each position is mapped to the pixel brightness value of the corresponding band, completing the conversion from feature space to image space, and outputting an initial high-resolution reconstruction result that can be directly displayed.

[0103] Furthermore, in this embodiment, the introduction of the boundary preservation cost module, which converts the boundary constraints into a spatial weight mask, specifically includes:

[0104] S4.2.1. Binarize the continuous edge lines in the boundary constraints to generate an initial edge mask that matches the spatial size of the current-level feature map. In the initial edge mask, the pixel value at the edge line position is 1, and the pixel value at the non-edge position is 0. This sub-step is used to convert the edge constraints into a binary label map. The continuous edge lines in the boundary constraints are line information stored in the form of a coordinate set, which needs to be converted into a grid-like mask of the same size as the feature map before it can participate in subsequent calculations. Binarization means that on a blank weight map, all pixel positions located on the edge lines are marked as first-class values, and all other positions not on the edge lines are marked as second-class values, forming a black and white binary image, where bright lines represent edge positions and dark backgrounds represent flat areas. This image is the initial edge mask.

[0105] S4.2.2 Perform Euclidean distance transformation on the initial edge mask to calculate the distance value from each pixel position to the nearest edge line in the current level feature map. This sub-step is used to quantify the proximity of each pixel to the edge. The binary mask only distinguishes between "edge" and "not edge" states, but does not provide fine information about the transition region. Euclidean distance transformation is an operation performed on the binary image. For each non-edge pixel, the straight-line distance between it and the nearest edge pixel is calculated. After traversing the entire image, each pixel position obtains a distance value. The closer to the edge, the smaller the value; the farther from the edge, the larger the value. This converts the binary mask into a continuous distance distribution map.

[0106] S4.2.3. Construct a spatial weighting function based on the distance value. The spatial weighting function is an exponential decay function with the distance value as the independent variable, such that pixels closer to the edge line receive larger weight values, and pixels farther from the edge line receive smaller weight values. This sub-step is used to convert the distance value into a weighting coefficient that can be used for weighting. The distance value itself is inversely proportional to the weight, that is, a smaller distance results in a larger weight, and a larger distance results in a smaller weight. A direct linear mapping would cause the weight to remain at a high level in a large area around the edge, which is not conducive to fine constraints. The exponential decay function is a non-linear mapping method. The weight value decreases rapidly exponentially with increasing distance, so that the weight is mainly concentrated in a narrow strip area near the edge line. When the distance is slightly farther, the weight quickly returns to zero, thereby ensuring that the constraint only acts on a local area around the edge and does not affect the flat area inside the edge.

[0107] S4.2.4. Apply the spatial weight function to the current-level feature map, weighting the gradient response of each pixel in the feature map with the weight value, and outputting the spatial weight mask after normalization. This spatial weight mask assigns high weights to edge regions and low weights to flat regions. This sub-step generates the final spatial weight mask and applies constraints. Gradient response refers to the magnitude of brightness change of each pixel in the feature map in space; edge regions should have higher gradient responses. Multiply the weight value calculated by the spatial weight function with the gradient response of the corresponding pixel in the current feature map, so that regions near the edge with significant gradient changes receive high weight values, regions near the edge with weak gradient changes receive intermediate weight values, and regions far from the edge receive low weight values. Then, normalization is performed to scale all values ​​to a uniform range, outputting the final spatial weight mask. This mask is directly used as the loss weight for each pixel position in subsequent loss calculations.

[0108] S5. Based on the radiometric statistics of the corrected image, the initial high-resolution image is linearly adjusted channel by channel to output a radiometrically accurate image.

[0109] This step is used for radiometric correction. While the image reconstructed by the super-resolution network improves spatial detail, the pixel brightness distribution of each band may systematically deviate from the original corrected image. This directly affects the accuracy of quantitative applications such as remote sensing image classification and change detection. Radiometric statistics refer to the distribution parameters such as the mean and standard deviation of each band in the corrected image. Channel-by-channel linear adjustment involves independently performing a linear transformation on each band of the initial high-resolution image to make its mean and standard deviation consistent with the corresponding bands in the corrected image. This brings the radiometric values ​​back to a range consistent with the original data without destroying the spatial detail reconstructed by the network, ensuring that the reconstruction results can be directly integrated into the standard product production process of the ground processing system.

[0110] To address the issue of radiometric drift in reconstructed images from deep learning, this invention designs the following channel-by-channel linear correction formula based on the radiometric statistical characteristics of the original corrected image. This formula achieves radiometric consistency between the reconstructed and original images while preserving spatial detail:

[0111] ,

[0112] in, For the initial high-resolution image in band b, Pixel value at; To correct the radiation fidelity image in band b, The pixel value at that location, and the spatial size are... completely consistent; The grayscale mean of band b in the initial high-resolution image is given. The initial image height and width; The standard deviation of band b of the initial high-resolution image; The mean and standard deviation of the grayscale values ​​in band b of the image were obtained statistically after radiometric calibration and atmospheric correction.

[0113] The above calculations are performed independently on each band of the initial high-resolution image, aligning the pixel value distribution of each band to the statistical level of the corrected image. This ensures that the spatial details obtained from the reconstruction are not changed, and that the radiometric values ​​are consistent with the original observations. This allows for direct integration into the standardized production process of the optical satellite ground processing system.

[0114] S6. Calculate the no-reference quality score and the full-reference quality score relative to the reference image of the radiometric image, dynamically adjust the coefficients of the boundary preservation cost and the geometric deformation penalty based on the weighted result of the two, and cyclically feed them back to the network parameter optimization until the score reaches the threshold, and output the final high-resolution reconstructed image.

[0115] This step is used for adaptive quality assessment and iterative optimization. No-reference quality scoring relies solely on the statistical characteristics of the image itself for quality assessment, requiring no external real-world reference. Full-reference quality scoring, on the other hand, requires pixel-by-pixel comparison between the reconstructed image and the original high-resolution reference image, calculating structural similarity and peak signal-to-noise ratio. In the early stages of iteration, the reconstruction quality generated by the network is relatively poor; at this point, full-reference scoring dominates, forcibly correcting the network output using real-world references. In the later stages of iteration, as the reconstruction quality stabilizes, excessive full-reference constraints can lead to overfitting and loss of perceptual realism. Therefore, the weight of no-reference scoring is gradually increased, allowing the network to focus more on the naturalness and sharpness of the image itself. Each iteration evaluates the current reconstruction quality based on the weighted sum of the two scores, reverse-calculates the adjustment amounts for the two cost coefficients, and updates them, enabling the network to move in a better direction in the next iteration. This cycle repeats until the quality score stabilizes and meets the target, outputting the final high-resolution reconstructed image.

[0116] Specifically, in this embodiment, the calculation of the no-reference quality score and the full-reference quality score relative to the reference image of the radiometrically faithful image, the dynamic adjustment of the coefficients of the boundary preservation cost and the geometric deformation penalty based on the weighted result of the two, and the iterative feedback to the network parameter optimization until the score reaches a threshold, and the output of the final high-resolution reconstructed image, specifically includes:

[0117] S6.1 Calculate a no-reference quality score based on the image's own statistical characteristics for the radiofidelity image. Simultaneously, calculate a weighted combination of peak signal-to-noise ratio (PSNR) and structural similarity index relative to the reference image for the full-reference quality score. This sub-step is used to evaluate the quality of the reconstructed image from two different dimensions. The no-reference quality score relies solely on the image's own statistical characteristics and does not require an external real reference. It assesses the image's perceptual quality by analyzing the naturalness of pixel value distribution, the correlation between adjacent pixels, and the richness of high-frequency details. The full-reference quality score requires a pixel-by-pixel comparison between the reconstructed image and the original high-resolution reference image. The PSNR reflects the overall magnitude of pixel brightness error, and the structural similarity index measures the similarity between the two images in terms of brightness, contrast, and structure. The two are weighted and combined to obtain a comprehensive full-reference score, which is used to evaluate the approximation between the reconstructed image and the real reference.

[0118] S6.2. The no-reference quality score and the full-reference quality score are weighted and summed according to a preset dynamic weight coefficient to obtain a comprehensive quality score. The dynamic weight coefficient adaptively adjusts with the number of iterations. In the early stages of iteration, the weight of the full-reference quality score is higher than that of the no-reference quality score, and in the later stages, their weights tend to be balanced. This sub-step is used to merge the two scores into a single optimization objective. The dynamic weight coefficient means that the proportion of the two scores in the comprehensive score is not fixed but automatically adjusted with the number of network training iterations. In the initial stage of network training, the gap between the reconstructed image and the real reference is large. At this time, the full-reference quality score can provide a clear optimization direction, so it is given a higher weight. As iterations proceed, the image quality gradually approaches the reference. Excessive full-reference constraints will limit the improvement of perceived realism, so its weight is gradually reduced, while the weight of the no-reference quality score is gradually increased, so that the network focuses more on the naturalness and sharpness of the image in later optimizations, ultimately achieving a relatively balanced state between the two scores.

[0119] To overcome the shortcomings of traditional fixed-weight quality evaluation methods that cannot adapt to different stages of network iteration, this invention designs the following comprehensive quality scoring formula with dynamic weights, achieving a smooth transition from reference supervision-led to perception-led quality:

[0120] ,

[0121] in, For the first The overall quality score for each iteration ranges from [0,1], with a larger value indicating better reconstruction quality. The full reference quality score for the kth round is given by PSNR(k), which is the peak signal-to-noise ratio between the radiometrically accurate image and the reference image in the kth round, and SSIM(k), which is the structural similarity index between the two. The weighting coefficients of 0.7 and 0.3 are preset according to the radiometric accuracy requirements of the remote sensing image. The no-reference quality score for the k-th round is calculated using natural scene statistical features, using the following formula: ,in The average gradient of the radiometric image. The mean and standard deviation of the average gradient of natural remote sensing images reflect the image's sharpness and naturalness. The dynamic weighting coefficients for the full reference quality score are adaptively adjusted with the iteration round k, and the calculation formula is as follows: ,in The initial weight is 0.8 by default. To predetermine the total number of iterations, in the initial stage of iteration... Approaching Later stages of iteration Approaching 0, it achieves a smooth transition of weights from full-reference dominance to no-reference dominance.

[0122] After each iteration, first calculate and Then calculate based on the current iteration round k. Substituting into the formula, we get Using its maximization as the optimization objective, the algorithm updates in reverse. and This enables adaptive optimization of network parameters.

[0123] S6.3. Taking the maximization of the comprehensive quality score as the optimization objective, calculate the first gradient of the comprehensive quality score with respect to the boundary preservation cost coefficient and the second gradient with respect to the geometric deformation penalty coefficient. Update the boundary preservation cost coefficient and the geometric deformation penalty coefficient according to the first gradient and the second gradient, respectively. This sub-step is used to adjust the regularization strength based on the feedback of the quality score. The comprehensive quality score is an overall evaluation of the network output quality; a higher score indicates better reconstruction quality. The gradient is a mathematical quantity that represents the sensitivity and direction of change of the comprehensive quality score when a parameter undergoes a small change. Calculate the gradient of the comprehensive score with respect to the two cost coefficients: the first gradient reflects the impact of adjusting the boundary preservation cost coefficient on the quality score, and the second gradient reflects the impact of adjusting the geometric deformation penalty coefficient on the quality score. If the gradient is positive, it means that increasing the coefficient helps improve quality, so the coefficient value is increased in that direction; if the gradient is negative, it means that decreasing the coefficient helps improve quality, so the coefficient value is decreased in that direction, so that the two coefficients gradually approach the optimal value.

[0124] S6.4. The updated boundary preservation cost coefficients and geometric deformation penalty coefficients are fed back into the hierarchical upsampling step to adjust the strength of the double regularization constraint. The feature extraction, hierarchical upsampling, and radiometric correction steps are then re-executed to generate an updated radiometrically accurate image. This sub-step is used to feed the adjusted coefficients back to the network front end to form a closed loop. The two updated cost coefficients are not merely recorded in memory but are actually involved in the network's forward computation, specifically inserted into the loss calculation stage of the hierarchical upsampling module. Increasing the coefficients corresponds to a stronger regularization penalty, forcing the network to pay more attention to edge preservation or geometric correction during subsequent forward inference; decreasing the coefficients corresponds to a relaxation of constraints, giving the network greater freedom in the corresponding dimensions. The network then re-executes a complete forward propagation with the adjusted coefficients, including the entire process of feature extraction, upsampling, and radiometric correction, generating a new round of radiometrically accurate images for use in the next round of quality evaluation.

[0125] S6.5. Repeat the above quality evaluation, coefficient update, and feedback iteration process until the rate of change of the comprehensive quality score is less than the preset convergence threshold. Output the radiometrically accurate image generated in the current round as the final high-resolution reconstructed image. This sub-step is used to determine when the iteration terminates and output the final result. Repeat the closed loop consisting of quality evaluation, coefficient update, and feedback iteration. The comprehensive quality score will change after each iteration. When the network parameters and cost coefficients tend to be optimal, the change in score will gradually decrease. Use the absolute value of the difference between the score of this round and the score of the previous round as the rate of change. If the rate of change is less than the preset convergence threshold, it means that the contribution of continued iteration to quality improvement is very small, and it is determined to be in a convergence state, and the iteration loop is terminated. Output the radiometrically accurate image generated in the current round as the final high-resolution reconstruction result. At this time, the image has reached a good level in terms of spatial detail, edge sharpness, and radiometric consistency.

[0126] Furthermore, in this embodiment, the dynamic weight coefficient is adaptively adjusted as the number of iterations increases, specifically including:

[0127] S6.2.1. Set the total number of iterations, define the weight coefficient of the full reference quality score as a variable that decreases with each iteration, and define the weight coefficient of the no-reference quality score as a variable that increases with each iteration, with the sum of the two always equal to a constant value. This sub-step is used to establish the basic rules for the change of the two weight coefficients. The total number of iterations is the pre-set total number of network training cycles, serving as a time reference for weight adjustment. The weight coefficient of the full reference quality score is defined as a variable that gradually decreases with each iteration, meaning that the contribution of this score to the overall quality score decreases as the iteration progresses. The weight coefficient of the no-reference quality score, on the other hand, gradually increases with each iteration. The sum of the two always remains a constant value, which is usually taken as one, ensuring that the overall quality score always remains within a fixed numerical range and does not cause an overall drift in the absolute value of the score due to changes in weight allocation.

[0128] S6.2.2 In the initial iteration, the weight coefficient of the full reference quality score is greater than that of the no-reference quality score, making network optimization primarily driven by the supervision signal of the reference image. As the number of iterations increases, the weight coefficient of the full reference quality score gradually decreases, while the weight coefficient of the no-reference quality score gradually increases. In the later stages of iteration, the weight coefficient of the no-reference quality score becomes greater than that of the full reference quality score, making network optimization primarily driven by the perceived quality of the image itself. This sub-step describes the dynamic evolution logic of the weight coefficients as the iteration process progresses. At the beginning of training, the reconstructed image output by the network deviates significantly from the real reference image. At this time, the full reference quality score has clear guiding significance and can provide accurate directional supervision for the network. Therefore, it is given a higher weight to enable the network to learn how to approximate the real reference. As the iteration progresses, the reconstructed image gradually approaches the reference image. Overemphasizing the full reference constraint will lead to an overly smooth reconstruction result or a loss of realism. At this time, its weight is gradually reduced, while the weight of the no-reference quality score is gradually increased, allowing the network to focus more on the visual naturalness and detail realism of the image in the later stages of learning.

[0129] S6.2.3 After each iteration completes the comprehensive quality score calculation, the two weight coefficients are updated based on the current iteration round, and the comprehensive quality score for the next iteration is calculated using the updated weight coefficients. This sub-step defines the temporal relationship of weight updates. In each iteration, the network first completes a full forward propagation to obtain the current radiometrically accurate image and calculates the comprehensive quality score. Then, based on the current iteration round number, the new values ​​of the two weight coefficients are recalculated. These two new values ​​are not used immediately in this round, but are used as weight parameters when calculating the comprehensive quality score in the next iteration. This ensures that quality evaluation and weight adjustment are decoupled temporally in each iteration. Weight updates are based on the results of the previous round but do not affect the score of the previous round; they only affect the optimization direction of subsequent iterations.

[0130] S6.2.4. When the number of iterations reaches the preset upper limit or the rate of change of the comprehensive quality score is lower than the preset convergence threshold multiple times consecutively, the cycle of weight coefficient update and network parameter optimization is terminated, and the current final high-resolution reconstructed image is output. This sub-step is used to set dual conditions for iteration termination. The preset upper limit is a hard constraint, specifying the maximum number of iterations the network can execute, preventing training from going on indefinitely due to slow convergence. The convergence condition is a soft constraint; when the rate of change of the comprehensive quality score is lower than the preset threshold multiple times consecutively, it indicates that the network has entered a stable state, and the quality improvement brought by continued iteration is extremely limited. Termination is triggered when either condition is met; the weight coefficients stop updating, the network parameters are no longer adjusted, and the reconstructed image generated in the current iteration is output as the final result. The dual conditions ensure that the network has sufficient opportunity to converge to the optimal state while avoiding the waste of computational resources caused by slow convergence.

[0131] Example 2

[0132] like Figure 2 As shown, this application provides an architecture diagram of a remote sensing image super-resolution reconstruction device for an optical satellite ground processing system, applied to the method described in Embodiment 1, including:

[0133] The data acquisition and prior knowledge construction module 210 is used to acquire multi-temporal optical satellite images and perform orientation correction. Based on gradient response and image pixel-level segmentation, it constructs a prior knowledge base containing boundary constraints, regional homogeneity constraints and perspective transformation matrices, and outputs the corrected image and the prior knowledge base.

[0134] The radiometric correction and sample generation module 220 is used to perform radiometric calibration and atmospheric correction on the corrected image, construct a degradation model based on the sensor diffusion function to generate training pair data, and synchronously feed the corrected image, the training pair data and the prior knowledge base into the network.

[0135] The feature extraction and fusion module 230 is used to utilize the boundary constraints modulate shallow features and the region homogeneous constraint clustering integration features in the prior knowledge base, output a multi-scale fusion feature tensor after channel recalibration, and transmit the perspective transformation matrix.

[0136] The hierarchical upsampling and regularization module 240 is used to perform hierarchical upsampling on the multi-scale fused feature tensor, and introduces dual regularization of boundary preservation cost derived from the boundary constraints and geometric deformation penalty derived from the perspective transformation matrix to generate an initial high-resolution image.

[0137] The radiation fidelity correction module 250 is used to linearly adjust the initial high-resolution image channel by channel based on the radiation statistics of the corrected image, and output a radiation fidelity image.

[0138] The quality evaluation and feedback iteration module 260 is used to calculate the no-reference quality score and the full-reference quality score relative to the reference image of the radiometric image. Based on the weighted result of the two, the coefficients of the boundary preservation cost and the geometric deformation penalty are dynamically adjusted and iteratively fed back to the network parameter optimization until the score reaches the threshold, and the final high-resolution reconstructed image is output.

[0139] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0140] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0141] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0142] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0143] It should be noted that a portion of the electronic device described in the above embodiments can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0144] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0145] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0146] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0147] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for super-resolution reconstruction of remote sensing images for optical satellite ground processing systems, characterized in that, Includes the following steps: Acquire multi-temporal optical satellite images and perform orientation correction. Construct a prior knowledge base containing boundary constraints, regional homogeneity constraints, and perspective transformation matrices based on gradient response and image pixel-level segmentation. Output the corrected image and the prior knowledge base. The corrected image is radiometrically calibrated and atmospherically corrected. A degradation model is constructed based on the sensor diffusion function to generate training pair data. The corrected image, the training pair data, and the prior knowledge base are synchronously fed into the network. The shallow features modulated by the boundary constraints in the prior knowledge base and the clustering integration features of the region homogeneity constraints are used to output a multi-scale feature tensor after channel recalibration and to pass the perspective transformation matrix. The multi-scale feature tensor is subjected to hierarchical upsampling, and dual regularization is introduced, which includes a boundary preservation cost derived from the boundary constraints and a geometric deformation penalty derived from the perspective transformation matrix, to generate an initial high-resolution image. The initial high-resolution image is linearly adjusted channel by channel based on the radiometric statistics of the corrected image to output a radiometrically faithful image; The no-reference quality score and the full-reference quality score relative to the reference image of the radiometric image are calculated. The coefficients of the boundary preservation cost and the geometric deformation penalty are dynamically adjusted based on the weighted result of the two, and the results are iteratively fed back to the network parameter optimization until the score reaches the threshold, and the final high-resolution reconstructed image is output.

2. The method according to claim 1, characterized in that, The prior knowledge base constructed based on gradient response and pixel-level image segmentation, including boundary constraints, region homogeneity constraints, and perspective transformation matrices, includes: Gradient magnitude calculation and non-maximum suppression are performed on the corrected image to extract edge response points and generate an initial edge response map; superpixel segmentation is performed on the corrected image to obtain a pixel-level segmentation map; Using the edge response point as the seed point, boundary shifting is performed along the gradient direction to spread the edge response point into a continuous edge line. The continuous edge line is then clipped and corrected according to the superpixel boundary in the pixel-level segmentation map to generate the boundary constraint. Using each superpixel region in the pixel-level segmentation map as the basic unit, the grayscale mean and texture response of pixels in each region are statistically analyzed to establish a region homogeneous descriptor and construct an adjacency relationship graph between adjacent superpixels. Based on the confidence propagation algorithm, the region attribute information is iteratively transmitted on the adjacency relationship graph. The transmission cost is used to quantify the difficulty of merging adjacent superpixel regions, realize the merging and optimization of homogeneous regions, and generate the region homogeneity constraint. Stable feature points are extracted within each superpixel region, feature point matching is performed across temporal images, a local perspective transformation model is fitted to the matching point pairs, and the perspective transformation matrix is ​​generated.

3. The method according to claim 2, characterized in that, The confidence-based propagation algorithm iteratively transmits region attribute information on the adjacency graph. The transmission cost is used to quantify the difficulty of merging adjacent superpixel regions, realizing the merging and optimization of homogeneous regions. Specifically, it includes: Each superpixel region in the adjacency graph is considered as a node, the node state is defined as its land cover category label, and the region merging cost between adjacent nodes is defined as the weighted sum of spectral feature differences and texture feature differences. Initialize the category label confidence at each node, perform forward message passing from the leaf node to the root node of the adjacency graph, and update the label confidence vector of each node according to the received neighbor node messages. After the forward propagation is completed, the backward message is traced back from the root node to the leaf node. Each node calculates the edge posterior probability by combining the forward and backward messages, and selects the category label corresponding to the maximum posterior probability as the final category label of the node. Based on the final category label of each node, adjacent superpixel regions with the same category label are merged, and the region homogeneity descriptor is recalculated using the merged region unit. The above message passing and region merging process is iteratively executed until the region label change rate in two adjacent iterations is lower than the preset convergence threshold, and the optimized region homogeneity constraint is output.

4. The method according to claim 1, characterized in that, The step of hierarchically upsampling the multi-scale fused feature tensor, introducing dual regularization of boundary preservation cost derived from the boundary constraints and geometric deformation penalty derived from the perspective transformation matrix, to generate an initial high-resolution image includes: The multi-scale fused feature tensor is used as the initial input and processed through multiple upsampling modules. In each upsampling module, the spatial resolution of the feature map is first doubled by a sub-pixel convolutional layer, and then the channel dimension feature is refined by a residual dense connection block to output the high-resolution feature map of that level. At the output of each upsampling module, a boundary preservation cost module is introduced to convert the boundary constraint into a spatial weight mask. The spatial weight mask is used to apply a weighted loss to the high-resolution feature map to constrain the pixel response of the edge region and the pixel response of the background region to maintain a preset contrast difference threshold. At the output of each upsampling module, a geometric deformation penalty module is introduced to decompose the perspective transformation matrix into rotation, scaling and translation components. The rotation, scaling and translation deviations of the current feature map spatial coordinates relative to the reference coordinate system are calculated respectively, and the weighted sum of the squares of the three is used as the geometric deformation penalty term to constrain the spatial deformation amplitude during the upsampling process. The final output of the multi-level upsampling module is mapped to the image space through a convolutional layer to generate the initial high-resolution image.

5. The method according to claim 4, characterized in that, The introduced boundary preservation cost module converts the boundary constraints into a spatial weight mask, specifically including: The continuous edge lines in the boundary constraints are binarized to generate an initial edge mask that matches the spatial size of the current level feature map. In the initial edge mask, the pixel value at the edge line position is 1, and the pixel value at the non-edge position is 0. Perform Euclidean distance transformation on the initial edge mask and calculate the distance from each pixel position in the current level feature map to the nearest edge line; A spatial weighting function is constructed based on the distance value. The spatial weighting function is an exponential decay function with the distance value as the independent variable, so that pixels closer to the edge line receive a larger weight value and pixels farther from the edge line receive a smaller weight value. The spatial weight function is applied to the current level feature map, and the gradient response of each pixel in the feature map is weighted by the weight value. After normalization, the spatial weight mask is output, so that the spatial weight mask assigns high weight to the edge region and low weight to the flat region.

6. The method according to claim 1, characterized in that, The process of utilizing the boundary constraints modulating shallow features and the region homogeneous constraint clustering integration features from the prior knowledge base, outputting a multi-scale fused feature tensor after channel recalibration, and transmitting the perspective transformation matrix specifically includes: The corrected image and the training paired data are input together into the front end of the deep convolutional neural network. The shallow texture features, mid-level structural features and deep semantic features under different receptive fields are extracted through the initial convolutional layer and are respectively denoted as the first-level features, the second-level features and the third-level features. The boundary constraints are converted into a boundary weight map with the same size as the first-level feature space. The boundary weight map is used to perform point-by-point weighted modulation on each pixel position of the first-level feature, so that the texture response of the edge region is enhanced and the texture response of the non-edge region is suppressed, and the modulated shallow edge enhancement feature is output. The homogeneous region labels in the homogeneous region constraint are mapped to the spatial location of the second-level feature. The feature vectors of all pixel locations belonging to the same homogeneous region label are pooled by mean to obtain the region prototype vector of each homogeneous region. Then, the region prototype vector is backfilled to all pixel locations of the corresponding region to make the features in the same region tend to be consistent, and the clustered and integrated mid-level region aggregation feature is output. The modulated shallow edge enhancement features, the clustered and integrated mid-level region aggregation features, and the third-level features are concatenated along the channel dimension and input into the channel attention module. The channel attention module performs global average pooling and global max pooling on each channel feature, generates activation weights for each channel through a shared multilayer perceptron, and recalibrates each concatenated channel using the activation weights to output the multi-scale fusion feature tensor. The perspective transformation matrix is ​​associated with the multi-scale fusion feature tensor as geometric constraint information and then synchronously transmitted to subsequent steps.

7. The method according to claim 6, characterized in that, The step of converting the boundary constraints into a boundary weight map with the same size as the first-level feature space specifically includes: Obtain the set of continuous edge line coordinates in the boundary constraint. Using the set of continuous edge line coordinates as an index, in a blank weight map with the same size as the first-level feature space, set the pixel value at the position corresponding to the edge line to a first preset value and set the pixel value at the position not on the edge line to a second preset value to generate a binarized boundary weight map. The binarized boundary weight map is subjected to Gaussian kernel convolution diffusion processing, so that the weight values ​​at the edge line position decay in a Gaussian distribution to both sides along the direction perpendicular to the edge line, thereby generating a diffused boundary weight map. The diffusion boundary weight map is normalized so that the weight values ​​of the entire map are distributed within a preset range, and the normalized boundary weight map is output. Using the normalized boundary weight map as the weight coefficient, the first-level features are multiplied and weighted channel by channel and pixel by pixel. The edge region obtains a higher weight coefficient than the flat region, thereby realizing the modulation of shallow texture features by boundary constraints.

8. The method according to claim 1, characterized in that, The calculation of the no-reference quality score and the full-reference quality score relative to the reference image of the radiometrically faithful image, the dynamic adjustment of the coefficients of the boundary preservation cost and the geometric deformation penalty based on the weighted result of the two, and the feedback to the network parameter optimization are cyclically fed back until the score reaches a threshold, and the final high-resolution reconstructed image is output, specifically including: A referenceless quality score based on the image's own statistical characteristics is calculated for the radiofidelity image, while a weighted combination of peak signal-to-noise ratio and structural similarity index is calculated for the radiofidelity image relative to the reference image as a full reference quality score. The no-reference quality score and the full-reference quality score are weighted and summed according to a preset dynamic weight coefficient to obtain a comprehensive quality score. The dynamic weight coefficient is adaptively adjusted as the number of iterations increases. In the early stage of the iteration, the weight of the full-reference quality score is higher than that of the no-reference quality score, and in the later stage of the iteration, the weights of the two tend to be balanced. With the goal of maximizing the overall quality score, the first gradient of the overall quality score with respect to the boundary preservation cost coefficient and the second gradient with respect to the geometric deformation penalty coefficient are calculated, and the boundary preservation cost coefficient and the geometric deformation penalty coefficient are updated according to the first gradient and the second gradient, respectively. The updated boundary preservation cost coefficient and the geometric deformation penalty coefficient are fed back into the hierarchical upsampling step to adjust the constraint strength of double regularization. The feature extraction, hierarchical upsampling and radiometric correction steps are re-executed to generate an updated radiometric image. Repeat the above quality evaluation, coefficient update and feedback iteration process until the rate of change of the comprehensive quality score is less than the preset convergence threshold, and output the radiometrically accurate image generated in the current round as the final high-resolution reconstructed image.

9. The method according to claim 8, characterized in that, The dynamic weighting coefficients are adaptively adjusted as the number of iterations increases, specifically including: Set the total number of iterations, define the weight coefficient of the full reference quality score as a variable that decreases with the number of iterations, define the weight coefficient of the no-reference quality score as a variable that increases with the number of iterations, and the sum of the two is always equal to a constant value. In the early stages of iteration, the weight coefficient of the full reference quality score is greater than that of the no-reference quality score, making the network optimization dominated by the supervision signal of the reference image. As the number of iterations increases, the weight coefficient of the full reference quality score gradually decreases, while the weight coefficient of the no-reference quality score gradually increases. In the later stages of iteration, the weight coefficient of the no-reference quality score is greater than that of the full reference quality score, making the network optimization dominated by the perceived quality of the image itself. After each iteration completes the calculation of the overall quality score, the two weight coefficients are updated according to the current iteration round, and the overall quality score of the next iteration is calculated using the updated weight coefficients. When the number of iterations reaches the preset upper limit or the rate of change of the comprehensive quality score is lower than the preset convergence threshold for several consecutive times, the cycle of weight coefficient update and network parameter optimization is terminated, and the current final high-resolution reconstructed image is output.

10. A remote sensing image super-resolution reconstruction device for an optical satellite ground processing system, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition and prior knowledge construction module is used to acquire multi-temporal optical satellite images and perform orientation correction. Based on gradient response and image pixel-level segmentation, it constructs a prior knowledge base containing boundary constraints, regional homogeneity constraints and perspective transformation matrices, and outputs the corrected image and the prior knowledge base. The radiometric correction and sample generation module is used to perform radiometric calibration and atmospheric correction on the corrected image, construct a degradation model based on the sensor diffusion function to generate training pair data, and synchronously feed the corrected image, the training pair data and the prior knowledge base into the network. The feature extraction and fusion module is used to utilize the boundary constraints modulate shallow features and the region homogeneous constraint clustering integration features in the prior knowledge base, output a multi-scale fused feature tensor after channel recalibration, and transmit the perspective transformation matrix; The hierarchical upsampling and regularization module is used to perform hierarchical upsampling on the multi-scale fused feature tensor, and introduces dual regularization of boundary preservation cost derived from the boundary constraints and geometric deformation penalty derived from the perspective transformation matrix to generate an initial high-resolution image. The radiation fidelity correction module is used to linearly adjust the initial high-resolution image channel by channel based on the radiation statistics of the corrected image, and output a radiation fidelity image. The quality evaluation and feedback iteration module is used to calculate the no-reference quality score and the full-reference quality score relative to the reference image of the radiometric image. Based on the weighted result of the two, the coefficients of the boundary preservation cost and the geometric deformation penalty are dynamically adjusted and iteratively fed back to the network parameter optimization until the score reaches the threshold, and the final high-resolution reconstructed image is output.

Images